diff --git a/.gitignore b/.gitignore old mode 100644 new mode 100755 diff --git a/README.md b/README.md index 9e2af39..0648e5d 100644 --- a/README.md +++ b/README.md @@ -1,29 +1,5 @@ - - -# Python WSI Preprocessing - -This project contains a variety of files for investigating image preprocessing using Python -with the aim of using deep learning to perform histopathology image classification of -whole slide images. - -See main tutorial [here](./docs/wsi-preprocessing-in-python/index.md). - -See main project at [https://github.com/CODAIT/deep-histopath](https://github.com/CODAIT/deep-histopath) -for more information. +Go to https://github.com/FAU-DLM/wsi_processing_pipelin for changes in this library. +All advances and new features will be committed there. diff --git a/deephistopath/wsi/slide.py b/deephistopath/wsi/slide.py deleted file mode 100644 index 3bde6c2..0000000 --- a/deephistopath/wsi/slide.py +++ /dev/null @@ -1,1027 +0,0 @@ -# ------------------------------------------------------------------------ -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# -# ------------------------------------------------------------------------ - -import glob -import math -import matplotlib.pyplot as plt -import multiprocessing -import numpy as np -import openslide -from openslide import OpenSlideError -import os -import PIL -from PIL import Image -import re -import sys -from deephistopath.wsi import util -from deephistopath.wsi.util import Time - -BASE_DIR = os.path.join(".", "data") -# BASE_DIR = os.path.join(os.sep, "Volumes", "BigData", "TUPAC") -TRAIN_PREFIX = "TUPAC-TR-" -SRC_TRAIN_DIR = os.path.join(BASE_DIR, "training_slides") -SRC_TRAIN_EXT = "svs" -DEST_TRAIN_SUFFIX = "" # Example: "train-" -DEST_TRAIN_EXT = "png" -SCALE_FACTOR = 32 -DEST_TRAIN_DIR = os.path.join(BASE_DIR, "training_" + DEST_TRAIN_EXT) -THUMBNAIL_SIZE = 300 -THUMBNAIL_EXT = "jpg" - -DEST_TRAIN_THUMBNAIL_DIR = os.path.join(BASE_DIR, "training_thumbnail_" + THUMBNAIL_EXT) - -FILTER_SUFFIX = "" # Example: "filter-" -FILTER_RESULT_TEXT = "filtered" -FILTER_DIR = os.path.join(BASE_DIR, "filter_" + DEST_TRAIN_EXT) -FILTER_THUMBNAIL_DIR = os.path.join(BASE_DIR, "filter_thumbnail_" + THUMBNAIL_EXT) -FILTER_PAGINATION_SIZE = 50 -FILTER_PAGINATE = True -FILTER_HTML_DIR = BASE_DIR - -TILE_SUMMARY_DIR = os.path.join(BASE_DIR, "tile_summary_" + DEST_TRAIN_EXT) -TILE_SUMMARY_ON_ORIGINAL_DIR = os.path.join(BASE_DIR, "tile_summary_on_original_" + DEST_TRAIN_EXT) -TILE_SUMMARY_SUFFIX = "tile_summary" -TILE_SUMMARY_THUMBNAIL_DIR = os.path.join(BASE_DIR, "tile_summary_thumbnail_" + THUMBNAIL_EXT) -TILE_SUMMARY_ON_ORIGINAL_THUMBNAIL_DIR = os.path.join(BASE_DIR, "tile_summary_on_original_thumbnail_" + THUMBNAIL_EXT) -TILE_SUMMARY_PAGINATION_SIZE = 50 -TILE_SUMMARY_PAGINATE = True -TILE_SUMMARY_HTML_DIR = BASE_DIR - -TILE_DATA_DIR = os.path.join(BASE_DIR, "tile_data") -TILE_DATA_SUFFIX = "tile_data" - -TOP_TILES_SUFFIX = "top_tile_summary" -TOP_TILES_DIR = os.path.join(BASE_DIR, TOP_TILES_SUFFIX + "_" + DEST_TRAIN_EXT) -TOP_TILES_THUMBNAIL_DIR = os.path.join(BASE_DIR, TOP_TILES_SUFFIX + "_thumbnail_" + THUMBNAIL_EXT) -TOP_TILES_ON_ORIGINAL_DIR = os.path.join(BASE_DIR, TOP_TILES_SUFFIX + "_on_original_" + DEST_TRAIN_EXT) -TOP_TILES_ON_ORIGINAL_THUMBNAIL_DIR = os.path.join(BASE_DIR, - TOP_TILES_SUFFIX + "_on_original_thumbnail_" + THUMBNAIL_EXT) - -TILE_DIR = os.path.join(BASE_DIR, "tiles_" + DEST_TRAIN_EXT) -TILE_SUFFIX = "tile" - -STATS_DIR = os.path.join(BASE_DIR, "svs_stats") - - -def open_slide(filename): - """ - Open a whole-slide image (*.svs, etc). - - Args: - filename: Name of the slide file. - - Returns: - An OpenSlide object representing a whole-slide image. - """ - try: - slide = openslide.open_slide(filename) - except OpenSlideError: - slide = None - except FileNotFoundError: - slide = None - return slide - - -def open_image(filename): - """ - Open an image (*.jpg, *.png, etc). - - Args: - filename: Name of the image file. - - returns: - A PIL.Image.Image object representing an image. - """ - image = Image.open(filename) - return image - - -def open_image_np(filename): - """ - Open an image (*.jpg, *.png, etc) as an RGB NumPy array. - - Args: - filename: Name of the image file. - - returns: - A NumPy representing an RGB image. - """ - pil_img = open_image(filename) - np_img = util.pil_to_np_rgb(pil_img) - return np_img - - -def get_training_slide_path(slide_number): - """ - Convert slide number to a path to the corresponding WSI training slide file. - - Example: - 5 -> ../data/training_slides/TUPAC-TR-005.svs - - Args: - slide_number: The slide number. - - Returns: - Path to the WSI training slide file. - """ - padded_sl_num = str(slide_number).zfill(3) - slide_filepath = os.path.join(SRC_TRAIN_DIR, TRAIN_PREFIX + padded_sl_num + "." + SRC_TRAIN_EXT) - return slide_filepath - - -def get_tile_image_path(tile): - """ - Obtain tile image path based on tile information such as row, column, row pixel position, column pixel position, - pixel width, and pixel height. - - Args: - tile: Tile object. - - Returns: - Path to image tile. - """ - t = tile - padded_sl_num = str(t.slide_num).zfill(3) - tile_path = os.path.join(TILE_DIR, padded_sl_num, - TRAIN_PREFIX + padded_sl_num + "-" + TILE_SUFFIX + "-r%d-c%d-x%d-y%d-w%d-h%d" % ( - t.r, t.c, t.o_c_s, t.o_r_s, t.o_c_e - t.o_c_s, t.o_r_e - t.o_r_s) + "." + DEST_TRAIN_EXT) - return tile_path - - -def get_tile_image_path_by_slide_row_col(slide_number, row, col): - """ - Obtain tile image path using wildcard lookup with slide number, row, and column. - - Args: - slide_number: The slide number. - row: The row. - col: The column. - - Returns: - Path to image tile. - """ - padded_sl_num = str(slide_number).zfill(3) - wilcard_path = os.path.join(TILE_DIR, padded_sl_num, - TRAIN_PREFIX + padded_sl_num + "-" + TILE_SUFFIX + "-r%d-c%d-*." % ( - row, col) + DEST_TRAIN_EXT) - img_path = glob.glob(wilcard_path)[0] - return img_path - - -def get_training_image_path(slide_number, large_w=None, large_h=None, small_w=None, small_h=None): - """ - Convert slide number and optional dimensions to a training image path. If no dimensions are supplied, - the corresponding file based on the slide number will be looked up in the file system using a wildcard. - - Example: - 5 -> ../data/training_png/TUPAC-TR-005-32x-49920x108288-1560x3384.png - - Args: - slide_number: The slide number. - large_w: Large image width. - large_h: Large image height. - small_w: Small image width. - small_h: Small image height. - - Returns: - Path to the image file. - """ - padded_sl_num = str(slide_number).zfill(3) - if large_w is None and large_h is None and small_w is None and small_h is None: - wildcard_path = os.path.join(DEST_TRAIN_DIR, TRAIN_PREFIX + padded_sl_num + "*." + DEST_TRAIN_EXT) - img_path = glob.glob(wildcard_path)[0] - else: - img_path = os.path.join(DEST_TRAIN_DIR, TRAIN_PREFIX + padded_sl_num + "-" + str( - SCALE_FACTOR) + "x-" + DEST_TRAIN_SUFFIX + str( - large_w) + "x" + str(large_h) + "-" + str(small_w) + "x" + str(small_h) + "." + DEST_TRAIN_EXT) - return img_path - - -def get_training_thumbnail_path(slide_number, large_w=None, large_h=None, small_w=None, small_h=None): - """ - Convert slide number and optional dimensions to a training thumbnail path. If no dimensions are - supplied, the corresponding file based on the slide number will be looked up in the file system using a wildcard. - - Example: - 5 -> ../data/training_thumbnail_jpg/TUPAC-TR-005-32x-49920x108288-1560x3384.jpg - - Args: - slide_number: The slide number. - large_w: Large image width. - large_h: Large image height. - small_w: Small image width. - small_h: Small image height. - - Returns: - Path to the thumbnail file. - """ - padded_sl_num = str(slide_number).zfill(3) - if large_w is None and large_h is None and small_w is None and small_h is None: - wilcard_path = os.path.join(DEST_TRAIN_THUMBNAIL_DIR, TRAIN_PREFIX + padded_sl_num + "*." + THUMBNAIL_EXT) - img_path = glob.glob(wilcard_path)[0] - else: - img_path = os.path.join(DEST_TRAIN_THUMBNAIL_DIR, TRAIN_PREFIX + padded_sl_num + "-" + str( - SCALE_FACTOR) + "x-" + DEST_TRAIN_SUFFIX + str( - large_w) + "x" + str(large_h) + "-" + str(small_w) + "x" + str(small_h) + "." + THUMBNAIL_EXT) - return img_path - - -def get_filter_image_path(slide_number, filter_number, filter_name_info): - """ - Convert slide number, filter number, and text to a path to a filter image file. - - Example: - 5, 1, "rgb" -> ../data/filter_png/TUPAC-TR-005-001-rgb.png - - Args: - slide_number: The slide number. - filter_number: The filter number. - filter_name_info: Descriptive text describing filter. - - Returns: - Path to the filter image file. - """ - dir = FILTER_DIR - if not os.path.exists(dir): - os.makedirs(dir) - img_path = os.path.join(dir, get_filter_image_filename(slide_number, filter_number, filter_name_info)) - return img_path - - -def get_filter_thumbnail_path(slide_number, filter_number, filter_name_info): - """ - Convert slide number, filter number, and text to a path to a filter thumbnail file. - - Example: - 5, 1, "rgb" -> ../data/filter_thumbnail_jpg/TUPAC-TR-005-001-rgb.jpg - - Args: - slide_number: The slide number. - filter_number: The filter number. - filter_name_info: Descriptive text describing filter. - - Returns: - Path to the filter thumbnail file. - """ - dir = FILTER_THUMBNAIL_DIR - if not os.path.exists(dir): - os.makedirs(dir) - img_path = os.path.join(dir, get_filter_image_filename(slide_number, filter_number, filter_name_info, thumbnail=True)) - return img_path - - -def get_filter_image_filename(slide_number, filter_number, filter_name_info, thumbnail=False): - """ - Convert slide number, filter number, and text to a filter file name. - - Example: - 5, 1, "rgb", False -> TUPAC-TR-005-001-rgb.png - 5, 1, "rgb", True -> TUPAC-TR-005-001-rgb.jpg - - Args: - slide_number: The slide number. - filter_number: The filter number. - filter_name_info: Descriptive text describing filter. - thumbnail: If True, produce thumbnail filename. - - Returns: - The filter image or thumbnail file name. - """ - if thumbnail: - ext = THUMBNAIL_EXT - else: - ext = DEST_TRAIN_EXT - padded_sl_num = str(slide_number).zfill(3) - padded_fi_num = str(filter_number).zfill(3) - img_filename = TRAIN_PREFIX + padded_sl_num + "-" + padded_fi_num + "-" + FILTER_SUFFIX + filter_name_info + "." + ext - return img_filename - - -def get_tile_summary_image_path(slide_number): - """ - Convert slide number to a path to a tile summary image file. - - Example: - 5 -> ../data/tile_summary_png/TUPAC-TR-005-tile_summary.png - - Args: - slide_number: The slide number. - - Returns: - Path to the tile summary image file. - """ - if not os.path.exists(TILE_SUMMARY_DIR): - os.makedirs(TILE_SUMMARY_DIR) - img_path = os.path.join(TILE_SUMMARY_DIR, get_tile_summary_image_filename(slide_number)) - return img_path - - -def get_tile_summary_thumbnail_path(slide_number): - """ - Convert slide number to a path to a tile summary thumbnail file. - - Example: - 5 -> ../data/tile_summary_thumbnail_jpg/TUPAC-TR-005-tile_summary.jpg - - Args: - slide_number: The slide number. - - Returns: - Path to the tile summary thumbnail file. - """ - if not os.path.exists(TILE_SUMMARY_THUMBNAIL_DIR): - os.makedirs(TILE_SUMMARY_THUMBNAIL_DIR) - img_path = os.path.join(TILE_SUMMARY_THUMBNAIL_DIR, get_tile_summary_image_filename(slide_number, thumbnail=True)) - return img_path - - -def get_tile_summary_on_original_image_path(slide_number): - """ - Convert slide number to a path to a tile summary on original image file. - - Example: - 5 -> ../data/tile_summary_on_original_png/TUPAC-TR-005-tile_summary.png - - Args: - slide_number: The slide number. - - Returns: - Path to the tile summary on original image file. - """ - if not os.path.exists(TILE_SUMMARY_ON_ORIGINAL_DIR): - os.makedirs(TILE_SUMMARY_ON_ORIGINAL_DIR) - img_path = os.path.join(TILE_SUMMARY_ON_ORIGINAL_DIR, get_tile_summary_image_filename(slide_number)) - return img_path - - -def get_tile_summary_on_original_thumbnail_path(slide_number): - """ - Convert slide number to a path to a tile summary on original thumbnail file. - - Example: - 5 -> ../data/tile_summary_on_original_thumbnail_jpg/TUPAC-TR-005-tile_summary.jpg - - Args: - slide_number: The slide number. - - Returns: - Path to the tile summary on original thumbnail file. - """ - if not os.path.exists(TILE_SUMMARY_ON_ORIGINAL_THUMBNAIL_DIR): - os.makedirs(TILE_SUMMARY_ON_ORIGINAL_THUMBNAIL_DIR) - img_path = os.path.join(TILE_SUMMARY_ON_ORIGINAL_THUMBNAIL_DIR, - get_tile_summary_image_filename(slide_number, thumbnail=True)) - return img_path - - -def get_top_tiles_on_original_image_path(slide_number): - """ - Convert slide number to a path to a top tiles on original image file. - - Example: - 5 -> ../data/top_tiles_on_original_png/TUPAC-TR-005-32x-49920x108288-1560x3384-top_tiles.png - - Args: - slide_number: The slide number. - - Returns: - Path to the top tiles on original image file. - """ - if not os.path.exists(TOP_TILES_ON_ORIGINAL_DIR): - os.makedirs(TOP_TILES_ON_ORIGINAL_DIR) - img_path = os.path.join(TOP_TILES_ON_ORIGINAL_DIR, get_top_tiles_image_filename(slide_number)) - return img_path - - -def get_top_tiles_on_original_thumbnail_path(slide_number): - """ - Convert slide number to a path to a top tiles on original thumbnail file. - - Example: - 5 -> ../data/top_tiles_on_original_thumbnail_jpg/TUPAC-TR-005-32x-49920x108288-1560x3384-top_tiles.jpg - - Args: - slide_number: The slide number. - - Returns: - Path to the top tiles on original thumbnail file. - """ - if not os.path.exists(TOP_TILES_ON_ORIGINAL_THUMBNAIL_DIR): - os.makedirs(TOP_TILES_ON_ORIGINAL_THUMBNAIL_DIR) - img_path = os.path.join(TOP_TILES_ON_ORIGINAL_THUMBNAIL_DIR, - get_top_tiles_image_filename(slide_number, thumbnail=True)) - return img_path - - -def get_tile_summary_image_filename(slide_number, thumbnail=False): - """ - Convert slide number to a tile summary image file name. - - Example: - 5, False -> TUPAC-TR-005-tile_summary.png - 5, True -> TUPAC-TR-005-tile_summary.jpg - - Args: - slide_number: The slide number. - thumbnail: If True, produce thumbnail filename. - - Returns: - The tile summary image file name. - """ - if thumbnail: - ext = THUMBNAIL_EXT - else: - ext = DEST_TRAIN_EXT - padded_sl_num = str(slide_number).zfill(3) - - training_img_path = get_training_image_path(slide_number) - large_w, large_h, small_w, small_h = parse_dimensions_from_image_filename(training_img_path) - img_filename = TRAIN_PREFIX + padded_sl_num + "-" + str(SCALE_FACTOR) + "x-" + str(large_w) + "x" + str( - large_h) + "-" + str(small_w) + "x" + str(small_h) + "-" + TILE_SUMMARY_SUFFIX + "." + ext - - return img_filename - - -def get_top_tiles_image_filename(slide_number, thumbnail=False): - """ - Convert slide number to a top tiles image file name. - - Example: - 5, False -> TUPAC-TR-005-32x-49920x108288-1560x3384-top_tiles.png - 5, True -> TUPAC-TR-005-32x-49920x108288-1560x3384-top_tiles.jpg - - Args: - slide_number: The slide number. - thumbnail: If True, produce thumbnail filename. - - Returns: - The top tiles image file name. - """ - if thumbnail: - ext = THUMBNAIL_EXT - else: - ext = DEST_TRAIN_EXT - padded_sl_num = str(slide_number).zfill(3) - - training_img_path = get_training_image_path(slide_number) - large_w, large_h, small_w, small_h = parse_dimensions_from_image_filename(training_img_path) - img_filename = TRAIN_PREFIX + padded_sl_num + "-" + str(SCALE_FACTOR) + "x-" + str(large_w) + "x" + str( - large_h) + "-" + str(small_w) + "x" + str(small_h) + "-" + TOP_TILES_SUFFIX + "." + ext - - return img_filename - - -def get_top_tiles_image_path(slide_number): - """ - Convert slide number to a path to a top tiles image file. - - Example: - 5 -> ../data/top_tiles_png/TUPAC-TR-005-32x-49920x108288-1560x3384-top_tiles.png - - Args: - slide_number: The slide number. - - Returns: - Path to the top tiles image file. - """ - if not os.path.exists(TOP_TILES_DIR): - os.makedirs(TOP_TILES_DIR) - img_path = os.path.join(TOP_TILES_DIR, get_top_tiles_image_filename(slide_number)) - return img_path - - -def get_top_tiles_thumbnail_path(slide_number): - """ - Convert slide number to a path to a tile summary thumbnail file. - - Example: - 5 -> ../data/top_tiles_thumbnail_jpg/TUPAC-TR-005-32x-49920x108288-1560x3384-top_tiles.jpg - Args: - slide_number: The slide number. - - Returns: - Path to the top tiles thumbnail file. - """ - if not os.path.exists(TOP_TILES_THUMBNAIL_DIR): - os.makedirs(TOP_TILES_THUMBNAIL_DIR) - img_path = os.path.join(TOP_TILES_THUMBNAIL_DIR, get_top_tiles_image_filename(slide_number, thumbnail=True)) - return img_path - - -def get_tile_data_filename(slide_number): - """ - Convert slide number to a tile data file name. - - Example: - 5 -> TUPAC-TR-005-32x-49920x108288-1560x3384-tile_data.csv - - Args: - slide_number: The slide number. - - Returns: - The tile data file name. - """ - padded_sl_num = str(slide_number).zfill(3) - - training_img_path = get_training_image_path(slide_number) - large_w, large_h, small_w, small_h = parse_dimensions_from_image_filename(training_img_path) - data_filename = TRAIN_PREFIX + padded_sl_num + "-" + str(SCALE_FACTOR) + "x-" + str(large_w) + "x" + str( - large_h) + "-" + str(small_w) + "x" + str(small_h) + "-" + TILE_DATA_SUFFIX + ".csv" - - return data_filename - - -def get_tile_data_path(slide_number): - """ - Convert slide number to a path to a tile data file. - - Example: - 5 -> ../data/tile_data/TUPAC-TR-005-32x-49920x108288-1560x3384-tile_data.csv - - Args: - slide_number: The slide number. - - Returns: - Path to the tile data file. - """ - if not os.path.exists(TILE_DATA_DIR): - os.makedirs(TILE_DATA_DIR) - file_path = os.path.join(TILE_DATA_DIR, get_tile_data_filename(slide_number)) - return file_path - - -def get_filter_image_result(slide_number): - """ - Convert slide number to the path to the file that is the final result of filtering. - - Example: - 5 -> ../data/filter_png/TUPAC-TR-005-32x-49920x108288-1560x3384-filtered.png - - Args: - slide_number: The slide number. - - Returns: - Path to the filter image file. - """ - padded_sl_num = str(slide_number).zfill(3) - training_img_path = get_training_image_path(slide_number) - large_w, large_h, small_w, small_h = parse_dimensions_from_image_filename(training_img_path) - img_path = os.path.join(FILTER_DIR, TRAIN_PREFIX + padded_sl_num + "-" + str( - SCALE_FACTOR) + "x-" + FILTER_SUFFIX + str(large_w) + "x" + str(large_h) + "-" + str(small_w) + "x" + str( - small_h) + "-" + FILTER_RESULT_TEXT + "." + DEST_TRAIN_EXT) - return img_path - - -def get_filter_thumbnail_result(slide_number): - """ - Convert slide number to the path to the file that is the final thumbnail result of filtering. - - Example: - 5 -> ../data/filter_thumbnail_jpg/TUPAC-TR-005-32x-49920x108288-1560x3384-filtered.jpg - - Args: - slide_number: The slide number. - - Returns: - Path to the filter thumbnail file. - """ - padded_sl_num = str(slide_number).zfill(3) - training_img_path = get_training_image_path(slide_number) - large_w, large_h, small_w, small_h = parse_dimensions_from_image_filename(training_img_path) - img_path = os.path.join(FILTER_THUMBNAIL_DIR, TRAIN_PREFIX + padded_sl_num + "-" + str( - SCALE_FACTOR) + "x-" + FILTER_SUFFIX + str(large_w) + "x" + str(large_h) + "-" + str(small_w) + "x" + str( - small_h) + "-" + FILTER_RESULT_TEXT + "." + THUMBNAIL_EXT) - return img_path - - -def parse_dimensions_from_image_filename(filename): - """ - Parse an image filename to extract the original width and height and the converted width and height. - - Example: - "TUPAC-TR-011-32x-97103x79079-3034x2471-tile_summary.png" -> (97103, 79079, 3034, 2471) - - Args: - filename: The image filename. - - Returns: - Tuple consisting of the original width, original height, the converted width, and the converted height. - """ - m = re.match(".*-([\d]*)x([\d]*)-([\d]*)x([\d]*).*\..*", filename) - large_w = int(m.group(1)) - large_h = int(m.group(2)) - small_w = int(m.group(3)) - small_h = int(m.group(4)) - return large_w, large_h, small_w, small_h - - -def small_to_large_mapping(small_pixel, large_dimensions): - """ - Map a scaled-down pixel width and height to the corresponding pixel of the original whole-slide image. - - Args: - small_pixel: The scaled-down width and height. - large_dimensions: The width and height of the original whole-slide image. - - Returns: - Tuple consisting of the scaled-up width and height. - """ - small_x, small_y = small_pixel - large_w, large_h = large_dimensions - large_x = round((large_w / SCALE_FACTOR) / math.floor(large_w / SCALE_FACTOR) * (SCALE_FACTOR * small_x)) - large_y = round((large_h / SCALE_FACTOR) / math.floor(large_h / SCALE_FACTOR) * (SCALE_FACTOR * small_y)) - return large_x, large_y - - -def training_slide_to_image(slide_number): - """ - Convert a WSI training slide to a saved scaled-down image in a format such as jpg or png. - - Args: - slide_number: The slide number. - """ - - img, large_w, large_h, new_w, new_h = slide_to_scaled_pil_image(slide_number) - - img_path = get_training_image_path(slide_number, large_w, large_h, new_w, new_h) - print("Saving image to: " + img_path) - if not os.path.exists(DEST_TRAIN_DIR): - os.makedirs(DEST_TRAIN_DIR) - img.save(img_path) - - thumbnail_path = get_training_thumbnail_path(slide_number, large_w, large_h, new_w, new_h) - save_thumbnail(img, THUMBNAIL_SIZE, thumbnail_path) - - -def slide_to_scaled_pil_image(slide_number): - """ - Convert a WSI training slide to a scaled-down PIL image. - - Args: - slide_number: The slide number. - - Returns: - Tuple consisting of scaled-down PIL image, original width, original height, new width, and new height. - """ - slide_filepath = get_training_slide_path(slide_number) - print("Opening Slide #%d: %s" % (slide_number, slide_filepath)) - slide = open_slide(slide_filepath) - - large_w, large_h = slide.dimensions - new_w = math.floor(large_w / SCALE_FACTOR) - new_h = math.floor(large_h / SCALE_FACTOR) - level = slide.get_best_level_for_downsample(SCALE_FACTOR) - whole_slide_image = slide.read_region((0, 0), level, slide.level_dimensions[level]) - whole_slide_image = whole_slide_image.convert("RGB") - img = whole_slide_image.resize((new_w, new_h), PIL.Image.BILINEAR) - return img, large_w, large_h, new_w, new_h - - -def slide_to_scaled_np_image(slide_number): - """ - Convert a WSI training slide to a scaled-down NumPy image. - - Args: - slide_number: The slide number. - - Returns: - Tuple consisting of scaled-down NumPy image, original width, original height, new width, and new height. - """ - pil_img, large_w, large_h, new_w, new_h = slide_to_scaled_pil_image(slide_number) - np_img = util.pil_to_np_rgb(pil_img) - return np_img, large_w, large_h, new_w, new_h - - -def show_slide(slide_number): - """ - Display a WSI slide on the screen, where the slide has been scaled down and converted to a PIL image. - - Args: - slide_number: The slide number. - """ - pil_img = slide_to_scaled_pil_image(slide_number)[0] - pil_img.show() - - -def save_thumbnail(pil_img, size, path, display_path=False): - """ - Save a thumbnail of a PIL image, specifying the maximum width or height of the thumbnail. - - Args: - pil_img: The PIL image to save as a thumbnail. - size: The maximum width or height of the thumbnail. - path: The path to the thumbnail. - display_path: If True, display thumbnail path in console. - """ - max_size = tuple(round(size * d / max(pil_img.size)) for d in pil_img.size) - img = pil_img.resize(max_size, PIL.Image.BILINEAR) - if display_path: - print("Saving thumbnail to: " + path) - dir = os.path.dirname(path) - if dir != '' and not os.path.exists(dir): - os.makedirs(dir) - img.save(path) - - -def get_num_training_slides(): - """ - Obtain the total number of WSI training slide images. - - Returns: - The total number of WSI training slide images. - """ - num_training_slides = len(glob.glob1(SRC_TRAIN_DIR, "*." + SRC_TRAIN_EXT)) - return num_training_slides - - -def training_slide_range_to_images(start_ind, end_ind): - """ - Convert a range of WSI training slides to smaller images (in a format such as jpg or png). - - Args: - start_ind: Starting index (inclusive). - end_ind: Ending index (inclusive). - - Returns: - The starting index and the ending index of the slides that were converted. - """ - for slide_num in range(start_ind, end_ind + 1): - training_slide_to_image(slide_num) - return (start_ind, end_ind) - - -def singleprocess_training_slides_to_images(): - """ - Convert all WSI training slides to smaller images using a single process. - """ - t = Time() - - num_train_images = get_num_training_slides() - training_slide_range_to_images(1, num_train_images) - - t.elapsed_display() - - -def multiprocess_training_slides_to_images(): - """ - Convert all WSI training slides to smaller images using multiple processes (one process per core). - Each process will process a range of slide numbers. - """ - timer = Time() - - # how many processes to use - num_processes = multiprocessing.cpu_count() - pool = multiprocessing.Pool(num_processes) - - num_train_images = get_num_training_slides() - if num_processes > num_train_images: - num_processes = num_train_images - images_per_process = num_train_images / num_processes - - print("Number of processes: " + str(num_processes)) - print("Number of training images: " + str(num_train_images)) - - # each task specifies a range of slides - tasks = [] - for num_process in range(1, num_processes + 1): - start_index = (num_process - 1) * images_per_process + 1 - end_index = num_process * images_per_process - start_index = int(start_index) - end_index = int(end_index) - tasks.append((start_index, end_index)) - if start_index == end_index: - print("Task #" + str(num_process) + ": Process slide " + str(start_index)) - else: - print("Task #" + str(num_process) + ": Process slides " + str(start_index) + " to " + str(end_index)) - - # start tasks - results = [] - for t in tasks: - results.append(pool.apply_async(training_slide_range_to_images, t)) - - for result in results: - (start_ind, end_ind) = result.get() - if start_ind == end_ind: - print("Done converting slide %d" % start_ind) - else: - print("Done converting slides %d through %d" % (start_ind, end_ind)) - - timer.elapsed_display() - - -def slide_stats(): - """ - Display statistics/graphs about training slides. - """ - t = Time() - - if not os.path.exists(STATS_DIR): - os.makedirs(STATS_DIR) - - num_train_images = get_num_training_slides() - slide_stats = [] - for slide_num in range(1, num_train_images + 1): - slide_filepath = get_training_slide_path(slide_num) - print("Opening Slide #%d: %s" % (slide_num, slide_filepath)) - slide = open_slide(slide_filepath) - (width, height) = slide.dimensions - print(" Dimensions: {:,d} x {:,d}".format(width, height)) - slide_stats.append((width, height)) - - max_width = 0 - max_height = 0 - min_width = sys.maxsize - min_height = sys.maxsize - total_width = 0 - total_height = 0 - total_size = 0 - which_max_width = 0 - which_max_height = 0 - which_min_width = 0 - which_min_height = 0 - max_size = 0 - min_size = sys.maxsize - which_max_size = 0 - which_min_size = 0 - for z in range(0, num_train_images): - (width, height) = slide_stats[z] - if width > max_width: - max_width = width - which_max_width = z + 1 - if width < min_width: - min_width = width - which_min_width = z + 1 - if height > max_height: - max_height = height - which_max_height = z + 1 - if height < min_height: - min_height = height - which_min_height = z + 1 - size = width * height - if size > max_size: - max_size = size - which_max_size = z + 1 - if size < min_size: - min_size = size - which_min_size = z + 1 - total_width = total_width + width - total_height = total_height + height - total_size = total_size + size - - avg_width = total_width / num_train_images - avg_height = total_height / num_train_images - avg_size = total_size / num_train_images - - stats_string = "" - stats_string += "%-11s {:14,d} pixels (slide #%d)".format(max_width) % ("Max width:", which_max_width) - stats_string += "\n%-11s {:14,d} pixels (slide #%d)".format(max_height) % ("Max height:", which_max_height) - stats_string += "\n%-11s {:14,d} pixels (slide #%d)".format(max_size) % ("Max size:", which_max_size) - stats_string += "\n%-11s {:14,d} pixels (slide #%d)".format(min_width) % ("Min width:", which_min_width) - stats_string += "\n%-11s {:14,d} pixels (slide #%d)".format(min_height) % ("Min height:", which_min_height) - stats_string += "\n%-11s {:14,d} pixels (slide #%d)".format(min_size) % ("Min size:", which_min_size) - stats_string += "\n%-11s {:14,d} pixels".format(round(avg_width)) % "Avg width:" - stats_string += "\n%-11s {:14,d} pixels".format(round(avg_height)) % "Avg height:" - stats_string += "\n%-11s {:14,d} pixels".format(round(avg_size)) % "Avg size:" - stats_string += "\n" - print(stats_string) - - stats_string += "\nslide number,width,height" - for i in range(0, len(slide_stats)): - (width, height) = slide_stats[i] - stats_string += "\n%d,%d,%d" % (i + 1, width, height) - stats_string += "\n" - - stats_file = open(os.path.join(STATS_DIR, "stats.txt"), "w") - stats_file.write(stats_string) - stats_file.close() - - t.elapsed_display() - - x, y = zip(*slide_stats) - colors = np.random.rand(num_train_images) - sizes = [10 for n in range(num_train_images)] - plt.scatter(x, y, s=sizes, c=colors, alpha=0.7) - plt.xlabel("width (pixels)") - plt.ylabel("height (pixels)") - plt.title("SVS Image Sizes") - plt.set_cmap("prism") - plt.tight_layout() - plt.savefig(os.path.join(STATS_DIR, "svs-image-sizes.png")) - plt.show() - - plt.clf() - plt.scatter(x, y, s=sizes, c=colors, alpha=0.7) - plt.xlabel("width (pixels)") - plt.ylabel("height (pixels)") - plt.title("SVS Image Sizes (Labeled with slide numbers)") - plt.set_cmap("prism") - for i in range(num_train_images): - snum = i + 1 - plt.annotate(str(snum), (x[i], y[i])) - plt.tight_layout() - plt.savefig(os.path.join(STATS_DIR, "svs-image-sizes-slide-numbers.png")) - plt.show() - - plt.clf() - area = [w * h / 1000000 for (w, h) in slide_stats] - plt.hist(area, bins=64) - plt.xlabel("width x height (M of pixels)") - plt.ylabel("# images") - plt.title("Distribution of image sizes in millions of pixels") - plt.tight_layout() - plt.savefig(os.path.join(STATS_DIR, "distribution-of-svs-image-sizes.png")) - plt.show() - - plt.clf() - whratio = [w / h for (w, h) in slide_stats] - plt.hist(whratio, bins=64) - plt.xlabel("width to height ratio") - plt.ylabel("# images") - plt.title("Image shapes (width to height)") - plt.tight_layout() - plt.savefig(os.path.join(STATS_DIR, "w-to-h.png")) - plt.show() - - plt.clf() - hwratio = [h / w for (w, h) in slide_stats] - plt.hist(hwratio, bins=64) - plt.xlabel("height to width ratio") - plt.ylabel("# images") - plt.title("Image shapes (height to width)") - plt.tight_layout() - plt.savefig(os.path.join(STATS_DIR, "h-to-w.png")) - plt.show() - - -def slide_info(display_all_properties=False): - """ - Display information (such as properties) about training images. - - Args: - display_all_properties: If True, display all available slide properties. - """ - t = Time() - - num_train_images = get_num_training_slides() - obj_pow_20_list = [] - obj_pow_40_list = [] - obj_pow_other_list = [] - for slide_num in range(1, num_train_images + 1): - slide_filepath = get_training_slide_path(slide_num) - print("\nOpening Slide #%d: %s" % (slide_num, slide_filepath)) - slide = open_slide(slide_filepath) - print("Level count: %d" % slide.level_count) - print("Level dimensions: " + str(slide.level_dimensions)) - print("Level downsamples: " + str(slide.level_downsamples)) - print("Dimensions: " + str(slide.dimensions)) - objective_power = int(slide.properties[openslide.PROPERTY_NAME_OBJECTIVE_POWER]) - print("Objective power: " + str(objective_power)) - if objective_power == 20: - obj_pow_20_list.append(slide_num) - elif objective_power == 40: - obj_pow_40_list.append(slide_num) - else: - obj_pow_other_list.append(slide_num) - print("Associated images:") - for ai_key in slide.associated_images.keys(): - print(" " + str(ai_key) + ": " + str(slide.associated_images.get(ai_key))) - print("Format: " + str(slide.detect_format(slide_filepath))) - if display_all_properties: - print("Properties:") - for prop_key in slide.properties.keys(): - print(" Property: " + str(prop_key) + ", value: " + str(slide.properties.get(prop_key))) - - print("\n\nSlide Magnifications:") - print(" 20x Slides: " + str(obj_pow_20_list)) - print(" 40x Slides: " + str(obj_pow_40_list)) - print(" ??x Slides: " + str(obj_pow_other_list) + "\n") - - t.elapsed_display() - - -# if __name__ == "__main__": - # show_slide(2) - # slide_info(display_all_properties=True) - # slide_stats() - - # training_slide_to_image(4) - # img_path = get_training_image_path(4) - # img = open_image(img_path) - # img.show() - - # slide_to_scaled_pil_image(5)[0].show() - # singleprocess_training_slides_to_images() - # multiprocess_training_slides_to_images() diff --git a/deephistopath/wsi/tiles.py b/deephistopath/wsi/tiles.py deleted file mode 100644 index 887aedd..0000000 --- a/deephistopath/wsi/tiles.py +++ /dev/null @@ -1,1959 +0,0 @@ -# ------------------------------------------------------------------------ -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# -# ------------------------------------------------------------------------ - -# To get around renderer issue on macOS going from Matplotlib image to NumPy image. -import matplotlib - -matplotlib.use('Agg') - -import colorsys -import math -import matplotlib.pyplot as plt -import multiprocessing -import numpy as np -import os -from PIL import Image, ImageDraw, ImageFont -from enum import Enum -from deephistopath.wsi import util -from deephistopath.wsi import filter -from deephistopath.wsi import slide -from deephistopath.wsi.util import Time - -TISSUE_HIGH_THRESH = 80 -TISSUE_LOW_THRESH = 10 - -ROW_TILE_SIZE = 1024 -COL_TILE_SIZE = 1024 -NUM_TOP_TILES = 50 - -DISPLAY_TILE_SUMMARY_LABELS = False -TILE_LABEL_TEXT_SIZE = 10 -LABEL_ALL_TILES_IN_TOP_TILE_SUMMARY = False -BORDER_ALL_TILES_IN_TOP_TILE_SUMMARY = False - -TILE_BORDER_SIZE = 2 # The size of the colored rectangular border around summary tiles. - -HIGH_COLOR = (0, 255, 0) -MEDIUM_COLOR = (255, 255, 0) -LOW_COLOR = (255, 165, 0) -NONE_COLOR = (255, 0, 0) - -FADED_THRESH_COLOR = (128, 255, 128) -FADED_MEDIUM_COLOR = (255, 255, 128) -FADED_LOW_COLOR = (255, 210, 128) -FADED_NONE_COLOR = (255, 128, 128) - -FONT_PATH = "/Library/Fonts/Arial Bold.ttf" -SUMMARY_TITLE_FONT_PATH = "/Library/Fonts/Courier New Bold.ttf" -SUMMARY_TITLE_TEXT_COLOR = (0, 0, 0) -SUMMARY_TITLE_TEXT_SIZE = 24 -SUMMARY_TILE_TEXT_COLOR = (255, 255, 255) -TILE_TEXT_COLOR = (0, 0, 0) -TILE_TEXT_SIZE = 36 -TILE_TEXT_BACKGROUND_COLOR = (255, 255, 255) -TILE_TEXT_W_BORDER = 5 -TILE_TEXT_H_BORDER = 4 - -HSV_PURPLE = 270 -HSV_PINK = 330 - - -def get_num_tiles(rows, cols, row_tile_size, col_tile_size): - """ - Obtain the number of vertical and horizontal tiles that an image can be divided into given a row tile size and - a column tile size. - - Args: - rows: Number of rows. - cols: Number of columns. - row_tile_size: Number of pixels in a tile row. - col_tile_size: Number of pixels in a tile column. - - Returns: - Tuple consisting of the number of vertical tiles and the number of horizontal tiles that the image can be divided - into given the row tile size and the column tile size. - """ - num_row_tiles = math.ceil(rows / row_tile_size) - num_col_tiles = math.ceil(cols / col_tile_size) - return num_row_tiles, num_col_tiles - - -def get_tile_indices(rows, cols, row_tile_size, col_tile_size): - """ - Obtain a list of tile coordinates (starting row, ending row, starting column, ending column, row number, column number). - - Args: - rows: Number of rows. - cols: Number of columns. - row_tile_size: Number of pixels in a tile row. - col_tile_size: Number of pixels in a tile column. - - Returns: - List of tuples representing tile coordinates consisting of starting row, ending row, - starting column, ending column, row number, column number. - """ - indices = list() - num_row_tiles, num_col_tiles = get_num_tiles(rows, cols, row_tile_size, col_tile_size) - for r in range(0, num_row_tiles): - start_r = r * row_tile_size - end_r = ((r + 1) * row_tile_size) if (r < num_row_tiles - 1) else rows - for c in range(0, num_col_tiles): - start_c = c * col_tile_size - end_c = ((c + 1) * col_tile_size) if (c < num_col_tiles - 1) else cols - indices.append((start_r, end_r, start_c, end_c, r + 1, c + 1)) - return indices - - -def create_summary_pil_img(np_img, title_area_height, row_tile_size, col_tile_size, num_row_tiles, num_col_tiles): - """ - Create a PIL summary image including top title area and right side and bottom padding. - - Args: - np_img: Image as a NumPy array. - title_area_height: Height of the title area at the top of the summary image. - row_tile_size: The tile size in rows. - col_tile_size: The tile size in columns. - num_row_tiles: The number of row tiles. - num_col_tiles: The number of column tiles. - - Returns: - Summary image as a PIL image. This image contains the image data specified by the np_img input and also has - potentially a top title area and right side and bottom padding. - """ - r = row_tile_size * num_row_tiles + title_area_height - c = col_tile_size * num_col_tiles - summary_img = np.zeros([r, c, np_img.shape[2]], dtype=np.uint8) - # add gray edges so that tile text does not get cut off - summary_img.fill(120) - # color title area white - summary_img[0:title_area_height, 0:summary_img.shape[1]].fill(255) - summary_img[title_area_height:np_img.shape[0] + title_area_height, 0:np_img.shape[1]] = np_img - summary = util.np_to_pil(summary_img) - return summary - - -def generate_tile_summaries(tile_sum, np_img, display=True, save_summary=False): - """ - Generate summary images/thumbnails showing a 'heatmap' representation of the tissue segmentation of all tiles. - - Args: - tile_sum: TileSummary object. - np_img: Image as a NumPy array. - display: If True, display tile summary to screen. - save_summary: If True, save tile summary images. - """ - z = 300 # height of area at top of summary slide - slide_num = tile_sum.slide_num - rows = tile_sum.scaled_h - cols = tile_sum.scaled_w - row_tile_size = tile_sum.scaled_tile_h - col_tile_size = tile_sum.scaled_tile_w - num_row_tiles, num_col_tiles = get_num_tiles(rows, cols, row_tile_size, col_tile_size) - summary = create_summary_pil_img(np_img, z, row_tile_size, col_tile_size, num_row_tiles, num_col_tiles) - draw = ImageDraw.Draw(summary) - - original_img_path = slide.get_training_image_path(slide_num) - np_orig = slide.open_image_np(original_img_path) - summary_orig = create_summary_pil_img(np_orig, z, row_tile_size, col_tile_size, num_row_tiles, num_col_tiles) - draw_orig = ImageDraw.Draw(summary_orig) - - for t in tile_sum.tiles: - border_color = tile_border_color(t.tissue_percentage) - tile_border(draw, t.r_s + z, t.r_e + z, t.c_s, t.c_e, border_color) - tile_border(draw_orig, t.r_s + z, t.r_e + z, t.c_s, t.c_e, border_color) - - summary_txt = summary_title(tile_sum) + "\n" + summary_stats(tile_sum) - - summary_font = ImageFont.truetype(SUMMARY_TITLE_FONT_PATH, size=SUMMARY_TITLE_TEXT_SIZE) - draw.text((5, 5), summary_txt, SUMMARY_TITLE_TEXT_COLOR, font=summary_font) - draw_orig.text((5, 5), summary_txt, SUMMARY_TITLE_TEXT_COLOR, font=summary_font) - - if DISPLAY_TILE_SUMMARY_LABELS: - count = 0 - for t in tile_sum.tiles: - count += 1 - label = "R%d\nC%d" % (t.r, t.c) - font = ImageFont.truetype(FONT_PATH, size=TILE_LABEL_TEXT_SIZE) - # drop shadow behind text - draw.text(((t.c_s + 3), (t.r_s + 3 + z)), label, (0, 0, 0), font=font) - draw_orig.text(((t.c_s + 3), (t.r_s + 3 + z)), label, (0, 0, 