title	emoji	colorFrom	colorTo	sdk	app_file	pinned
Cellcounter	🔬	blue	gray	gradio	app.py	false

Cellcounter

Live demo: https://sallamsaka-cellcounter.hf.space

Cellcounter predicts a cell density map from microscopy images using a U-Net model and a total cell count from the sum of the density map.

---

Web App Summary

• Input: an image (preferably an actual microscopy image)
• Output:
  • Predicted density map
  • Predicted cell count

How to use: click a gallery image or upload your own image and inference runs immediately.

---

Brief Methodology

• Model: U-Net style CNN trained to output a density map using PyTorch.
• Training data: 2018 Data Science Bowl nuclei dataset (Kaggle). 80% of the stage1 training data was used for training and 20% for validation.
• Label construction: for each cell mask, I computed a centroid and placed a normalized Gaussian around each centroid so that each cell mask image array sums to one and the sum of all cell masks is the total number of cells.
• Data Augmentation: Doubled the training set with contrast-inverted images. Training was performed on cropped out 256x256 patches of images to cover each image.
• Loss incorporated the following factors:
  • Pixel-level value deviation from target density
  • Shape of predicted densities at different scales
  • Absolute error of total counts
• Inference: sliding-window prediction on 256×256 patches, then stitch patches into a full-image density map.

---

Results

I also compared a count-only training loss vs a multi-term (pixel + shape + count) loss.

Count-only loss

• Patch MAE: 2.615
• Full-image MAE: 6.366

Multi-term (pixel/shape/count) loss

• Patch MAE: 3.021
• Full-image MAE: 5.209

Takeaway: count-only training did slightly better on patches, but the multi-term loss produced better full-image behavior after stitching probably due to the spatial information provided by the predicted density.

---

Examples

• Example 1:

• Example 2:

---

Limitations

The model performs best on:

• cells that contrast well with the background
• uniform cell sizes/appearance

The model seems to struggle with:

• overlapping / crowded cells
• large variation in cell size
• low contrast images
• variation in transparency of cells
• colorful / complex looking cells

The Kaggle dataset was not designed specifically for density prediction, so, some inconsistencies in the data such as fractions of cells being considered a whole cell leads to inconsistencies in training.

---

Repository structure

.
├── app.py
├── best.pt
├── demo_images/
├── screenshots/
├── requirements.txt
├── cell count.ipynb
├── count_loss_trial.ipynb
└── README.md

---

Citations / credits

• Allen Goodman, Anne Carpenter, Elizabeth Park, jlefman-nvidia, Josette_BoozAllen, Kyle, Maggie, Nilofer, Peter Sedivec, and Will Cukierski. 2018 Data Science Bowl . https://kaggle.com/competitions/data-science-bowl-2018, 2018. Kaggle.
• ChatGPT (learning aid, guidance, troubleshooting, etc)