Graphene Image Segmentation Workflow

This repository contains a set of Jupyter notebooks that implement the full pipeline for preparing graphene microscopy images and their segmentation masks, starting from raw images, through ilastik probability maps, to UNet-ready multi-label masks.

Overview

1. Image normalization
   Notebook: tem_normalization_clahe_minmax.ipynb

   • Prepares raw microscopy images for downstream analysis.
   • Includes intensity normalization (clahe, minmax)
   • Ensures reproducibility and standardization of input data.

2. Probability maps post-processing
   Notebook: ilastik_h5_to_mask.ipynb

   • Converts ilastik probability map outputs (.h5) into binary masks.
   • Supports user-defined thresholds and class selection.
   • Automatically scans directories for .h5 files and matching raw images.
   • Exports masks in PNG and NumPy (.npy) formats.
   • Generates overlays and probability heatmaps for quality control.

3. Label masks post-processing (UNet dataset preparation)
   Notebook: mask_postprocessing_for_unet.ipynb

   • Processes manually refined ilastik label exports (*_Labels.h5).
   • Fills unlabeled pixels (class 0) using nearest-neighbor interpolation.
   • Applies morphological smoothing to reduce edge artifacts (dilution, erosion)
   • Aligns raw images with masks, builds one-hot multi-label stacks (e.g., GFM, carbonhole, undefined).
   • Saves paired datasets as images (.png) and compressed masks (.npz), ready for supervised training.
   • Performs sanity checks to ensure non-overlapping classes and consistent pixel counts.

Folder structure

Each notebook defines its own input and output paths, but the workflow is organized around:

• notebooks/ — Jupyter notebooks for each stage of the pipeline.
• data/ — raw images and intermediate outputs.
• masks_output/, net_input_images/, net_input_masks/ — generated binary and multi-label masks.

Requirements

• Python 3.8+
• Dependencies: numpy, matplotlib, h5py, scikit-image, scipy, opencv-python, Pillow

Usage

Open each notebook in sequence to reproduce the workflow:

1. Run image normalization.
2. Run probability maps post-processing for ilastik .h5 outputs.
3. Run label masks post-processing for refined labels and dataset preparation.

---

This pipeline ensures reproducibility, cleanliness, and quality control in preparing microscopy image datasets for deep learning applications.