0), font=font) - - draw.text(((t.c_s + 2), (t.r_s + 2 + z)), label, SUMMARY_TILE_TEXT_COLOR, font=font) - draw_orig.text(((t.c_s + 2), (t.r_s + 2 + z)), label, SUMMARY_TILE_TEXT_COLOR, font=font) - - if display: - summary.show() - summary_orig.show() - if save_summary: - save_tile_summary_image(summary, slide_num) - save_tile_summary_on_original_image(summary_orig, slide_num) - - -def generate_top_tile_summaries(tile_sum, np_img, display=True, save_summary=False, show_top_stats=True, - label_all_tiles=LABEL_ALL_TILES_IN_TOP_TILE_SUMMARY, - border_all_tiles=BORDER_ALL_TILES_IN_TOP_TILE_SUMMARY): - """ - Generate summary images/thumbnails showing the top tiles ranked by score. - - Args: - tile_sum: TileSummary object. - np_img: Image as a NumPy array. - display: If True, display top tiles to screen. - save_summary: If True, save top tiles images. - show_top_stats: If True, append top tile score stats to image. - label_all_tiles: If True, label all tiles. If False, label only top tiles. - """ - z = 300 # height of area at top of summary slide - slide_num = tile_sum.slide_num - rows = tile_sum.scaled_h - cols = tile_sum.scaled_w - row_tile_size = tile_sum.scaled_tile_h - col_tile_size = tile_sum.scaled_tile_w - num_row_tiles, num_col_tiles = get_num_tiles(rows, cols, row_tile_size, col_tile_size) - summary = create_summary_pil_img(np_img, z, row_tile_size, col_tile_size, num_row_tiles, num_col_tiles) - draw = ImageDraw.Draw(summary) - - original_img_path = slide.get_training_image_path(slide_num) - np_orig = slide.open_image_np(original_img_path) - summary_orig = create_summary_pil_img(np_orig, z, row_tile_size, col_tile_size, num_row_tiles, num_col_tiles) - draw_orig = ImageDraw.Draw(summary_orig) - - if border_all_tiles: - for t in tile_sum.tiles: - border_color = faded_tile_border_color(t.tissue_percentage) - tile_border(draw, t.r_s + z, t.r_e + z, t.c_s, t.c_e, border_color, border_size=1) - tile_border(draw_orig, t.r_s + z, t.r_e + z, t.c_s, t.c_e, border_color, border_size=1) - - tbs = TILE_BORDER_SIZE - top_tiles = tile_sum.top_tiles() - for t in top_tiles: - border_color = tile_border_color(t.tissue_percentage) - tile_border(draw, t.r_s + z, t.r_e + z, t.c_s, t.c_e, border_color) - tile_border(draw_orig, t.r_s + z, t.r_e + z, t.c_s, t.c_e, border_color) - if border_all_tiles: - tile_border(draw, t.r_s + z + tbs, t.r_e + z - tbs, t.c_s + tbs, t.c_e - tbs, (0, 0, 0)) - tile_border(draw_orig, t.r_s + z + tbs, t.r_e + z - tbs, t.c_s + tbs, t.c_e - tbs, (0, 0, 0)) - - summary_title = "Slide %03d Top Tile Summary:" % slide_num - summary_txt = summary_title + "\n" + summary_stats(tile_sum) - - summary_font = ImageFont.truetype(SUMMARY_TITLE_FONT_PATH, size=SUMMARY_TITLE_TEXT_SIZE) - draw.text((5, 5), summary_txt, SUMMARY_TITLE_TEXT_COLOR, font=summary_font) - draw_orig.text((5, 5), summary_txt, SUMMARY_TITLE_TEXT_COLOR, font=summary_font) - - tiles_to_label = tile_sum.tiles if label_all_tiles else top_tiles - h_offset = TILE_BORDER_SIZE + 2 - v_offset = TILE_BORDER_SIZE - h_ds_offset = TILE_BORDER_SIZE + 3 - v_ds_offset = TILE_BORDER_SIZE + 1 - for t in tiles_to_label: - label = "R%d\nC%d" % (t.r, t.c) - font = ImageFont.truetype(FONT_PATH, size=TILE_LABEL_TEXT_SIZE) - # drop shadow behind text - draw.text(((t.c_s + h_ds_offset), (t.r_s + v_ds_offset + z)), label, (0, 0, 0), font=font) - draw_orig.text(((t.c_s + h_ds_offset), (t.r_s + v_ds_offset + z)), label, (0, 0, 0), font=font) - - draw.text(((t.c_s + h_offset), (t.r_s + v_offset + z)), label, SUMMARY_TILE_TEXT_COLOR, font=font) - draw_orig.text(((t.c_s + h_offset), (t.r_s + v_offset + z)), label, SUMMARY_TILE_TEXT_COLOR, font=font) - - if show_top_stats: - summary = add_tile_stats_to_top_tile_summary(summary, top_tiles, z) - summary_orig = add_tile_stats_to_top_tile_summary(summary_orig, top_tiles, z) - - if display: - summary.show() - summary_orig.show() - if save_summary: - save_top_tiles_image(summary, slide_num) - save_top_tiles_on_original_image(summary_orig, slide_num) - - -def add_tile_stats_to_top_tile_summary(pil_img, tiles, z): - np_sum = util.pil_to_np_rgb(pil_img) - sum_r, sum_c, sum_ch = np_sum.shape - np_stats = np_tile_stat_img(tiles) - st_r, st_c, _ = np_stats.shape - combo_c = sum_c + st_c - combo_r = max(sum_r, st_r + z) - combo = np.zeros([combo_r, combo_c, sum_ch], dtype=np.uint8) - combo.fill(255) - combo[0:sum_r, 0:sum_c] = np_sum - combo[z:st_r + z, sum_c:sum_c + st_c] = np_stats - result = util.np_to_pil(combo) - return result - - -def np_tile_stat_img(tiles): - """ - Generate tile scoring statistics for a list of tiles and return the result as a NumPy array image. - - Args: - tiles: List of tiles (such as top tiles) - - Returns: - Tile scoring statistics converted into an NumPy array image. - """ - tt = sorted(tiles, key=lambda t: (t.r, t.c), reverse=False) - tile_stats = "Tile Score Statistics:\n" - count = 0 - for t in tt: - if count > 0: - tile_stats += "\n" - count += 1 - tup = (t.r, t.c, t.rank, t.tissue_percentage, t.color_factor, t.s_and_v_factor, t.quantity_factor, t.score) - tile_stats += "R%03d C%03d #%003d TP:%6.2f%% CF:%4.0f SVF:%4.2f QF:%4.2f S:%0.4f" % tup - np_stats = np_text(tile_stats, font_path=SUMMARY_TITLE_FONT_PATH, font_size=14) - return np_stats - - -def tile_border_color(tissue_percentage): - """ - Obtain the corresponding tile border color for a particular tile tissue percentage. - - Args: - tissue_percentage: The tile tissue percentage - - Returns: - The tile border color corresponding to the tile tissue percentage. - """ - if tissue_percentage >= TISSUE_HIGH_THRESH: - border_color = HIGH_COLOR - elif (tissue_percentage >= TISSUE_LOW_THRESH) and (tissue_percentage < TISSUE_HIGH_THRESH): - border_color = MEDIUM_COLOR - elif (tissue_percentage > 0) and (tissue_percentage < TISSUE_LOW_THRESH): - border_color = LOW_COLOR - else: - border_color = NONE_COLOR - return border_color - - -def faded_tile_border_color(tissue_percentage): - """ - Obtain the corresponding faded tile border color for a particular tile tissue percentage. - - Args: - tissue_percentage: The tile tissue percentage - - Returns: - The faded tile border color corresponding to the tile tissue percentage. - """ - if tissue_percentage >= TISSUE_HIGH_THRESH: - border_color = FADED_THRESH_COLOR - elif (tissue_percentage >= TISSUE_LOW_THRESH) and (tissue_percentage < TISSUE_HIGH_THRESH): - border_color = FADED_MEDIUM_COLOR - elif (tissue_percentage > 0) and (tissue_percentage < TISSUE_LOW_THRESH): - border_color = FADED_LOW_COLOR - else: - border_color = FADED_NONE_COLOR - return border_color - - -def summary_title(tile_summary): - """ - Obtain tile summary title. - - Args: - tile_summary: TileSummary object. - - Returns: - The tile summary title. - """ - return "Slide %03d Tile Summary:" % tile_summary.slide_num - - -def summary_stats(tile_summary): - """ - Obtain various stats about the slide tiles. - - Args: - tile_summary: TileSummary object. - - Returns: - Various stats about the slide tiles as a string. - """ - return "Original Dimensions: %dx%d\n" % (tile_summary.orig_w, tile_summary.orig_h) + \ - "Original Tile Size: %dx%d\n" % (tile_summary.orig_tile_w, tile_summary.orig_tile_h) + \ - "Scale Factor: 1/%dx\n" % tile_summary.scale_factor + \ - "Scaled Dimensions: %dx%d\n" % (tile_summary.scaled_w, tile_summary.scaled_h) + \ - "Scaled Tile Size: %dx%d\n" % (tile_summary.scaled_tile_w, tile_summary.scaled_tile_w) + \ - "Total Mask: %3.2f%%, Total Tissue: %3.2f%%\n" % ( - tile_summary.mask_percentage(), tile_summary.tissue_percentage) + \ - "Tiles: %dx%d = %d\n" % (tile_summary.num_col_tiles, tile_summary.num_row_tiles, tile_summary.count) + \ - " %5d (%5.2f%%) tiles >=%d%% tissue\n" % ( - tile_summary.high, tile_summary.high / tile_summary.count * 100, TISSUE_HIGH_THRESH) + \ - " %5d (%5.2f%%) tiles >=%d%% and <%d%% tissue\n" % ( - tile_summary.medium, tile_summary.medium / tile_summary.count * 100, TISSUE_LOW_THRESH, - TISSUE_HIGH_THRESH) + \ - " %5d (%5.2f%%) tiles >0%% and <%d%% tissue\n" % ( - tile_summary.low, tile_summary.low / tile_summary.count * 100, TISSUE_LOW_THRESH) + \ - " %5d (%5.2f%%) tiles =0%% tissue" % (tile_summary.none, tile_summary.none / tile_summary.count * 100) - - -def tile_border(draw, r_s, r_e, c_s, c_e, color, border_size=TILE_BORDER_SIZE): - """ - Draw a border around a tile with width TILE_BORDER_SIZE. - - Args: - draw: Draw object for drawing on PIL image. - r_s: Row starting pixel. - r_e: Row ending pixel. - c_s: Column starting pixel. - c_e: Column ending pixel. - color: Color of the border. - border_size: Width of tile border in pixels. - """ - for x in range(0, border_size): - draw.rectangle([(c_s + x, r_s + x), (c_e - 1 - x, r_e - 1 - x)], outline=color) - - -def save_tile_summary_image(pil_img, slide_num): - """ - Save a tile summary image and thumbnail to the file system. - - Args: - pil_img: Image as a PIL Image. - slide_num: The slide number. - """ - t = Time() - filepath = slide.get_tile_summary_image_path(slide_num) - pil_img.save(filepath) - print("%-20s | Time: %-14s Name: %s" % ("Save Tile Sum", str(t.elapsed()), filepath)) - - t = Time() - thumbnail_filepath = slide.get_tile_summary_thumbnail_path(slide_num) - slide.save_thumbnail(pil_img, slide.THUMBNAIL_SIZE, thumbnail_filepath) - print("%-20s | Time: %-14s Name: %s" % ("Save Tile Sum Thumb", str(t.elapsed()), thumbnail_filepath)) - - -def save_top_tiles_image(pil_img, slide_num): - """ - Save a top tiles image and thumbnail to the file system. - - Args: - pil_img: Image as a PIL Image. - slide_num: The slide number. - """ - t = Time() - filepath = slide.get_top_tiles_image_path(slide_num) - pil_img.save(filepath) - print("%-20s | Time: %-14s Name: %s" % ("Save Top Tiles Image", str(t.elapsed()), filepath)) - - t = Time() - thumbnail_filepath = slide.get_top_tiles_thumbnail_path(slide_num) - slide.save_thumbnail(pil_img, slide.THUMBNAIL_SIZE, thumbnail_filepath) - print("%-20s | Time: %-14s Name: %s" % ("Save Top Tiles Thumb", str(t.elapsed()), thumbnail_filepath)) - - -def save_tile_summary_on_original_image(pil_img, slide_num): - """ - Save a tile summary on original image and thumbnail to the file system. - - Args: - pil_img: Image as a PIL Image. - slide_num: The slide number. - """ - t = Time() - filepath = slide.get_tile_summary_on_original_image_path(slide_num) - pil_img.save(filepath) - print("%-20s | Time: %-14s Name: %s" % ("Save Tile Sum Orig", str(t.elapsed()), filepath)) - - t = Time() - thumbnail_filepath = slide.get_tile_summary_on_original_thumbnail_path(slide_num) - slide.save_thumbnail(pil_img, slide.THUMBNAIL_SIZE, thumbnail_filepath) - print( - "%-20s | Time: %-14s Name: %s" % ("Save Tile Sum Orig T", str(t.elapsed()), thumbnail_filepath)) - - -def save_top_tiles_on_original_image(pil_img, slide_num): - """ - Save a top tiles on original image and thumbnail to the file system. - - Args: - pil_img: Image as a PIL Image. - slide_num: The slide number. - """ - t = Time() - filepath = slide.get_top_tiles_on_original_image_path(slide_num) - pil_img.save(filepath) - print("%-20s | Time: %-14s Name: %s" % ("Save Top Orig", str(t.elapsed()), filepath)) - - t = Time() - thumbnail_filepath = slide.get_top_tiles_on_original_thumbnail_path(slide_num) - slide.save_thumbnail(pil_img, slide.THUMBNAIL_SIZE, thumbnail_filepath) - print( - "%-20s | Time: %-14s Name: %s" % ("Save Top Orig Thumb", str(t.elapsed()), thumbnail_filepath)) - - -def summary_and_tiles(slide_num, display=True, save_summary=False, save_data=True, save_top_tiles=True): - """ - Generate tile summary and top tiles for slide. - - Args: - slide_num: The slide number. - display: If True, display tile summary to screen. - save_summary: If True, save tile summary images. - save_data: If True, save tile data to csv file. - save_top_tiles: If True, save top tiles to files. - - """ - img_path = slide.get_filter_image_result(slide_num) - np_img = slide.open_image_np(img_path) - - tile_sum = score_tiles(slide_num, np_img) - if save_data: - save_tile_data(tile_sum) - generate_tile_summaries(tile_sum, np_img, display=display, save_summary=save_summary) - generate_top_tile_summaries(tile_sum, np_img, display=display, save_summary=save_summary) - if save_top_tiles: - for tile in tile_sum.top_tiles(): - tile.save_tile() - return tile_sum - - -def save_tile_data(tile_summary): - """ - Save tile data to csv file. - - Args - tile_summary: TimeSummary object. - """ - - time = Time() - - csv = summary_title(tile_summary) + "\n" + summary_stats(tile_summary) - - csv += "\n\n\nTile Num,Row,Column,Tissue %,Tissue Quantity,Col Start,Row Start,Col End,Row End,Col Size,Row Size," + \ - "Original Col Start,Original Row Start,Original Col End,Original Row End,Original Col Size,Original Row Size," + \ - "Color Factor,S and V Factor,Quantity Factor,Score\n" - - for t in tile_summary.tiles: - line = "%d,%d,%d,%4.2f,%s,%d,%d,%d,%d,%d,%d,%d,%d,%d,%d,%d,%d,%4.0f,%4.2f,%4.2f,%0.4f\n" % ( - t.tile_num, t.r, t.c, t.tissue_percentage, t.tissue_quantity().name, t.c_s, t.r_s, t.c_e, t.r_e, t.c_e - t.c_s, - t.r_e - t.r_s, t.o_c_s, t.o_r_s, t.o_c_e, t.o_r_e, t.o_c_e - t.o_c_s, t.o_r_e - t.o_r_s, t.color_factor, - t.s_and_v_factor, t.quantity_factor, t.score) - csv += line - - data_path = slide.get_tile_data_path(tile_summary.slide_num) - csv_file = open(data_path, "w") - csv_file.write(csv) - csv_file.close() - - print("%-20s | Time: %-14s Name: %s" % ("Save Tile Data", str(time.elapsed()), data_path)) - - -def tile_to_pil_tile(tile): - """ - Convert tile information into the corresponding tile as a PIL image read from the whole-slide image file. - - Args: - tile: Tile object. - - Return: - Tile as a PIL image. - """ - t = tile - slide_filepath = slide.get_training_slide_path(t.slide_num) - s = slide.open_slide(slide_filepath) - - x, y = t.o_c_s, t.o_r_s - w, h = t.o_c_e - t.o_c_s, t.o_r_e - t.o_r_s - tile_region = s.read_region((x, y), 0, (w, h)) - # RGBA to RGB - pil_img = tile_region.convert("RGB") - return pil_img - - -def tile_to_np_tile(tile): - """ - Convert tile information into the corresponding tile as a NumPy image read from the whole-slide image file. - - Args: - tile: Tile object. - - Return: - Tile as a NumPy image. - """ - pil_img = tile_to_pil_tile(tile) - np_img = util.pil_to_np_rgb(pil_img) - return np_img - - -def save_display_tile(tile, save=True, display=False): - """ - Save and/or display a tile image. - - Args: - tile: Tile object. - save: If True, save tile image. - display: If True, dispaly tile image. - """ - tile_pil_img = tile_to_pil_tile(tile) - - if save: - t = Time() - img_path = slide.get_tile_image_path(tile) - dir = os.path.dirname(img_path) - if not os.path.exists(dir): - os.makedirs(dir) - tile_pil_img.save(img_path) - print("%-20s | Time: %-14s Name: %s" % ("Save Tile", str(t.elapsed()), img_path)) - - if display: - tile_pil_img.show() - - -def score_tiles(slide_num, np_img=None, dimensions=None, small_tile_in_tile=False): - """ - Score all tiles for a slide and return the results in a TileSummary object. - - Args: - slide_num: The slide number. - np_img: Optional image as a NumPy array. - dimensions: Optional tuple consisting of (original width, original height, new width, new height). Used for dynamic - tile retrieval. - small_tile_in_tile: If True, include the small NumPy image in the Tile objects. - - Returns: - TileSummary object which includes a list of Tile objects containing information about each tile. - """ - if dimensions is None: - img_path = slide.get_filter_image_result(slide_num) - o_w, o_h, w, h = slide.parse_dimensions_from_image_filename(img_path) - else: - o_w, o_h, w, h = dimensions - - if np_img is None: - np_img = slide.open_image_np(img_path) - - row_tile_size = round(ROW_TILE_SIZE / slide.SCALE_FACTOR) # use round? - col_tile_size = round(COL_TILE_SIZE / slide.SCALE_FACTOR) # use round? - - num_row_tiles, num_col_tiles = get_num_tiles(h, w, row_tile_size, col_tile_size) - - tile_sum = TileSummary(slide_num=slide_num, - orig_w=o_w, - orig_h=o_h, - orig_tile_w=COL_TILE_SIZE, - orig_tile_h=ROW_TILE_SIZE, - scaled_w=w, - scaled_h=h, - scaled_tile_w=col_tile_size, - scaled_tile_h=row_tile_size, - tissue_percentage=filter.tissue_percent(np_img), - num_col_tiles=num_col_tiles, - num_row_tiles=num_row_tiles) - - count = 0 - high = 0 - medium = 0 - low = 0 - none = 0 - tile_indices = get_tile_indices(h, w, row_tile_size, col_tile_size) - for t in tile_indices: - count += 1 # tile_num - r_s, r_e, c_s, c_e, r, c = t - np_tile = np_img[r_s:r_e, c_s:c_e] - t_p = filter.tissue_percent(np_tile) - amount = tissue_quantity(t_p) - if amount == TissueQuantity.HIGH: - high += 1 - elif amount == TissueQuantity.MEDIUM: - medium += 1 - elif amount == TissueQuantity.LOW: - low += 1 - elif amount == TissueQuantity.NONE: - none += 1 - o_c_s, o_r_s = slide.small_to_large_mapping((c_s, r_s), (o_w, o_h)) - o_c_e, o_r_e = slide.small_to_large_mapping((c_e, r_e), (o_w, o_h)) - - # pixel adjustment in case tile dimension too large (for example, 1025 instead of 1024) - if (o_c_e - o_c_s) > COL_TILE_SIZE: - o_c_e -= 1 - if (o_r_e - o_r_s) > ROW_TILE_SIZE: - o_r_e -= 1 - - score, color_factor, s_and_v_factor, quantity_factor = score_tile(np_tile, t_p, slide_num, r, c) - - np_scaled_tile = np_tile if small_tile_in_tile else None - tile = Tile(tile_sum, slide_num, np_scaled_tile, count, r, c, r_s, r_e, c_s, c_e, o_r_s, o_r_e, o_c_s, - o_c_e, t_p, color_factor, s_and_v_factor, quantity_factor, score) - tile_sum.tiles.append(tile) - - tile_sum.count = count - tile_sum.high = high - tile_sum.medium = medium - tile_sum.low = low - tile_sum.none = none - - tiles_by_score = tile_sum.tiles_by_score() - rank = 0 - for t in tiles_by_score: - rank += 1 - t.rank = rank - - return tile_sum - - -def score_tile(np_tile, tissue_percent, slide_num, row, col): - """ - Score tile based on tissue percentage, color factor, saturation/value factor, and tissue quantity factor. - - Args: - np_tile: Tile as NumPy array. - tissue_percent: The percentage of the tile judged to be tissue. - slide_num: Slide number. - row: Tile row. - col: Tile column. - - Returns tuple consisting of score, color factor, saturation/value factor, and tissue quantity factor. - """ - color_factor = hsv_purple_pink_factor(np_tile) - s_and_v_factor = hsv_saturation_and_value_factor(np_tile) - amount = tissue_quantity(tissue_percent) - quantity_factor = tissue_quantity_factor(amount) - combined_factor = color_factor * s_and_v_factor * quantity_factor - score = (tissue_percent ** 2) * np.log(1 + combined_factor) / 1000.0 - # scale score to between 0 and 1 - score = 1.0 - (10.0 / (10.0 + score)) - return score, color_factor, s_and_v_factor, quantity_factor - - -def tissue_quantity_factor(amount): - """ - Obtain a scoring factor based on the quantity of tissue in a tile. - - Args: - amount: Tissue amount as a TissueQuantity enum value. - - Returns: - Scoring factor based on the tile tissue quantity. - """ - if amount == TissueQuantity.HIGH: - quantity_factor = 1.0 - elif amount == TissueQuantity.MEDIUM: - quantity_factor = 0.2 - elif amount == TissueQuantity.LOW: - quantity_factor = 0.1 - else: - quantity_factor = 0.0 - return quantity_factor - - -def tissue_quantity(tissue_percentage): - """ - Obtain TissueQuantity enum member (HIGH, MEDIUM, LOW, or NONE) for corresponding tissue percentage. - - Args: - tissue_percentage: The tile tissue percentage. - - Returns: - TissueQuantity enum member (HIGH, MEDIUM, LOW, or NONE). - """ - if tissue_percentage >= TISSUE_HIGH_THRESH: - return TissueQuantity.HIGH - elif (tissue_percentage >= TISSUE_LOW_THRESH) and (tissue_percentage < TISSUE_HIGH_THRESH): - return TissueQuantity.MEDIUM - elif (tissue_percentage > 0) and (tissue_percentage < TISSUE_LOW_THRESH): - return TissueQuantity.LOW - else: - return TissueQuantity.NONE - - -def image_list_to_tiles(image_num_list, display=False, save_summary=True, save_data=True, save_top_tiles=True): - """ - Generate tile summaries and tiles for a list of images. - - Args: - image_num_list: List of image numbers. - display: If True, display tile summary images to screen. - save_summary: If True, save tile summary images. - save_data: If True, save tile data to csv file. - save_top_tiles: If True, save top tiles to files. - """ - tile_summaries_dict = dict() - for slide_num in image_num_list: - tile_summary = summary_and_tiles(slide_num, display, save_summary, save_data, save_top_tiles) - tile_summaries_dict[slide_num] = tile_summary - return image_num_list, tile_summaries_dict - - -def image_range_to_tiles(start_ind, end_ind, display=False, save_summary=True, save_data=True, save_top_tiles=True): - """ - Generate tile summaries and tiles for a range of images. - - Args: - start_ind: Starting index (inclusive). - end_ind: Ending index (inclusive). - display: If True, display tile summary images to screen. - save_summary: If True, save tile summary images. - save_data: If True, save tile data to csv file. - save_top_tiles: If True, save top tiles to files. - """ - image_num_list = list() - tile_summaries_dict = dict() - for slide_num in range(start_ind, end_ind + 1): - tile_summary = summary_and_tiles(slide_num, display, save_summary, save_data, save_top_tiles) - image_num_list.append(slide_num) - tile_summaries_dict[slide_num] = tile_summary - return image_num_list, tile_summaries_dict - - -def singleprocess_filtered_images_to_tiles(display=False, save_summary=True, save_data=True, save_top_tiles=True, - html=True, image_num_list=None): - """ - Generate tile summaries and tiles for training images using a single process. - - Args: - display: If True, display tile summary images to screen. - save_summary: If True, save tile summary images. - save_data: If True, save tile data to csv file. - save_top_tiles: If True, save top tiles to files. - html: If True, generate HTML page to display tiled images - image_num_list: Optionally specify a list of image slide numbers. - """ - t = Time() - print("Generating tile summaries\n") - - if image_num_list is not None: - image_num_list, tile_summaries_dict = image_list_to_tiles(image_num_list, display, save_summary, save_data, - save_top_tiles) - else: - num_training_slides = slide.get_num_training_slides() - image_num_list, tile_summaries_dict = image_range_to_tiles(1, num_training_slides, display, save_summary, save_data, - save_top_tiles) - - print("Time to generate tile summaries: %s\n" % str(t.elapsed())) - - if html: - generate_tiled_html_result(image_num_list, tile_summaries_dict, save_data) - - -def multiprocess_filtered_images_to_tiles(display=False, save_summary=True, save_data=True, save_top_tiles=True, - html=True, image_num_list=None): - """ - Generate tile summaries and tiles for all training images using multiple processes (one process per core). - - Args: - display: If True, display images to screen (multiprocessed display not recommended). - save_summary: If True, save tile summary images. - save_data: If True, save tile data to csv file. - save_top_tiles: If True, save top tiles to files. - html: If True, generate HTML page to display tiled images. - image_num_list: Optionally specify a list of image slide numbers. - """ - timer = Time() - print("Generating tile summaries (multiprocess)\n") - - if save_summary and not os.path.exists(slide.TILE_SUMMARY_DIR): - os.makedirs(slide.TILE_SUMMARY_DIR) - - # how many processes to use - num_processes = multiprocessing.cpu_count() - pool = multiprocessing.Pool(num_processes) - - if image_num_list is not None: - num_train_images = len(image_num_list) - else: - num_train_images = slide.get_num_training_slides() - if num_processes > num_train_images: - num_processes = num_train_images - images_per_process = num_train_images / num_processes - - print("Number of processes: " + str(num_processes)) - print("Number of training images: " + str(num_train_images)) - - tasks = [] - for num_process in range(1, num_processes + 1): - start_index = (num_process - 1) * images_per_process + 1 - end_index = num_process * images_per_process - start_index = int(start_index) - end_index = int(end_index) - if image_num_list is not None: - sublist = image_num_list[start_index - 1:end_index] - tasks.append((sublist, display, save_summary, save_data, save_top_tiles)) - print("Task #" + str(num_process) + ": Process slides " + str(sublist)) - else: - tasks.append((start_index, end_index, display, save_summary, save_data, save_top_tiles)) - if start_index == end_index: - print("Task #" + str(num_process) + ": Process slide " + str(start_index)) - else: - print("Task #" + str(num_process) + ": Process slides " + str(start_index) + " to " + str(end_index)) - - # start tasks - results = [] - for t in tasks: - if image_num_list is not None: - results.append(pool.apply_async(image_list_to_tiles, t)) - else: - results.append(pool.apply_async(image_range_to_tiles, t)) - - slide_nums = list() - tile_summaries_dict = dict() - for result in results: - image_nums, tile_summaries = result.get() - slide_nums.extend(image_nums) - tile_summaries_dict.update(tile_summaries) - print("Done tiling slides: %s" % image_nums) - - if html: - generate_tiled_html_result(slide_nums, tile_summaries_dict, save_data) - - print("Time to generate tile previews (multiprocess): %s\n" % str(timer.elapsed())) - - -def image_row(slide_num, tile_summary, data_link): - """ - Generate HTML for viewing a tiled image. - - Args: - slide_num: The slide number. - tile_summary: TileSummary object. - data_link: If True, add link to tile data csv file. - - Returns: - HTML table row for viewing a tiled image. - """ - orig_img = slide.get_training_image_path(slide_num) - orig_thumb = slide.get_training_thumbnail_path(slide_num) - filt_img = slide.get_filter_image_result(slide_num) - filt_thumb = slide.get_filter_thumbnail_result(slide_num) - sum_img = slide.get_tile_summary_image_path(slide_num) - sum_thumb = slide.get_tile_summary_thumbnail_path(slide_num) - osum_img = slide.get_tile_summary_on_original_image_path(slide_num) - osum_thumb = slide.get_tile_summary_on_original_thumbnail_path(slide_num) - top_img = slide.get_top_tiles_image_path(slide_num) - top_thumb = slide.get_top_tiles_thumbnail_path(slide_num) - otop_img = slide.get_top_tiles_on_original_image_path(slide_num) - otop_thumb = slide.get_top_tiles_on_original_thumbnail_path(slide_num) - html = " \n" + \ - " \n" + \ - " S%03d Original
\n" % (orig_img, slide_num) + \ - " \n" % (orig_thumb) + \ - " \n" + \ - " \n" + \ - " \n" + \ - " S%03d Filtered
\n" % (filt_img, slide_num) + \ - " \n" % (filt_thumb) + \ - " \n" + \ - " \n" - - html += " \n" + \ - " S%03d Tiles
\n" % (sum_img, slide_num) + \ - " \n" % (sum_thumb) + \ - " \n" + \ - " \n" - - html += " \n" + \ - " S%03d Tiles
\n" % (osum_img, slide_num) + \ - " \n" % (osum_thumb) + \ - " \n" + \ - " \n" - - html += " \n" - if data_link: - html += "
S%03d Tile Summary\n" % slide_num + \ - " (Data)
\n" % slide.get_tile_data_path(slide_num) - else: - html += "
S%03d Tile Summary
\n" % slide_num - - html += "
\n" + \ - " %s\n" % summary_stats(tile_summary).replace("\n", "
\n ") + \ - "
\n" + \ - " \n" - - html += " \n" + \ - " S%03d Top Tiles
\n" % (top_img, slide_num) + \ - " \n" % (top_thumb) + \ - " \n" + \ - " \n" - - html += " \n" + \ - " S%03d Top Tiles
\n" % (otop_img, slide_num) + \ - " \n" % (otop_thumb) + \ - " \n" + \ - " \n" - - top_tiles = tile_summary.top_tiles() - num_tiles = len(top_tiles) - score_num = 0 - for t in top_tiles: - score_num += 1 - t.tile_num = score_num - # sort top tiles by rows and columns to make them easier to locate on HTML page - top_tiles = sorted(top_tiles, key=lambda t: (t.r, t.c), reverse=False) - - html += " \n" + \ - "
S%03d Top %d Tile Scores
\n" % (slide_num, num_tiles) + \ - "
\n" - - html += " \n" - MAX_TILES_PER_ROW = 15 - num_cols = math.ceil(num_tiles / MAX_TILES_PER_ROW) - num_rows = num_tiles if num_tiles < MAX_TILES_PER_ROW else MAX_TILES_PER_ROW - for row in range(num_rows): - html += " \n" - for col in range(num_cols): - html += " \n" - html += " \n" - html += "
" - tile_num = row + (col * num_rows) + 1 - if tile_num <= num_tiles: - t = top_tiles[tile_num - 1] - label = "R%03d C%03d %0.4f (#%02d)" % (t.r, t.c, t.score, t.tile_num) - tile_img_path = slide.get_tile_image_path(t) - html += "%s" % (tile_img_path, label) - else: - html += " " - html += "
\n" - - html += "
\n" - html += " \n" - - html += " \n" - return html - - -def generate_tiled_html_result(slide_nums, tile_summaries_dict, data_link): - """ - Generate HTML to view the tiled images. - - Args: - slide_nums: List of slide numbers. - tile_summaries_dict: Dictionary of TileSummary objects keyed by slide number. - data_link: If True, add link to tile data csv file. - """ - slide_nums = sorted(slide_nums) - if not slide.TILE_SUMMARY_PAGINATE: - html = "" - html += filter.html_header("Tiles") - - html += " \n" - for slide_num in slide_nums: - html += image_row(slide_num, data_link) - html += "
\n" - - html += filter.html_footer() - text_file = open(os.path.join(slide.TILE_SUMMARY_HTML_DIR, "tiles.html"), "w") - text_file.write(html) - text_file.close() - else: - total_len = len(slide_nums) - page_size = slide.TILE_SUMMARY_PAGINATION_SIZE - num_pages = math.ceil(total_len / page_size) - for page_num in range(1, num_pages + 1): - start_index = (page_num - 1) * page_size - end_index = (page_num * page_size) if (page_num < num_pages) else total_len - page_slide_nums = slide_nums[start_index:end_index] - - html = "" - html += filter.html_header("Tiles, Page %d" % page_num) - - html += "
" - if page_num > 1: - if page_num == 2: - html += "< " - else: - html += "< " % (page_num - 1) - html += "Page %d" % page_num - if page_num < num_pages: - html += " > " % (page_num + 1) - html += "
\n" - - html += " \n" - for slide_num in page_slide_nums: - tile_summary = tile_summaries_dict[slide_num] - html += image_row(slide_num, tile_summary, data_link) - html += "
\n" - - html += filter.html_footer() - if page_num == 1: - text_file = open(os.path.join(slide.TILE_SUMMARY_HTML_DIR, "tiles.html"), "w") - else: - text_file = open(os.path.join(slide.TILE_SUMMARY_HTML_DIR, "tiles-%d.html" % page_num), "w") - text_file.write(html) - text_file.close() - - -def np_hsv_hue_histogram(h): - """ - Create Matplotlib histogram of hue values for an HSV image and return the histogram as a NumPy array image. - - Args: - h: Hue values as a 1-dimensional int NumPy array (scaled 0 to 360) - - Returns: - Matplotlib histogram of hue values converted to a NumPy array image. - """ - figure = plt.figure() - canvas = figure.canvas - _, _, patches = plt.hist(h, bins=360) - plt.title("HSV Hue Histogram, mean=%3.1f, std=%3.1f" % (np.mean(h), np.std(h))) - - bin_num = 0 - for patch in patches: - rgb_color = colorsys.hsv_to_rgb(bin_num / 360.0, 1, 1) - patch.set_facecolor(rgb_color) - bin_num += 1 - - canvas.draw() - w, h = canvas.get_width_height() - np_hist = np.fromstring(canvas.get_renderer().tostring_rgb(), dtype=np.uint8).reshape(h, w, 3) - plt.close(figure) - util.np_info(np_hist) - return np_hist - - -def np_histogram(data, title, bins="auto"): - """ - Create Matplotlib histogram and return it as a NumPy array image. - - Args: - data: Data to plot in the histogram. - title: Title of the histogram. - bins: Number of histogram bins, "auto" by default. - - Returns: - Matplotlib histogram as a NumPy array image. - """ - figure = plt.figure() - canvas = figure.canvas - plt.hist(data, bins=bins) - plt.title(title) - - canvas.draw() - w, h = canvas.get_width_height() - np_hist = np.fromstring(canvas.get_renderer().tostring_rgb(), dtype=np.uint8).reshape(h, w, 3) - plt.close(figure) - util.np_info(np_hist) - return np_hist - - -def np_hsv_saturation_histogram(s): - """ - Create Matplotlib histogram of saturation values for an HSV image and return the histogram as a NumPy array image. - - Args: - s: Saturation values as a 1-dimensional float NumPy array - - Returns: - Matplotlib histogram of saturation values converted to a NumPy array image. - """ - title = "HSV Saturation Histogram, mean=%.2f, std=%.2f" % (np.mean(s), np.std(s)) - return np_histogram(s, title) - - -def np_hsv_value_histogram(v): - """ - Create Matplotlib histogram of value values for an HSV image and return the histogram as a NumPy array image. - - Args: - v: Value values as a 1-dimensional float NumPy array - - Returns: - Matplotlib histogram of saturation values converted to a NumPy array image. - """ - title = "HSV Value Histogram, mean=%.2f, std=%.2f" % (np.mean(v), np.std(v)) - return np_histogram(v, title) - - -def np_rgb_channel_histogram(rgb, ch_num, ch_name): - """ - Create Matplotlib histogram of an RGB channel for an RGB image and return the histogram as a NumPy array image. - - Args: - rgb: Image as RGB NumPy array. - ch_num: Which channel (0=red, 1=green, 2=blue) - ch_name: Channel name ("R", "G", "B") - - Returns: - Matplotlib histogram of RGB channel converted to a NumPy array image. - """ - - ch = rgb[:, :, ch_num] - ch = ch.flatten() - title = "RGB %s Histogram, mean=%.2f, std=%.2f" % (ch_name, np.mean(ch), np.std(ch)) - return np_histogram(ch, title, bins=256) - - -def np_rgb_r_histogram(rgb): - """ - Obtain RGB R channel histogram as a NumPy array image. - - Args: - rgb: Image as RGB NumPy array. - - Returns: - Histogram of RGB R channel as a NumPy array image. - """ - hist = np_rgb_channel_histogram(rgb, 0, "R") - return hist - - -def np_rgb_g_histogram(rgb): - """ - Obtain RGB G channel histogram as a NumPy array image. - - Args: - rgb: Image as RGB NumPy array. - - Returns: - Histogram of RGB G channel as a NumPy array image. - """ - hist = np_rgb_channel_histogram(rgb, 1, "G") - return hist - - -def np_rgb_b_histogram(rgb): - """ - Obtain RGB B channel histogram as a NumPy array image. - - Args: - rgb: Image as RGB NumPy array. - - Returns: - Histogram of RGB B channel as a NumPy array image. - """ - hist = np_rgb_channel_histogram(rgb, 2, "B") - return hist - - -def pil_hue_histogram(h): - """ - Create Matplotlib histogram of hue values for an HSV image and return the histogram as a PIL image. - - Args: - h: Hue values as a 1-dimensional int NumPy array (scaled 0 to 360) - - Returns: - Matplotlib histogram of hue values converted to a PIL image. - """ - np_hist = np_hsv_hue_histogram(h) - pil_hist = util.np_to_pil(np_hist) - return pil_hist - - -def display_image_with_hsv_hue_histogram(np_rgb, text=None, scale_up=False): - """ - Display an image with its corresponding hue histogram. - - Args: - np_rgb: RGB image tile as a NumPy array - text: Optional text to display above image - scale_up: If True, scale up image to display by slide.SCALE_FACTOR - """ - hsv = filter.filter_rgb_to_hsv(np_rgb) - h = filter.filter_hsv_to_h(hsv) - np_hist = np_hsv_hue_histogram(h) - hist_r, hist_c, _ = np_hist.shape - - if scale_up: - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=1) - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=0) - - img_r, img_c, img_ch = np_rgb.shape - if text is not None: - np_t = np_text(text) - t_r, t_c, _ = np_t.shape - t_i_c = max(t_c, img_c) - t_i_r = t_r + img_r - t_i = np.zeros([t_i_r, t_i_c, img_ch], dtype=np.uint8) - t_i.fill(255) - t_i[0:t_r, 0:t_c] = np_t - t_i[t_r:t_r + img_r, 0:img_c] = np_rgb - np_rgb = t_i # for simplicity assign title+image to image - img_r, img_c, img_ch = np_rgb.shape - - r = max(img_r, hist_r) - c = img_c + hist_c - combo = np.zeros([r, c, img_ch], dtype=np.uint8) - combo.fill(255) - combo[0:img_r, 0:img_c] = np_rgb - combo[0:hist_r, img_c:c] = np_hist - pil_combo = util.np_to_pil(combo) - pil_combo.show() - - -def display_image(np_rgb, text=None, scale_up=False): - """ - Display an image with optional text above image. - - Args: - np_rgb: RGB image tile as a NumPy array - text: Optional text to display above image - scale_up: If True, scale up image to display by slide.SCALE_FACTOR - """ - if scale_up: - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=1) - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=0) - - img_r, img_c, img_ch = np_rgb.shape - if text is not None: - np_t = np_text(text) - t_r, t_c, _ = np_t.shape - t_i_c = max(t_c, img_c) - t_i_r = t_r + img_r - t_i = np.zeros([t_i_r, t_i_c, img_ch], dtype=np.uint8) - t_i.fill(255) - t_i[0:t_r, 0:t_c] = np_t - t_i[t_r:t_r + img_r, 0:img_c] = np_rgb - np_rgb = t_i - - pil_img = util.np_to_pil(np_rgb) - pil_img.show() - - -def display_image_with_hsv_histograms(np_rgb, text=None, scale_up=False): - """ - Display an image with its corresponding HSV hue, saturation, and value histograms. - - Args: - np_rgb: RGB image tile as a NumPy array - text: Optional text to display above image - scale_up: If True, scale up image to display by slide.SCALE_FACTOR - """ - hsv = filter.filter_rgb_to_hsv(np_rgb) - np_h = np_hsv_hue_histogram(filter.filter_hsv_to_h(hsv)) - np_s = np_hsv_saturation_histogram(filter.filter_hsv_to_s(hsv)) - np_v = np_hsv_value_histogram(filter.filter_hsv_to_v(hsv)) - h_r, h_c, _ = np_h.shape - s_r, s_c, _ = np_s.shape - v_r, v_c, _ = np_v.shape - - if scale_up: - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=1) - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=0) - - img_r, img_c, img_ch = np_rgb.shape - if text is not None: - np_t = np_text(text) - t_r, t_c, _ = np_t.shape - t_i_c = max(t_c, img_c) - t_i_r = t_r + img_r - t_i = np.zeros([t_i_r, t_i_c, img_ch], dtype=np.uint8) - t_i.fill(255) - t_i[0:t_r, 0:t_c] = np_t - t_i[t_r:t_r + img_r, 0:img_c] = np_rgb - np_rgb = t_i # for simplicity assign title+image to image - img_r, img_c, img_ch = np_rgb.shape - - hists_c = max(h_c, s_c, v_c) - hists_r = h_r + s_r + v_r - hists = np.zeros([hists_r, hists_c, img_ch], dtype=np.uint8) - - hists[0:h_r, 0:h_c] = np_h - hists[h_r:h_r + s_r, 0:s_c] = np_s - hists[h_r + s_r:h_r + s_r + v_r, 0:v_c] = np_v - - r = max(img_r, hists_r) - c = img_c + hists_c - combo = np.zeros([r, c, img_ch], dtype=np.uint8) - combo.fill(255) - combo[0:img_r, 0:img_c] = np_rgb - combo[0:hists_r, img_c:c] = hists - pil_combo = util.np_to_pil(combo) - pil_combo.show() - - -def display_image_with_rgb_histograms(np_rgb, text=None, scale_up=False): - """ - Display an image with its corresponding RGB histograms. - - Args: - np_rgb: RGB image tile as a NumPy array - text: Optional text to display above image - scale_up: If True, scale up image to display by slide.SCALE_FACTOR - """ - np_r = np_rgb_r_histogram(np_rgb) - np_g = np_rgb_g_histogram(np_rgb) - np_b = np_rgb_b_histogram(np_rgb) - r_r, r_c, _ = np_r.shape - g_r, g_c, _ = np_g.shape - b_r, b_c, _ = np_b.shape - - if scale_up: - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=1) - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=0) - - img_r, img_c, img_ch = np_rgb.shape - if text is not None: - np_t = np_text(text) - t_r, t_c, _ = np_t.shape - t_i_c = max(t_c, img_c) - t_i_r = t_r + img_r - t_i = np.zeros([t_i_r, t_i_c, img_ch], dtype=np.uint8) - t_i.fill(255) - t_i[0:t_r, 0:t_c] = np_t - t_i[t_r:t_r + img_r, 0:img_c] = np_rgb - np_rgb = t_i # for simplicity assign title+image to image - img_r, img_c, img_ch = np_rgb.shape - - hists_c = max(r_c, g_c, b_c) - hists_r = r_r + g_r + b_r - hists = np.zeros([hists_r, hists_c, img_ch], dtype=np.uint8) - - hists[0:r_r, 0:r_c] = np_r - hists[r_r:r_r + g_r, 0:g_c] = np_g - hists[r_r + g_r:r_r + g_r + b_r, 0:b_c] = np_b - - r = max(img_r, hists_r) - c = img_c + hists_c - combo = np.zeros([r, c, img_ch], dtype=np.uint8) - combo.fill(255) - combo[0:img_r, 0:img_c] = np_rgb - combo[0:hists_r, img_c:c] = hists - pil_combo = util.np_to_pil(combo) - pil_combo.show() - - -def pil_text(text, w_border=TILE_TEXT_W_BORDER, h_border=TILE_TEXT_H_BORDER, font_path=FONT_PATH, - font_size=TILE_TEXT_SIZE, text_color=TILE_TEXT_COLOR, background=TILE_TEXT_BACKGROUND_COLOR): - """ - Obtain a PIL image representation of text. - - Args: - text: The text to convert to