UNet for Histopathology Cell Segmentation

This project applies a UNet architecture for semantic segmentation of nuclei in histopathology patches. The core objective is to evaluate how well synthetically generated histology images support downstream cell segmentation, particularly in low-data or domain-shift scenarios.

We use the Lizard-v2 dataset and predict both segmentation masks and nuclear subtype annotations, spanning six clinically relevant cell types:

• Epithelial
• Lymphocyte
• Plasma
• Eosinophil
• Neutrophil
• Connective Tissue

---

File Overview

• main.py — Performs segmentation on prepared patches, computes Dice scores, and saves visualized prediction masks.
• prepare_patches.ipynb — Converts raw Lizard-v2 data into model-ready patches with matching labels.

---

Expected folder structure

Make sure your data is organized like this:

../data/lizard_v2/
├── images/train/
├── classes/train/
└── instances/train/

---

Usage

1. Run prepare_patches.ipynb to preprocess and patchify Lizard-v2 images.
2. Then execute main.py to train the UNet model, perform segmentation, and evaluate performance (e.g., Dice coefficient).

---

Results

We evaluate segmentation performance using per-class Dice scores and provide qualitative comparisons of predicted vs ground-truth masks for both real and synthetic samples.