an image. - w_border: Tile text width border (left and right). - h_border: Tile text height border (top and bottom). - font_path: Path to font. - font_size: Size of font. - text_color: Tile text color. - background: Tile background color. - - Returns: - PIL image representing the text. - """ - - font = ImageFont.truetype(font_path, font_size) - x, y = ImageDraw.Draw(Image.new("RGB", (1, 1), background)).textsize(text, font) - image = Image.new("RGB", (x + 2 * w_border, y + 2 * h_border), background) - draw = ImageDraw.Draw(image) - draw.text((w_border, h_border), text, text_color, font=font) - return image - - -def np_text(text, w_border=TILE_TEXT_W_BORDER, h_border=TILE_TEXT_H_BORDER, font_path=FONT_PATH, - font_size=TILE_TEXT_SIZE, text_color=TILE_TEXT_COLOR, background=TILE_TEXT_BACKGROUND_COLOR): - """ - Obtain a NumPy array image representation of text. - - Args: - text: The text to convert to an image. - w_border: Tile text width border (left and right). - h_border: Tile text height border (top and bottom). - font_path: Path to font. - font_size: Size of font. - text_color: Tile text color. - background: Tile background color. - - Returns: - NumPy array representing the text. - """ - pil_img = pil_text(text, w_border, h_border, font_path, font_size, - text_color, background) - np_img = util.pil_to_np_rgb(pil_img) - return np_img - - -def display_tile(tile, rgb_histograms=True, hsv_histograms=True): - """ - Display a tile with its corresponding RGB and HSV histograms. - - Args: - tile: The Tile object. - rgb_histograms: If True, display RGB histograms. - hsv_histograms: If True, display HSV histograms. - """ - - text = "S%03d R%03d C%03d\n" % (tile.slide_num, tile.r, tile.c) - text += "Score:%4.2f Tissue:%5.2f%% CF:%2.0f SVF:%4.2f QF:%4.2f\n" % ( - tile.score, tile.tissue_percentage, tile.color_factor, tile.s_and_v_factor, tile.quantity_factor) - text += "Rank #%d of %d" % (tile.rank, tile.tile_summary.num_tiles()) - - np_scaled_tile = tile.get_np_scaled_tile() - if np_scaled_tile is not None: - small_text = text + "\n \nSmall Tile (%d x %d)" % (np_scaled_tile.shape[1], np_scaled_tile.shape[0]) - if rgb_histograms and hsv_histograms: - display_image_with_rgb_and_hsv_histograms(np_scaled_tile, small_text, scale_up=True) - elif rgb_histograms: - display_image_with_rgb_histograms(np_scaled_tile, small_text, scale_up=True) - elif hsv_histograms: - display_image_with_hsv_histograms(np_scaled_tile, small_text, scale_up=True) - else: - display_image(np_scaled_tile, small_text, scale_up=True) - - np_tile = tile.get_np_tile() - text += " based on small tile\n \nLarge Tile (%d x %d)" % (np_tile.shape[1], np_tile.shape[0]) - if rgb_histograms and hsv_histograms: - display_image_with_rgb_and_hsv_histograms(np_tile, text) - elif rgb_histograms: - display_image_with_rgb_histograms(np_tile, text) - elif hsv_histograms: - display_image_with_hsv_histograms(np_tile, text) - else: - display_image(np_tile, text) - - -def display_image_with_rgb_and_hsv_histograms(np_rgb, text=None, scale_up=False): - """ - Display a tile with its corresponding RGB and HSV histograms. - - Args: - np_rgb: RGB image tile as a NumPy array - text: Optional text to display above image - scale_up: If True, scale up image to display by slide.SCALE_FACTOR - """ - hsv = filter.filter_rgb_to_hsv(np_rgb) - np_r = np_rgb_r_histogram(np_rgb) - np_g = np_rgb_g_histogram(np_rgb) - np_b = np_rgb_b_histogram(np_rgb) - np_h = np_hsv_hue_histogram(filter.filter_hsv_to_h(hsv)) - np_s = np_hsv_saturation_histogram(filter.filter_hsv_to_s(hsv)) - np_v = np_hsv_value_histogram(filter.filter_hsv_to_v(hsv)) - - r_r, r_c, _ = np_r.shape - g_r, g_c, _ = np_g.shape - b_r, b_c, _ = np_b.shape - h_r, h_c, _ = np_h.shape - s_r, s_c, _ = np_s.shape - v_r, v_c, _ = np_v.shape - - if scale_up: - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=1) - np_rgb = np.repeat(np_rgb, slide.SCALE_FACTOR, axis=0) - - img_r, img_c, img_ch = np_rgb.shape - if text is not None: - np_t = np_text(text) - t_r, t_c, _ = np_t.shape - t_i_c = max(t_c, img_c) - t_i_r = t_r + img_r - t_i = np.zeros([t_i_r, t_i_c, img_ch], dtype=np.uint8) - t_i.fill(255) - t_i[0:t_r, 0:t_c] = np_t - t_i[t_r:t_r + img_r, 0:img_c] = np_rgb - np_rgb = t_i # for simplicity assign title+image to image - img_r, img_c, img_ch = np_rgb.shape - - rgb_hists_c = max(r_c, g_c, b_c) - rgb_hists_r = r_r + g_r + b_r - rgb_hists = np.zeros([rgb_hists_r, rgb_hists_c, img_ch], dtype=np.uint8) - rgb_hists[0:r_r, 0:r_c] = np_r - rgb_hists[r_r:r_r + g_r, 0:g_c] = np_g - rgb_hists[r_r + g_r:r_r + g_r + b_r, 0:b_c] = np_b - - hsv_hists_c = max(h_c, s_c, v_c) - hsv_hists_r = h_r + s_r + v_r - hsv_hists = np.zeros([hsv_hists_r, hsv_hists_c, img_ch], dtype=np.uint8) - hsv_hists[0:h_r, 0:h_c] = np_h - hsv_hists[h_r:h_r + s_r, 0:s_c] = np_s - hsv_hists[h_r + s_r:h_r + s_r + v_r, 0:v_c] = np_v - - r = max(img_r, rgb_hists_r, hsv_hists_r) - c = img_c + rgb_hists_c + hsv_hists_c - combo = np.zeros([r, c, img_ch], dtype=np.uint8) - combo.fill(255) - combo[0:img_r, 0:img_c] = np_rgb - combo[0:rgb_hists_r, img_c:img_c + rgb_hists_c] = rgb_hists - combo[0:hsv_hists_r, img_c + rgb_hists_c:c] = hsv_hists - pil_combo = util.np_to_pil(combo) - pil_combo.show() - - -def rgb_to_hues(rgb): - """ - Convert RGB NumPy array to 1-dimensional array of hue values (HSV H values in degrees). - - Args: - rgb: RGB image as a NumPy array - - Returns: - 1-dimensional array of hue values in degrees - """ - hsv = filter.filter_rgb_to_hsv(rgb, display_np_info=False) - h = filter.filter_hsv_to_h(hsv, display_np_info=False) - return h - - -def hsv_saturation_and_value_factor(rgb): - """ - Function to reduce scores of tiles with narrow HSV saturations and values since saturation and value standard - deviations should be relatively broad if the tile contains significant tissue. - - Example of a blurred tile that should not be ranked as a top tile: - ../data/tiles_png/006/TUPAC-TR-006-tile-r58-c3-x2048-y58369-w1024-h1024.png - - Args: - rgb: RGB image as a NumPy array - - Returns: - Saturation and value factor, where 1 is no effect and less than 1 means the standard deviations of saturation and - value are relatively small. - """ - hsv = filter.filter_rgb_to_hsv(rgb, display_np_info=False) - s = filter.filter_hsv_to_s(hsv) - v = filter.filter_hsv_to_v(hsv) - s_std = np.std(s) - v_std = np.std(v) - if s_std < 0.05 and v_std < 0.05: - factor = 0.4 - elif s_std < 0.05: - factor = 0.7 - elif v_std < 0.05: - factor = 0.7 - else: - factor = 1 - - factor = factor ** 2 - return factor - - -def hsv_purple_deviation(hsv_hues): - """ - Obtain the deviation from the HSV hue for purple. - - Args: - hsv_hues: NumPy array of HSV hue values. - - Returns: - The HSV purple deviation. - """ - purple_deviation = np.sqrt(np.mean(np.abs(hsv_hues - HSV_PURPLE) ** 2)) - return purple_deviation - - -def hsv_pink_deviation(hsv_hues): - """ - Obtain the deviation from the HSV hue for pink. - - Args: - hsv_hues: NumPy array of HSV hue values. - - Returns: - The HSV pink deviation. - """ - pink_deviation = np.sqrt(np.mean(np.abs(hsv_hues - HSV_PINK) ** 2)) - return pink_deviation - - -def hsv_purple_pink_factor(rgb): - """ - Compute scoring factor based on purple and pink HSV hue deviations and degree to which a narrowed hue color range - average is purple versus pink. - - Args: - rgb: Image an NumPy array. - - Returns: - Factor that favors purple (hematoxylin stained) tissue over pink (eosin stained) tissue. - """ - hues = rgb_to_hues(rgb) - hues = hues[hues >= 260] # exclude hues under 260 - hues = hues[hues <= 340] # exclude hues over 340 - if len(hues) == 0: - return 0 # if no hues between 260 and 340, then not purple or pink - pu_dev = hsv_purple_deviation(hues) - pi_dev = hsv_pink_deviation(hues) - avg_factor = (340 - np.average(hues)) ** 2 - - if pu_dev == 0: # avoid divide by zero if tile has no tissue - return 0 - - factor = pi_dev / pu_dev * avg_factor - return factor - - -def hsv_purple_vs_pink_average_factor(rgb, tissue_percentage): - """ - Function to favor purple (hematoxylin) over pink (eosin) staining based on the distance of the HSV hue average - from purple and pink. - - Args: - rgb: Image as RGB NumPy array - tissue_percentage: Amount of tissue on the tile - - Returns: - Factor, where >1 to boost purple slide scores, <1 to reduce pink slide scores, or 1 no effect. - """ - - factor = 1 - # only applies to slides with a high quantity of tissue - if tissue_percentage < TISSUE_HIGH_THRESH: - return factor - - hues = rgb_to_hues(rgb) - hues = hues[hues >= 200] # Remove hues under 200 - if len(hues) == 0: - return factor - avg = np.average(hues) - # pil_hue_histogram(hues).show() - - pu = HSV_PURPLE - avg - pi = HSV_PINK - avg - pupi = pu + pi - # print("Av: %4d, Pu: %4d, Pi: %4d, PuPi: %4d" % (avg, pu, pi, pupi)) - # Av: 250, Pu: 20, Pi: 80, PuPi: 100 - # Av: 260, Pu: 10, Pi: 70, PuPi: 80 - # Av: 270, Pu: 0, Pi: 60, PuPi: 60 ** PURPLE - # Av: 280, Pu: -10, Pi: 50, PuPi: 40 - # Av: 290, Pu: -20, Pi: 40, PuPi: 20 - # Av: 300, Pu: -30, Pi: 30, PuPi: 0 - # Av: 310, Pu: -40, Pi: 20, PuPi: -20 - # Av: 320, Pu: -50, Pi: 10, PuPi: -40 - # Av: 330, Pu: -60, Pi: 0, PuPi: -60 ** PINK - # Av: 340, Pu: -70, Pi: -10, PuPi: -80 - # Av: 350, Pu: -80, Pi: -20, PuPi: -100 - - if pupi > 30: - factor *= 1.2 - if pupi < -30: - factor *= .8 - if pupi > 0: - factor *= 1.2 - if pupi > 50: - factor *= 1.2 - if pupi < -60: - factor *= .8 - - return factor - - -class TileSummary: - """ - Class for tile summary information. - """ - - slide_num = None - orig_w = None - orig_h = None - orig_tile_w = None - orig_tile_h = None - scale_factor = slide.SCALE_FACTOR - scaled_w = None - scaled_h = None - scaled_tile_w = None - scaled_tile_h = None - mask_percentage = None - num_row_tiles = None - num_col_tiles = None - - count = 0 - high = 0 - medium = 0 - low = 0 - none = 0 - - def __init__(self, slide_num, orig_w, orig_h, orig_tile_w, orig_tile_h, scaled_w, scaled_h, scaled_tile_w, - scaled_tile_h, tissue_percentage, num_col_tiles, num_row_tiles): - self.slide_num = slide_num - self.orig_w = orig_w - self.orig_h = orig_h - self.orig_tile_w = orig_tile_w - self.orig_tile_h = orig_tile_h - self.scaled_w = scaled_w - self.scaled_h = scaled_h - self.scaled_tile_w = scaled_tile_w - self.scaled_tile_h = scaled_tile_h - self.tissue_percentage = tissue_percentage - self.num_col_tiles = num_col_tiles - self.num_row_tiles = num_row_tiles - self.tiles = [] - - def __str__(self): - return summary_title(self) + "\n" + summary_stats(self) - - def mask_percentage(self): - """ - Obtain the percentage of the slide that is masked. - - Returns: - The amount of the slide that is masked as a percentage. - """ - return 100 - self.tissue_percentage - - def num_tiles(self): - """ - Retrieve the total number of tiles. - - Returns: - The total number of tiles (number of rows * number of columns). - """ - return self.num_row_tiles * self.num_col_tiles - - def tiles_by_tissue_percentage(self): - """ - Retrieve the tiles ranked by tissue percentage. - - Returns: - List of the tiles ranked by tissue percentage. - """ - sorted_list = sorted(self.tiles, key=lambda t: t.tissue_percentage, reverse=True) - return sorted_list - - def tiles_by_score(self): - """ - Retrieve the tiles ranked by score. - - Returns: - List of the tiles ranked by score. - """ - sorted_list = sorted(self.tiles, key=lambda t: t.score, reverse=True) - return sorted_list - - def top_tiles(self): - """ - Retrieve the top-scoring tiles. - - Returns: - List of the top-scoring tiles. - """ - sorted_tiles = self.tiles_by_score() - top_tiles = sorted_tiles[:NUM_TOP_TILES] - return top_tiles - - def get_tile(self, row, col): - """ - Retrieve tile by row and column. - - Args: - row: The row - col: The column - - Returns: - Corresponding Tile object. - """ - tile_index = (row - 1) * self.num_col_tiles + (col - 1) - tile = self.tiles[tile_index] - return tile - - def display_summaries(self): - """ - Display summary images. - """ - summary_and_tiles(self.slide_num, display=True, save_summary=False, save_data=False, save_top_tiles=False) - - -class Tile: - """ - Class for information about a tile. - """ - - def __init__(self, tile_summary, slide_num, np_scaled_tile, tile_num, r, c, r_s, r_e, c_s, c_e, o_r_s, o_r_e, o_c_s, - o_c_e, t_p, color_factor, s_and_v_factor, quantity_factor, score): - self.tile_summary = tile_summary - self.slide_num = slide_num - self.np_scaled_tile = np_scaled_tile - self.tile_num = tile_num - self.r = r - self.c = c - self.r_s = r_s - self.r_e = r_e - self.c_s = c_s - self.c_e = c_e - self.o_r_s = o_r_s - self.o_r_e = o_r_e - self.o_c_s = o_c_s - self.o_c_e = o_c_e - self.tissue_percentage = t_p - self.color_factor = color_factor - self.s_and_v_factor = s_and_v_factor - self.quantity_factor = quantity_factor - self.score = score - - def __str__(self): - return "[Tile #%d, Row #%d, Column #%d, Tissue %4.2f%%, Score %0.4f]" % ( - self.tile_num, self.r, self.c, self.tissue_percentage, self.score) - - def __repr__(self): - return "\n" + self.__str__() - - def mask_percentage(self): - return 100 - self.tissue_percentage - - def tissue_quantity(self): - return tissue_quantity(self.tissue_percentage) - - def get_pil_tile(self): - return tile_to_pil_tile(self) - - def get_np_tile(self): - return tile_to_np_tile(self) - - def save_tile(self): - save_display_tile(self, save=True, display=False) - - def display_tile(self): - save_display_tile(self, save=False, display=True) - - def display_with_histograms(self): - display_tile(self, rgb_histograms=True, hsv_histograms=True) - - def get_np_scaled_tile(self): - return self.np_scaled_tile - - def get_pil_scaled_tile(self): - return util.np_to_pil(self.np_scaled_tile) - - -class TissueQuantity(Enum): - NONE = 0 - LOW = 1 - MEDIUM = 2 - HIGH = 3 - - -def dynamic_tiles(slide_num, small_tile_in_tile=False): - """ - Generate tile summary with top tiles using original WSI training slide without intermediate image files saved to - file system. - - Args: - slide_num: The slide number. - small_tile_in_tile: If True, include the small NumPy image in the Tile objects. - - Returns: - TileSummary object with list of top Tile objects. The actual tile images are not retrieved until the - Tile get_tile() methods are called. - """ - np_img, large_w, large_h, small_w, small_h = slide.slide_to_scaled_np_image(slide_num) - filt_np_img = filter.apply_image_filters(np_img) - tile_summary = score_tiles(slide_num, filt_np_img, (large_w, large_h, small_w, small_h), small_tile_in_tile) - return tile_summary - - -def dynamic_tile(slide_num, row, col, small_tile_in_tile=False): - """ - Generate a single tile dynamically based on slide number, row, and column. If more than one tile needs to be - retrieved dynamically, dynamic_tiles() should be used. - - Args: - slide_num: The slide number. - row: The row. - col: The column. - small_tile_in_tile: If True, include the small NumPy image in the Tile objects. - - Returns: - Tile tile object. - """ - tile_summary = dynamic_tiles(slide_num, small_tile_in_tile) - tile = tile_summary.get_tile(row, col) - return tile - -# if __name__ == "__main__": - # tile = dynamic_tile(2, 29, 16, True) - # tile.display_with_histograms() - - # singleprocess_filtered_images_to_tiles() - # multiprocess_filtered_images_to_tiles() diff --git a/docs/wsi-preprocessing-in-python/_layouts/default.html b/docs/wsi-preprocessing-in-python/_layouts/default.html deleted file mode 100644 index cddd070..0000000 --- a/docs/wsi-preprocessing-in-python/_layouts/default.html +++ /dev/null @@ -1 +0,0 @@ -{{ content }} diff --git a/docs/wsi-preprocessing-in-python/images/127-rgb-after-filters.png b/docs/wsi-preprocessing-in-python/images/127-rgb-after-filters.png deleted file mode 100644 index 5ba0ddb..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/127-rgb-after-filters.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/127-rgb.png b/docs/wsi-preprocessing-in-python/images/127-rgb.png deleted file mode 100644 index 76a2383..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/127-rgb.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-2.png b/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-2.png deleted file mode 100644 index 80c94b7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-2.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-avoid-overmask-rem-small-obj.png b/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-avoid-overmask-rem-small-obj.png deleted file mode 100644 index 8be2ba0..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-avoid-overmask-rem-small-obj.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-overmask-rem-small-obj.png b/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-overmask-rem-small-obj.png deleted file mode 100644 index 9aa1ed8..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch-overmask-rem-small-obj.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch.png b/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch.png deleted file mode 100644 index 80c94b7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/21-avoid-overmask-green-ch.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch-avoid-overmask-rem-small-obj.png b/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch-avoid-overmask-rem-small-obj.png deleted file mode 100644 index da51436..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch-avoid-overmask-rem-small-obj.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch-overmask-rem-small-obj.png b/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch-overmask-rem-small-obj.png deleted file mode 100644 index bb4c7cf..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch-overmask-rem-small-obj.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch.png b/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch.png deleted file mode 100644 index 5901473..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/21-overmask-green-ch.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/21-rgb.png b/docs/wsi-preprocessing-in-python/images/21-rgb.png deleted file mode 100644 index ba06b04..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/21-rgb.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/337-001.png b/docs/wsi-preprocessing-in-python/images/337-001.png deleted file mode 100644 index 7a86cfe..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/337-001.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/337-002.png b/docs/wsi-preprocessing-in-python/images/337-002.png deleted file mode 100644 index 91c76e4..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/337-002.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/337-003.png b/docs/wsi-preprocessing-in-python/images/337-003.png deleted file mode 100644 index 9c59bd6..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/337-003.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/337-004.png b/docs/wsi-preprocessing-in-python/images/337-004.png deleted file mode 100644 index 0872986..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/337-004.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/337-005.png b/docs/wsi-preprocessing-in-python/images/337-005.png deleted file mode 100644 index f725bf1..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/337-005.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/337-006.png b/docs/wsi-preprocessing-in-python/images/337-006.png deleted file mode 100644 index a8ddfb2..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/337-006.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/337-007.png b/docs/wsi-preprocessing-in-python/images/337-007.png deleted file mode 100644 index e08d3a8..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/337-007.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/337-008.png b/docs/wsi-preprocessing-in-python/images/337-008.png deleted file mode 100644 index 0aeeca9..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/337-008.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/424-rgb.png b/docs/wsi-preprocessing-in-python/images/424-rgb.png deleted file mode 100644 index 4ac79a1..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/424-rgb.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/498-rgb-after-filters.png b/docs/wsi-preprocessing-in-python/images/498-rgb-after-filters.png deleted file mode 100644 index 6eebd1c..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/498-rgb-after-filters.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/498-rgb.png b/docs/wsi-preprocessing-in-python/images/498-rgb.png deleted file mode 100644 index 84fe3f0..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/498-rgb.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/5-steps.png b/docs/wsi-preprocessing-in-python/images/5-steps.png deleted file mode 100644 index c3c1f20..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/5-steps.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/TUPAC-TR-002-tile-r34-c34-x33793-y33799-w1024-h1024.png b/docs/wsi-preprocessing-in-python/images/TUPAC-TR-002-tile-r34-c34-x33793-y33799-w1024-h1024.png deleted file mode 100644 index f21777a..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/TUPAC-TR-002-tile-r34-c34-x33793-y33799-w1024-h1024.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/TUPAC-TR-002-tile-r35-c37-x36865-y34823-w1024-h1024.png b/docs/wsi-preprocessing-in-python/images/TUPAC-TR-002-tile-r35-c37-x36865-y34823-w1024-h1024.png deleted file mode 100644 index 5a9169c..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/TUPAC-TR-002-tile-r35-c37-x36865-y34823-w1024-h1024.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/adaptive-equalization.png b/docs/wsi-preprocessing-in-python/images/adaptive-equalization.png deleted file mode 100644 index 12ef7dc..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/adaptive-equalization.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/basic-threshold.png b/docs/wsi-preprocessing-in-python/images/basic-threshold.png deleted file mode 100644 index b38941a..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/basic-threshold.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-closing-20.png b/docs/wsi-preprocessing-in-python/images/binary-closing-20.png deleted file mode 100644 index bee169d..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-closing-20.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-closing-5.png b/docs/wsi-preprocessing-in-python/images/binary-closing-5.png deleted file mode 100644 index cbc58b8..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-closing-5.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-dilation-20.png b/docs/wsi-preprocessing-in-python/images/binary-dilation-20.png deleted file mode 100644 index b301b15..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-dilation-20.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-dilation-5.png b/docs/wsi-preprocessing-in-python/images/binary-dilation-5.png deleted file mode 100644 index 30903c5..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-dilation-5.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-erosion-20.png b/docs/wsi-preprocessing-in-python/images/binary-erosion-20.png deleted file mode 100644 index 499d33e..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-erosion-20.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-erosion-5.png b/docs/wsi-preprocessing-in-python/images/binary-erosion-5.png deleted file mode 100644 index 7501409..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-erosion-5.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-erosion-no-grays.png b/docs/wsi-preprocessing-in-python/images/binary-erosion-no-grays.png deleted file mode 100644 index a861ff3..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-erosion-no-grays.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-erosion-original.png b/docs/wsi-preprocessing-in-python/images/binary-erosion-original.png deleted file mode 100644 index 9658135..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-erosion-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-opening-20.png b/docs/wsi-preprocessing-in-python/images/binary-opening-20.png deleted file mode 100644 index d6db36f..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-opening-20.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/binary-opening-5.png b/docs/wsi-preprocessing-in-python/images/binary-opening-5.png deleted file mode 100644 index 173ed7b..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/binary-opening-5.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/blue-filter.png b/docs/wsi-preprocessing-in-python/images/blue-filter.png deleted file mode 100644 index 93d8385..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/blue-filter.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/blue-original.png b/docs/wsi-preprocessing-in-python/images/blue-original.png deleted file mode 100644 index c4a6e5e..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/blue-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/blue-pen-filter.png b/docs/wsi-preprocessing-in-python/images/blue-pen-filter.png deleted file mode 100644 index 082488d..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/blue-pen-filter.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/blue-pen.png b/docs/wsi-preprocessing-in-python/images/blue-pen.png deleted file mode 100644 index d25aec7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/blue-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/blue.png b/docs/wsi-preprocessing-in-python/images/blue.png deleted file mode 100644 index a4569ff..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/blue.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/canny-original-cropped.png b/docs/wsi-preprocessing-in-python/images/canny-original-cropped.png deleted file mode 100644 index d5a042d..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/canny-original-cropped.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/canny-original-with-inverse-mask.png b/docs/wsi-preprocessing-in-python/images/canny-original-with-inverse-mask.png deleted file mode 100644 index 610af40..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/canny-original-with-inverse-mask.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/canny-original.png b/docs/wsi-preprocessing-in-python/images/canny-original.png deleted file mode 100644 index 9033075..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/canny-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/canny.png b/docs/wsi-preprocessing-in-python/images/canny.png deleted file mode 100644 index fa92c1c..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/canny.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/color-histograms-large-tile.png b/docs/wsi-preprocessing-in-python/images/color-histograms-large-tile.png deleted file mode 100644 index 5b05cce..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/color-histograms-large-tile.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/color-histograms-small-tile.png b/docs/wsi-preprocessing-in-python/images/color-histograms-small-tile.png deleted file mode 100644 index 7f5e78b..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/color-histograms-small-tile.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-blue-pen.png b/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-blue-pen.png deleted file mode 100644 index 3b6bf35..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-blue-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-green-pen-no-blue-pen.png b/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-green-pen-no-blue-pen.png deleted file mode 100644 index f8f8248..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-green-pen-no-blue-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-green-pen.png b/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-green-pen.png deleted file mode 100644 index a1b710e..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-no-green-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-original-with-no-green-pen-no-blue-pen.png b/docs/wsi-preprocessing-in-python/images/combine-pen-filters-original-with-no-green-pen-no-blue-pen.png deleted file mode 100644 index 653a593..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-original-with-no-green-pen-no-blue-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-original.png b/docs/wsi-preprocessing-in-python/images/combine-pen-filters-original.png deleted file mode 100644 index 4100395..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pen-filters-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pens-background-mask.png b/docs/wsi-preprocessing-in-python/images/combine-pens-background-mask.png deleted file mode 100644 index 095d873..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pens-background-mask.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pens-background-original-with-inverse-mask.png b/docs/wsi-preprocessing-in-python/images/combine-pens-background-original-with-inverse-mask.png deleted file mode 100644 index ef96f4c..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pens-background-original-with-inverse-mask.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pens-background-original-with-mask.png b/docs/wsi-preprocessing-in-python/images/combine-pens-background-original-with-mask.png deleted file mode 100644 index d7dd065..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pens-background-original-with-mask.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/combine-pens-background-original.png b/docs/wsi-preprocessing-in-python/images/combine-pens-background-original.png deleted file mode 100644 index 664ebee..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/combine-pens-background-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/complement.png b/docs/wsi-preprocessing-in-python/images/complement.png deleted file mode 100644 index cbeb858..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/complement.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/contrast-stretching.png b/docs/wsi-preprocessing-in-python/images/contrast-stretching.png deleted file mode 100644 index 4237717..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/contrast-stretching.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/display-image-with-text.png b/docs/wsi-preprocessing-in-python/images/display-image-with-text.png deleted file mode 100644 index e1a22a7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/display-image-with-text.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/distribution-of-svs-image-sizes.png b/docs/wsi-preprocessing-in-python/images/distribution-of-svs-image-sizes.png deleted file mode 100644 index 93c444d..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/distribution-of-svs-image-sizes.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/entropy-grayscale.png b/docs/wsi-preprocessing-in-python/images/entropy-grayscale.png deleted file mode 100644 index 2f03147..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/entropy-grayscale.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/entropy-original-entropy-mask.png b/docs/wsi-preprocessing-in-python/images/entropy-original-entropy-mask.png deleted file mode 100644 index ac311dc..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/entropy-original-entropy-mask.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/entropy-original-inverse-entropy-mask.png b/docs/wsi-preprocessing-in-python/images/entropy-original-inverse-entropy-mask.png deleted file mode 100644 index e68969e..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/entropy-original-inverse-entropy-mask.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/entropy-original.png b/docs/wsi-preprocessing-in-python/images/entropy-original.png deleted file mode 100644 index 35ce056..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/entropy-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/entropy.png b/docs/wsi-preprocessing-in-python/images/entropy.png deleted file mode 100644 index 4cd08cc..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/entropy.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/eosin-channel.png b/docs/wsi-preprocessing-in-python/images/eosin-channel.png deleted file mode 100644 index fcf88f3..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/eosin-channel.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/fill-holes-remove-small-holes-100.png b/docs/wsi-preprocessing-in-python/images/fill-holes-remove-small-holes-100.png deleted file mode 100644 index d783883..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/fill-holes-remove-small-holes-100.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/fill-holes-remove-small-holes-10000.png b/docs/wsi-preprocessing-in-python/images/fill-holes-remove-small-holes-10000.png deleted file mode 100644 index 625bbe7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/fill-holes-remove-small-holes-10000.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/fill-holes.png b/docs/wsi-preprocessing-in-python/images/fill-holes.png deleted file mode 100644 index aea7cff..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/fill-holes.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/filter-example.png b/docs/wsi-preprocessing-in-python/images/filter-example.png deleted file mode 100644 index 4cfe631..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/filter-example.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/filters-001-008.png b/docs/wsi-preprocessing-in-python/images/filters-001-008.png deleted file mode 100644 index 9ed3c8b..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/filters-001-008.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/grays-filter.png b/docs/wsi-preprocessing-in-python/images/grays-filter.png deleted file mode 100644 index 497ee72..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/grays-filter.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/grayscale.png b/docs/wsi-preprocessing-in-python/images/grayscale.png deleted file mode 100644 index b8694e4..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/grayscale.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/green-channel-filter.png b/docs/wsi-preprocessing-in-python/images/green-channel-filter.png deleted file mode 100644 index ef76ab7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/green-channel-filter.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/green-filter.png b/docs/wsi-preprocessing-in-python/images/green-filter.png deleted file mode 100644 index b5ec735..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/green-filter.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/green-original.png b/docs/wsi-preprocessing-in-python/images/green-original.png deleted file mode 100644 index 60d675e..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/green-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/green-pen-filter.png b/docs/wsi-preprocessing-in-python/images/green-pen-filter.png deleted file mode 100644 index 2224764..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/green-pen-filter.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/green-pen.png b/docs/wsi-preprocessing-in-python/images/green-pen.png deleted file mode 100644 index 07745db..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/green-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/green.png b/docs/wsi-preprocessing-in-python/images/green.png deleted file mode 100644 index 4037fc5..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/green.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/hematoxylin-channel.png b/docs/wsi-preprocessing-in-python/images/hematoxylin-channel.png deleted file mode 100644 index 51183f4..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/hematoxylin-channel.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/histogram-equalization.png b/docs/wsi-preprocessing-in-python/images/histogram-equalization.png deleted file mode 100644 index 8e5ddc8..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/histogram-equalization.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/hsv-hue-histogram.png b/docs/wsi-preprocessing-in-python/images/hsv-hue-histogram.png deleted file mode 100644 index 3bbc559..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/hsv-hue-histogram.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/hysteresis-threshold.png b/docs/wsi-preprocessing-in-python/images/hysteresis-threshold.png deleted file mode 100644 index fde68e7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/hysteresis-threshold.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/kmeans-original.png b/docs/wsi-preprocessing-in-python/images/kmeans-original.png deleted file mode 100644 index c4dc401..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/kmeans-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/kmeans-segmentation-after-otsu.png b/docs/wsi-preprocessing-in-python/images/kmeans-segmentation-after-otsu.png deleted file mode 100644 index cfb27e1..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/kmeans-segmentation-after-otsu.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/kmeans-segmentation.png b/docs/wsi-preprocessing-in-python/images/kmeans-segmentation.png deleted file mode 100644 index 534ae18..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/kmeans-segmentation.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/not-blue-pen.png b/docs/wsi-preprocessing-in-python/images/not-blue-pen.png deleted file mode 100644 index 5bf8e7b..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/not-blue-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/not-blue.png b/docs/wsi-preprocessing-in-python/images/not-blue.png deleted file mode 100644 index 5935c7e..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/not-blue.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/not-green-pen.png b/docs/wsi-preprocessing-in-python/images/not-green-pen.png deleted file mode 100644 index 19307ee..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/not-green-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/not-green.png b/docs/wsi-preprocessing-in-python/images/not-green.png deleted file mode 100644 index c0f92c7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/not-green.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/not-red-pen.png b/docs/wsi-preprocessing-in-python/images/not-red-pen.png deleted file mode 100644 index dda9840..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/not-red-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/not-red.png b/docs/wsi-preprocessing-in-python/images/not-red.png deleted file mode 100644 index e8f9bc1..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/not-red.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/openslide-available-slides.png b/docs/wsi-preprocessing-in-python/images/openslide-available-slides.png deleted file mode 100644 index d83a5ca..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/openslide-available-slides.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/openslide-whole-slide-image-zoomed.png b/docs/wsi-preprocessing-in-python/images/openslide-whole-slide-image-zoomed.png deleted file mode 100644 index 1a9f2e6..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/openslide-whole-slide-image-zoomed.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/openslide-whole-slide-image.png b/docs/wsi-preprocessing-in-python/images/openslide-whole-slide-image.png deleted file mode 100644 index 43738a8..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/openslide-whole-slide-image.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/optional-tile-labels.png b/docs/wsi-preprocessing-in-python/images/optional-tile-labels.png deleted file mode 100644 index 3128257..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/optional-tile-labels.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/otsu-mask.png b/docs/wsi-preprocessing-in-python/images/otsu-mask.png deleted file mode 100644 index ebbe237..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/otsu-mask.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/otsu-threshold.png b/docs/wsi-preprocessing-in-python/images/otsu-threshold.png deleted file mode 100644 index f75614d..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/otsu-threshold.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/pink-and-purple-slide.png b/docs/wsi-preprocessing-in-python/images/pink-and-purple-slide.png deleted file mode 100644 index 6a137d4..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/pink-and-purple-slide.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/purple-slide.png b/docs/wsi-preprocessing-in-python/images/purple-slide.png deleted file mode 100644 index beb45b5..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/purple-slide.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/rag-thresh-1.png b/docs/wsi-preprocessing-in-python/images/rag-thresh-1.png deleted file mode 100644 index 62af1e9..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/rag-thresh-1.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/rag-thresh-20.png b/docs/wsi-preprocessing-in-python/images/rag-thresh-20.png deleted file mode 100644 index f2740ec..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/rag-thresh-20.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/rag-thresh-9.png b/docs/wsi-preprocessing-in-python/images/rag-thresh-9.png deleted file mode 100644 index 86a9a51..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/rag-thresh-9.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/rag-thresh-original.png b/docs/wsi-preprocessing-in-python/images/rag-thresh-original.png deleted file mode 100644 index fe2ade4..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/rag-thresh-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/red-filter.png b/docs/wsi-preprocessing-in-python/images/red-filter.png deleted file mode 100644 index bd8b818..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/red-filter.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/red-pen-filter.png b/docs/wsi-preprocessing-in-python/images/red-pen-filter.png deleted file mode 100644 index b887feb..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/red-pen-filter.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/red-pen-slides-filters.png b/docs/wsi-preprocessing-in-python/images/red-pen-slides-filters.png deleted file mode 100644 index 7af55ed..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/red-pen-slides-filters.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/red-pen.png b/docs/wsi-preprocessing-in-python/images/red-pen.png deleted file mode 100644 index 2572242..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/red-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/red.png b/docs/wsi-preprocessing-in-python/images/red.png deleted file mode 100644 index 3791756..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/red.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/remove-more-green-more-gray.png b/docs/wsi-preprocessing-in-python/images/remove-more-green-more-gray.png deleted file mode 100644 index a558446..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/remove-more-green-more-gray.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/remove-small-holes-100.png b/docs/wsi-preprocessing-in-python/images/remove-small-holes-100.png deleted file mode 100644 index 2678518..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/remove-small-holes-100.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/remove-small-holes-10000.png b/docs/wsi-preprocessing-in-python/images/remove-small-holes-10000.png deleted file mode 100644 index 8bc03e8..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/remove-small-holes-10000.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/remove-small-objects-100.png b/docs/wsi-preprocessing-in-python/images/remove-small-objects-100.png deleted file mode 100644 index be3f6af..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/remove-small-objects-100.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/remove-small-objects-10000.png b/docs/wsi-preprocessing-in-python/images/remove-small-objects-10000.png deleted file mode 100644 index f6b4917..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/remove-small-objects-10000.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/scoring-formula.png b/docs/wsi-preprocessing-in-python/images/scoring-formula.png deleted file mode 100644 index 1985382..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/scoring-formula.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-rgb-hsv.png b/docs/wsi-preprocessing-in-python/images/slide-2-rgb-hsv.png deleted file mode 100644 index 621c60a..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-rgb-hsv.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-30.png b/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-30.png deleted file mode 100644 index ff14d34..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-30.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-31.png b/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-31.png deleted file mode 100644 index 3c7955e..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-31.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-32.png b/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-32.png deleted file mode 100644 index 4a55279..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-row-25-col-32.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-tile-tissue-heatmap-original.png b/docs/wsi-preprocessing-in-python/images/slide-2-tile-tissue-heatmap-original.png deleted file mode 100644 index efd0df4..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-tile-tissue-heatmap-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-tile-tissue-heatmap.png b/docs/wsi-preprocessing-in-python/images/slide-2-tile-tissue-heatmap.png deleted file mode 100644 index 8f62ece..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-tile-tissue-heatmap.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-tissue-percentage-tile-1000.png b/docs/wsi-preprocessing-in-python/images/slide-2-tissue-percentage-tile-1000.png deleted file mode 100644 index d48b2af..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-tissue-percentage-tile-1000.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-tissue-percentage-tile-1500.png b/docs/wsi-preprocessing-in-python/images/slide-2-tissue-percentage-tile-1500.png deleted file mode 100644 index 9c30610..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-tissue-percentage-tile-1500.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-borders.png b/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-borders.png deleted file mode 100644 index 5aa35bc..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-borders.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-labels-borders.png b/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-labels-borders.png deleted file mode 100644 index c2b6b00..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-labels-borders.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-labels.png b/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-labels.png deleted file mode 100644 index 98874af..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-top-tile-labels.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-top-tiles-original.png b/docs/wsi-preprocessing-in-python/images/slide-2-top-tiles-original.png deleted file mode 100644 index 2ddac94..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-top-tiles-original.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-2-top-tiles.png b/docs/wsi-preprocessing-in-python/images/slide-2-top-tiles.png deleted file mode 100644 index b1d50ec..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-2-top-tiles.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-4-rgb.png b/docs/wsi-preprocessing-in-python/images/slide-4-rgb.png deleted file mode 100644 index 89e8cc7..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-4-rgb.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-4-top-tile-1.png b/docs/wsi-preprocessing-in-python/images/slide-4-top-tile-1.png deleted file mode 100644 index 2c6e3c2..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-4-top-tile-1.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-4-top-tile-2.png b/docs/wsi-preprocessing-in-python/images/slide-4-top-tile-2.png deleted file mode 100644 index 42a91a5..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-4-top-tile-2.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-pen.png b/docs/wsi-preprocessing-in-python/images/slide-pen.png deleted file mode 100644 index d8f1acd..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-pen.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/slide-scan.png b/docs/wsi-preprocessing-in-python/images/slide-scan.png deleted file mode 100644 index 5bf0dcf..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/slide-scan.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/svs-image-sizes.png b/docs/wsi-preprocessing-in-python/images/svs-image-sizes.png deleted file mode 100644 index 6f1e83f..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/svs-image-sizes.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/tile-data.png b/docs/wsi-preprocessing-in-python/images/tile-data.png deleted file mode 100644 index 18e4d2f..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/tile-data.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/tiles-page.png b/docs/wsi-preprocessing-in-python/images/tiles-page.png deleted file mode 100644 index d037803..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/tiles-page.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/images/wsi-example.png b/docs/wsi-preprocessing-in-python/images/wsi-example.png deleted file mode 100644 index 671697e..0000000 Binary files a/docs/wsi-preprocessing-in-python/images/wsi-example.png and /dev/null differ diff --git a/docs/wsi-preprocessing-in-python/index.md b/docs/wsi-preprocessing-in-python/index.md deleted file mode 100644 index 456829d..0000000 --- a/docs/wsi-preprocessing-in-python/index.md +++ /dev/null @@ -1,2775 +0,0 @@ ---- -layout: default ---- - - -* Table of contents. -{:toc} - - -# Whole-Slide Image Preprocessing in Python - - -## Project Introduction - -The primary goal of the [Tumor Proliferation Assessment Challenge 2016 (TUPAC16)](http://tupac.tue-image.nl/) is to -develop algorithms to automatically predict breast cancer tumor proliferation scores. In this challenge, the training -set consists of 500 whole-slide images which are scored (1, 2, or 3) by pathologists based on mitosis -counts. A higher proliferation score indicates a worse prognosis since higher tumor proliferation rates are -correlated with worse outcomes. The tissue samples are stained with hematoxylin and eosin (H&E). - -One of our first approaches to this challenge was to apply deep learning to breast cancer whole-slide images, -following an approach similar to the process used by Ertosun and Rubin in -[Automated Grading of Gliomas using Deep Learning in Digital Pathology Images: A modular approach with ensemble of -convolutional neural networks](https://web.stanford.edu/group/rubinlab/pubs/2243353.pdf). One important part of the -technique described by Ertosun and Rubin involves image preprocessing, where large whole-slide images are divided into -tiles and only tiles that consist of at least 90% tissue are further analyzed. Tissue is determined by hysteresis -thresholding on the grayscale image complement. - -The three TUPAC16 challenge tasks were won by Paeng et al, described in -[A Unified Framework for Tumor Proliferation Score Prediction in Breast -Histopathology](https://pdfs.semanticscholar.org/7d9b/ccac7a9a850cc84a980e5abeaeac2aef94e6.pdf). In their technique, -identification of tissue regions in whole-slide images is done using Otsu thresholding, morphological operations, and -binary dilation. - -Tissue identification in whole-slide images can be an important precursor to deep learning. Deep learning is -computationally expensive and medical whole-slide images are enormous. Typically, a large portion of a slide isn't -useful, such as the background, shadows, water, smudges, and pen marks. We can use preprocessing to -rapidly reduce the quantity and increase the quality of the image data to be analyzed. This -can lead to faster, more accurate model training. - -In this tutorial, we will take a look at whole-slide image processing and will describe various filters -that can be used to increase the accuracy of tissue identification. -After determining a useful set of filters for tissue segmentation, we'll divide slides into tiles and determine sets -of tiles that typically represent good tissue samples. - -The solution should demonstrate high performance, flexibility, and accuracy. Filters should be easy to combine, -chain, and modify. Tile scoring should be easy to modify for accurate tile selection. The solution should offer -the ability to view filter, tile, and score results across large, unique datasets. The solution should also have -the ability to work in a batch mode, where all image files and intermediary files are written to the file system, -and in a dynamic mode, where high-scoring tissue tiles can be retrieved from the original WSI files without requiring -any intermediary files. - -In summary, we will scale down whole-slide images, apply filters to these scaled-down images for tissue segmentation, -break the slides into tiles, score the tiles, and then retrieve the top tiles based on their scores. - -| **5 Steps** | -| -------------------- | -| ![5 Steps](images/5-steps.png "5 Steps") | - - -### Setup - -This project makes heavy use of Python3. Python is an ideal language for image processing. -OpenSlide is utilized for reading WSI files. Pillow is used for basic image manipulation in Python. -NumPy is used for fast, concise, powerful processing of images as NumPy arrays. Scikit-image is heavily used for -a wide variety of image functionality, such as morphology, thresholding, and edge detection. - -Some quick setup steps on macOS follow. - -Install a package manager such as [Homebrew](https://brew.sh/). - - /usr/bin/ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)" - -Install [Python3](https://www.python.org/). - - brew install python3 - -Install [OpenSlide](http://openslide.org/). -Note that OpenSlide is licensed under the [LGPL 2.1 -License](https://raw.githubusercontent.com/openslide/openslide/master/lgpl-2.1.txt). - - brew install openslide - -Next, we can install a variety of useful Python packages using the [pip3](https://pip.pypa.io/en/stable/) -package manager. These packages include: -[matplotlib](https://pypi.python.org/pypi/matplotlib/), -[numpy](https://pypi.python.org/pypi/numpy), -[openslide-python](https://pypi.python.org/pypi/openslide-python), -[Pillow](https://pypi.org/project/Pillow/), -[scikit-image](https://pypi.python.org/pypi/scikit-image), -[scikit-learn](https://pypi.python.org/pypi/scikit-learn), -and [scipy](https://pypi.python.org/pypi/scipy). - - pip3 install -U matplotlib numpy openslide-python Pillow scikit-image scikit-learn scipy - -We utilize scikit-image filters (hysteresis thresholding) in this tutorial that are not present in the -latest released version of scikit-image at the time of this writing (0.13.1). We can install scikit-image -from source, as described in the README at [https://github.com/scikit-image/scikit-image](https://github.com/scikit-image/scikit-image). - - git clone https://github.com/scikit-image/scikit-image.git - cd scikit-image - pip3 install -r requirements.txt - pip3 install . - - -### Whole Slide Imaging Background - -A whole-slide image is a digital representation of a microscopic slide, typically at a very high level of magnification -such as 20x or 40x. As a result of this high magnification, whole slide images are typically very large in size. -The maximum file size for a single whole-slide image in our training dataset was 3.4 GB, with an average over 1 GB. - -| **WSI Example Slide** | -| -------------------- | -| ![WSI Example Slide](images/wsi-example.png "WSI Example Slide") | - - -A whole-slide image is created by a microscope that scans a slide and combines smaller images into a large image. -Techniques include combining scanned square tiles into a whole-slide image and combining scanned strips -into a resulting whole-slide image. Occasionally, the smaller constituent images can be -visually discerned, as in the shaded area at the top of the slide seen below. - -| **Combining Smaller Images into a Whole-Slide Image** | -| -------------------- | -| ![Combining Smaller Images into a Whole-Slide Image](images/slide-scan.png "Combining Smaller Images into a Whole-Slide Image") | - - -A fairly unusual feature of whole-slide images is the very large image size. -For our training dataset of 500 images, the width varied from 19,920 pixels to 198,220 pixels, -with an average of 101,688 pixels. The height varied from 13,347 pixels to 256,256 pixels, -with an average of 73,154 pixels. The image total pixel sizes varied from -369,356,640 to 35,621,634,048 pixels, with an average of -7,670,709,628 pixels. The 500 training images take up a total of 525 GB of storage space. - -| **Training Image Sizes** | -| -------------------- | -| ![Training Image Sizes](images/svs-image-sizes.png "Training Image Sizes") | - - -Here we see a histogram distribution of the training image sizes in megapixels. - -| **Distribution of Images Based on Number of Pixels** | -| -------------------- | -| ![Distribution of Image Sizes](images/distribution-of-svs-image-sizes.png "Distribution of Image Sizes") | - - -The [OpenSlide](http://openslide.org/) project can be used to read a variety of whole-slide -image formats, including the [Aperio *.svs slide format](http://openslide.org/formats/aperio/) -of our training image set. This is a pyramidal, tiled format, where the massive slide is composed of -a large number of constituent tiles. - -To use the OpenSlide Python interface to view whole slide images, we can clone the -[OpenSlide Python interface from GitHub](https://github.com/openslide/openslide-python) -and utilize the included DeepZoom `deepzoom_multiserver.py` script. - - git clone https://github.com/openslide/openslide-python.git - cd openslide-python/examples/deepzoom - python3 deepzoom_multiserver.py -Q 100 WSI_DIRECTORY - -The `deepzoom_multiserver.py` script starts a web interface on port 5000 and displays -the image files at the specified file system location (the `WSI_DIRECTORY` value above, -which could be a location such as `~/git/python-wsi-preprocessing/data/`). If image -files exist in subdirectories, they will also be displayed in the list of available -slides. - -If this viewing application is installed on a server that also hosts the whole-slide image repository, this -offers a convenient mechanism for users to view the slides without requiring local storage space. - -| **OpenSlide Available Slides** | -| -------------------- | -| ![OpenSlide Available Slides](images/openslide-available-slides.png "OpenSlide Available Slides") | - - -Here we can see the initial view of one of the whole-slide images viewed in a web browser. - -| **OpenSlide Whole Slide Image** | -| -------------------- | -| ![OpenSlide Whole Slide Image](images/openslide-whole-slide-image.png "OpenSlide Whole Slide Image") | - - -Using this web interface, the whole-slide image can be zoomed to the highest magnification, revealing fine details at -the tile level. Zooming and scrolling operations make it relatively easy to visually peruse the whole slide image. - -| **OpenSlide Whole Slide Image Zoomed** | -| -------------------- | -| ![OpenSlide Whole Slide Image Zoomed](images/openslide-whole-slide-image-zoomed.png "OpenSlide Whole Slide Image Zoomed") | - - -## Scale Down Images - -To develop a set of filters that can be applied to an entire set of large whole-slide images, two of the first issues -we are confronted with are the size of the data and the format of the data. As mentioned, for our training dataset, -the average `svs` file size is over 1 GB and we have 500 total images. Additionally, the `svs` format is a fairly unusual -format which typically can't be visually displayed by default by common applications and operating systems. Therefore, we will -develop some code to overcome these important issues. Using OpenSlide and Python, we'll convert the training dataset to -smaller images in a common format, thus reformulating a big data problem as a small data problem. Before filtering -at the entire slide level, we will shrink the width and height down by a factor of 32x, which means we can perform -filtering on 1/1024th the image data. Converting 500 `svs` files to `png` files at 1/32 scale takes -approximately 12 minutes on a typical MacBook Pro using the code described below. - -In the `wsi/slide.py` file, we have many functions that can be used in relation to the original `svs` images. Of -particular importance are the following functions: - - open_slide() - show_slide() - slide_info(display_all_properties=True) - slide_stats() - training_slide_to_image() - singleprocess_training_slides_to_images() - multiprocess_training_slides_to_images() - -The `open_slide()` function uses OpenSlide to read in an `svs` file. The `show_slide()` function opens a WSI `svs` file -and displays a scaled-down version of the slide to the screen. The `slide_info()` function displays metadata -associated with all `svs` files. The `slide_stats()` function looks at all images and summarizes size information -about the set of slides. It also generates a variety of charts for a visual representation of the slide statistics. -The `training_slide_to_image()` function converts a single `svs` slide to a smaller image in a more common format such as -`jpg` or `png`. The `singleprocess_training_slides_to_images()` function converts all `svs` slides to smaller images, -and the `multiprocess_training_slides_to_images()` function uses multiple processes (1 process per core) to -speed up the slide conversion process. For the last three functions, when an image is saved, a thumbnail image is also -saved. By default, the thumbnail has a maximum height or width of 300 pixels and is `jpg` format. - -One of the first actions we can take to become more familiar with the training dataset is to have a look at the metadata -associated with each image, which we can do with the `slide_info()` function. Here we can see a sample of this -metadata for Slide #1: - -``` -Opening Slide #1: /Volumes/BigData/TUPAC/training_slides/TUPAC-TR-001.svs -Level count: 5 -Level dimensions: ((130304, 247552), (32576, 61888), (8144, 15472), (2036, 3868), (1018, 1934)) -Level downsamples: (1.0, 4.0, 16.0, 64.0, 128.0) -Dimensions: (130304, 247552) -Objective power: 40 -Associated images: - macro: - thumbnail: -Format: aperio -Properties: - Property: aperio.AppMag, value: 40 - Property: aperio.MPP, value: 0.16437 - Property: openslide.comment, value: Aperio Image Library v11.0.37 -130304x247552 (256x256) JPEG/RGB Q=40;Mirax Digital Slide|AppMag = 40|MPP = 0.16437 - Property: openslide.level-count, value: 5 - Property: openslide.level[0].downsample, value: 1 - Property: openslide.level[0].height, value: 247552 - Property: openslide.level[0].tile-height, value: 256 - Property: openslide.level[0].tile-width, value: 256 - Property: openslide.level[0].width, value: 130304 - Property: openslide.level[1].downsample, value: 4 - Property: openslide.level[1].height, value: 61888 - Property: openslide.level[1].tile-height, value: 256 - Property: openslide.level[1].tile-width, value: 256 - Property: openslide.level[1].width, value: 32576 - Property: openslide.level[2].downsample, value: 16 - Property: openslide.level[2].height, value: 15472 - Property: openslide.level[2].tile-height, value: 256 - Property: openslide.level[2].tile-width, value: 256 - Property: openslide.level[2].width, value: 8144 - Property: openslide.level[3].downsample, value: 64 - Property: openslide.level[3].height, value: 3868 - Property: openslide.level[3].tile-height, value: 256 - Property: openslide.level[3].tile-width, value: 256 - Property: openslide.level[3].width, value: 2036 - Property: openslide.level[4].downsample, value: 128 - Property: openslide.level[4].height, value: 1934 - Property: openslide.level[4].tile-height, value: 256 - Property: openslide.level[4].tile-width, value: 256 - Property: openslide.level[4].width, value: 1018 - Property: openslide.mpp-x, value: 0.16436999999999999 - Property: openslide.mpp-y, value: 0.16436999999999999 - Property: openslide.objective-power, value: 40 - Property: openslide.quickhash-1, value: 0e0631ade42ae3384aaa727ce2e36a8272fe67039c513e17dccfdd592f6040cb - Property: openslide.vendor, value: aperio - Property: tiff.ImageDescription, value: Aperio Image Library v11.0.37 -130304x247552 (256x256) JPEG/RGB Q=40;Mirax Digital Slide|AppMag = 40|MPP = 0.16437 - Property: tiff.ResolutionUnit, value: inch -``` - -The most important metadata for our purposes is that the slide has a width of 130,304 pixels and a height of -247,552 pixels. Note that these values are displayed as width followed by height. For most of our image processing, -we will be using NumPy arrays, where rows (height) are followed by columns (width). - -If we visually look over the metadata associated with other images in the training dataset, we see that the slides -are not consistent in their various properties such as the number of levels contained in the `svs` files. The metadata -implies that the dataset comes from a variety of sources. The variability in the slides, especially regarding -issues such as H&E staining and pen marks on the slides, needs to be considered during our filter development. - -If we call the `slide_stats()` function, in addition to the charts, we obtain a table of pixel statistics, shown -below. - -**Training Images Statistics**
- -| Attribute | Size | Slide # | -| ---------- | --------------------- | ------- | -| Max width | 198,220 pixels | 10 | -| Max height | 256,256 pixels | 387 | -| Max size | 35,621,634,048 pixels | 387 | -| Min width | 19,920 pixels | 112 | -| Min height | 13,347 pixels | 108 | -| Min size | 369,356,640 pixels | 112 | -| Avg width | 101,688 pixels | | -| Avg height | 73,154 pixels | | -| Avg size | 7,670,709,629 pixels | | - - -The `wsi/slide.py` file contains constants that can be used to control various image conversion settings. For example, -the `SCALE_FACTOR` constant controls the factor by which the slides will be scaled down. Its default value is 32, -meaning that the height and width will be scaled down by a factor of 32. This means that when we perform filtering, -it will be performed on an image 1/1024th the size of the original high-resolution image. -The `DEST_TRAIN_EXT` constant controls the output format. We will use the default format, `png`. - -Using macOS, the following conversion times using -`singleprocess_training_slides_to_images()` and `multiprocess_training_slides_to_images()` -on the 500 image training set were obtained: - -**Training Image Dataset Conversion Times**
- -| Format | Processes | Time | -| ------ | -------------- | ------ | -| jpg | single process | 26m09s | -| jpg | multi process | 10m21s | -| png | single process | 42m59s | -| png | multi process | 11m58s | - - -After calling `multiprocess_training_slides_to_images()` using the `png` format, we have 500 scaled-down -whole-slide images in lossless `png` format that we will examine in greater detail in relation to our filters. - - -### Image Saving, Displaying, and Conversions - -In order to load, save, and display images, we use the Python [Pillow](https://pillow.readthedocs.io/en/4.3.x/) -package. In particular, we make use of the Image module, which contains an Image class used to represent an image. -The `wsi/slide.py` file contains an `open_image()` function to open an image stored in the file system. -The `get_training_image_path()` function takes a slide number and returns the path to the corresponding training image -file, meaning the scaled-down `png` file that we created by calling `multiprocess_training_slides_to_images()`. - -If we want to convert a single `svs` WSI file to a scaled-down `png` (without converting all `svs` files), -open that `png` image file as a PIL Image, and display the image to the screen, we can do the following. - -``` -slide.training_slide_to_image(4) -img_path = slide.get_training_image_path(4) -img = slide.open_image(img_path) -img.show() -``` - -To mathematically manipulate the images, we use NumPy arrays. The `wsi/util.py` file contains a -`pil_to_np_rgb()` function that converts a PIL Image to a 3-dimensional NumPy array in RGB format. The first dimension -represents the number of rows, the second dimension represents the number of columns, and the third dimension -represents the channel (red, green, and blue). - -``` -rgb = util.pil_to_np_rgb(img) -``` - -For convenience, the `display_img()` function can be used to display a NumPy array image. Text can be added to -the displayed image, which can be very useful when visually comparing the results of multiple filters. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -``` - -| **Display Image with Text** | -| -------------------- | -| ![Display Image with Text](images/display-image-with-text.png "Display Image with Text") | - - -When performing operations on NumPy arrays, functions in the `wsi/filter.py` file will often utilize the -`util.np_info()` function to display information about the NumPy array and the amount of time required to perform the -operation. For example, the above call to `pil_to_np_rgb()` internally calls `np_info()`: - -``` -t = Time() -rgb = np.asarray(pil_img) -np_info(rgb, "RGB", t.elapsed()) -return rgb -``` - -This call to `np_info()` generates console output such as the following: - -``` -RGB | Time: 0:00:00.162484 Type: uint8 Shape: (1385, 1810, 3) -``` - -We see that the PIL-to-NumPy array conversion took 0.16s. The type of the NumPy array is `uint8`, which means -that each pixel is represented by a red, green, and blue unsigned integer value from 0 to 255. The image has a height of -1385 pixels, a width of 1810 pixels, and three channels (representing red, green, and blue). - -We can obtain additional information about NumPy arrays by setting the `util.ADDITIONAL_NP_STATS` constant to `True`. -If we rerun the above code with `ADDITIONAL_NP_STATS = True`, we see the following: - -``` -RGB | Time: 0:00:00.157696 Min: 2.00 Max: 255.00 Mean: 182.62 Binary: F Type: uint8 Shape: (1385, 1810, 3) -``` - -The minimum value is 2, the maximum value is 255, the mean value is 182.62, and binary is false, meaning that the -image is not a binary image. A binary image is an image that consists of only two values (True or False, 1.0 or 0.0, -255 or 0). Binary images are produced by actions such as thresholding. - -When interacting with NumPy image processing code, the information provided by `np_info()` can be extremely useful. -For example, some functions return boolean NumPy arrays, other functions return float NumPy arrays, and other -functions may return `uint8` NumPy arrays. Before performing actions on a NumPy array, it's usually necessary to know -the data type of the array and the nature of the data in that array. For performance reasons, normally -`ADDITIONAL_NP_STATS` should be set to `False`. - -The `wsi/util.py` file contains an `np_to_pil()` function that converts a NumPy array to a PIL Image. - -If we have a PIL Image, saving the image to the file system can be accomplished by calling the Image's `save()` -function. - -``` -img.save(path) -``` - - -## Apply Filters for Tissue Segmentation - -Next, we will investigate image filters and will determine a set of filters that can be utilized for effective -tissue segmentation with our dataset. -We will mask out non-tissue by setting non-tissue pixels to 0 for their red, green, and blue channels. For our -particular dataset, our mask will AND together a green channel mask, a grays mask, a red pen mask, a green pen mask, -and a blue pen mask. Following this, we will mask out small objects from the images. - -The filtering approach that we develop here has several benefits. All relevant filters are centralized in a single -file, `wsi/filter.py`, for convenience. Filters return results in a standard format and the returned datatype can -easily be changed (`boolean`, `uint8`, `float`). Critical filter debug information (shape, type, processing time, etc) -is output to the console. Filter results can be easily viewed across the entire dataset or subsets of the dataset. -Multiprocessing is used for increased performance. Additionally, filters can easily be combined, strung together, -or otherwise modified. - -To filter our scaled-down 500 `png` image training set and generate 4,500 `png` filter preview images and 4,500 `jpg` thumbnails -takes about 23m30s on my MacBook Pro. Filtering the 500 image training set without saving files takes approximately -6 minutes. - -### Filters - -Let's take a look at several ways that our images can be filtered. Filters are represented by functions -in the `wsi/filter.py` file and have `filter_` prepended to the function names. - - -#### RGB to Grayscale - -A very common task in image processing is to convert an RGB image to a grayscale image. In this process, the three -color channels are replaced by a single grayscale channel. The grayscale pixel value is computed by combining the -red, green, and blue values in set percentages. The `filter_rgb_to_grayscale()` function multiplies the red value by -21.25%, the green value by 71.54%, and the blue value by 7.21%, and these values are added together to obtain the -grayscale pixel value. - -Although the PIL Image `convert("L")` function can also be used to convert an RGB image to a grayscale image, we -will instead use the `filter_rgb_to_grayscale()` function, since having a reference to the RGB image as a NumPy array -can often be very useful during image processing. - -Below, we'll open a slide as a PIL Image, convert this to an RGB NumPy array, and then convert this to a grayscale -NumPy array. - - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -grayscale = filter.filter_rgb_to_grayscale(rgb) -util.display_img(grayscale, "Grayscale") -``` - -Here we see the displayed grayscale image. - -| **Grayscale Filter** | -| -------------------- | -| ![Grayscale Filter](images/grayscale.png "Grayscale Filter") | - - -In the console, we see that the grayscale image is a two-dimensional NumPy array, since the 3 color channels have -been combined into a single grayscale channel. The data type is `uint8` and pixels are represented by integer -values between 0 and 255. - - -``` -RGB | Time: 0:00:00.159974 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.101953 Type: uint8 Shape: (1385, 1810) -``` - - -#### Complement - -In our whole-slide image training set, the slide backgrounds are illuminated by white light, which means that a `uint8` -pixel in the background of a grayscale image is usually close to or equal to 255. However, conceptually and -mathematically it is often useful to have background values close to or equal to 0. For example, this is useful in -thresholding, where we might ask if a pixel value is above a particular threshold value. This can also be useful in -masking out a background of 0 values from an image. - -The `filter_complement()` function inverts the values and thus the colors in the NumPy array representation of an image. -Below, we use the `filter_complement()` function to invert the previously obtained grayscale image. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -grayscale = filter.filter_rgb_to_grayscale(rgb) -complement = filter.filter_complement(grayscale) -util.display_img(complement, "Complement") -``` - -| **Complement Filter** | -| -------------------- | -| ![Complement Filter](images/complement.png "Complement Filter") | - - -In the console output, we see that computing the complement is a very fast operation. - -``` -RGB | Time: 0:00:00.177398 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.105015 Type: uint8 Shape: (1385, 1810) -Complement | Time: 0:00:00.001439 Type: uint8 Shape: (1385, 1810) -``` - - -#### Thresholding - - -##### Basic Threshold - -With basic thresholding, a binary image is generated, where each value in the resulting NumPy array indicates -whether the corresponding pixel in the original image is above a particular threshold value. So, a -pixel with a value of 160 with a threshold of 150 would generate a True (or 255, or 1.0), and a pixel with a value -of 140 with a threshold of 150 would generate a False (or 0, or 0.0). - -Here, we apply a basic threshold with a threshold value of 100 to the grayscale complement of the original image. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -grayscale = filter.filter_rgb_to_grayscale(rgb) -complement = filter.filter_complement(grayscale) -thresh = filter.filter_threshold(complement, threshold=100) -util.display_img(thresh, "Threshold") -``` - -The result is a binary image where pixel values that were above 100 are shown in white and pixel values that were 100 or -lower are shown in black. - -| **Basic Threshold Filter** | -| -------------------- | -| ![Basic Threshold Filter](images/basic-threshold.png "Basic Threshold Filter") | - - -In the console output, we see that basic thresholding is a very fast operation. - -``` -RGB | Time: 0:00:00.164464 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.102431 Type: uint8 Shape: (1385, 1810) -Complement | Time: 0:00:00.001397 Type: uint8 Shape: (1385, 1810) -Threshold | Time: 0:00:00.001456 Type: bool Shape: (1385, 1810) -``` - - -##### Hysteresis Threshold - -Hysteresis thresholding is a two-level threshold. The top-level threshold is treated in a similar fashion as basic -thresholding. The bottom-level threshold must be exceeded and must be connected to the top-level threshold. This -processes typically results in much better thresholding than basic thresholding. Reasonable values for the top -and bottom thresholds for images can be determined through experimentation. - -The `filter_hysteresis_threshold()` function uses default bottom and top threshold values of 50 and 100. The -default array output type from this function is `uint8`. Since the output of this function is a binary image, the -values in the output array will be either 255 or 0. The output type of this function can be specified using the -`output_type` parameter. Note that when performing masking, it is typically more useful to have a NumPy array of -boolean values. - -Here, we perform a hysteresis threshold on the complement of the grayscale image. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -grayscale = filter.filter_rgb_to_grayscale(rgb) -complement = filter.filter_complement(grayscale) -hyst = filter.filter_hysteresis_threshold(complement) -util.display_img(hyst, "Hysteresis Threshold") -``` - -In the displayed image, the result is a binary image. All pixel values are either white (255) or black (0). -The red display text in the corner can be ignored since it is for informational purposes only and is not present when -we save the images to the file system. - -Notice that the shadow area along the top edge of the slide makes it through the hysteresis threshold filter even -though conceptually it is background and should not be treated as tissue. - -| **Hysteresis Threshold Filter** | -| -------------------- | -| ![Hysteresis Threshold Filter](images/hysteresis-threshold.png "Hysteresis Threshold Filter") | - - -Here we see the console output from our filter operations. - -``` -RGB | Time: 0:00:00.167947 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.109109 Type: uint8 Shape: (1385, 1810) -Complement | Time: 0:00:00.001453 Type: uint8 Shape: (1385, 1810) -Hysteresis Threshold | Time: 0:00:00.079292 Type: uint8 Shape: (1385, 1810) -``` - - -##### Otsu Threshold - -Thresholding using Otsu's method is another popular thresholding technique. This technique was used in the image -processing described in [A Unified Framework for Tumor Proliferation Score Prediction in Breast -Histopathology](https://pdfs.semanticscholar.org/7d9b/ccac7a9a850cc84a980e5abeaeac2aef94e6.pdf). This technique is -described in more detail at -[https://en.wikipedia.org/wiki/Otsu%27s_method](https://en.wikipedia.org/wiki/Otsu%27s_method). - -Let's try Otsu's method on the complement image as we did when demonstrating hysteresis thresholding. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -grayscale = filter.filter_rgb_to_grayscale(rgb) -complement = filter.filter_complement(grayscale) -otsu = filter.filter_otsu_threshold(complement) -util.display_img(otsu, "Otsu Threshold") -``` - - -In the resulting image, we see that Otsu's method generates roughly similar results as hysteresis thresholding. -However, Otsu's method is less aggressive in terms of what it lets through for the tissue in the upper left -area of the slide. The background shadow area at the top of the slide is passed through the -filter in a similar fashion as hysteresis thresholding. Most of the slides in the training set do not have such a -pronounced shadow area, but it would be nice to have an image processing solution that treats the shadow area as -background. - -| **Otsu Threshold Filter** | -| -------------------- | -| ![Otsu Threshold Filter](images/otsu-threshold.png "Otsu Threshold Filter") | - - -In terms of performance, thresholding using Otsu's method is very fast, as we see in the console output. - -``` -RGB | Time: 0:00:00.166855 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.111960 Type: uint8 Shape: (1385, 1810) -Complement | Time: 0:00:00.001746 Type: uint8 Shape: (1385, 1810) -Otsu Threshold | Time: 0:00:00.014615 Type: uint8 Shape: (1385, 1810) -``` - - -#### Contrast - -For an image, suppose we have a histogram of the number of pixels (intensity on y-axis) plotted against the range -of possible pixel values (x-axis, 0 to 255). Contrast is a measure of the difference in intensities. An image with -low contrast is typically dull and details are not clearly seen visually. An image with high contrast is typically -sharp and details can clearly be discerned. Increasing the contrast in an image can be used to bring out various details -in the image. - - -##### Contrast Stretching - -One form of increasing the contrast in an image is contrast stretching. Suppose that all intensities in an image occur -between 100 and 150 on a scale from 0 to 255. If we rescale the intensities so that 100 now corresponds to 0 and -150 corresponds to 255 and we linearly rescale the intensities between these points, we have increased the contrast -in the image and differences in detail can more clearly be seen. This is contrast stretching. - -As an example, here we perform contrast stretching with a low pixel value of 100 and a high pixel value of 200 on -the complement of the grayscale image. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -grayscale = filter.filter_rgb_to_grayscale(rgb) -complement = filter.filter_complement(grayscale) -contrast_stretch = filter.filter_contrast_stretch(complement, low=100, high=200) -util.display_img(contrast_stretch, "Contrast Stretch") -``` - -This can be used to visually inspect details in the previous intensity range of 100 to 200, since the image filter has -spread out this range across the full spectrum. - - -| **Contrast Stretching Filter** | -| -------------------- | -| ![Contrast Stretching Filter](images/contrast-stretching.png "Contrast Stretching Filter") | - - -Here we see the console output from this set of filters. - -``` -RGB | Time: 0:00:00.171582 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.110818 Type: uint8 Shape: (1385, 1810) -Complement | Time: 0:00:00.002410 Type: uint8 Shape: (1385, 1810) -Contrast Stretch | Time: 0:00:00.058357 Type: uint8 Shape: (1385, 1810) -``` - - -##### Histogram Equalization - -Histogram equalization is another technique that can be used to increase contrast in an image. However, unlike -contrast stretching, which has a linear distribution of the resulting intensities, the histogram equalization -transformation is based on probabilities and is non-linear. For more information about histogram equalization, please -see [https://en.wikipedia.org/wiki/Histogram_equalization](https://en.wikipedia.org/wiki/Histogram_equalization). - -As an example, here we display the grayscale image. We increase contrast in the grayscale image using histogram -equalization and display the resulting image. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -grayscale = filter.filter_rgb_to_grayscale(rgb) -util.display_img(grayscale, "Grayscale") -hist_equ = filter.filter_histogram_equalization(grayscale) -util.display_img(hist_equ, "Histogram Equalization") -``` - -Comparing the grayscale image and the image after histogram equalization, we see that contrast in the image has been -increased. - -| **Grayscale Filter** | **Histogram Equalization Filter** | -| -------------------- | --------------------------------- | -| ![Grayscale Filter](images/grayscale.png "Grayscale Filter") | ![Histogram Equalization Filter](images/histogram-equalization.png "Histogram Equalization Filter") | - - -Console output following histogram equalization is shown here. - -``` -RGB | Time: 0:00:00.175498 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.110181 Type: uint8 Shape: (1385, 1810) -Hist Equalization | Time: 0:00:00.116568 Type: uint8 Shape: (1385, 1810) -``` - - -##### Adaptive Equalization - -Rather than applying a single transformation to all pixels in an image, adaptive histogram equalization applies -transformations to local regions in an image. As a result, adaptive equalization allows contrast to be enhanced to -different extents in different regions based on the regions' intensity histograms. For more information about adaptive -equalization, please see -[https://en.wikipedia.org/wiki/Adaptive_histogram_equalization](https://en.wikipedia.org/wiki/Adaptive_histogram_equalization). - -The `filter_adaptive_equalization()` function utilizes the scikit-image contrast limited adaptive histogram -equalization (CLAHE) implementation. Below, we apply adaptive equalization to the grayscale image and display both -the grayscale image and the image after adaptive equalization for comparison. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -grayscale = filter.filter_rgb_to_grayscale(rgb) -util.display_img(grayscale, "Grayscale") -adaptive_equ = filter.filter_adaptive_equalization(grayscale) -util.display_img(adaptive_equ, "Adaptive Equalization") -``` - -| **Grayscale Filter** | **Adaptive Equalization Filter** | -| -------------------- | --------------------------------- | -| ![Grayscale Filter](images/grayscale.png "Grayscale Filter") | ![Adaptive Equalization Filter](images/adaptive-equalization.png "Adaptive Equalization Filter") | - - -In the console output, we can see that adaptive equalization is more compute-intensive than constrast stretching and -histogram equalization. - -``` -RGB | Time: 0:00:00.167076 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.106797 Type: uint8 Shape: (1385, 1810) -Adapt Equalization | Time: 0:00:00.223172 Type: uint8 Shape: (1385, 1810) -``` - - -#### Color - -The WSI tissue samples in the training dataset have been H&E stained. Eosin stains basic structures such as -most cytoplasm proteins with a pink tone. Hematoxylin stains acidic structures such as DNA and RNA with a purple -tone. This means that cells tend to be stained pink, and particular areas of the cells such as the nuclei tend to be -stained purple. However, note that appearance can vary greatly based on the types of cells that are stained and the -amounts of stain applied. - -As an example of staining differences, below we see a slide that has pink and purple staining next to another slide -where all tissue appears purple. - -| **Pink and Purple Slide** | **Purple Slide** | -| -------------------- | --------------------------------- | -| ![Pink and Purple Slide](images/pink-and-purple-slide.png "Pink and Purple Slide") | ![Purple Slide](images/purple-slide.png "Purple Slide") | - - -Another factor regarding color is that many slides have been marked with red, green, and blue pens. Whereas in general -we would like our filters to include pink and purple colors, since these typically indicate stained tissue, we would -like our filters to exclude red, green, and blue colors, since these typically indicate pen marks on the slides which -are not tissue. - -Below, we see an example of a slide that has been marked with red pen and some green pen. - -| **Slide Marked with Red and Green Pen** | -| -------------------- | -| ![Slide Marked with Red and Green Pen](images/slide-pen.png "Slide Marked with Red and Green Pen") | - - -Developing color filters that can be used to filter tissue areas can be fairly challenging for a variety of reasons, -including: - -1. Filters need to be general enough to work across all slides in the dataset. -2. Filters should handle issues such as variations in shadows and lighting. -3. The amount of H&E (purple and pink) staining can vary greatly from slide to slide. -4. Pen mark colors (red, green, and blue) vary due to issues such as lighting and pen marks over tissue. -5. There can be color overlap between stained tissue and pen marks, so we need to balance how aggressively stain -colors are inclusively filtered and how pen colors are exclusively filtered. - - -##### RGB to HED - -The scikit-image `skimage.color` package features an `rgb2hed()` function that performs color deconvolution on the -original RGB image to create HED (Hematoxylin, Eosin, Diaminobenzidine) channels. The `filter_rgb_to_hed()` function -encapsulates `rgb2hed()`. The `filter_hed_to_hematoxylin()` and `filter_hed_to_eosin()` functions read the hematoxylin -and eosin channels and rescale the resulting 2-dimensional NumPy arrays (for example, 0 to 255 for `uint8`) -to increase contrast. - -Here, we'll convert the RGB image to an HED image. We'll then obtain the hematoxylin and eosin channels and display -the resulting images. - -``` -img_path = slide.get_training_image_path(4) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -hed = filter.filter_rgb_to_hed(rgb) -hema = filter.filter_hed_to_hematoxylin(hed) -util.display_img(hema, "Hematoxylin Channel") -eosin = filter.filter_hed_to_eosin(hed) -util.display_img(eosin, "Eosin Channel") -``` - -Notice that the hematoxylin channel does fairly well at detecting the purple areas of the original slide, -which could potentially be used to narrow in on tissue with cell nuclei and thus on regions that can be inspected for -mitoses. Both the hematoxylin and eosin channel filters include the background in the resulting image, which is -rather unfortunate in terms of differentiating tissue from non-tissue. Also, notice in the eosin channel that the red -pen is considered to be part of the eosin stain spectrum. - - -| **Hematoxylin Channel** | **Eosin Channel** | -| -------------------- | --------------------------------- | -| ![Hematoxylin Channel](images/hematoxylin-channel.png "Hematoxylin Channel") | ![Eosin Channel](images/eosin-channel.png "Eosin Channel") | - - -Console output: - -``` -RGB | Time: 0:00:00.397570 Type: uint8 Shape: (2594, 2945, 3) -RGB to HED | Time: 0:00:01.322220 Type: uint8 Shape: (2594, 2945, 3) -HED to Hematoxylin | Time: 0:00:00.136750 Type: uint8 Shape: (2594, 2945) -HED to Eosin | Time: 0:00:00.086537 Type: uint8 Shape: (2594, 2945) -``` - - -##### Green Channel Filter - -If we look at an RGB color wheel, we see that purple and pink are next to each other. On the other side of color wheel, -we have yellow and green. Since green is one of our 3 NumPy array RGB color channels, filtering out pixels that have a -high green channel value can be one way to potentially filter out parts of the slide that are not pink or purple. This -includes the white background, since white also has a high green channel value along with high red and blue channel -values. - -We'll use the default green threshold value of 200 for the `filter_green_channel()` function, meaning that any pixels -with green channel values of 200 or greater will be rejected. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -not_green = filter.filter_green_channel(rgb) -util.display_img(not_green, "Green Channel Filter") -``` - -The green channel filter does a decent job of differentiating the tissue from the white background. However, notice -that the shadow area at the top of the slide is not excluded by the filter. - -| **Original Slide** | **Green Channel Filter** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/display-image-with-text.png "Original Slide") | ![Green Channel Filter](images/green-channel-filter.png "Green Channel Filter") | - -A filter such as the green channel filter most likely would be used in conjunction with other filters for masking -purposes. As a result, the default output type for the green channel filter is `bool`, as we see in the console -output. If another output type is desired, this can be set with the function's `output_type` parameter. - -``` -RGB | Time: 0:00:00.169249 Type: uint8 Shape: (1385, 1810, 3) -Filter Green Channel | Time: 0:00:00.005618 Type: bool Shape: (1385, 1810) -``` - - -##### Grays Filter - -Next, let's utilize a filter that can filter out the annoying shadow area at the top of slide #2. Notice that the -shadow area consists of a gradient of dark-to-light grays. A gray pixel has red, green, and blue channel values that -are close together. The `filter_grays()` function filters out pixels that have red, blue, and green values that -are within a certain tolerance of each other. The default tolerance for `filter_grays()` is 15. The grays filter -also filters out white and black pixels, since they have similar red, green, and blue values. - -Here, we run the grays filter on the original RGB image. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -not_grays = filter.filter_grays(rgb) -util.display_img(not_grays, "Grays Filter") -``` - -Notice that in addition to filtering out the white background, the grays filter has indeed filtered out the shadow -area at the top of the slide. - -| **Original Slide** | **Grays Filter** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/display-image-with-text.png "Original Slide") | ![Grays Filter](images/grays-filter.png "Grays Filter") | - -Like the green channel filter, the default type of the returned array is `bool` since the grays filter will typically -be used in combination with other filters. Since the grays filter is fast, it offers a potentially -low-cost way to filter out shadows from the slides during preprocessing. - -``` -RGB | Time: 0:00:00.169642 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.082075 Type: bool Shape: (1385, 1810) -``` - - -##### Red Filter - -Next, let's turn our attention to filtering out shades of red, which can be used to filter out a significant amount of -the red pen color. The red pen consists of a wide variety of closely related red shades. Certain shades are -reddish, others are maroonish, and others are pinkish, for example. These color gradations are a result of a variety of -factors, such as the amount of ink, lighting, shadowing, and tissue under the pen marks. - -The `filter_red()` function filters out reddish colors through a red channel lower threshold value, a green channel -upper threshold value, and a blue channel upper threshold value. The generated mask is based on a pixel being above -the red channel threshold value and below the green and blue channel threshold values. One way to determine these -values is to display the slide image in a web browser and use a tool such as the Chrome ColorPick Eyedropper to -click on a red pen pixel to determine the approximate red, green, and blue values. - -In this example with slide #4, we use a red threshold value of 150, a green threshold value of 80, and a blue -threshold value of 90. In addition, to help us visualize the filter results, we apply the red filter to the -original RGB image as a mask and also apply the inverse of the red filter to the original image as a mask. - -``` -img_path = slide.get_training_image_path(4) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -not_red = filter.filter_red(rgb, red_lower_thresh=150, green_upper_thresh=80, blue_upper_thresh=90, display_np_info=True) -util.display_img(not_red, "Red Filter (150, 80, 90)") -util.display_img(util.mask_rgb(rgb, not_red), "Not Red") -util.display_img(util.mask_rgb(rgb, ~not_red), "Red") -``` - -We see that the red filter filters out much of the red pen. - -| **Original Slide** | **Red Filter** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/slide-4-rgb.png "Original Slide") | ![Red Filter](images/red-filter.png "Red Filter") | - - -Applying the red filter and the inverse of the red filter as masks to the original image, we see that our threshold -values did quite well at filtering out a large amount of the red pen. - -| **Not Red** | **Red** | -| -------------------- | --------------------------------- | -| ![Not Red](images/not-red.png "Not Red") | ![Red](images/red.png "Red") | - - -Here we see the console output from the above image filtering: - -``` -RGB | Time: 0:00:00.404069 Type: uint8 Shape: (2594, 2945, 3) -Filter Red | Time: 0:00:00.034864 Type: bool Shape: (2594, 2945) -Mask RGB | Time: 0:00:00.053997 Type: uint8 Shape: (2594, 2945, 3) -Mask RGB | Time: 0:00:00.022750 Type: uint8 Shape: (2594, 2945, 3) -``` - - -##### Red Pen Filter - -Next, let's turn our attention to a more inclusive red pen filter that handles more shades of red. Since the -`filter_red()` function returns a boolean array result, we can combine multiple sets of `filter_red()` threshold -values (`red_lower_thresh`, `green_upper_thresh`, `blue_upper_thresh`) using boolean operators such as `&`. We can -determine these values using a color picker tool such as the Chrome ColorPick Eyedropper, as mentioned previously. -In addition to determining various shades of red pen on a single slide, shades of red pen from other slides should be -identified and included. Note that we need to be careful with pinkish shades of red due to the similarity of these -shades to eosin staining. - -Using the color picker technique, the `filter_red_pen()` function utilizes the following sets of red threshold values. - -``` -result = filter_red(rgb, red_lower_thresh=150, green_upper_thresh=80, blue_upper_thresh=90) & \ - filter_red(rgb, red_lower_thresh=110, green_upper_thresh=20, blue_upper_thresh=30) & \ - filter_red(rgb, red_lower_thresh=185, green_upper_thresh=65, blue_upper_thresh=105) & \ - filter_red(rgb, red_lower_thresh=195, green_upper_thresh=85, blue_upper_thresh=125) & \ - filter_red(rgb, red_lower_thresh=220, green_upper_thresh=115, blue_upper_thresh=145) & \ - filter_red(rgb, red_lower_thresh=125, green_upper_thresh=40, blue_upper_thresh=70) & \ - filter_red(rgb, red_lower_thresh=200, green_upper_thresh=120, blue_upper_thresh=150) & \ - filter_red(rgb, red_lower_thresh=100, green_upper_thresh=50, blue_upper_thresh=65) & \ - filter_red(rgb, red_lower_thresh=85, green_upper_thresh=25, blue_upper_thresh=45) -``` - -Let's apply the red pen filter to slide #4. - -``` -img_path = slide.get_training_image_path(4) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -not_red_pen = filter.filter_red_pen(rgb) -util.display_img(not_red_pen, "Red Pen Filter") -util.display_img(util.mask_rgb(rgb, not_red_pen), "Not Red Pen") -util.display_img(util.mask_rgb(rgb, ~not_red_pen), "Red Pen") -``` - -| **Original Slide** | **Red Pen Filter** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/slide-4-rgb.png "Original Slide") | ![Red Pen Filter](images/red-pen-filter.png "Red Pen Filter") | - -Compared with using a single set of red threshold values, we can see that the red pen filter is significantly -more inclusive in terms of the shades of red that are accepted. As a result, more red pen is filtered. However, notice -that some of the pinkish-red from eosin-stained tissue is also included as a result of this more aggressive filtering. - - -| **Not Red Pen** | **Red Pen** | -| -------------------- | --------------------------------- | -| ![Not Red Pen](images/not-red-pen.png "Not Red Pen") | ![Red Pen](images/red-pen.png "Red Pen") | - - -Even though the red pen filter ANDs nine sets of red filter results together, we see that the performance is excellent. - -``` -RGB | Time: 0:00:00.392082 Type: uint8 Shape: (2594, 2945, 3) -Filter Red Pen | Time: 0:00:00.251170 Type: bool Shape: (2594, 2945) -Mask RGB | Time: 0:00:00.037256 Type: uint8 Shape: (2594, 2945, 3) -Mask RGB | Time: 0:00:00.026589 Type: uint8 Shape: (2594, 2945, 3) -``` - -##### Blue Filter - -If we visually examine the 500 slides in the training dataset, we see that several of the slides have been marked -with blue pen. Rather than blue lines, many of the blue marks consist of blue dots surrounding particular areas of -interest on the slides, although this is not always the case. Some of the slides also have blue pen lines. Once again, -the blue pen marks consist of several gradations of blue. - -We'll start by creating a filter to filter out blue. The `filter_blue()` function operates in a similar way as the -`filter_red()` function. It takes a red channel upper threshold value, a green channel upper threshold value, and -a blue channel lower threshold value. The generated mask is based on a pixel being below the red channel threshold -value, below the green channel threshold value, and above the blue channel threshold value. - -Once again, we'll apply the results of the blue filter and the inverse of the blue filter as masks to the original -RGB image to help visualize the filter results. - -``` -img_path = slide.get_training_image_path(241) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -not_blue = filter.filter_blue(rgb, red_upper_thresh=130, green_upper_thresh=155, blue_lower_thresh=180, display_np_info=True) -util.display_img(not_blue, "Blue Filter (130, 155, 180)") -util.display_img(util.mask_rgb(rgb, not_blue), "Not Blue") -util.display_img(util.mask_rgb(rgb, ~not_blue), "Blue") -``` - -We see that a lot of the blue pen has been filtered out. - -| **Original Slide** | **Blue Filter** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/blue-original.png "Original Slide") | ![Blue Filter](images/blue-filter.png "Blue Filter") | - - -| **Not Blue** | **Blue** | -| -------------------- | --------------------------------- | -| ![Not Blue](images/not-blue.png "Not Blue") | ![Blue](images/blue.png "Blue") | - - -Console output: - -``` -RGB | Time: 0:00:00.432772 Type: uint8 Shape: (2058, 3240, 3) -Filter Blue | Time: 0:00:00.029066 Type: bool Shape: (2058, 3240) -Mask RGB | Time: 0:00:00.038966 Type: uint8 Shape: (2058, 3240, 3) -Mask RGB | Time: 0:00:00.021153 Type: uint8 Shape: (2058, 3240, 3) -``` - - -##### Blue Pen Filter - -In `filter_blue_pen()`, we AND together various blue shade ranges using `filter_blue()` with -sets of red, green, and blue threshold values to create a blue pen filter that filters out various shades of blue. - -``` -result = filter_blue(rgb, red_upper_thresh=60, green_upper_thresh=120, blue_lower_thresh=190) & \ - filter_blue(rgb, red_upper_thresh=120, green_upper_thresh=170, blue_lower_thresh=200) & \ - filter_blue(rgb, red_upper_thresh=175, green_upper_thresh=210, blue_lower_thresh=230) & \ - filter_blue(rgb, red_upper_thresh=145, green_upper_thresh=180, blue_lower_thresh=210) & \ - filter_blue(rgb, red_upper_thresh=37, green_upper_thresh=95, blue_lower_thresh=160) & \ - filter_blue(rgb, red_upper_thresh=30, green_upper_thresh=65, blue_lower_thresh=130) & \ - filter_blue(rgb, red_upper_thresh=130, green_upper_thresh=155, blue_lower_thresh=180) & \ - filter_blue(rgb, red_upper_thresh=40, green_upper_thresh=35, blue_lower_thresh=85) & \ - filter_blue(rgb, red_upper_thresh=30, green_upper_thresh=20, blue_lower_thresh=65) & \ - filter_blue(rgb, red_upper_thresh=90, green_upper_thresh=90, blue_lower_thresh=140) & \ - filter_blue(rgb, red_upper_thresh=60, green_upper_thresh=60, blue_lower_thresh=120) & \ - filter_blue(rgb, red_upper_thresh=110, green_upper_thresh=110, blue_lower_thresh=175) -``` - -We apply the filter and its inverse to the original slide to help us visualize the results. - -``` -img_path = slide.get_training_image_path(241) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -not_blue_pen = filter.filter_blue_pen(rgb) -util.display_img(not_blue_pen, "Blue Pen Filter") -util.display_img(util.mask_rgb(rgb, not_blue_pen), "Not Blue Pen") -util.display_img(util.mask_rgb(rgb, ~not_blue_pen), "Blue Pen") -``` - -For this slide, we see that `filter_blue_pen()` filters out more blue than the previous `filter_blue()` example. - -| **Original Slide** | **Blue Pen Filter** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/blue-original.png "Original Slide") | ![Blue Pen Filter](images/blue-pen-filter.png "Blue Pen Filter") | - - -| **Not Blue Pen** | **Blue Pen** | -| -------------------- | --------------------------------- | -| ![Not Blue Pen](images/not-blue-pen.png "Not Blue Pen") | ![Blue Pen](images/blue-pen.png "Blue Pen") | - - -We see from the console output that the blue pen filter is quite fast. - -``` -RGB | Time: 0:00:00.348514 Type: uint8 Shape: (2058, 3240, 3) -Filter Blue Pen | Time: 0:00:00.288286 Type: bool Shape: (2058, 3240) -Mask RGB | Time: 0:00:00.033348 Type: uint8 Shape: (2058, 3240, 3) -Mask RGB | Time: 0:00:00.019622 Type: uint8 Shape: (2058, 3240, 3) -``` - -As an aside, we can quantify the differences in filtering between the `filter_blue()` and `filter_blue_pen()` -results. - -``` -not_blue = filter.filter_blue(rgb, red_upper_thresh=130, green_upper_thresh=155, blue_lower_thresh=180, display_np_info=True) -not_blue_pen = filter.filter_blue_pen(rgb) -print("filter_blue: " + filter.mask_percentage_text(filter.mask_percent(not_blue))) -print("filter_blue_pen: " + filter.mask_percentage_text(filter.mask_percent(not_blue_pen))) -``` - -The `filter_blue()` example filters out 0.45% of the slide pixels and the `filter_blue_pen()` example filters out -0.69% of the slide pixels. - -``` -filter_blue: 0.45% -filter_blue_pen: 0.69% -``` - -##### Green Filter - -We utilize the `filter_green()` function to filter green color shades. Using a color picker tool, -if we examine the green pen marks on the slides, the green and blue channel -values for pixels appear to track together. As a result of this, this function has a red channel upper -threshold value, a green channel lower threshold value, and a blue channel lower threshold value. - -``` -img_path = slide.get_training_image_path(51) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -not_green = filter.filter_green(rgb, red_upper_thresh=150, green_lower_thresh=160, blue_lower_thresh=140, display_np_info=True) -util.display_img(not_green, "Green Filter (150, 160, 140)") -util.display_img(util.mask_rgb(rgb, not_green), "Not Green") -util.display_img(util.mask_rgb(rgb, ~not_green), "Green") -``` - -Using a red upper threshold of 150, a green lower threshold of 160, and a blue lower threshold of 140, we see that the -much of the green ink above the background is filtered out, but most of the green ink above the tissue is not filtered -out. - -| **Original Slide** | **Green Filter** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/green-original.png "Original Slide") | ![Green Filter](images/green-filter.png "Green Filter") | - - -| **Not Green** | **Green** | -| -------------------- | --------------------------------- | -| ![Not Green](images/not-green.png "Not Green") | ![Green](images/green.png "Green") | - - -Console output: - -``` -RGB | Time: 0:00:00.611914 Type: uint8 Shape: (2291, 3839, 3) -Filter Green | Time: 0:00:00.077429 Type: bool Shape: (2291, 3839) -Mask RGB | Time: 0:00:00.049026 Type: uint8 Shape: (2291, 3839, 3) -Mask RGB | Time: 0:00:00.027211 Type: uint8 Shape: (2291, 3839, 3) -``` - -##### Green Pen Filter - -To handle the green pen shades, the `filter_green_pen()` function combines different shade results using sets of -red, green, and blue threshold values passed to the `filter_green()` function. - -``` -result = filter_green(rgb, red_upper_thresh=150, green_lower_thresh=160, blue_lower_thresh=140) & \ - filter_green(rgb, red_upper_thresh=70, green_lower_thresh=110, blue_lower_thresh=110) & \ - filter_green(rgb, red_upper_thresh=45, green_lower_thresh=115, blue_lower_thresh=100) & \ - filter_green(rgb, red_upper_thresh=30, green_lower_thresh=75, blue_lower_thresh=60) & \ - filter_green(rgb, red_upper_thresh=195, green_lower_thresh=220, blue_lower_thresh=210) & \ - filter_green(rgb, red_upper_thresh=225, green_lower_thresh=230, blue_lower_thresh=225) & \ - filter_green(rgb, red_upper_thresh=170, green_lower_thresh=210, blue_lower_thresh=200) & \ - filter_green(rgb, red_upper_thresh=20, green_lower_thresh=30, blue_lower_thresh=20) & \ - filter_green(rgb, red_upper_thresh=50, green_lower_thresh=60, blue_lower_thresh=40) & \ - filter_green(rgb, red_upper_thresh=30, green_lower_thresh=50, blue_lower_thresh=35) & \ - filter_green(rgb, red_upper_thresh=65, green_lower_thresh=70, blue_lower_thresh=60) & \ - filter_green(rgb, red_upper_thresh=100, green_lower_thresh=110, blue_lower_thresh=105) & \ - filter_green(rgb, red_upper_thresh=165, green_lower_thresh=180, blue_lower_thresh=180) & \ - filter_green(rgb, red_upper_thresh=140, green_lower_thresh=140, blue_lower_thresh=150) & \ - filter_green(rgb, red_upper_thresh=185, green_lower_thresh=195, blue_lower_thresh=195) -``` - -If we apply the green pen filter, we see that it includes most of the green shades above the tissue in slide 51. - -``` -img_path = slide.get_training_image_path(51) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "RGB") -not_green_pen = filter.filter_green_pen(rgb) -util.display_img(not_green_pen, "Green Pen Filter") -util.display_img(util.mask_rgb(rgb, not_green_pen), "Not Green Pen") -util.display_img(util.mask_rgb(rgb, ~not_green_pen), "Green Pen") -``` - -| **Original Slide** | **Green Pen Filter** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/green-original.png "Original Slide") | ![Green Pen Filter](images/green-pen-filter.png "Green Pen Filter") | - - -| **Not Green Pen** | **Green Pen** | -| -------------------- | --------------------------------- | -| ![Not Green Pen](images/not-green-pen.png "Not Green Pen") | ![Green Pen](images/green-pen.png "Green Pen") | - - -Like the other pen filters, the green pen filter's performance is quite good. - -``` -RGB | Time: 0:00:00.540223 Type: uint8 Shape: (2291, 3839, 3) -Filter Green Pen | Time: 0:00:00.487728 Type: bool Shape: (2291, 3839) -Mask RGB | Time: 0:00:00.044024 Type: uint8 Shape: (2291, 3839, 3) -Mask RGB | Time: 0:00:00.022867 Type: uint8 Shape: (2291, 3839, 3) -``` - - -##### K-Means Segmentation - -The scikit-image library contains functionality that allows for image segmentation using k-means clustering based -on location and color. This allows regions of similarly colored pixels to be grouped together. These regions are -colored based on the average color of the pixels in the individual regions. This could potentially be used to filter -regions based on their colors, where we could filter on pink shades for eosin-stained tissue and purple shades for -hematoxylin-stained tissue. - -The `filter_kmeans_segmentation()` function has a default value of 800 segments. We'll increase this to 3000 using -the `n_segments` parameter. In the example below, we'll perform k-means segmentation on the original image. In -addition, we'll create a threshold using Otsu's method and apply the resulting mask to the original image. We'll then -perform k-means segmentation on that image. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -kmeans_seg = filter.filter_kmeans_segmentation(rgb, n_segments=3000) -util.display_img(kmeans_seg, "K-Means Segmentation", bg=True) -otsu_mask = util.mask_rgb(rgb, filter.filter_otsu_threshold(filter.filter_complement(filter.filter_rgb_to_grayscale(rgb)), output_type="bool")) -util.display_img(otsu_mask, "Image after Otsu Mask", bg=True) -kmeans_seg_otsu = filter.filter_kmeans_segmentation(otsu_mask, n_segments=3000) -util.display_img(kmeans_seg_otsu, "K-Means Segmentation after Otsu Mask", bg=True) -``` - - -| **Original Slide** | **K-Means Segmentation** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/kmeans-original.png "Original Slide") | ![K-Means Segmentation](images/kmeans-segmentation.png "K-Means Segmentation") | - - -| **Image after Otsu Mask** | **K-Means Segmentation after Otsu Mask** | -| -------------------- | --------------------------------- | -| ![Image after Otsu Mask](images/otsu-mask.png "Image after Otsu Mask") | ![K-Means Segmentation after Otsu Mask](images/kmeans-segmentation-after-otsu.png "K-Means Segmentation after Otsu Mask") | - - -Note that there are a couple practical difficulties in terms of implementing automated tissue detection using k-means -segmentation. To begin with, due to the variation in tissue stain colors across the image dataset, it can be difficult -to filter on "pinkish" and "purplish" colors across all the slides. In addition, the k-means segmentation technique -is very computationally expensive, as we can see in the console output. The compute time increases with the number -of segments. For 3000 segments, we have a filter time of ~20 seconds, whereas all operations that we have seen up to -this point are subsecond. If we use the default value of 800 segments, compute time for the k-means segmentation filter -is ~7 seconds. - -``` -RGB | Time: 0:00:00.172848 Type: uint8 Shape: (1385, 1810, 3) -K-Means Segmentation | Time: 0:00:20.238886 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.076287 Type: uint8 Shape: (1385, 1810) -Complement | Time: 0:00:00.000374 Type: uint8 Shape: (1385, 1810) -Otsu Threshold | Time: 0:00:00.013864 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.008522 Type: uint8 Shape: (1385, 1810, 3) -K-Means Segmentation | Time: 0:00:20.130044 Type: uint8 Shape: (1385, 1810, 3) -``` - ---- - -The sci-kit image library also makes it possible to combine similarly colored regions. One way to do this with the -k-means segmentation results is to build a region adjacency graph (RAG) and combine regions based on a threshold value. -The `filter_rag_threshold()` function performs k-means segmentation, builds the RAG, and allows us to pass in the RAG -threshold value. - -Here, we perform k-means segmentation, build a RAG, and apply different RAG thresholds to combine similar regions. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -rag_thresh = filter.filter_rag_threshold(rgb) -util.display_img(rag_thresh, "RAG Threshold (9)", bg=True) -rag_thresh = filter.filter_rag_threshold(rgb, threshold=1) -util.display_img(rag_thresh, "RAG Threshold (1)", bg=True) -rag_thresh = filter.filter_rag_threshold(rgb, threshold=20) -util.display_img(rag_thresh, "RAG Threshold (20)", bg=True) -``` - -| **Original Slide** | **RAG Threshold = 9** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/rag-thresh-original.png "Original Slide") | ![RAG Threshold = 9](images/rag-thresh-9.png "RAG Threshold = 9") | - - -| **RAG Threshold = 1** | **RAG Threshold = 20** | -| -------------------- | --------------------------------- | -| ![RAG Threshold = 1](images/rag-thresh-1.png "RAG Threshold = 1") | ![RAG Threshold = 20](images/rag-thresh-20.png "RAG Threshold = 20") | - - -Even using the default 800 number of segments for the k-means segmentation, we see that this technique is very -computationally expensive. - -``` -RGB | Time: 0:00:00.462239 Type: uint8 Shape: (1385, 1810, 3) -RAG Threshold | Time: 0:00:24.677776 Type: uint8 Shape: (1385, 1810, 3) -RAG Threshold | Time: 0:00:26.683581 Type: uint8 Shape: (1385, 1810, 3) -RAG Threshold | Time: 0:00:23.774296 Type: uint8 Shape: (1385, 1810, 3) -``` - - -##### RGB to HSV - -Comparing hematoxylin and eosin staining can be challenging in the RGB color space. One way to simplify -this comparison is to convert to a different color space such as HSV (Hue-Saturation-Value). -The scikit-image `skimage.color` package features an `rgb2hsv()` function that converts an RGB image -to an HSV image. The `filter_rgb_to_hsv()` function wraps this scikit-image function. -In the HSV color model, the hue is represented by 360 degrees. Purple has a hue of 270 and -pink has a hue of 330. We discuss hematoxylin and eosin stain comparison in our later discussion -of tile scoring, where we favor hematoxylin-stained tissue over eosin-stained tissue. - -As an example, in the `wsi/tiles.py` file, the `display_image_with_rgb_and_hsv_histograms()` -function takes in an image as a NumPy array in RGB color space and displays the image -along with its RGB and HSV histograms. Internally, this function utilizes the `filter_rgb_to_hsv()` -function. - - -``` -# To get around renderer issue on OSX going from Matplotlib image to NumPy image. -import matplotlib -matplotlib.use('Agg') - -from deephistopath.wsi import slide -from deephistopath.wsi import tiles -from deephistopath.wsi import util - -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -tiles.display_image_with_rgb_and_hsv_histograms(rgb) -``` - -Here we see slide #2 along with its RGB and HSV histograms. Notice that the HSV hue histogram -columns have additionally been colored based on their corresponding hue values to aid in -visual inspection. - -| **Slide 2 RGB and HSV Histograms** | -| -------------------- | -| ![Slide 2 RGB and HSV Histograms](images/slide-2-rgb-hsv.png "Slide 2 RGB and HSV Histograms") | - - -#### Morphology - -Information about image morphology can be found at -[https://en.wikipedia.org/wiki/Mathematical_morphology](https://en.wikipedia.org/wiki/Mathematical_morphology). -The primary morphology operators are erosion, dilation, opening, and closing. With erosion, pixels along the edges -of an object are removed. With dilation, pixels along the edges of an object are added. Opening is erosion followed -by dilation. Closing is dilation followed by erosion. With morphology operators, a structuring element (such as -a square, circle, cross, etc) is passed along the edges of the objects to perform the operations. Morphology operators -are typically performed on binary and grayscale images. In our examples, we apply morphology operators to binary -images (2-dimensional arrays of 2 values, such as True/False, 1.0/0.0, and 255/0). - - -##### Erosion - -Let's have a look at an erosion example. -We create a binary image by calling the `filter_grays()` function on the original RGB image. The -`filter_binary_erosion()` function uses a disk as the structuring element that erodes the edges of the -"No Grays" binary image. We demonstrate erosion with disk structuring elements of radius 5 and radius 20. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -no_grays = filter.filter_grays(rgb, output_type="bool") -util.display_img(no_grays, "No Grays", bg=True) -bin_erosion_5 = filter.filter_binary_erosion(no_grays, disk_size=5) -util.display_img(bin_erosion_5, "Binary Erosion (5)", bg=True) -bin_erosion_20 = filter.filter_binary_erosion(no_grays, disk_size=20) -util.display_img(bin_erosion_20, "Binary Erosion (20)", bg=True) -``` - -| **Original Slide** | **No Grays** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/binary-erosion-original.png "Original Slide") | ![No Grays](images/binary-erosion-no-grays.png "No Grays") | - - -| **Binary Erosion (disk_size = 5)** | **Binary Erosion (disk_size = 20)** | -| -------------------- | --------------------------------- | -| ![Binary Erosion (disk_size = 5)](images/binary-erosion-5.png "Binary Erosion (disk_size = 5)") | ![Binary Erosion (disk_size = 20)](images/binary-erosion-20.png "Binary Erosion (disk_size = 20)") | - - -Notice that increasing the structuring element radius increases the compute time. - -``` -RGB | Time: 0:00:00.171309 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.086484 Type: bool Shape: (1385, 1810) -Binary Erosion | Time: 0:00:00.167290 Type: uint8 Shape: (1385, 1810) -Binary Erosion | Time: 0:00:00.765442 Type: uint8 Shape: (1385, 1810) -``` - - -##### Dilation - -The `filter_binary_dilation()` function utilizes a disk structuring element in a similar manner as the corresponding -erosion function. We'll utilize the same "No Grays" binary image from the previous example and dilate the image -utilizing a disk radius of 5 pixels followed by a disk radius of 20 pixels. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -no_grays = filter.filter_grays(rgb, output_type="bool") -util.display_img(no_grays, "No Grays", bg=True) -bin_dilation_5 = filter.filter_binary_dilation(no_grays, disk_size=5) -util.display_img(bin_dilation_5, "Binary Dilation (5)", bg=True) -bin_dilation_20 = filter.filter_binary_dilation(no_grays, disk_size=20) -util.display_img(bin_dilation_20, "Binary Dilation (20)", bg=True) -``` - -We see that dilation expands the edges of the binary image as opposed to the erosion, which shrinks the edges. - -| **Binary Dilation (disk_size = 5)** | **Binary Dilation (disk_size = 20)** | -| -------------------- | --------------------------------- | -| ![Binary Dilation (disk_size = 5)](images/binary-dilation-5.png "Binary Dilation (disk_size = 5)") | ![Binary Dilation (disk_size = 20)](images/binary-dilation-20.png "Binary Dilation (disk_size = 20)") | - - -Console output: - -``` -RGB | Time: 0:00:00.176491 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.081817 Type: bool Shape: (1385, 1810) -Binary Dilation | Time: 0:00:00.096302 Type: uint8 Shape: (1385, 1810) -Binary Dilation | Time: 0:00:00.538761 Type: uint8 Shape: (1385, 1810) -``` - - -##### Opening - -As mentioned, opening is erosion followed by dilation. Opening can be used to remove small foreground objects. - - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -no_grays = filter.filter_grays(rgb, output_type="bool") -util.display_img(no_grays, "No Grays", bg=True) -bin_opening_5 = filter.filter_binary_opening(no_grays, disk_size=5) -util.display_img(bin_opening_5, "Binary Opening (5)", bg=True) -bin_opening_20 = filter.filter_binary_opening(no_grays, disk_size=20) -util.display_img(bin_opening_20, "Binary Opening (20)", bg=True) -``` - -| **Binary Opening (disk_size = 5)** | **Binary Opening (disk_size = 20)** | -| -------------------- | --------------------------------- | -| ![Binary Opening (disk_size = 5)](images/binary-opening-5.png "Binary Opening (disk_size = 5)") | ![Binary Opening (disk_size = 20)](images/binary-opening-20.png "Binary Opening (disk_size = 20)") | - - -Opening is a fairly expensive operation, since it is an erosion followed by a dilation. The compute time increases -with the size of the structuring element. The 5-pixel disk radius for the structuring element results in a 0.25s -operation, whereas the 20-pixel disk radius results in a 2.45s operation. - -``` -RGB | Time: 0:00:00.169241 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.085474 Type: bool Shape: (1385, 1810) -Binary Opening | Time: 0:00:00.248629 Type: uint8 Shape: (1385, 1810) -Binary Opening | Time: 0:00:02.452089 Type: uint8 Shape: (1385, 1810) -``` - - -##### Closing - -Closing is a dilation followed by an erosion. Closing can be used to remove small background holes. - - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -no_grays = filter.filter_grays(rgb, output_type="bool") -util.display_img(no_grays, "No Grays", bg=True) -bin_closing_5 = filter.filter_binary_closing(no_grays, disk_size=5) -util.display_img(bin_closing_5, "Binary Closing (5)", bg=True) -bin_closing_20 = filter.filter_binary_closing(no_grays, disk_size=20) -util.display_img(bin_closing_20, "Binary Closing (20)", bg=True) -``` - -| **Binary Closing (disk_size = 5)** | **Binary Closing (disk_size = 20)** | -| -------------------- | --------------------------------- | -| ![Binary Closing (disk_size = 5)](images/binary-closing-5.png "Binary Closing (disk_size = 5)") | ![Binary Closing (disk_size = 20)](images/binary-closing-20.png "Binary Closing (disk_size = 20)") | - - -Like opening, closing is a fairly expensive operation since it performs both a dilation and an erosion. Compute time -increases with structuring element size. - -``` -RGB | Time: 0:00:00.179190 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.079992 Type: bool Shape: (1385, 1810) -Binary Closing | Time: 0:00:00.241882 Type: uint8 Shape: (1385, 1810) -Binary Closing | Time: 0:00:02.592515 Type: uint8 Shape: (1385, 1810) -``` - - -##### Remove Small Objects - -The scikit-image `remove_small_objects()` function removes objects less than a particular minimum size. The -`filter_remove_small_objects()` function wraps this and adds additional functionality. This can be useful for -removing small islands of noise from images. We'll demonstrate it here with two sizes, 100 pixels and 10,000 pixels, -and we'll perform this on the "No Grays" binary image. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -no_grays = filter.filter_grays(rgb, output_type="bool") -util.display_img(no_grays, "No Grays", bg=True) -remove_small_100 = filter.filter_remove_small_objects(no_grays, min_size=100) -util.display_img(remove_small_100, "Remove Small Objects (100)", bg=True) -remove_small_10000 = filter.filter_remove_small_objects(no_grays, min_size=10000) -util.display_img(remove_small_10000, "Remove Small Objects (10000)", bg=True) -``` - -Notice in the "No Grays" binary image that we see lots of scattered, small objects. - -| **Original Slide** | **No Grays** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/binary-erosion-original.png "Original Slide") | ![No Grays](images/binary-erosion-no-grays.png "No Grays") | - - -After removing small objects with a connected size less than 100 pixels, we see that the smallest objects have been -removed from the binary image. With a minimum size of 10,000 pixels, we see that many larger objects have also been -removed from the binary image. - -| **Remove Small Objects (100)** | **Remove Small Objects (10000)** | -| -------------------- | --------------------------------- | -| ![Remove Small Objects (100)](images/remove-small-objects-100.png "Remove Small Objects (100)") | ![Remove Small Objects (10000)](images/remove-small-objects-10000.png "Remove Small Objects (10000)") | - - -The performance of the filters to remove small objects is quite fast. - -``` -RGB | Time: 0:00:00.177367 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.081827 Type: bool Shape: (1385, 1810) -Remove Small Objs | Time: 0:00:00.053734 Type: uint8 Shape: (1385, 1810) -Remove Small Objs | Time: 0:00:00.044924 Type: uint8 Shape: (1385, 1810) -``` - - -##### Remove Small Holes - -The scikit-image `remove_small_holes()` function is similar to the `remove_small_objects()` function except it removes -holes rather than objects from binary images. Here we demonstrate this using the `filter_remove_small_holes()` -function with sizes of 100 pixels and 10,000 pixels. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -no_grays = filter.filter_grays(rgb, output_type="bool") -util.display_img(no_grays, "No Grays", bg=True) -remove_small_100 = filter.filter_remove_small_holes(no_grays, min_size=100) -util.display_img(remove_small_100, "Remove Small Holes (100)", bg=True) -remove_small_10000 = filter.filter_remove_small_holes(no_grays, min_size=10000) -util.display_img(remove_small_10000, "Remove Small Holes (10000)", bg=True) -``` - -Notice that using a minimum size of 10,000 removes more holes than a size of 100, as we would expect. - -| **Remove Small Holes (100)** | **Remove Small Holes (10000)** | -| -------------------- | --------------------------------- | -| ![Remove Small Holes (100)](images/remove-small-holes-100.png "Remove Small Holes (100)") | ![Remove Small Holes (10000)](images/remove-small-holes-10000.png "Remove Small Holes (10000)") | - - -Console output: - -``` -RGB | Time: 0:00:00.171669 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.081116 Type: bool Shape: (1385, 1810) -Remove Small Holes | Time: 0:00:00.043491 Type: uint8 Shape: (1385, 1810) -Remove Small Holes | Time: 0:00:00.044550 Type: uint8 Shape: (1385, 1810) -``` - - -##### Fill Holes - -The scikit-image `binary_fill_holes()` function is similar to the `remove_small_holes()` function. Using its default -settings, it generates results similar but typically not identical to `remove_small_holes()` with a high minimum -size value. - -Here, we'll display the result of `filter_binary_fill_holes()` on the image after gray shades have been removed. After -this, we'll perform exclusive-or operations to look at the differences between "Fill Holes" and "Remove Small Holes" -with size values of 100 and 10,000. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -no_grays = filter.filter_grays(rgb, output_type="bool") -fill_holes = filter.filter_binary_fill_holes(no_grays) -util.display_img(fill_holes, "Fill Holes", bg=True) - -remove_holes_100 = filter.filter_remove_small_holes(no_grays, min_size=100, output_type="bool") -util.display_img(fill_holes ^ remove_holes_100, "Differences between Fill Holes and Remove Small Holes (100)", bg=True) - -remove_holes_10000 = filter.filter_remove_small_holes(no_grays, min_size=10000, output_type="bool") -util.display_img(fill_holes ^ remove_holes_10000, "Differences between Fill Holes and Remove Small Holes (10000)", bg=True) - -``` - -| **Original Slide** | **Fill Holes** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/binary-erosion-original.png "Original Slide") | ![Fill Holes](images/fill-holes.png "Fill Holes") | - - -In this example, increasing the minimum small hole size results in less differences between "Fill Holes" and -"Remove Small Holes". - -| **Differences between Fill Holes and Remove Small Holes (100)** | **Differences between Fill Holes and Remove Small Holes (10000)** | -| -------------------- | --------------------------------- | -| ![Differences between Fill Holes and Remove Small Holes (100)](images/fill-holes-remove-small-holes-100.png "Differences between Fill Holes and Remove Small Holes (100)") | ![Differences between Fill Holes and Remove Small Holes (10000)](images/fill-holes-remove-small-holes-10000.png "Differences between Fill Holes and Remove Small Holes (10000)") | - - -Console output: - -``` -RGB | Time: 0:00:00.176696 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.082582 Type: bool Shape: (1385, 1810) -Binary Fill Holes | Time: 0:00:00.069583 Type: bool Shape: (1385, 1810) -Remove Small Holes | Time: 0:00:00.046232 Type: bool Shape: (1385, 1810) -Remove Small Holes | Time: 0:00:00.044539 Type: bool Shape: (1385, 1810) -``` - - -#### Entropy - -The scikit-image `entropy()` function allows us to filter images based on complexity. Since areas such as slide -backgrounds are less complex than area of interest such as cell nuclei, filtering on entropy offers interesting -possibilities for tissue identification. - -Here, we use the `filter_entropy()` function to filter the grayscale image based on entropy. We display -the resulting binary image. After that, we mask the original image with the entropy mask and the inverse of the entropy -mask. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original") -gray = filter.filter_rgb_to_grayscale(rgb) -util.display_img(gray, "Grayscale") -entropy = filter.filter_entropy(gray, output_type="bool") -util.display_img(entropy, "Entropy") -util.display_img(util.mask_rgb(rgb, entropy), "Original with Entropy Mask") -util.display_img(util.mask_rgb(rgb, ~entropy), "Original with Inverse of Entropy Mask") -``` - -| **Original Slide** | **Grayscale** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/entropy-original.png "Original Slide") | ![Grayscale](images/entropy-grayscale.png "Grayscale") | - - -| **Entropy Filter** | -| ------------------ | -| ![Entropy Filter](images/entropy.png "Entropy Filter") | - - -The results of the original image with the inverse of the entropy mask are particularly interesting. Notice that much -of the white background including the shadow region at the top of the slide has been filtered out. Additionally, notice -that for the stained regions, a significant amount of the pink eosin-stained area has been filtered out while a -smaller proportion of the purple-stained hemotoxylin area has been filtered out. This makes sense since hemotoxylin -stains regions such as cell nuclei, which are structures with significant complexity. Therefore, entropy seems -like a potential tool that could be used to identify regions of interest where mitoses are occurring. - - -| **Original with Entropy Mask** | **Original with Inverse of Entropy Mask** | -| -------------------- | --------------------------------- | -| ![Original with Entropy Mask](images/entropy-original-entropy-mask.png "Original with Entropy Mask") | ![Original with Inverse of Entropy Mask](images/entropy-original-inverse-entropy-mask.png "Original with Inverse of Entropy Mask") | - - -A drawback of using entropy is that its computation is significant. The entropy filter takes over 3 seconds to run -in this example. - -``` -RGB | Time: 0:00:00.177166 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.116245 Type: uint8 Shape: (1385, 1810) -Entropy | Time: 0:00:03.306786 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.010422 Type: uint8 Shape: (1385, 1810, 3) -Mask RGB | Time: 0:00:00.006140 Type: uint8 Shape: (1385, 1810, 3) -``` - - -#### Canny Edge Detection - -Edges in images are areas where there is typically a significant, abrupt change in image brightness. -The Canny edge detection algorithm is implemented in sci-kit image. More information about -edge detection can be found at [https://en.wikipedia.org/wiki/Edge_detection](https://en.wikipedia.org/wiki/Edge_detection). -More information about Canny edge detection can be found at -[https://en.wikipedia.org/wiki/Canny_edge_detector](https://en.wikipedia.org/wiki/Canny_edge_detector). - -The sci-kit image `canny()` function returns a binary edge map for the detected edges in an input image. In the -example below, we call `filter_canny()` on the grayscale image and display the resulting Canny edges. -After this, we crop a 600x600 area of the original slide and display it. We apply the inverse of the -canny mask to the cropped original slide area and display it for comparison. - -``` -img_path = slide.get_training_image_path(2) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original", bg=True) -gray = filter.filter_rgb_to_grayscale(rgb) -canny = filter.filter_canny(gray, output_type="bool") -util.display_img(canny, "Canny", bg=True) -rgb_crop = rgb[300:900, 300:900] -canny_crop = canny[300:900, 300:900] -util.display_img(rgb_crop, "Original", size=24, bg=True) -util.display_img(util.mask_rgb(rgb_crop, ~canny_crop), "Original with ~Canny Mask", size=24, bg=True) -``` - -| **Original** | **Canny Edges** | -| -------------------- | --------------------------------- | -| ![Original](images/canny-original.png "Original") | ![Canny Edges](images/canny.png "Canny Edges") | - - -By applying the inverse of the canny edge mask to the original image, the detected edges are colored black. This -visually accentuates the different structures in the slide. - -| **Cropped Original** | **Cropped Original with Inverse Canny Edges Mask** | -| -------------------- | --------------------------------- | -| ![Cropped Original](images/canny-original-cropped.png "Cropped Original") | ![Cropped Original with Inverse Canny Edges Mask](images/canny-original-with-inverse-mask.png "Cropped Original with Inverse Canny Edges Mask") | - - -In the console output, we see that Canny edge detection is fairly expensive, since its computation took over 1 second. - -``` -RGB | Time: 0:00:00.174458 Type: uint8 Shape: (1385, 1810, 3) -Gray | Time: 0:00:00.116023 Type: uint8 Shape: (1385, 1810) -Canny Edges | Time: 0:00:01.017241 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.001443 Type: uint8 Shape: (600, 600, 3) -``` - - -### Combining Filters - -Since our image filters utilize NumPy arrays, it is straightforward to combine our filters. For example, when -we have filters that return boolean images for masking, we can use standard boolean algebra on our arrays to perform -operations such as AND, OR, XOR, and NOT. We can also run filters on the results of other filters. - -As an example, here we run our green pen and blue pen filters on the original RGB image to filter out the green and -blue pen marks on the slide. We combine the resulting masks with a boolean AND (&) operation. We display the resulting -mask and this mask applied to the original image, masking out the green and blue pen marks from the image. - -``` -img_path = slide.get_training_image_path(74) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original") -no_green_pen = filter.filter_green_pen(rgb) -util.display_img(no_green_pen, "No Green Pen") -no_blue_pen = filter.filter_blue_pen(rgb) -util.display_img(no_blue_pen, "No Blue Pen") -no_gp_bp = no_green_pen & no_blue_pen -util.display_img(no_gp_bp, "No Green Pen, No Blue Pen") -util.display_img(util.mask_rgb(rgb, no_gp_bp), "Original with No Green Pen, No Blue Pen") -``` - -| **Original Slide** | -| -------------------- | -| ![Original Slide](images/combine-pen-filters-original.png "Original Slide") | - -| **No Green Pen** | **No Blue Pen** | -| -------------------- | --------------------------------- | -| ![No Green Pen](images/combine-pen-filters-no-green-pen.png "No Green Pen") | ![No Blue Pen](images/combine-pen-filters-no-blue-pen.png "No Blue Pen") | - -| **No Green Pen, No Blue Pen** | **Original with No Green Pen, No Blue Pen** | -| -------------------- | --------------------------------- | -| ![No Green Pen, No Blue Pen](images/combine-pen-filters-no-green-pen-no-blue-pen.png "No Green Pen, No Blue Pen") | ![Original with No Green Pen, No Blue Pen](images/combine-pen-filters-original-with-no-green-pen-no-blue-pen.png "Original with No Green Pen, No Blue Pen") | - - -Console Output: - -``` -RGB | Time: 0:00:00.525283 Type: uint8 Shape: (2592, 3509, 3) -Filter Green Pen | Time: 0:00:00.562343 Type: bool Shape: (2592, 3509) -Filter Blue Pen | Time: 0:00:00.414910 Type: bool Shape: (2592, 3509) -Mask RGB | Time: 0:00:00.054763 Type: uint8 Shape: (2592, 3509, 3) -``` - - ---- - -Let's try another combination of filters that should give us fairly good tissue segmentation for this slide, -where the slide background and blue and green pen marks are removed. We can do this for this slide by ANDing -together the "No Grays" filter, the "Green Channel" filter, the "No Green Pen" filter, and the "No Blue Pen" filter. -In addition, we can use our "Remove Small Objects" filter to remove small islands from the mask. We display the -resulting mask. We apply this mask and the inverse of the mask to the original image to visually see which parts of the -slide are passed through and which parts are masked out. - -``` -img_path = slide.get_training_image_path(74) -img = slide.open_image(img_path) -rgb = util.pil_to_np_rgb(img) -util.display_img(rgb, "Original") -mask = filter.filter_grays(rgb) & filter.filter_green_channel(rgb) & filter.filter_green_pen(rgb) & filter.filter_blue_pen(rgb) -mask = filter.filter_remove_small_objects(mask, min_size=100, output_type="bool") -util.display_img(mask, "No Grays, Green Channel, No Green Pen, No Blue Pen, No Small Objects") -util.display_img(util.mask_rgb(rgb, mask), "Original with No Grays, Green Channel, No Green Pen, No Blue Pen, No Small Objects") -util.display_img(util.mask_rgb(rgb, ~mask), "Original with Inverse Mask") -``` - -| **Original Slide** | **No Grays, Green Channel, No Green Pen, No Blue Pen, No Small Objects** | -| -------------------- | --------------------------------- | -| ![Original Slide](images/combine-pens-background-original.png "Original Slide") | ![No Grays, Green Channel, No Green Pen, No Blue Pen, No Small Objects](images/combine-pens-background-mask.png "No Grays, Green Channel, No Green Pen, No Blue Pen, No Small Objects") | - - -We see that this combination does a good job at allowing us to filter the most relevant tissue sections of this slide. - -| **Original with No Grays, Green Channel, No Green Pen, No Blue Pen, No Small Objects** | **Original with Inverse Mask** | -| -------------------- | --------------------------------- | -| ![Original with No Grays, Green Channel, No Green Pen, No Blue Pen, No Small Objects](images/combine-pens-background-original-with-mask.png "Original with No Grays, Green Channel, No Green Pen, No Blue Pen, No Small Objects") | ![Original with Inverse Mask](images/combine-pens-background-original-with-inverse-mask.png "Original with Inverse Mask") | - - -Console Output: - -``` -RGB | Time: 0:00:00.496920 Type: uint8 Shape: (2592, 3509, 3) -Filter Grays | Time: 0:00:00.361576 Type: bool Shape: (2592, 3509) -Filter Green Channel | Time: 0:00:00.020190 Type: bool Shape: (2592, 3509) -Filter Green Pen | Time: 0:00:00.488955 Type: bool Shape: (2592, 3509) -Filter Blue Pen | Time: 0:00:00.369501 Type: bool Shape: (2592, 3509) -Remove Small Objs | Time: 0:00:00.178179 Type: bool Shape: (2592, 3509) -Mask RGB | Time: 0:00:00.047400 Type: uint8 Shape: (2592, 3509, 3) -Mask RGB | Time: 0:00:00.048710 Type: uint8 Shape: (2592, 3509, 3) -``` - - ---- - -In the `wsi/filter.py` file, the `apply_filters_to_image(slide_num, save=True, display=False)` function is the -primary way we apply a set of filters to an image with the goal of identifying the tissue in the slide. This -function allows us to see the results of each filter and the combined results of different filters. If the -`save` parameter is `True`, the various filter results will be saved to the file system. If the `display` -parameter is `True`, the filter results will be displayed on the screen. The function returns a tuple consisting of -the resulting NumPy array image and a dictionary of information that is used elsewhere for generating an HTML page -to view the various filter results for multiple slides, as we will see later. - -The `apply_filters_to_image()` function calls the `apply_image_filters()` function, which creates green channel, grays, -red pen, green pen, and blue pen masks and combines these into a single mask using boolean ANDs. -After this, small objects are removed from the mask. - -``` -mask_not_green = filter_green_channel(rgb) -mask_not_gray = filter_grays(rgb) -mask_no_red_pen = filter_red_pen(rgb) -mask_no_green_pen = filter_green_pen(rgb) -mask_no_blue_pen = filter_blue_pen(rgb) -mask_gray_green_pens = mask_not_gray & mask_not_green & mask_no_red_pen & mask_no_green_pen & mask_no_blue_pen -mask_remove_small = filter_remove_small_objects(mask_gray_green_pens, min_size=500, output_type="bool") -``` - -After each of the above masks is created, it is applied to the original image and the resulting image is saved -to the file system, displayed to the screen, or both. - -Let's try this function out. In this example, we run `apply_filters_to_image()` on slide #337 and display the results -to the screen. - -``` -filter.apply_filters_to_image(337, display=True, save=False) -``` - -Note that this function utilizes the scaled-down `png` image for slide #337. If we have not generated `png` images for -all the slides (typically by calling `slide.multiprocess_training_slides_to_images()`), we can generate the individual -scaled-down `png` image and then apply the filters to this image. - -``` -slide.training_slide_to_image(337) -filter.apply_filters_to_image(337, display=True, save=False) -``` - -Here, we see the original slide #337 and the green channel filter applied to it. The original slide is marked as 0.12% -masked because a small number of pixels in the original image are black (0 values for the red, green, and blue -channels). Notice that the green channel filter -with a default threshold of 200 removes most of the white background but only a relatively small fraction of the green -pen. The green channel filter masks 72.60% of the original slide. - -| **Slide 337, F001** | **Slide 337, F002** | -| -------------------- | --------------------------------- | -| ![Slide 337, F001](images/337-001.png "Slide 337, F001") | ![Slide 337, F002](images/337-002.png "Slide 337, F002") | - - -Here, we see the results of the grays filter and the red pen filter. For this slide, the grays filter masks 68.01% of -the slide, which is actually less than the green channel filter. The red pen filter masks only 0.18% of the slide, -which makes sense since there are no red pen marks on the slide. - -| **Slide 337, F003** | **Slide 337, F004** | -| -------------------- | --------------------------------- | -| ![Slide 337, F003](images/337-003.png "Slide 337, F003") | ![Slide 337, F004](images/337-004.png "Slide 337, F004") | - - -The green pen filter masks 3.81% of the slide. Visually, we see that it does a decent job of masking out the green -pen marks on the slide. The blue pen filter masks 0.12% of the slide, which is accurate since there are no blue pen -marks on the slide. - -| **Slide 337, F005** | **Slide 337, F006** | -| -------------------- | --------------------------------- | -| ![Slide 337, F005](images/337-005.png "Slide 337, F005") | ![Slide 337, F006](images/337-006.png "Slide 337, F006") | - - -Combining the above filters with a boolean AND results in 74.57% masking. Cleaning up these results by remove small -objects results in a masking of 76.11%. This potentially gives a good tissue segmentation that we can use for deep -learning. - -| **Slide 337, F007** | **Slide 337, F008** | -| -------------------- | --------------------------------- | -| ![Slide 337, F007](images/337-007.png "Slide 337, F007") | ![Slide 337, F008](images/337-008.png "Slide 337, F008") | - - -In the console, we see the slide #337 processing time takes ~12.6s in this example. The filtering is only a relatively -small fraction of this time. If we set `display` to `False`, processing only takes ~2.3s. - -``` -Processing slide #337 -RGB | Time: 0:00:00.568235 Type: uint8 Shape: (2515, 3149, 3) -Filter Green Channel | Time: 0:00:00.017670 Type: bool Shape: (2515, 3149) -Mask RGB | Time: 0:00:00.037547 Type: uint8 Shape: (2515, 3149, 3) -Filter Grays | Time: 0:00:00.323861 Type: bool Shape: (2515, 3149) -Mask RGB | Time: 0:00:00.032874 Type: uint8 Shape: (2515, 3149, 3) -Filter Red Pen | Time: 0:00:00.253547 Type: bool Shape: (2515, 3149) -Mask RGB | Time: 0:00:00.035073 Type: uint8 Shape: (2515, 3149, 3) -Filter Green Pen | Time: 0:00:00.395172 Type: bool Shape: (2515, 3149) -Mask RGB | Time: 0:00:00.032597 Type: uint8 Shape: (2515, 3149, 3) -Filter Blue Pen | Time: 0:00:00.314914 Type: bool Shape: (2515, 3149) -Mask RGB | Time: 0:00:00.034853 Type: uint8 Shape: (2515, 3149, 3) -Mask RGB | Time: 0:00:00.034556 Type: uint8 Shape: (2515, 3149, 3) -Remove Small Objs | Time: 0:00:00.160241 Type: bool Shape: (2515, 3149) -Mask RGB | Time: 0:00:00.030854 Type: uint8 Shape: (2515, 3149, 3) -Slide #337 processing time: 0:00:12.576835 -``` - -Since `apply_filters_to_image()` returns the resulting image as a NumPy array, we can perform further processing on -the image. If we look at the `apply_filters_to_image()` results for slide #337, we can see that some grayish greenish -pen marks remain on the slide. We can filter some of these out using our `filter_green()` function with different -threshold values and our `filter_grays()` function with an increased tolerance value. - -We'll compare the results by cropping two regions of the slide before and after this additional processing and -displaying all four of these regions together. - -``` -rgb, _ = filter.apply_filters_to_image(337, display=False, save=False) - -not_greenish = filter.filter_green(rgb, red_upper_thresh=125, green_lower_thresh=30, blue_lower_thresh=30, display_np_info=True) -not_grayish = filter.filter_grays(rgb, tolerance=30) -rgb_new = util.mask_rgb(rgb, not_greenish & not_grayish) - -row1 = np.concatenate((rgb[1200:1800, 150:750], rgb[1150:1750, 2050:2650]), axis=1) -row2 = np.concatenate((rgb_new[1200:1800, 150:750], rgb_new[1150:1750, 2050:2650]), axis=1) -result = np.concatenate((row1, row2), axis=0) -util.display_img(result) -``` - -After the additional processing, we see that the pen marks in the displayed regions have been significantly reduced. - -| **Remove More Green and More Gray** | -| -------------------- | -| ![Remove More Green and More Gray](images/remove-more-green-more-gray.png "Remove More Green and More Gray") | - - -As another example, here we can see a summary of filters applied to a slide by `apply_filters_to_image()` and the -resulting masked image. - -| **Filter Example** | -| ------------------ | -| ![Filter Example](images/filter-example.png "Filter Example") | - - -### Applying Filters to Multiple Images - -When designing our set of tissue-selecting filters, one very important requirement is the ability to visually inspect -the filter results across multiple slides. Ideally we should easily be able to alternate between displaying the -results for a single image, a select subset of our training image dataset, and our entire dataset. Additionally, -multiprocessing can result in a significant performance boost, so we should be able to multiprocess our image -processing if desired. - -A simple, powerful way to visually inspect our filter results is to generate an HTML page for a set of images. - -The following functions in `wsi/filter.py` can be used to apply filters to multiple images: - -``` -apply_filters_to_image_list(image_num_list, save, display) -apply_filters_to_image_range(start_ind, end_ind, save, display) -singleprocess_apply_filters_to_images(save=True, display=False, html=True, image_num_list=None) -multiprocess_apply_filters_to_images(save=True, display=False, html=True, image_num_list=None) - -``` - -The `apply_filters_to_image_list()` function takes a list of image numbers for processing. It does not generate an -HTML page but it does generate information that can be used by other functions to generate an HTML page. -The `save` parameter if `True` will save the generated images to the file system. If the `display` parameter -is `True`, the generated images will be displayed to the screen. If several slides are being processed, -`display` should be set to False. - -The `apply_filters_to_image_range()` function is similar to `apply_filters_to_image_list()` except than rather than -taking a list of image numbers, it takes a starting index number and ending index number for the slides in the -training set. Like `apply_filters_to_image_list()`, the `apply_filters_to_image_range()` function has `save` and -`display` parameters. - -The `singleprocess_apply_filters_to_images()` and `multiprocess_apply_filters_to_images()` functions are the -primary functions that should be called to apply filters to multiple images. Both of these functions feature `save` -and `display` parameters. The additional `html` parameter if `True` generates an HTML page for displaying the filter -results on the image set. The `singleprocess_apply_filters_to_images()` and `multiprocess_apply_filters_to_images()` -functions also feature an `image_num_list` parameter which specifies a list of image numbers that should be -processed. If `image_num_list` is not supplied, all training images are processed. - -As an example, let's use a single process to apply our filters to images 1, 2, and 3. We can accomplish this with -the following: - -``` -filter.singleprocess_apply_filters_to_images(image_num_list=[1, 2, 3]) -``` - -In addition to saving the filtered images to the file system, this creates a `filters.html` file that displays all the -filtered slide images. -If we open the `filters.html` file in a browser, we can see 8 images displayed for each slide. Each separate slide -is displayed as a separate row. Here, we see the filter results for slides #1, #2, and #3 displayed in a browser. - -| **Filters for Slides 1, 2, 3** | -| -------------------- | -| ![Filters for Slides 1, 2, 3](images/filters-001-008.png "Filters for Slides 1, 2, 3") | - - -To apply all filters to all images in the training set using multiprocessing, we can utilize the -`multiprocess_apply_filters_to_images()` function. Since there are 9 generated images per slide -(8 of which are shown in the HTML summary) and 500 slides, this results in a total of 4,500 images -and 4,500 thumbnails. Generating `png` images and `jpg` thumbnails, this takes about 24 minutes on -my MacBook Pro. - -``` -filter.multiprocess_apply_filters_to_images() -``` - -If we display the `filters.html` file in a browser, we see that the filter results for the first 50 slides are -displayed. By default, results are paginated at 50 slides per page. Pagination can be turned on and off using the -`FILTER_PAGINATE` constant. The pagination size can be adjusted using the `FILTER_PAGINATION_SIZE` constant. - -One useful action we can take is to group similar slides into categories. For example, -we can group slides into sets that have red, green, and blue pen marks on them. - -``` -red_pen_slides = [4, 15, 24, 48, 63, 67, 115, 117, 122, 130, 135, 165, 166, 185, 209, 237, 245, 249, 279, 281, 282, 289, 336, 349, 357, 380, 450, 482] -green_pen_slides = [51, 74, 84, 86, 125, 180, 200, 337, 359, 360, 375, 382, 431] -blue_pen_slides = [7, 28, 74, 107, 130, 140, 157, 174, 200, 221, 241, 318, 340, 355, 394, 410, 414, 457, 499] -``` - -We can run our filters on the list of red pen slides in the following manner: - -``` -filter.multiprocess_apply_filters_to_images(image_num_list=red_pen_slides) -``` - -In this way, we can make tweaks to specific filters or combinations of specific filters and see how these changes apply -to the subset of relevant training images without requiring reprocessing of the entire training dataset. - -| **Red Pen Slides with Filter Results** | -| -------------------- | -| ![Red Pen Slides with Filter Results](images/red-pen-slides-filters.png "Red Pen Slides with Filter Results") | - - -### Overmask Avoidance - -When developing filters and filter settings to perform tissue segmentation on the entire training -set, we have to deal with a great amount of variation in the slide samples. To begin with, some slides have a large -amount of tissue on them, while other slides only have a minimal amount of tissue. There is a great deal of -variation in tissue staining. We also need to deal with additional issues such as pen marks and shadows on some of -the slides. - -Slide #498 is an example of a slide with a large tissue sample. After filtering, the slide is 46% masked. - -| **Slide with Large Tissue Sample** | **Slide with Large Tissue Sample after Filtering** | -| -- | -- | -| ![Slide with Large Tissue Sample](images/498-rgb.png "Slide with Large Tissue Sample") | ![Slide with Large Tissue Sample after Filtering](images/498-rgb-after-filters.png "Slide with Large Tissue Sample after Filtering") | - - -Slide #127 is an example of a small tissue sample. After filtering, the slide is 93% masked. With such a small tissue -sample to begin with, we need to be careful that our filters don't overmask this slide. - -| **Slide with Small Tissue Sample** | **Slide with Small Tissue Sample after Filtering** | -| -- | -- | -| ![Slide with Small Tissue Sample](images/127-rgb.png "Slide with Small Tissue Sample") | ![Slide with Small Tissue Sample after Filtering](images/127-rgb-after-filters.png "Slide with Small Tissue Sample after Filtering") | - - -Being aggressive in our filtering may generate excellent results for many of the slides but may -result in overmasking of other slides, where the amount of non-tissue masking is too high. For example, if 99% of -a slide is masked, it has been overmasked. - -Avoiding overmasking across the entire training dataset can be difficult. For example, suppose we have a slide that -has only a proportionaly small amount of tissue on it to start, say 10%. If this particular tissue sample has been -poorly stained so that it is perhaps a light purplish grayish color, applying our grays or green channel filters might -result in a significant portion of the tissue being masked out. This could also potentially result in small -islands of non-masked tissue, and since we utilize a filter to remove small objects, this could result in the -further masking out of additional tissue regions. In such a situation, masking of 95% to 100% of the slide is possible. - -Slide #424 has a small tissue sample and its staining is a very faint lavender color. Slide #424 is -at risk for overmasking with our given combination of filters. - -| **Slide with Small Tissue Sample and Faint Staining** | -| -- | -| ![Slide with Small Tissue Sample and Faint Staining](images/424-rgb.png "Slide with Small Tissue Sample and Faint Staining") | - - -Therefore, rather than having fixed settings, we can automatically have our filters tweak parameter values to avoid -overmasking if desired. As examples, the `filter_green_channel()` and `filter_remove_small_objects()` functions have -this ability. If masking exceeds a certain overmasking threshold, a parameter value can be changed to lower -the amount of masking until the masking is below the overmasking threshold. - -``` -filter.filter_green_channel(np_img, green_thresh=200, avoid_overmask=True, overmask_thresh=90, output_type="bool") -filter.filter_remove_small_objects(np_img, min_size=3000, avoid_overmask=True, overmask_thresh=95, output_type="uint8") -``` - -For the `filter_green_channel()` function, if a `green_thresh` value of 200 results in masking over 90%, the -function will try with a higher `green_thresh` value (228) and the masking level will be checked. This will continue -until the masking doesn't exceed the overmask threshold of 90% or the threshold is 255. - -For the `filter_remove_small_objects()` function, if a `min_size` value of 3000 results in a masking level over 95%, -the function will try with a lower `min_size` value (1500) and the masking level will be checked. These `min_size` -reductions will continue until the masking level isn't over 95% or the minimum size is 0. For the image filtering -specified in `apply_image_filters`, a starting `min_size` value of 500 for `filter_remove_small_objects()` is used. - -Examining our full set of images using `multiprocess_apply_filters_to_images()`, we can identify slides that are -at risk for overmasking. We can create a list of these slide numbers and use `multiprocess_apply_filters_to_images()` -with this list of slide numbers to generate the `filters.html` page that allows us to visually inspect the filters -applied to this set of slides. - -``` -overmasked_slides = [1, 21, 29, 37, 43, 88, 116, 126, 127, 142, 145, 173, 196, 220, 225, 234, 238, 284, 292, 294, 304, - 316, 401, 403, 424, 448, 452, 472, 494] -filter.multiprocess_apply_filters_to_images(image_num_list=overmasked_slides) -``` - -Let's have a look at how we reduce overmasking on slide 21, which is a slide that has very faint staining. - -| **Slide 21** | -| -------------------- | -| ![Slide 21](images/21-rgb.png "Slide 21") | - - -We'll run our filters on slide #21. - -``` -filter.singleprocess_apply_filters_to_images(image_num_list=[21]) -``` - -If we set the `filter_green_channel()` and `filter_remove_small_objects()` `avoid_overmask` parameters to False, -97.69% of the original image is masked by the "green channel" filter and 99.92% of the original image is -masked by the subsequent "remove small objects" filter. This is significant overmasking. - -| **Overmasked by Green Channel Filter (97.69%)** | **Overmasked by Remove Small Objects Filter (99.92%)** | -| -- | -- | -| ![Overmasked by Green Channel Filter (97.69%)](images/21-overmask-green-ch.png "Overmasked by Green Channel Filter (97.69%)") | ![Overmasked by Remove Small Objects Filter (99.92%)](images/21-overmask-green-ch-overmask-rem-small-obj.png "Overmasked by Remove Small Objects Filter (99.92%)") - -If we set `avoid_overmask` to True for `filter_remove_small_objects()`, we see that the "remove small objects" -filter does not perform any further masking since the 97.69% masking from the previous "green channel" filter -already exceeds its overmasking threshold of 95%. - -| **Overmasked by Green Channel Filter (97.69%)** | **Avoid Overmask by Remove Small Objects Filter (97.69%)** | -| -- | -- | -| ![Overmasked by Green Channel Filter (97.69%)](images/21-overmask-green-ch.png "Overmasked by Green Channel Filter (97.69%)") | ![Avoid Overmask by Remove Small Objects Filter (97.69%)](images/21-overmask-green-ch-avoid-overmask-rem-small-obj.png "Avoid Overmask by Remove Small Objects Filter (97.69%)") - - -If we set `avoid_overmask` back to False for `filter_remove_small_objects()` and we set `avoid_overmask` to True for -`filter_green_channel()`, we see that 87.91% of the original image is masked by the "green channel" filter (under -the 90% overmasking threshold for the filter) and 97.40% of the image is masked by the subsequent -"remove small objects" filter. - -| **Avoid Overmask by Green Channel Filter (87.91%)** | **Overmask by Remove Small Objects Filter (97.40%)** | -| -- | -- | -| ![Avoid Overmask by Green Channel Filter (87.91%)](images/21-avoid-overmask-green-ch.png "Avoid Overmask by Green Channel Filter (87.91%)") | ![Overmask by Remove Small Objects Filter (97.40%)](images/21-avoid-overmask-green-ch-overmask-rem-small-obj.png "Overmask by Remove Small Objects Filter (97.40%)") - - -If we set `avoid_overmask` to True for both `filter_green_channel()` and `filter_remove_small_objects()`, we see that -the resulting masking after the "remove small objects" filter has been reduced to 94.88%, which is under its -overmasking threshold of 95%. - -| **Avoid Overmask by Green Channel Filter (87.91%)** | **Avoid Overmask by Remove Small Objects Filter (94.88%)** | -| -- | -- | -| ![Avoid Overmask by Green Channel Filter (87.91%)](images/21-avoid-overmask-green-ch-2.png "Avoid Overmask by Green Channel Filter (87.91%)") | ![Avoid Overmask by Remove Small Objects Filter (94.88%)](images/21-avoid-overmask-green-ch-avoid-overmask-rem-small-obj.png "Avoid Overmask by Remove Small Objects Filter (94.88%)") - - -Thus, in this example we've reduced the masking from 99.92% to 94.88%. - -We can see the filter adjustments being made in the console output. - -``` -Processing slide #21 -RGB | Time: 0:00:00.095414 Type: uint8 Shape: (1496, 1576, 3) -Save Image | Time: 0:00:00.617039 Name: ../data/filter_png/TUPAC-TR-021-001-rgb.png -Save Thumbnail | Time: 0:00:00.019557 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-001-rgb.jpg -Mask percentage 97.69% >= overmask threshold 90.00% for Remove Green Channel green_thresh=200, so try 228 -Filter Green Channel | Time: 0:00:00.005335 Type: bool Shape: (1496, 1576) -Filter Green Channel | Time: 0:00:00.010499 Type: bool Shape: (1496, 1576) -Mask RGB | Time: 0:00:00.009980 Type: uint8 Shape: (1496, 1576, 3) -Save Image | Time: 0:00:00.322629 Name: ../data/filter_png/TUPAC-TR-021-002-rgb-not-green.png -Save Thumbnail | Time: 0:00:00.018244 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-002-rgb-not-green.jpg -Filter Grays | Time: 0:00:00.072200 Type: bool Shape: (1496, 1576) -Mask RGB | Time: 0:00:00.010461 Type: uint8 Shape: (1496, 1576, 3) -Save Image | Time: 0:00:00.295995 Name: ../data/filter_png/TUPAC-TR-021-003-rgb-not-gray.png -Save Thumbnail | Time: 0:00:00.017668 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-003-rgb-not-gray.jpg -Filter Red Pen | Time: 0:00:00.055296 Type: bool Shape: (1496, 1576) -Mask RGB | Time: 0:00:00.008704 Type: uint8 Shape: (1496, 1576, 3) -Save Image | Time: 0:00:00.595753 Name: ../data/filter_png/TUPAC-TR-021-004-rgb-no-red-pen.png -Save Thumbnail | Time: 0:00:00.016758 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-004-rgb-no-red-pen.jpg -Filter Green Pen | Time: 0:00:00.088633 Type: bool Shape: (1496, 1576) -Mask RGB | Time: 0:00:00.008860 Type: uint8 Shape: (1496, 1576, 3) -Save Image | Time: 0:00:00.585474 Name: ../data/filter_png/TUPAC-TR-021-005-rgb-no-green-pen.png -Save Thumbnail | Time: 0:00:00.016964 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-005-rgb-no-green-pen.jpg -Filter Blue Pen | Time: 0:00:00.069669 Type: bool Shape: (1496, 1576) -Mask RGB | Time: 0:00:00.009665 Type: uint8 Shape: (1496, 1576, 3) -Save Image | Time: 0:00:00.589634 Name: ../data/filter_png/TUPAC-TR-021-006-rgb-no-blue-pen.png -Save Thumbnail | Time: 0:00:00.016736 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-006-rgb-no-blue-pen.jpg -Mask RGB | Time: 0:00:00.009115 Type: uint8 Shape: (1496, 1576, 3) -Save Image | Time: 0:00:00.294103 Name: ../data/filter_png/TUPAC-TR-021-007-rgb-no-gray-no-green-no-pens.png -Save Thumbnail | Time: 0:00:00.017540 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-007-rgb-no-gray-no-green-no-pens.jpg -Mask percentage 97.40% >= overmask threshold 95.00% for Remove Small Objs size 500, so try 250 -Mask percentage 96.83% >= overmask threshold 95.00% for Remove Small Objs size 250, so try 125 -Mask percentage 95.87% >= overmask threshold 95.00% for Remove Small Objs size 125, so try 62 -Remove Small Objs | Time: 0:00:00.031198 Type: bool Shape: (1496, 1576) -Remove Small Objs | Time: 0:00:00.062300 Type: bool Shape: (1496, 1576) -Remove Small Objs | Time: 0:00:00.095616 Type: bool Shape: (1496, 1576) -Remove Small Objs | Time: 0:00:00.128008 Type: bool Shape: (1496, 1576) -Mask RGB | Time: 0:00:00.007214 Type: uint8 Shape: (1496, 1576, 3) -Save Image | Time: 0:00:00.235025 Name: ../data/filter_png/TUPAC-TR-021-008-rgb-not-green-not-gray-no-pens-remove-small.png -Save Thumbnail | Time: 0:00:00.016905 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-008-rgb-not-green-not-gray-no-pens-remove-small.jpg -Save Image | Time: 0:00:00.232206 Name: ../data/filter_png/TUPAC-TR-021-32x-50432x47872-1576x1496-filtered.png -Save Thumbnail | Time: 0:00:00.017285 Name: ../data/filter_thumbnail_jpg/TUPAC-TR-021-32x-50432x47872-1576x1496-filtered.jpg -Slide #021 processing time: 0:00:04.596086 - -``` - - -## Tiles - -Following our filtering, we should have fairly good tissue segmentation for our dataset, -where non-tissue pixels have been masked out from our 1/32x scaled-down slide images. At this -stage, we break our images into tile regions. Tiling code is located in the `wsi/tiles.py` -file. - -For visualization, the tissue percentage of each tile is color-coded in a similar fashion -to a heat map. Tiles with 80% or more tissue are green, tiles less than 80% tissue and greater -or equal to 10% tissue are yellow, tiles less than 10% tissue and greater than 0% tissue are -orange, and tiles with 0% tissue are red. - -The heat map threshold values can be adjusted by modifying the `TISSUE_HIGH_THRESH` and -`TISSUE_LOW_THRESH` constants in `wsi/tiles.py`, which have default values of 80 and 10 -respectively. Heat map colors can be adjusted by modifying the `HIGH_COLOR`, `MEDIUM_COLOR`, -`LOW_COLOR`, and `NONE_COLOR` constants. The heat map border size can be adjusted using the -`TILE_BORDER_SIZE` constant, which has a default value of 2. -Tile sizes are specified according to number of pixels in the original WSI files. The -default `ROW_TILE_SIZE` and `COL_TILE_SIZE` values are 1,024 pixels. - -To generate and display tiles for a single slide, we utilize the `summary_and_tiles()` function, -which generates tile summaries and returns the top scoring tiles for a slide. We discuss -tile scoring in a later section. - -Let's generate tile tissue heat map summaries for slide #2 and display the summaries to the screen. - -``` -tiles.summary_and_tiles(2, display=True, save_summary=True, save_data=False, save_top_tiles=False) -``` - -Here, we see the tile tissue segmentation heat map summaries that are generated. The heat maps are -displayed on the masked image and the original image to allow for comparison. - -| **Tissue Heat Map** | **Tissue Heat Map on Original** | -| ------------------------ | ------------------------------------ | -| ![Tissue Heat Map](images/slide-2-tile-tissue-heatmap.png "Tissue Heat Map") | ![Tissue Heat Map on Original](images/slide-2-tile-tissue-heatmap-original.png "Tissue Heat Map on Original") | - -We see a variety of slide statistics displayed on the tile summaries. We see that slide #2 -has dimensions of 57,922x44,329. After scaling down the slide width and height by 1/32x, we have a -`png` image with dimensions 1,810x1,385. Breaking this image down into 32x32 tiles, we have 57 rows -and 44 columns, making a total of 2,508 tiles. Using our tissue segmentation filtering algorithms, -we have 1,283 tiles with high tissue percentages (>=80%), 397 tiles with medium tissue percentages -(>=10% and <80%), 102 tiles with low tissue percentages (>0% and <10%), and 726 tiles with no tissue -(0%). - -| Characteristic | Result | -| ------------------- | ------------- | -| Original Dimensions | 57,922x44,329 | -| Original Tile Size | 1,024x1,024 | -| Scale Factor | 1/32x | -| Scaled Dimensions | 1,810x1,385 | -| Scaled Tile Size | 32x32 | -| Total Mask | 41.60% | -| Total Tissue | 58.40% | -| Tiles | 57x44 = 2,508 | -| | 1,283 (51.16%) tiles >=80% tissue | -| | 397 (15.83%) tiles >=10% and <80% tissue | -| | 102 ( 4.07%) tiles >0% and <10% tissue | -| | 726 (28.95%) tiles =0% tissue | - - -Often it can be useful to know the exact row and column of a particular tile or tiles. If the -`DISPLAY_TILE_SUMMARY_LABELS` constant is set to True, the row and column of each tile is -output on the tile summaries. Generating the tile labels is fairly time-consuming, so usually -`DISPLAY_TILE_SUMMARY_LABELS` should be set to False for performance. - -| **Optional Tile Labels** | -| -------------------- | -| ![Optional Tile Labels](images/optional-tile-labels.png "Optional Tile Labels") | - - -## Tile Scoring - -In order to selectively choose how "good" a tile is compared to other tiles, we assign scores to -tiles based on tissue percentage and color characteristics. To determine the "best" tiles, we -sort based on score and return the top scoring tiles. We generate top tile summaries based on the -top scoring tiles, in a similar fashion as the tissue percentage summaries. - -The `score_tile()` function assigns a score to a tile based on the tissue percentage and various -color characteristics of the tile. The scoring formula utilized by `score_tile()` can be summarized -as follows. - -| **Scoring Formula** | -| -------------------- | -| ![Scoring Formula](images/scoring-formula.png "Scoring Formula") | - -The scoring formula generates good results for the images in the dataset and was developed through -experimentation with the training dataset. The *tissuepercent* is emphasized by squaring its value. -The *colorfactor* value is used to weigh hematoxylin staining heavier than eosin staining. Utilizing -the HSV color model, broad saturation and value distributions are given more weight by the -*saturationvaluefactor*. The *quantityfactor* value utilizes the tissue percentage to give more weight -to tiles with more tissue. Note that if *colorfactor*, *saturationvaluefactor*, or -*quantityfactor* evaluate to 0, the *score* will be 0. The *score* is scaled to a value from -0.0 to 1.0. - -During our discussion of color staining, we mentioned that tissue with hematoxylin staining is most -likely preferable to eosin staining. Hematoxylin stains acidic structures such as DNA and RNA with -a purple tone, while eosin stains basic structures such as cytoplasm proteins with a pink tone. -Let's discuss how we can more heavily score tiles with hematoxylin staining over eosin staining. - -Differentiating purplish shades from pinkish shades can be difficult using the RGB color space -(see [https://en.wikipedia.org/wiki/RGB_color_space](https://en.wikipedia.org/wiki/RGB_color_space)). -Therefore, to compute our *colorfactor* value, we first convert our tile in RGB color space -to HSV color space (see [https://en.wikipedia.org/wiki/HSL_and_HSV](https://en.wikipedia.org/wiki/HSL_and_HSV)). -HSV stands for Hue-Saturation-Value. In this color model, the hue is represented as a degree value -on a circle. Purple has a hue of 270 degrees and pink has a hue of 330 -degrees. We remove all hues less than 260 and greater than 340. Next, we compute the deviation from -purple (270) and the deviation from pink (330). We compute an average factor which is the squared -difference of 340 and the hue average. The *colorfactor* is computed as the pink deviation times -the average factor divided by the purple deviation. - -Let's have a closer look at a 32x32 tile and its accompanying HSV hue histogram. Note that in order -to properly convert a matplotlib chart image (the histogram) to a NumPy image on macOS, we currently -need to include a call to `matplotlib.use('Agg')`. -One way we can obtain a particular tile for analysis is to call -the `dynamic_tile()` function, which we describe in more detail later. Here, we obtain -the tile at the 29th row and 16th column on slide #2. Setting the `small_tile_in_tile` parameter -to `True` means that the scaled-down 32x32 tile is included in the returned Tile object. -The `display_image_with_hsv_hue_histogram()` function is used to display the small tile and its hue -histogram. - -``` -# To get around renderer issue on macOS going from Matplotlib image to NumPy image. -import matplotlib -matplotlib.use('Agg') -from deephistopath.wsi import tiles - -tile = tiles.dynamic_tile(2, 29, 16, True) -tiles.display_image_with_hsv_hue_histogram(tile.get_np_scaled_tile(), scale_up=True) -``` - -Here we see the 32x32 slide with its accompanying hue histogram. For convenience, colors have -been added to the histogram. -Also, notice that the non-tissue masked-out pixels have a peak at 0 degrees. - -| **Tile HSV Hue Histogram** | -| -------------------- | -| ![Tile HSV Hue Histogram](images/hsv-hue-histogram.png "Tile HSV Hue Histogram") | - - -For convenience, the `Tile` class has a `display_with_histograms()` function that can be used -to display histograms for both the RGB and HSV color spaces. If the scaled-down small tile is -included in the Tile object (using the `dynamic_tile()` `small_tile_in_tile` parameter with a -value of `True`), histograms will be displayed for both the small tile and the large tile. - -``` -import matplotlib -matplotlib.use('Agg') -from deephistopath.wsi import tiles - -tile = tiles.dynamic_tile(2, 29, 16, True) -tile.display_with_histograms(); -``` - -Here we see RGB and HSV histograms for the scaled-down tile at slide 2, row 29, column 16. We -see its score and tissue percentage. This tile's score was ranked 734 out of -a total of 2,508 tiles on this slide. - -| **Small Tile Color Histograms** | -| -------------------- | -| ![Small Tile Color Histograms](images/color-histograms-small-tile.png "Small Tile Color Histograms") | - - -Here we see RGB and HSV histograms for the full-sized 1,024x1,024 tile at slide 2, row 29, -column 16. Notice that the small tile pixels offer a reasonable approximation of the colors -present on the large tile. Also, notice that the masked-out pixels in the small tissue -correspond fairly accurately with the non-tissue regions of the large tile. - -| **Large Tile Color Histograms** | -| -------------------- | -| ![Large Tile Color Histograms](images/color-histograms-large-tile.png "Large Tile Color Histograms") | - - -If the `save_data` parameter of the `summary_and_tiles()` function is set to `True`, detailed data about -the slide tiles are saved in `csv` format. - -``` -tiles.summary_and_tiles(2, display=True, save_summary=True, save_data=True, save_top_tiles=False) -``` - -For slide #2, this generates a `TUPAC-TR-002-32x-57922x44329-1810x1385-tile_data.csv` file. - -| **Tile Data** | -| ------------- | -| ![Tile Data](images/tile-data.png "Tile Data") | - - -In addition to the tile tissue heat map summaries, the `summary_and_tiles()` function generates -top tile summaries. By default it highlights the top 50 scoring tiles. The number of top tiles can be -controlled by the `NUM_TOP_TILES` constant. - -``` -tiles.summary_and_tiles(2, display=True, save_summary=True, save_data=False, save_top_tiles=False) -``` - -Here we see the top tile summary on the masked image for slide #2. Notice that tiles with high -tissue percentages and hematoxylin-stained tissue are favored over tiles with low tissue -percentages and eosin-stained tissue. Notice that statistics about the top 50 scoring tiles are -displayed to the right of the image. - -| **Top Tiles** | -| ------------- | -| ![Top Tiles](images/slide-2-top-tiles.png "Top Tiles") | - - -For visual inspection, the top tile summary is also generated over the original slide image, as -we see here. - -| **Top Tiles on Original** | -| ------------------------- | -| ![Top Tiles on Original](images/slide-2-top-tiles-original.png "Top Tiles on Original") | - - -When analyzing top tile results, it can be useful to see the tissue percentage heat map -of surrounding tiles. This can be accomplished by setting the `BORDER_ALL_TILES_IN_TOP_TILE_SUMMARY` -constant to `True`. Likewise, it can useful to see the row and column coordinates of all tiles, -which can be accomplished using the `LABEL_ALL_TILES_IN_TOP_TILE_SUMMARY` constant with a value of -`True`. - -| **Top Tile Borders** | **Top Tile Labels** | -| -------------------- | -------------------- | -| ![Top Tile Borders](images/slide-2-top-tile-borders.png "Top Tile Borders") | ![Top Tile Labels](images/slide-2-top-tile-labels.png "Top Tile Labels") | - - -Here we see a section of a top tile summary that features both the tile tissue heat map and the -row and column labels. - -| **Top Tile Labels and Borders** | -| ------------------------- | -| ![Top Tile Labels and Borders](images/slide-2-top-tile-labels-borders.png "Top Tile Labels and Borders") | - -## Top Tile Retrieval - -Top tiles can be saved as files in batch mode or retrieved dynamically. In batch mode, -tiling, scoring, and saving the 1,000 tissue percentage heat map summaries (2 per image), -the 1,000 top tile summaries (2 per image), the 2,000 thumbnails, and 25,000 1Kx1K tiles -(50 per image) takes approximately 2 hours. - -If the `save_top_tiles` parameter of the `summary_and_tiles()` function is set to `True`, -the top-ranking tiles for the specified slide will be saved to the file system. - -``` -tiles.summary_and_tiles(2, display=True, save_summary=True, save_data=False, save_top_tiles=True) -``` - -In general, it is recommended that the user utilize the `singleprocess_filtered_images_to_tiles()` -and `multiprocess_filtered_images_to_tiles()` functions in `wsi/tiles.py`. These functions -generate convenient HTML pages for investigating the tiles generated for a slide set. The -`multiprocess_filtered_images_to_tiles()` utilizes multiprocessing for added performance. If -no `image_num_list` parameter is provided, all images in the dataset will be processed. - -Here, we generate the top 50 tiles for slides #1, #2, and #3. - -``` -tiles.multiprocess_filtered_images_to_tiles(image_num_list=[1, 2, 3]) -``` - -On the generated `tiles.html` page, we see the original slide images, the images after filtering, -the tissue percentage heat map summaries on the filtered images and the original images, tile summary -data including links to the generated `csv` file for each slide, the top tile summaries on the -filtered images and the original images, and links to the top 50 tile files for each slide. - -| **Tiles Page** | -| ------------- | -| ![Tiles Page](images/tiles-page.png "Tiles Page") | - - -The full-size 1,024x1,024 tiles can be investigated using the top tile links. Here we see the -two top-scoring tiles on slide 2 at row 34, column 34 and row 35, column 37. - -| **Slide #1, Top Tile #1** | **Slide #1, Top Tile #2** | -| ------------------------ | ------------------------------------ | -| ![Slide #1, Top Tile #1](images/TUPAC-TR-002-tile-r34-c34-x33793-y33799-w1024-h1024.png "Slide #1, Top Tile #1") | ![Slide #1, Top Tile #2](images/TUPAC-TR-002-tile-r35-c37-x36865-y34823-w1024-h1024.png "Slide #1, Top Tile #2") | - - -Tiles can also be retrieved dynamically. In dynamic tile retrieval, slides are scaled down, -filtered, tiled, and scored all in-memory. The top tiles can then be retrieved from the -original WSI file and stored in-memory. No intermediate files are written to the file system -during dynamic tile retrieval. - -Here, we dynamically obtain a `TileSummary` object by calling `dynamic_tiles()` for -slide #2. We obtain the top-scoring tiles from `tile_summary`, outputting status -information about each tile. The status information includes the tile number, the row -number, the column number, the tissue percentage, and the tile score. - -``` -tile_summary = tiles.dynamic_tiles(2) -top_tiles = tile_summary.top_tiles() -for t in top_tiles: - print(t) -``` - -In the console output, we see that the original `svs` file is opened, the slide is -scaled down, and our series of filters is run on the scaled-down image. After that, -the tiles are scored, and we see status information about the top 50 tiles for -the slide. - -``` -Opening Slide #2: ../data/training_slides/TUPAC-TR-002.svs -RGB | Time: 0:00:00.007339 Type: uint8 Shape: (1385, 1810, 3) -Filter Green Channel | Time: 0:00:00.005135 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.007973 Type: uint8 Shape: (1385, 1810, 3) -Filter Grays | Time: 0:00:00.073780 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.008114 Type: uint8 Shape: (1385, 1810, 3) -Filter Red Pen | Time: 0:00:00.066007 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.007925 Type: uint8 Shape: (1385, 1810, 3) -Filter Green Pen | Time: 0:00:00.105854 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.008034 Type: uint8 Shape: (1385, 1810, 3) -Filter Blue Pen | Time: 0:00:00.087092 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.007963 Type: uint8 Shape: (1385, 1810, 3) -Mask RGB | Time: 0:00:00.007807 Type: uint8 Shape: (1385, 1810, 3) -Remove Small Objs | Time: 0:00:00.034308 Type: bool Shape: (1385, 1810) -Mask RGB | Time: 0:00:00.007814 Type: uint8 Shape: (1385, 1810, 3) -[Tile #1915, Row #34, Column #34, Tissue 100.00%, Score 0.8824] -[Tile #1975, Row #35, Column #37, Tissue 100.00%, Score 0.8816] -[Tile #1974, Row #35, Column #36, Tissue 99.90%, Score 0.8811] -[Tile #500, Row #9, Column #44, Tissue 99.32%, Score 0.8797] -[Tile #814, Row #15, Column #16, Tissue 99.22%, Score 0.8795] -[Tile #1916, Row #34, Column #35, Tissue 100.00%, Score 0.8789] -[Tile #1956, Row #35, Column #18, Tissue 99.51%, Score 0.8784] -[Tile #1667, Row #30, Column #14, Tissue 98.63%, Score 0.8783] -[Tile #1839, Row #33, Column #15, Tissue 99.51%, Score 0.8782] -[Tile #1725, Row #31, Column #15, Tissue 99.61%, Score 0.8781] -[Tile #2061, Row #37, Column #9, Tissue 98.54%, Score 0.8779] -[Tile #724, Row #13, Column #40, Tissue 99.90%, Score 0.8778] -[Tile #1840, Row #33, Column #16, Tissue 99.22%, Score 0.8777] -[Tile #758, Row #14, Column #17, Tissue 99.41%, Score 0.8775] -[Tile #1722, Row #31, Column #12, Tissue 98.24%, Score 0.8771] -[Tile #722, Row #13, Column #38, Tissue 99.51%, Score 0.8769] -[Tile #1803, Row #32, Column #36, Tissue 99.22%, Score 0.8769] -[Tile #446, Row #8, Column #47, Tissue 100.00%, Score 0.8768] -[Tile #988, Row #18, Column #19, Tissue 99.61%, Score 0.8767] -[Tile #2135, Row #38, Column #26, Tissue 99.80%, Score 0.8767] -[Tile #704, Row #13, Column #20, Tissue 99.61%, Score 0.8767] -[Tile #816, Row #15, Column #18, Tissue 99.41%, Score 0.8766] -[Tile #1180, Row #21, Column #40, Tissue 99.90%, Score 0.8765] -[Tile #1178, Row #21, Column #38, Tissue 99.80%, Score 0.8765] -[Tile #1042, Row #19, Column #16, Tissue 99.71%, Score 0.8764] -[Tile #1783, Row #32, Column #16, Tissue 99.80%, Score 0.8764] -[Tile #1978, Row #35, Column #40, Tissue 100.00%, Score 0.8763] -[Tile #832, Row #15, Column #34, Tissue 99.61%, Score 0.8762] -[Tile #1901, Row #34, Column #20, Tissue 99.90%, Score 0.8759] -[Tile #701, Row #13, Column #17, Tissue 99.80%, Score 0.8758] -[Tile #817, Row #15, Column #19, Tissue 99.32%, Score 0.8757] -[Tile #2023, Row #36, Column #28, Tissue 100.00%, Score 0.8754] -[Tile #775, Row #14, Column #34, Tissue 99.51%, Score 0.8754] -[Tile #1592, Row #28, Column #53, Tissue 100.00%, Score 0.8753] -[Tile #702, Row #13, Column #18, Tissue 99.22%, Score 0.8753] -[Tile #759, Row #14, Column #18, Tissue 99.51%, Score 0.8752] -[Tile #1117, Row #20, Column #34, Tissue 99.90%, Score 0.8751] -[Tile #1907, Row #34, Column #26, Tissue 99.32%, Score 0.8750] -[Tile #1781, Row #32, Column #14, Tissue 99.61%, Score 0.8749] -[Tile #2250, Row #40, Column #27, Tissue 99.61%, Score 0.8749] -[Tile #1902, Row #34, Column #21, Tissue 99.90%, Score 0.8749] -[Tile #2014, Row #36, Column #19, Tissue 99.22%, Score 0.8749] -[Tile #2013, Row #36, Column #18, Tissue 99.51%, Score 0.8747] -[Tile #1175, Row #21, Column #35, Tissue 99.71%, Score 0.8746] -[Tile #760, Row #14, Column #19, Tissue 99.22%, Score 0.8746] -[Tile #779, Row #14, Column #38, Tissue 99.32%, Score 0.8745] -[Tile #1863, Row #33, Column #39, Tissue 99.71%, Score 0.8745] -[Tile #1899, Row #34, Column #18, Tissue 99.51%, Score 0.8745] -[Tile #778, Row #14, Column #37, Tissue 99.90%, Score 0.8743] -[Tile #1724, Row #31, Column #14, Tissue 99.51%, Score 0.8741] -``` - -If we'd like to obtain each tile as a NumPy array, we can do -so by calling the `get_np_tile()` function on the `Tile` -object. - -``` -tile_summary = tiles.dynamic_tiles(2) -top_tiles = tile_summary.top_tiles() -for t in top_tiles: - print(t) - np_tile = t.get_np_tile() -``` - -As a further example, here we dynamically retrieve the tiles -for slide #4 and display the top 2 tiles along with their -RGB and HSV histograms. - -``` -tile_summary = tiles.dynamic_tiles(4) -top = tile_summary.top_tiles()[:2] -for t in top: - t.display_with_histograms() -``` - -| **Slide #4, Top Tile #1** | **Slide #4, Top Tile #2** | -| ------------------------ | ------------------------------------ | -| ![Slide #4, Top Tile #1](images/slide-4-top-tile-1.png "Slide #4, Top Tile #1") | ![Slide #4, Top Tile #2](images/slide-4-top-tile-2.png "Slide #4, Top Tile #2") | - - -Next, we dynamically retrieve the tiles for slide #2. We -display (not shown) the tile tissue heat map and top tile summaries and -then obtain the tiles ordered by tissue percentage. -We display the 1,000th and 1,500th tiles by tissue percentage. - -``` -tile_summary = tiles.dynamic_tiles(2) -tile_summary.display_summaries() -ts = tile_summary.tiles_by_tissue_percentage() -ts[999].display_with_histograms() -ts[1499].display_with_histograms() -``` - -Here we see the 1,000th and 1,500th tiles ordered by tissue percentage for slide #2. -Note that the displayed tile rank information is based on score rather than -tissue percentage alone. - -| **Slide #2, Tissue Percentage #1000** | **Slide #2, Tissue Percentage #1500** | -| ------------------------ | ------------------------------------ | -| ![Slide #2, Tissue Percentage #1000](images/slide-2-tissue-percentage-tile-1000.png "Slide #2, Tissue Percentage #1000") | ![Slide #2, Tissue Percentage #1500](images/slide-2-tissue-percentage-tile-1500.png "Slide #2, Tissue Percentage #1500") | - - -Tiles can be retrieved based on position. Here, we display the tiles at row 25, column 30 and row 25, column 31 on slide #2. - -``` -tile_summary = tiles.dynamic_tiles(2) -tile_summary.get_tile(25, 30).display_tile() -tile_summary.get_tile(25, 31).display_tile() -``` - -| **Slide #2, Row #25, Column #30** | **Slide #2, Row #25, Column #31** | -| ------------------------ | ------------------------------------ | -| ![Slide #2, Row #25, Column #30](images/slide-2-row-25-col-30.png "Slide #2, Row #25, Column #30") | ![Slide #2, Row #25, Column #31](images/slide-2-row-25-col-31.png "Slide #2, Row #25, Column #31") | - -If an individual tile is required, the `dynamic_tile()` function can be used. - -``` -tiles.dynamic_tile(2, 25, 32).display_tile() -``` - -| **Slide #2, Row #25, Column #32** | -| --------------------------------- | -| ![Slide #2, Row #25, Column #32](images/slide-2-row-25-col-32.png "Slide #2, Row #25, Column #32") | - -If multiple tiles need to be retrieved dynamically, for performance reasons `dynamic_tiles()` is -preferable to `dynamic_tile()`. - - -## Summary - -In this tutorial, we've taken a look at how Python, in particular with packages such as NumPy and -scikit-image, can be used for tissue segmentation in whole-slide images. In order to efficiently process -images in our dataset, we utilized OpenSlide to scale down the slides. Using NumPy arrays, we -investigated a wide variety of image filters and settled on a combination and series of filters that -demonstrated fast, acceptably accurate tissue segmentation for our dataset. Following this, we divided -the filtered images into tiles and scored the tiles based on tissue percentage and color characteristics -such as the degree of hematoxylin staining versus eosin staining. We then demonstrated how we can -retrieve the top-scoring tiles which have high tissue percentages and preferred staining characteristics. -We saw how whole-slide images could be processed in batches or dynamically. Scaling, filtering, -tiling, scoring, and saving the top tiles can be accomplished in batch mode using multiprocessing in -the following manner. - -``` -slide.multiprocess_training_slides_to_images() -filter.multiprocess_apply_filters_to_images() -tiles.multiprocess_filtered_images_to_tiles() -``` - -The above code generates HTML filter and tile pages which simplify visual -inspection of the image processing and the final tile results. - -Since the average number of pixels per whole-slide image is 7,670,709,629 and we have reduced -the data to the top 50 1,024x1,024 pixel tiles, we have reduced the raw image data down by a -factor of 146x while identifying tiles that have significant potential for further useful -analysis. diff --git a/example.ipynb b/example.ipynb new file mode 100755 index 0000000..91b6fe5 --- /dev/null +++ b/example.ipynb @@ -0,0 +1,227 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "scrolled": true + }, + "outputs": [], + "source": [ + "%reload_ext autoreload\n", + "%autoreload 2\n", + "%matplotlib inline\n", + "#%matplotlib notebook\n", + "import numpy as np\n", + "from tqdm import tqdm_notebook as tqdm\n", + "import pathlib\n", + "from pathlib import Path\n", + "Path.ls = lambda x: [p for p in list(x.iterdir()) if '.ipynb_checkpoints' not in p.name]\n", + "import pandas as pd\n", + "import PIL\n", + "from wsi import slide, filter, tiles, util\n", + "\n", + "\n", + "base_path = Path('/home/Deep_Learner/shared/Datasets/Hypophysenadenome/')\n", + "wsis_path = base_path/'wsis_experimenting'\n", + "rois_path = base_path/'rois_experimenting'\n", + "tiles_path = base_path/'tiles_experimenting'\n", + "wsis_path.mkdir(exist_ok=True)\n", + "tiles_path.mkdir(exist_ok=True)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "# WSIs" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## Process one WSI" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "scrolled": true + }, + "outputs": [], + "source": [ + "p = wsis_path.ls()[0];p\n", + "df = tiles.WsiOrROIToTiles(wsiPath=p, \n", + " tilesFolderPath=tiles_path, \n", + " tile_height=256,tile_width=256, \n", + " tile_naming_func=tiles.get_wsi_name_from_path_pituitary_adenoma_entities,\n", + " save_tiles=False, \n", + " tile_score_thresh = 0.55)\n", + "df.head()" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## Process multiple WSIs in parallel" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "scrolled": true + }, + "outputs": [], + "source": [ + "df = tiles.WsiOrROIToTilesMultithreaded(wsis_path.ls()[:2], \n", + " tiles_path, \n", + " 256, \n", + " 256, \n", + " tile_naming_func=tiles.get_wsi_name_from_path_pituitary_adenoma_entities, \n", + " save_tiles=False)\n", + "df.head()" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## Extract one tile directly from a WSI at a specific level" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "p = wsis_path.ls()[0];p" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "s = slide.open_slide(p)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "s.level_dimensions" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "tiles.ExtractTileFromWSI(p, x=0,y=0, width=1232,height=840,level=5)" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "# ROIs" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## Process one ROI" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "p = rois_path.ls()[0];p\n", + "\n", + "tiles.WsiOrROIToTiles(wsiPath=p, \n", + " tilesFolderPath=tiles_path, \n", + " tile_height=1024,\n", + " tile_width=1024, \n", + " is_wsi=False, \n", + " tile_naming_func=tiles.get_roi_name_from_path_pituitary_adenoma_entities, \n", + " save_tiles=True)\n", + "\n", + "util.show_multiple_images_big(tiles_path.ls()[:2])" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## Extract one tile directly from a ROI" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "p = rois_path.ls()[0];p\n", + "\n", + "img = PIL.Image.open(p)\n", + "\n", + "print(img.size)\n", + "\n", + "tiles.ExtractTileFromPILImage(p, 1000, 1000, 512, 512)" + ] + } + ], + "metadata": { + "kernelspec": { + "display_name": "DLM Py3", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.7.3" + }, + "toc": { + "base_numbering": 1, + "nav_menu": {}, + "number_sections": true, + "sideBar": true, + "skip_h1_title": false, + "title_cell": "Table of Contents", + "title_sidebar": "Contents", + "toc_cell": false, + "toc_position": { + "height": "calc(100% - 180px)", + "left": "10px", + "top": "150px", + "width": "258px" + }, + "toc_section_display": true, + "toc_window_display": true + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/deephistopath/wsi/filter.py b/wsi/filter.py old mode 100644 new mode 100755 similarity index 98% rename from deephistopath/wsi/filter.py rename to wsi/filter.py index 1eaf388..7b9b480 --- a/deephistopath/wsi/filter.py +++ b/wsi/filter.py @@ -14,6 +14,8 @@ # # ------------------------------------------------------------------------ +from pathlib import Path +import PIL import math import multiprocessing import numpy as np @@ -27,9 +29,23 @@ import skimage.morphology as sk_morphology import skimage.segmentation as sk_segmentation -from deephistopath.wsi import slide -from deephistopath.wsi import util -from deephistopath.wsi.util import Time +from wsi import slide, util, tiles +from wsi.util import Time + + + + +def filter_img(img_pil:PIL.Image.Image) -> PIL.Image.Image: + """ + + """ + img_np = util.pil_to_np_rgb(img_pil) + grayscale_np = filter_rgb_to_grayscale(img_np) + complement_np = filter_complement(grayscale_np) + otsu_np = filter_otsu_threshold(complement_np).astype(np.bool) + filtered_img_np = util.mask_rgb(img_np, otsu_np) + return util.np_to_pil(filtered_img_np) + def filter_rgb_to_grayscale(np_img, output_type="uint8"): @@ -45,12 +61,12 @@ def filter_rgb_to_grayscale(np_img, output_type="uint8"): Returns: Grayscale image as NumPy array with shape (h, w). """ - t = Time() + #t = Time() # Another common RGB ratio possibility: [0.299, 0.587, 0.114] grayscale = np.dot(np_img[..., :3], [0.2125, 0.7154, 0.0721]) if output_type != "float": grayscale = grayscale.astype("uint8") - util.np_info(grayscale, "Gray", t.elapsed()) + #util.np_info(grayscale, "Gray", t.elapsed()) return grayscale @@ -65,12 +81,12 @@ def filter_complement(np_img, output_type="uint8"): Returns: Complement image as Numpy array. """ - t = Time() + #t = Time() if output_type == "float": complement = 1.0 - np_img else: complement = 255 - np_img - util.np_info(complement, "Complement", t.elapsed()) + #util.np_info(complement, "Complement", t.elapsed()) return complement @@ -110,7 +126,7 @@ def filter_otsu_threshold(np_img, output_type="uint8"): Returns: NumPy array (bool, float, or uint8) where True, 1.0, and 255 represent a pixel above Otsu threshold. """ - t = Time() + #t = Time() otsu_thresh_value = sk_filters.threshold_otsu(np_img) otsu = (np_img > otsu_thresh_value) if output_type == "bool": @@ -119,7 +135,7 @@ def filter_otsu_threshold(np_img, output_type="uint8"): otsu = otsu.astype(float) else: otsu = otsu.astype("uint8") * 255 - util.np_info(otsu, "Otsu Threshold", t.elapsed()) + #util.np_info(otsu, "Otsu Threshold", t.elapsed()) return otsu diff --git a/wsi/openslide_overwrite.py b/wsi/openslide_overwrite.py new file mode 100755 index 0000000..bed26be --- /dev/null +++ b/wsi/openslide_overwrite.py @@ -0,0 +1,47 @@ +import sys +import PIL + +def _load_image(buf, size): + '''buf must be a buffer.''' + + # Load entire buffer at once if possible + MAX_PIXELS_PER_LOAD = (1 << 29) - 1 + # Otherwise, use chunks smaller than the maximum to reduce memory + # requirements + PIXELS_PER_LOAD = 1 << 26 + + def do_load(buf, size): + '''buf can be a string, but should be a ctypes buffer to avoid an + extra copy in the caller.''' + # First reorder the bytes in a pixel from native-endian aRGB to + # big-endian RGBa to work around limitations in RGBa loader + rawmode = (sys.byteorder == 'little') and 'BGRA' or 'ARGB' + buf = PIL.Image.frombuffer('RGBA', size, buf, 'raw', rawmode, 0, 1) + # Image.tobytes() is named tostring() in Pillow 1.x and PIL + buf = (getattr(buf, 'tobytes', None) or buf.tostring)() + # Now load the image as RGBA, undoing premultiplication + return PIL.Image.frombuffer('RGBA', size, buf, 'raw', 'RGBa', 0, 1) + + # Fast path for small buffers + w, h = size + if w * h <= MAX_PIXELS_PER_LOAD: + return do_load(buf, size) + + # Load in chunks to avoid OverflowError in PIL.Image.frombuffer() + # https://github.com/python-pillow/Pillow/issues/1475 + if w > PIXELS_PER_LOAD: + # We could support this, but it seems like overkill + raise ValueError('Width %d is too large (maximum %d)' % + (w, PIXELS_PER_LOAD)) + rows_per_load = PIXELS_PER_LOAD // w + img = PIL.Image.new('RGBA', (w, h)) + for y in range(0, h, rows_per_load): + rows = min(h - y, rows_per_load) + if sys.version[0] == '2': + chunk = buffer(buf, 4 * y * w, 4 * rows * w) + else: + # PIL.Image.frombuffer() won't take a memoryview or + # bytearray, so we can't avoid copying + chunk = memoryview(buf)[y * w:(y + rows) * w].tobytes() + img.paste(do_load(chunk, (w, rows)), (0, y)) + return img diff --git a/wsi/slide.py b/wsi/slide.py new file mode 100755 index 0000000..b802c51 --- /dev/null +++ b/wsi/slide.py @@ -0,0 +1,114 @@ +# ------------------------------------------------------------------------ +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# +# ------------------------------------------------------------------------ + +import pathlib +from pathlib import Path +import glob +import math +import matplotlib.pyplot as plt +import multiprocessing +import numpy as np +import openslide +from openslide import OpenSlideError +import os +import PIL +from PIL import Image +import re +import sys +from wsi import util, tiles +from wsi.util import Time +from typing import List, Callable, Union + + +def open_slide(path:Union[str, pathlib.Path]): + """ + Open a whole-slide image (*.svs,*.ndpi, etc). + + Args: + path: Path to the slide file. + + Returns: + An OpenSlide object representing a whole-slide image. + """ + #try: + slide = openslide.open_slide(str(path)) + #except OpenSlideError: + # slide = None + #except FileNotFoundError: + # slide = None + return slide + + +def open_image(filename): + """ + Open an image (*.jpg, *.png, etc). + + Args: + filename: Name of the image file. + + returns: + A PIL.Image.Image object representing an image. + """ + image = Image.open(filename) + return image + + +def open_image_np(filename): + """ + Open an image (*.jpg, *.png, etc) as an RGB NumPy array. + + Args: + filename: Name of the image file. + + returns: + A NumPy representing an RGB image. + """ + pil_img = open_image(filename) + np_img = util.pil_to_np_rgb(pil_img) + return np_img + + +def small_to_large_mapping(small_pixel, large_dimensions, scale_factor): + """ + Map a scaled-down pixel width and height to the corresponding pixel of the original whole-slide image. + + Args: + small_pixel: The scaled-down width and height. + large_dimensions: The width and height of the original whole-slide image. + + Returns: + Tuple consisting of the scaled-up width and height. + """ + small_x, small_y = small_pixel + large_w, large_h = large_dimensions + large_x = round((large_w / scale_factor) / math.floor(large_w / scale_factor) * (scale_factor * small_x)) + large_y = round((large_h / scale_factor) / math.floor(large_h / scale_factor) * (scale_factor * small_y)) + return large_x, large_y + + +def get_conversion_factor(wsi_path:pathlib.Path, level:int)->float: + """ + Arguments: + wsi_path: path to a whole-slide image + level: level of the whole-slide image, 0 means highest resolution, with every level the resolution halves + Result: + returns a conversion factor, to convert pixel size into micrometer + """ + sl = open_slide(wsi_path) + mpp_x = float(sl.properties.get('openslide.mpp-x')) + mpp_y = float(sl.properties.get('openslide.mpp-y')) + assert mpp_x==mpp_y + return mpp_x*2**level \ No newline at end of file diff --git a/wsi/tiles.py b/wsi/tiles.py new file mode 100755 index 0000000..0c855b6 --- /dev/null +++ b/wsi/tiles.py @@ -0,0 +1,1219 @@ +# ------------------------------------------------------------------------ +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# +# ------------------------------------------------------------------------ + +# To get around renderer issue on macOS going from Matplotlib image to NumPy image. +import matplotlib + +matplotlib.use('Agg') + +import PIL +import pathlib +from pathlib import Path +import colorsys +import math +import matplotlib.pyplot as plt +import multiprocessing +import numpy +import numpy as np +import os +import PIL +from PIL import Image, ImageDraw, ImageFont +from enum import Enum +from wsi import util, filter, slide +from wsi import openslide_overwrite +from wsi.util import Time +import openslide +import multiprocessing +from typing import List, Callable, Union, Dict, Tuple, Union +from tqdm import tqdm_notebook as tqdm +import pandas +import pandas as pd +import warnings + + + +TISSUE_HIGH_THRESH = 80 +TISSUE_LOW_THRESH = 10 +HSV_PURPLE = 270 +HSV_PINK = 330 + +############################# classes ######################################### + + + +class TileSummary: + """ + Class for tile summary information. + """ + + wsi_path = None + is_wsi = None + tiles_folder_path = None + orig_w = None + orig_h = None + orig_tile_w = None + orig_tile_h = None + scale_factor = None + scaled_w = None + scaled_h = None + scaled_tile_w = None + scaled_tile_h = None + mask_percentage = None + num_row_tiles = None + num_col_tiles = None + tile_score_thresh = None + level = None + best_level_for_downsample = None + real_scale_factor = None + + count = 0 + high = 0 + medium = 0 + low = 0 + none = 0 + + def __init__(self, + wsi_path, + is_wsi, + tiles_folder_path, + orig_w, + orig_h, + orig_tile_w, + orig_tile_h, + scale_factor, + scaled_w, + scaled_h, + scaled_tile_w, + scaled_tile_h, + tissue_percentage, + num_col_tiles, + num_row_tiles, + tile_score_thresh, + level, + best_level_for_downsample, + real_scale_factor): + self.wsi_path = wsi_path + self.is_wsi = is_wsi + self.tiles_folder_path = tiles_folder_path + self.orig_w = orig_w + self.orig_h = orig_h + self.orig_tile_w = orig_tile_w + self.orig_tile_h = orig_tile_h + self.scale_factor = scale_factor + self.scaled_w = scaled_w + self.scaled_h = scaled_h + self.scaled_tile_w = scaled_tile_w + self.scaled_tile_h = scaled_tile_h + self.tissue_percentage = tissue_percentage + self.num_col_tiles = num_col_tiles + self.num_row_tiles = num_row_tiles + self.tile_score_thresh = tile_score_thresh + self.level = level + self.best_level_for_downsample = best_level_for_downsample + self.real_scale_factor = real_scale_factor + self.tiles = [] + + def __str__(self): + return summary_title(self) + "\n" + summary_stats(self) + + def mask_percentage(self): + """ + Obtain the percentage of the slide that is masked. + + Returns: + The amount of the slide that is masked as a percentage. + """ + return 100 - self.tissue_percentage + + def num_tiles(self): + """ + Retrieve the total number of tiles. + + Returns: + The total number of tiles (number of rows * number of columns). + """ + return self.num_row_tiles * self.num_col_tiles + + def tiles_by_tissue_percentage(self): + """ + Retrieve the tiles ranked by tissue percentage. + + Returns: + List of the tiles ranked by tissue percentage. + """ + sorted_list = sorted(self.tiles, key=lambda t: t.tissue_percentage, reverse=True) + return sorted_list + + def tiles_by_score(self): + """ + Retrieve the tiles ranked by score. + + Returns: + List of the tiles ranked by score. + """ + sorted_list = sorted(self.tiles, key=lambda t: t.score, reverse=True) + return sorted_list + + def get_tile(self, row, col): + """ + Retrieve tile by row and column. + + Args: + row: The row + col: The column + + Returns: + Corresponding Tile object. + """ + tile_index = (row - 1) * self.num_col_tiles + (col - 1) + tile = self.tiles[tile_index] + return tile + + def top_tiles(self): + """ + Retrieve only the tiles that pass scoring. + + Returns: + List of the top-scoring tiles. + """ + sorted_tiles = self.tiles_by_score() + top_tiles = [tile for tile in sorted_tiles + if self.check_tile(tile)] + print(f'{self.wsi_path}: Number of tiles that will be saved/all possible tiles: {len(top_tiles)}/{len(sorted_tiles)}') + return top_tiles + + def check_tile(self, tile): + width = tile.o_c_e - tile.o_c_s + height = tile.o_r_e - tile.o_r_s + return tile.score > self.tile_score_thresh and width >= 0.7*self.orig_tile_w and height >= 0.7*self.orig_tile_h + + +class Tile: + """ + Class for information about a tile. + """ + tile_summary = None + wsi_path = None + is_wsi = None + tiles_folder_path = None + np_scaled_tile = None + tile_num = None + r = None + c = None + r_s = None + r_e = None + c_s = None + c_e = None + o_r_s = None + o_r_e = None + o_c_s = None + o_c_e = None + t_p = None + color_factor = None + s_and_v_factor = None + quantity_factor = None + score = None + tile_naming_func = None + level = None + best_level_for_downsample = None + real_scale_factor = None + + def __init__(self, + tile_summary, + wsi_path, + is_wsi, + tiles_folder_path, + np_scaled_tile, + tile_num, + r, + c, + r_s, + r_e, + c_s, + c_e, + o_r_s, + o_r_e, + o_c_s, + o_c_e, + t_p, + color_factor, + s_and_v_factor, + quantity_factor, + score, + tile_naming_func, + level, + best_level_for_downsample, + real_scale_factor): + self.tile_summary = tile_summary + self.wsi_path = wsi_path + self.is_wsi = is_wsi + self.tiles_folder_path = tiles_folder_path + self.np_scaled_tile = np_scaled_tile + self.tile_num = tile_num + self.r = r + self.c = c + self.r_s = r_s + self.r_e = r_e + self.c_s = c_s + self.c_e = c_e + self.o_r_s = o_r_s + self.o_r_e = o_r_e + self.o_c_s = o_c_s + self.o_c_e = o_c_e + self.tissue_percentage = t_p + self.color_factor = color_factor + self.s_and_v_factor = s_and_v_factor + self.quantity_factor = quantity_factor + self.score = score + self.tile_naming_func = tile_naming_func + self.level = level + self.best_level_for_downsample = best_level_for_downsample + self.real_scale_factor = real_scale_factor + + def __str__(self): + return "[Tile #%d, Row #%d, Column #%d, Tissue %4.2f%%, Score %0.4f]" % ( + self.tile_num, self.r, self.c, self.tissue_percentage, self.score) + + def __repr__(self): + return "\n" + self.__str__() + + def mask_percentage(self): + return 100 - self.tissue_percentage + + def tissue_quantity(self): + return tissue_quantity(self.tissue_percentage) + + def get_pil_tile(self): + return tile_to_pil_tile(self, self.is_wsi) + + def get_np_tile(self): + return tile_to_np_tile(self) + + def save_tile(self): + save_display_tile(self, save=True, display=False, is_wsi=self.is_wsi) + + def display_tile(self): + save_display_tile(self, save=False, display=True, is_wsi=self.is_wsi) + + def display_with_histograms(self): + display_tile(self, rgb_histograms=True, hsv_histograms=True) + + def get_np_scaled_tile(self): + return self.np_scaled_tile + + def get_pil_scaled_tile(self): + return util.np_to_pil(self.np_scaled_tile) + + def get_width(self): + return self.o_c_e - self.o_c_s + + def get_height(self): + return self.o_r_e - self.o_r_s + + def get_x(self): + """ + upper left x coordinate + """ + return self.o_c_s + + def get_y(self): + """ + upper left x coordinate + """ + return self.o_r_s + + def get_path(self)->pathlib.Path: + return pathlib.Path(get_tile_image_path(self)) + + def get_name(self)->str: + return pathlib.Path(get_tile_image_path(self)).name + + + +class TissueQuantity(Enum): + NONE = 0 + LOW = 1 + MEDIUM = 2 + HIGH = 3 + + + +############################# functions ######################################### + +def show_np_with_bboxes(img:numpy.ndarray, bboxes:List[numpy.ndarray], figsize:tuple=(10,10)): + """ + Arguments: + img: img as numpy array + bboxes: List of bounding boxes where each bbox is a numpy array: + array([ x-upper-left, y-upper-left, width, height]) + e.g. array([ 50., 211., 17., 19.]) + """ + # Create figure and axes + fig,ax = plt.subplots(1,1,figsize=figsize) + # Display the image + ax.imshow(img) + # Create a Rectangle patch for each bbox + for b in bboxes: + rect = matplotlib.patches.Rectangle((b[0],b[1]),b[2],b[3],linewidth=1,edgecolor='r',facecolor='none') + # Add the patch to the Axes + ax.add_patch(rect) + plt.show() + +def show_wsi_with_marked_tiles(figsize:Tuple[int] = (10,10), + scale_factor:int = 32, + tilesummary:TileSummary=None, + wsi_path:pathlib.Path=None, + df_tiles:pandas.DataFrame=None, + level:int = 0): + """ + Either provide a TileSummary object or wsi_path, df_tiles and level. + + Loads a whole slide image, scales it down, converts it into a numpy array and shows it with a grid overlay for all tiles + that passed scoring to visualize which tiles e.g. "tiles.WsiOrROIToTilesMultithreaded" calculated as worthy to keep. + Arguments: + figsize: Size of the plotted matplotlib figure containing the image. + scale_factor: The larger, the faster this method works, but the plotted image has less resolution. + tilesummary: a TileSummary object of one wsi + wsi_path: Path to a whole-slide image + df_tiles: A pandas dataframe from e.g. "tiles.WsiOrROIToTilesMultithreaded" with spacial information about all tiles + level: The level that was specified in e.g. "tiles.WsiOrROIToTilesMultithreaded". 0 means highest magnification. + """ + if tilesummary != None: + wsi_pil, large_w, large_h, new_w, new_h, best_level_for_downsample = wsi_to_scaled_pil_image(tilesummary.wsi_path, + scale_factor=tilesummary.scale_factor, + level=tilesummary.level) + wsi_np = util.pil_to_np_rgb(wsi_pil) + boxes =[] + for tile in tilesummary.top_tiles(): + box = np.array([tile.get_x(), tile.get_y(), tile.get_width(), tile.get_height()])/scale_factor + boxes.append(box) + show_np_with_bboxes(wsi_np, boxes, figsize) + + else: + wsi_pil, large_w, large_h, new_w, new_h, best_level_for_downsample = wsi_to_scaled_pil_image(wsi_path, + scale_factor=scale_factor, + level=level) + wsi_np = util.pil_to_np_rgb(wsi_pil) + boxes =[] + for index, row in df_tiles.iterrows(): + if row['wsi_path'] == wsi_path: + box = np.array([row['x_upper_left'], row['y_upper_left'], row['pixels_width'], row['pixels_height']])/scale_factor + boxes.append(box) + + show_np_with_bboxes(wsi_np, boxes, figsize) + + +def scoring_function_1(tissue_percent, combined_factor): + """ + use this, if you want tissue with lots of cells (lots of hematoxylin stained tissue) + """ + return tissue_percent * combined_factor / 1000.0 + +def scoring_function_2(tissue_percent, combined_factor): + """ + use this, if you mostly care that there is any tissue in the tile + """ + return (tissue_percent ** 2) * np.log(1 + combined_factor) / 1000.0 + + +def ExtractTileFromWSI(path:Union[str, pathlib.Path], x:int, y:int, width:int, height:int, level:int)-> PIL.Image: + """ + Args: + path: path to wsi + x: x-coordinate of the upper left pixel. The method assumes, that you know the dimensions of your specified level. + y: y-coordinate of the upper left pixel. The method assumes, that you know the dimensions of your specified level. + width: tile width + height: tile height + level: Level of the WSI you want to extract the tile from. 0 means highest resolution. + + Return: + tile as PIL.Image as RGB + """ + s = slide.open_slide(str(path)) + tile_region = s.read_region((x, y), level, (width, height)) + # RGBA to RGB + pil_img = tile_region.convert("RGB") + return pil_img + +def ExtractTileFromPILImage(path:Union[str, pathlib.Path], x:int, y:int, width:int, height:int)-> PIL.Image: + """ + Args: + path: path to PIL Image + x: x-coordinate of the upper left pixel + y: y-coordinate of the upper left pixel + width: tile width + height: tile height + + Return: + tile as PIL.Image as RGB + """ + pil_img = PIL.Image.open(path) + pil_img = pil_img.crop((x, y, x+width, y+height)) + return pil_img + +def get_roi_name_from_path_pituitary_adenoma_entities(roi_path): + path = Path(roi_path) + split = path.stem.split('-') + if split[2] == 'HE': + return f'{split[0]}-{split[1]}-{split[2]}-{split[3]}-{split[4]}' + else: + return f'{split[0]}-{split[1]}-{split[2]}-{split[3]}-{split[4]}-{split[5]}' + +def get_wsi_name_from_path_pituitary_adenoma_entities(wsi_path): + path = Path(wsi_path) + split = path.stem.split('-') + return f'{split[0]}-{split[1]}-{split[2]}-{split[3]}' + + +def WsiOrROIToTiles(wsiPath:pathlib.Path, + tilesFolderPath:pathlib.Path, + tile_height:int, + tile_width:int, + tile_naming_func:Callable, + tile_score_thresh:float = 0.55, + tile_scoring_function = scoring_function_1, + is_wsi:bool = True, + level = 0, + save_tiles:bool = False, + return_as_tilesummary_object = False)-> Union[TileSummary, pandas.DataFrame]: + """ + There is currently a bug with levels above 0. Tiles do not get scored correctly and empty tiles will pass scoring. + + Calculates tile coordinates and returns them in a pandas dataframe. If save_tiles == True the tiles will also be extracted + and saved from the WSI or ROI (ROI is assumed to be a "normal" image format like .png). + + Arguments: + wsiPath: Path to a WSI or ROI + tilesFolderPath: The folder where the extracted tiles will be saved (only needed if save_tiles=True). + tileHeigth: Number of pixels tile height. + tileWidth: Number of pixels tile width. + tile_score_thresh: Tiles with a score higher than the number from "tileScoringFunction" will be saved. + tileScoringFunction: Function to score one tile to determine if it should be saved or not. + is_wsi: if true, a WSI format like .ndpi is assumed, if false, a format like png is assumed (ROI) + tile_naming_func: A function, that takes a pathlib.Path to the WSI or ROI as an argument and returns a string. + This string will then be used as part of the name for the tile (plus some specific tile information and + the file format .png, whick is generated by this library). + level: Level of the WSI you want to extract the tile from. 0 means highest resolution. + save_tiles: if True the tiles will be extracted and saved to {tilesFolderPath} + return_as_tilesummary_object: return_as_tilesummary_object: Set this to true, if you + want the TileSummary object and not a pandas dataframe. + Return: + if return_as_tilesummary_object == True: + a TileSummary object will be returned + else: + pandas dataframe with coloumns: ['tile_name','wsi_path','level','x_upper_left','y_upper_left','pixels_width','pixels_height'] + """ + if(not is_wsi and level != 0): + raise ValueError("Specifiying a level only makes sense when extracting tiles from WSIs. Just leave the default value.") + if(tilesFolderPath is None and save_tiles == True): + raise ValueError("You should specify a {tilesFolderPath}") + + print(f"Starting to process {str(wsiPath)}") + if(is_wsi): + scale_factor = 32 + else: + scale_factor = 1 + ### against DecompressionBombWarning + #mage.MAX_IMAGE_PIXELS = 10000000000000 + openslide.lowlevel._load_image = openslide_overwrite._load_image + if(is_wsi): + img_pil, original_width, original_height, scaled_width, scaled_height, best_level_for_downsample = wsi_to_scaled_pil_image(wsiPath, scale_factor, level) + else: + img_pil = Image.open(wsiPath) + original_width = scaled_width = img_pil.width + original_height = scaled_height = img_pil.height + best_level_for_downsample = 0 + + img_pil_filtered = filter.filter_img(img_pil) + tilesummary = create_tilesummary(wsiPath, + is_wsi, + tilesFolderPath, + img_pil, + img_pil_filtered, + original_width, + original_height, + scaled_width, + scaled_height, + tile_height, + tile_width, + scale_factor, + tile_score_thresh, + tile_scoring_function, + tile_naming_func, + level, + best_level_for_downsample) + + if(save_tiles): + for tile in tilesummary.top_tiles(): + tile.save_tile() + + if return_as_tilesummary_object: + return tilesummary + + else: + rows_list = [] + for tile in tilesummary.top_tiles(): + row = {'tile_name':tile.get_name(), + 'wsi_path':tile.wsi_path, + 'level':tile.level, + 'x_upper_left':tile.get_x(), + 'y_upper_left':tile.get_y(), + 'pixels_width':tile.get_width(), + 'pixels_height':tile.get_height()} + rows_list.append(row) + + if(len(rows_list) == 0): + return pd.DataFrame(columns=['tile_name','wsi_path', \ + 'level','x_upper_left','y_upper_left','pixels_width','pixels_height']) + else: + return pd.DataFrame(rows_list).set_index('tile_name', inplace=False) + + +def WsiOrROIToTilesMultithreaded(wsiPaths:List[pathlib.Path], + tilesFolderPath:pathlib.Path, + tileHeight:int, + tileWidth:int, + tile_naming_func:Callable, + tile_score_thresh:float = 0.55, + tileScoringFunction = scoring_function_1, + is_wsi = True, + level = 0, + save_tiles:bool = False, + return_as_tilesummary_object = False)-> Union[List[TileSummary], pandas.DataFrame]: + """ + The method WsiOrROIToTiles for a list of WSIs/ROIs in parallel on multiple threads. + + Arguments: + wsiPaths: A list of paths to the WSIs or ROIs + tilesFolderPath: The folder where the extracted tiles will be saved (only needed if save_tiles=True). + tileHeigth: Number of pixels tile height. + tileWidth: Number of pixels tile width. + tile_score_thresh: Tiles with a score higher than the number from "tileScoringFunction" will be saved. + tileScoringFunction: Function to score one tile to determine if it should be saved or not. + is_wsi: if true, a WSI format like .ndpi is assumed, if false, a format like png is assumed (ROI) + tile_naming_func: A function, that takes a pathlib.Path to the WSI or ROI as an argument and returns a string. + This string will then be used as part of the name for the tile (plus some specific tile information and + the file format .png, whick is generated by this library). + level: Level of the WSI you want to extract the tile from. 0 means highest resolution. + save_tiles: if True the tiles will be extracted and saved to {tilesFolderPath} + return_as_tilesummary_object: Set this to true, if you want the TileSummary object and not a pandas dataframe. + Return: + if return_as_tilesummary_object == True: + a List of TileSummary objects will be returned + else: + pandas dataframe with coloumns: ['tile_name','wsi_path','level','x_upper_left','y_upper_left','pixels_width','pixels_height'] + """ + + pbar = tqdm(total=len(wsiPaths)) + results = [] + def update(res): + results.append(res) + pbar.update() + + with multiprocessing.Pool() as pool: + for p in wsiPaths: + pool.apply_async(WsiOrROIToTiles, + args=(p, + tilesFolderPath, + tileHeight, + tileWidth, + tile_naming_func, + tile_score_thresh, + tileScoringFunction, + is_wsi, + level, + save_tiles, + return_as_tilesummary_object), + callback=update) + + + pool.close() + pool.join() + + if return_as_tilesummary_object: + return results + else: + merged_df = None + for res in tqdm(results): + if merged_df is None: + merged_df = res + else: + merged_df = merged_df.append(res, sort=False) + + return merged_df.drop_duplicates(inplace=False) + + +def wsi_to_scaled_pil_image(wsi_filepath:pathlib.Path, scale_factor = 32, level = 0): + """ + Convert a WSI training slide to a PIL image. + + Args: + + Returns: + + """ + #wsi = openslide.open_slide(str(wsi_filepath)) + #large_w, large_h = wsi.dimensions + #new_w = math.floor(large_w / scale_factor) + #new_h = math.floor(large_h / scale_factor) + #level = wsi.get_best_level_for_downsample(scale_factor) + #img = wsi.read_region((0, 0), level, wsi.level_dimensions[level]) + #img = img.convert("RGB") + #if(scale_factor > 1): + # img = img.resize((new_w, new_h), PIL.Image.BILINEAR) + #return img, large_w, large_h, new_w, new_h + + wsi = openslide.open_slide(str(wsi_filepath)) + large_w, large_h = wsi.level_dimensions[level] + best_level_for_downsample = wsi.get_best_level_for_downsample(scale_factor) + new_w, new_h = wsi.level_dimensions[best_level_for_downsample] + img = wsi.read_region((0, 0), best_level_for_downsample, wsi.level_dimensions[best_level_for_downsample]) + img = img.convert("RGB") + return img, large_w, large_h, new_w, new_h, best_level_for_downsample + + + +def create_tilesummary(wsiPath, + is_wsi, + tilesFolderPath, + img_pil:PIL.Image.Image, + img_pil_filtered:PIL.Image.Image, + wsi_original_width:int, + wsi_original_height:int, + wsi_scaled_width:int, + wsi_scaled_height:int, + tile_height:int, + tile_width:int, + scale_factor:int, + tile_score_thresh:float, + tile_scoring_function, + tile_naming_func, + level:int, + best_level_for_downsample:int = 0)->TileSummary: + """ + + Args: + + """ + np_img = util.pil_to_np_rgb(img_pil) + np_img_filtered = util.pil_to_np_rgb(img_pil_filtered) + + tile_sum = score_tiles(np_img, + np_img_filtered, + wsiPath, + is_wsi, + tilesFolderPath, + tile_height, + tile_width, + scale_factor, + wsi_original_width, + wsi_original_height, + wsi_scaled_width, + wsi_scaled_height, + tile_score_thresh, + tile_scoring_function, + tile_naming_func, + level, + best_level_for_downsample) + + return tile_sum + + +def get_num_tiles(rows, cols, row_tile_size, col_tile_size): + """ + Obtain the number of vertical and horizontal tiles that an image can be divided into given a row tile size and + a column tile size. + + Args: + rows: Number of rows. + cols: Number of columns. + row_tile_size: Number of pixels in a tile row. + col_tile_size: Number of pixels in a tile column. + + Returns: + Tuple consisting of the number of vertical tiles and the number of horizontal tiles that the image can be divided + into given the row tile size and the column tile size. + """ + num_row_tiles = math.ceil(rows / row_tile_size) + num_col_tiles = math.ceil(cols / col_tile_size) + return num_row_tiles, num_col_tiles + + +def get_tile_indices(rows, cols, row_tile_size, col_tile_size): + """ + Obtain a list of tile coordinates (starting row, ending row, starting column, ending column, row number, column number). + + Args: + rows: Number of rows. + cols: Number of columns. + row_tile_size: Number of pixels in a tile row. + col_tile_size: Number of pixels in a tile column. + + Returns: + List of tuples representing tile coordinates consisting of starting row, ending row, + starting column, ending column, row number, column number. + """ + indices = list() + num_row_tiles, num_col_tiles = get_num_tiles(rows, cols, row_tile_size, col_tile_size) + for r in range(0, num_row_tiles): + start_r = r * row_tile_size + end_r = ((r + 1) * row_tile_size) if (r < num_row_tiles - 1) else rows + for c in range(0, num_col_tiles): + start_c = c * col_tile_size + end_c = ((c + 1) * col_tile_size) if (c < num_col_tiles - 1) else cols + indices.append((start_r, end_r, start_c, end_c, r + 1, c + 1)) + return indices + + + +def tile_to_pil_tile(tile:Tile, is_wsi:bool): + """ + Convert tile information into the corresponding tile as a PIL image read from the whole-slide image file. + + Args: + tile: Tile object. + is_wsi: if true, a WSI format like .ndpi is assumed, if false, a format like png is assumed (ROI) + + Return: + Tile as a PIL image. + """ + #x, y = tile.o_c_s, tile.o_r_s + #width, height = tile.o_c_e - tile.o_c_s, tile.o_r_e - tile.o_r_s + x = tile.get_x() + y = tile.get_y() + width = tile.get_width() + height = tile.get_height() + if(is_wsi): + pil_img = ExtractTileFromWSI(tile.wsi_path, x, y, width, height, tile.level) + else: + pil_img = ExtractTileFromPILImage(tile.wsi_path, x, y, width, height) + return pil_img + + +def tile_to_np_tile(tile, is_wsi:bool): + """ + Convert tile information into the corresponding tile as a NumPy image read from the whole-slide image file. + + Args: + tile: Tile object. + is_wsi: if true, a WSI format like .ndpi is assumed, if false, a format like png is assumed (ROI) + + Return: + Tile as a NumPy image. + """ + pil_img = tile_to_pil_tile(tile, is_wsi) + np_img = util.pil_to_np_rgb(pil_img) + return np_img + + + +def get_tile_image_path(tile:Tile): + """ + Obtain tile image path based on tile information such as row, column, row pixel position, column pixel position, + pixel width, and pixel height. + + Args: + tile: Tile object. + + Returns: + Path to image tile. + """ + t = tile + if tile.tiles_folder_path is None: + return os.path.join(tile.tile_naming_func(tile.wsi_path) + "-" + 'tile' + "-r%d-c%d-x%d-y%d-w%d-h%d" % ( + t.r, t.c, t.o_c_s, t.o_r_s, t.o_c_e - t.o_c_s, t.o_r_e - t.o_r_s) + "." + 'png') + else: + return os.path.join(tile.tiles_folder_path, + tile.tile_naming_func(tile.wsi_path) + "-" + 'tile' + "-r%d-c%d-x%d-y%d-w%d-h%d" % ( + t.r, t.c, t.o_c_s, t.o_r_s, t.o_c_e - t.o_c_s, t.o_r_e - t.o_r_s) + "." + 'png') + + + +def save_display_tile(tile, save, display, is_wsi:bool): + """ + Save and/or display a tile image. + + Args: + tile: Tile object. + save: If True, save tile image. + display: If True, dispaly tile image. + """ + tile_pil_img = tile_to_pil_tile(tile, is_wsi) + + if save: + t = Time() + img_path = get_tile_image_path(tile) + dir = os.path.dirname(img_path) + if not os.path.exists(dir): + os.makedirs(dir) + tile_pil_img.save(img_path) + #print("%-20s | Time: %-14s Name: %s" % ("Save Tile", str(t.elapsed()), img_path)) + + if display: + tile_pil_img.show() + + + +def score_tiles(img_np:np.array, + img_np_filtered:np.array, + wsi_path:pathlib.Path, + is_wsi:bool, + tilesFolderPath:pathlib.Path, + tile_height:int, + tile_width:int, + scale_factor:int, + wsi_original_width:int, + wsi_original_height:int, + wsi_scaled_width:int, + wsi_scaled_height:int, + tile_score_thresh:float, + tile_scoring_function, + tile_naming_func, + level:int, + best_level_for_downsample:int) -> TileSummary: + """ + Score all tiles for a slide and return the results in a TileSummary object. + + Args: + + Returns: + TileSummary object which includes a list of Tile objects containing information about each tile. + """ + #img_path = slide.get_filter_image_result(slide_num) + #o_w, o_h, w, h = slide.parse_dimensions_from_image_filename(img_path) + #np_img = slide.open_image_np(img_path) + + #tile_height_scaled = round(tile_height / scale_factor) # use round? + #tile_width_scaled = round(tile_width / scale_factor) # use round? + + real_scale_factor = int(math.pow(2,best_level_for_downsample-level)) + tile_height_scaled = round(tile_height / real_scale_factor) # use round? + tile_width_scaled = round(tile_width / real_scale_factor) # use round? + + num_row_tiles, num_col_tiles = get_num_tiles(wsi_scaled_height, + wsi_scaled_width, + tile_height_scaled, + tile_width_scaled) + + tile_sum = TileSummary(wsi_path=wsi_path, + is_wsi=is_wsi, + tiles_folder_path=tilesFolderPath, + orig_w=wsi_original_width, + orig_h=wsi_original_height, + orig_tile_w=tile_width, + orig_tile_h=tile_height, + scale_factor=scale_factor, + scaled_w=wsi_scaled_width, + scaled_h=wsi_scaled_height, + scaled_tile_w=tile_width_scaled, + scaled_tile_h=tile_height_scaled, + tissue_percentage=filter.tissue_percent(img_np_filtered), + num_col_tiles=num_col_tiles, + num_row_tiles=num_row_tiles, + tile_score_thresh=tile_score_thresh, + level=level, + best_level_for_downsample=best_level_for_downsample, + real_scale_factor=real_scale_factor) + + + count = 0 + high = 0 + medium = 0 + low = 0 + none = 0 + tile_indices = get_tile_indices(wsi_scaled_height, wsi_scaled_width, tile_height_scaled, tile_width_scaled) + for t in tile_indices: + count += 1 # tile_num + r_s, r_e, c_s, c_e, r, c = t + np_tile = img_np_filtered[r_s:r_e, c_s:c_e] + t_p = filter.tissue_percent(np_tile) + amount = tissue_quantity(t_p) + if amount == TissueQuantity.HIGH: + high += 1 + elif amount == TissueQuantity.MEDIUM: + medium += 1 + elif amount == TissueQuantity.LOW: + low += 1 + elif amount == TissueQuantity.NONE: + none += 1 + + o_c_s, o_r_s = slide.small_to_large_mapping((c_s, r_s), (wsi_original_width, wsi_original_height), real_scale_factor) + #print("o_c_s: " + str(o_c_s)) + #print("o_r_s: " + str(o_r_s)) + o_c_e, o_r_e = slide.small_to_large_mapping((c_e, r_e), (wsi_original_width, wsi_original_height), real_scale_factor) + #print("o_c_e: " + str(o_c_e)) + #print("o_r_e: " + str(o_r_e)) + + # pixel adjustment in case tile dimension too large (for example, 1025 instead of 1024) + if (o_c_e - o_c_s) > tile_width: + o_c_e -= 1 + if (o_r_e - o_r_s) > tile_height: + o_r_e -= 1 + + score, color_factor, s_and_v_factor, quantity_factor = score_tile(np_tile, t_p, r, c, tile_scoring_function) + + np_tile #if small_tile_in_tile else None + + tile = Tile(tile_sum, wsi_path, is_wsi, tilesFolderPath, np_tile, count, r, c, r_s, r_e, c_s, c_e, o_r_s, o_r_e, o_c_s, + o_c_e, t_p, color_factor, s_and_v_factor, quantity_factor, score, tile_naming_func, level, + best_level_for_downsample, real_scale_factor) + tile_sum.tiles.append(tile) + + tile_sum.count = count + tile_sum.high = high + tile_sum.medium = medium + tile_sum.low = low + tile_sum.none = none + + tiles_by_score = tile_sum.tiles_by_score() + rank = 0 + for t in tiles_by_score: + rank += 1 + t.rank = rank + + return tile_sum + + + +def score_tile(np_tile, tissue_percent, row, col, scoring_function): + """ + Score tile based on tissue percentage, color factor, saturation/value factor, and tissue quantity factor. + + Args: + np_tile: Tile as NumPy array. + tissue_percent: The percentage of the tile judged to be tissue. + slide_num: Slide number. + row: Tile row. + col: Tile column. + + Returns tuple consisting of score, color factor, saturation/value factor, and tissue quantity factor. + """ + color_factor = hsv_purple_pink_factor(np_tile) + s_and_v_factor = hsv_saturation_and_value_factor(np_tile) + amount = tissue_quantity(tissue_percent) + quantity_factor = tissue_quantity_factor(amount) + combined_factor = color_factor * s_and_v_factor + score = scoring_function(tissue_percent, combined_factor) + + #if combined_factor != 0.0 or tissue_percent != 0.0: + # print(f'before: {score}') + + # scale score to between 0 and 1 + score = 1.0 - (10.0 / (10.0 + score)) + + #if combined_factor != 0.0 or tissue_percent != 0.0: + # print(f'after: {score}') + + return score, color_factor, s_and_v_factor, quantity_factor + +def tissue_quantity_factor(amount): + """ + Obtain a scoring factor based on the quantity of tissue in a tile. + + Args: + amount: Tissue amount as a TissueQuantity enum value. + + Returns: + Scoring factor based on the tile tissue quantity. + """ + if amount == TissueQuantity.HIGH: + quantity_factor = 1.0 + elif amount == TissueQuantity.MEDIUM: + quantity_factor = 0.2 + elif amount == TissueQuantity.LOW: + quantity_factor = 0.1 + else: + quantity_factor = 0.0 + return quantity_factor + + +def tissue_quantity(tissue_percentage): + """ + Obtain TissueQuantity enum member (HIGH, MEDIUM, LOW, or NONE) for corresponding tissue percentage. + + Args: + tissue_percentage: The tile tissue percentage. + + Returns: + TissueQuantity enum member (HIGH, MEDIUM, LOW, or NONE). + """ + if tissue_percentage >= TISSUE_HIGH_THRESH: + return TissueQuantity.HIGH + elif (tissue_percentage >= TISSUE_LOW_THRESH) and (tissue_percentage < TISSUE_HIGH_THRESH): + return TissueQuantity.MEDIUM + elif (tissue_percentage > 0) and (tissue_percentage < TISSUE_LOW_THRESH): + return TissueQuantity.LOW + else: + return TissueQuantity.NONE + + + +def rgb_to_hues(rgb): + """ + Convert RGB NumPy array to 1-dimensional array of hue values (HSV H values in degrees). + + Args: + rgb: RGB image as a NumPy array + + Returns: + 1-dimensional array of hue values in degrees + """ + hsv = filter.filter_rgb_to_hsv(rgb, display_np_info=False) + h = filter.filter_hsv_to_h(hsv, display_np_info=False) + return h + + +def hsv_saturation_and_value_factor(rgb): + """ + Function to reduce scores of tiles with narrow HSV saturations and values since saturation and value standard + deviations should be relatively broad if the tile contains significant tissue. + + Example of a blurred tile that should not be ranked as a top tile: + ../data/tiles_png/006/TUPAC-TR-006-tile-r58-c3-x2048-y58369-w1024-h1024.png + + Args: + rgb: RGB image as a NumPy array + + Returns: + Saturation and value factor, where 1 is no effect and less than 1 means the standard deviations of saturation and + value are relatively small. + """ + hsv = filter.filter_rgb_to_hsv(rgb, display_np_info=False) + s = filter.filter_hsv_to_s(hsv) + v = filter.filter_hsv_to_v(hsv) + s_std = np.std(s) + v_std = np.std(v) + if s_std < 0.05 and v_std < 0.05: + factor = 0.4 + elif s_std < 0.05: + factor = 0.7 + elif v_std < 0.05: + factor = 0.7 + else: + factor = 1 + + factor = factor ** 2 + return factor + + +def hsv_purple_deviation(hsv_hues): + """ + Obtain the deviation from the HSV hue for purple. + + Args: + hsv_hues: NumPy array of HSV hue values. + + Returns: + The HSV purple deviation. + """ + purple_deviation = np.sqrt(np.mean(np.abs(hsv_hues - HSV_PURPLE) ** 2)) + return purple_deviation + + +def hsv_pink_deviation(hsv_hues): + """ + Obtain the deviation from the HSV hue for pink. + + Args: + hsv_hues: NumPy array of HSV hue values. + + Returns: + The HSV pink deviation. + """ + pink_deviation = np.sqrt(np.mean(np.abs(hsv_hues - HSV_PINK) ** 2)) + return pink_deviation + + +def hsv_purple_pink_factor(rgb): + """ + Compute scoring factor based on purple and pink HSV hue deviations and degree to which a narrowed hue color range + average is purple versus pink. + + Args: + rgb: Image an NumPy array. + + Returns: + Factor that favors purple (hematoxylin stained) tissue over pink (eosin stained) tissue. + """ + hues = rgb_to_hues(rgb) + hues = hues[hues >= 260] # exclude hues under 260 + hues = hues[hues <= 340] # exclude hues over 340 + if len(hues) == 0: + return 0 # if no hues between 260 and 340, then not purple or pink + pu_dev = hsv_purple_deviation(hues) + pi_dev = hsv_pink_deviation(hues) + avg_factor = (340 - np.average(hues)) ** 2 + + if pu_dev == 0: # avoid divide by zero if tile has no tissue + return 0 + + factor = pi_dev / pu_dev * avg_factor + return factor + + +def hsv_purple_vs_pink_average_factor(rgb, tissue_percentage): + """ + Function to favor purple (hematoxylin) over pink (eosin) staining based on the distance of the HSV hue average + from purple and pink. + + Args: + rgb: Image as RGB NumPy array + tissue_percentage: Amount of tissue on the tile + + Returns: + Factor, where >1 to boost purple slide scores, <1 to reduce pink slide scores, or 1 no effect. + """ + + factor = 1 + # only applies to slides with a high quantity of tissue + if tissue_percentage < TISSUE_HIGH_THRESH: + return factor + + hues = rgb_to_hues(rgb) + hues = hues[hues >= 200] # Remove hues under 200 + if len(hues) == 0: + return factor + avg = np.average(hues) + # pil_hue_histogram(hues).show() + + pu = HSV_PURPLE - avg + pi = HSV_PINK - avg + pupi = pu + pi + # print("Av: %4d, Pu: %4d, Pi: %4d, PuPi: %4d" % (avg, pu, pi, pupi)) + # Av: 250, Pu: 20, Pi: 80, PuPi: 100 + # Av: 260, Pu: 10, Pi: 70, PuPi: 80 + # Av: 270, Pu: 0, Pi: 60, PuPi: 60 ** PURPLE + # Av: 280, Pu: -10, Pi: 50, PuPi: 40 + # Av: 290, Pu: -20, Pi: 40, PuPi: 20 + # Av: 300, Pu: -30, Pi: 30, PuPi: 0 + # Av: 310, Pu: -40, Pi: 20, PuPi: -20 + # Av: 320, Pu: -50, Pi: 10, PuPi: -40 + # Av: 330, Pu: -60, Pi: 0, PuPi: -60 ** PINK + # Av: 340, Pu: -70, Pi: -10, PuPi: -80 + # Av: 350, Pu: -80, Pi: -20, PuPi: -100 + + if pupi > 30: + factor *= 1.2 + if pupi < -30: + factor *= .8 + if pupi > 0: + factor *= 1.2 + if pupi > 50: + factor *= 1.2 + if pupi < -60: + factor *= .8 + + return factor \ No newline at end of file diff --git a/deephistopath/wsi/util.py b/wsi/util.py old mode 100644 new mode 100755 similarity index 83% rename from deephistopath/wsi/util.py rename to wsi/util.py index c2e3a9c..38ab6f5 --- a/deephistopath/wsi/util.py +++ b/wsi/util.py @@ -14,14 +14,43 @@ # # ------------------------------------------------------------------------ +import pathlib +from pathlib import Path import datetime import numpy as np from PIL import Image, ImageDraw, ImageFont +import matplotlib.pyplot as plt +import matplotlib.image as mpimg +import fastai # If True, display additional NumPy array stats (min, max, mean, is_binary). ADDITIONAL_NP_STATS = False + +def show_np(np): + return util.np_to_pil(np) + +def show_multiple_images(paths:list, rows = 3, figsize=(128, 64)): + """ + Args: + paths: A list of paths to images. + """ + imgs = [fastai.vision.open_image(p) for p in paths] + fastai.vision.show_all(imgs=imgs, r=rows, figsize=figsize) + +def show_multiple_images_big(paths:list, axis_off:bool = False): + """ + Args: + paths: A list of paths to images. + """ + for p in paths: + plt.imshow(mpimg.imread(str(p))) + if(axis_off): + plt.axis('off') + plt.show() + + def pil_to_np_rgb(pil_img): """ Convert a PIL Image to a NumPy array. @@ -34,9 +63,9 @@ def pil_to_np_rgb(pil_img): Returns: The PIL image converted to a NumPy array. """ - t = Time() + #t = Time() rgb = np.asarray(pil_img) - np_info(rgb, "RGB", t.elapsed()) + #np_info(rgb, "RGB", t.elapsed()) return rgb @@ -124,9 +153,9 @@ def mask_rgb(rgb, mask): Returns: NumPy array representing an RGB image with mask applied. """ - t = Time() + #t = Time() result = rgb * np.dstack([mask, mask, mask]) - np_info(result, "Mask RGB", t.elapsed()) + #np_info(result, "Mask RGB", t.elapsed()) return result