A Graph Attention Network Approach for Histopathology Patch Classification

Link to report (PDF)

Overview:

This project contains the PyTorch implementation for a Graph Attention Network (GAT) framework for classifying colorectal cancer patches from the PathMNIST dataset. We explored both an End-to-End (raw pixel) and a Hybrid Autoencoder approach to node embeddings.

Requirements:

To set up the environment, run:

pip install torch torchvision
pip install torch-geometric
pip install medmnist
pip install pandas numpy matplotlib scikit-learn

Hardware:

The code was run on Windows 11 v24H2, using an Nvidia RTX 5070 GPU (12 GB of VRAM) with an AMD Ryzen 7 9800x3D processor, and 31.1 GB of standard RAM (6000 MT/s).

Repository Structure:

├── train_ae.ipynb              # defines and trains the Autoencoder used for the Hybrid approach.
├── train_model_gat_ae.ipynb    # uses the model from train_ae.ipynb to train the Hybrid GAT.
├── train_model_mlp.ipynb       # trains a simple MLP for baseline comparison.
├── train_model_simple.ipynb    # trains the End-to-End GAT.
├── visualize_graphs.ipynb      # visualizes the attention scores
├── download_data.ipynb         # contains the downloading data and graph creations.
├── GATModels.py                # contains the GAT model definitions. 
├── AutoEncoder.py              # contains the Autoencoder definition.
├── util/                       # contains various utility files used in graph creation and dataset representation. 
└── README.md                   # project documentation

Usage:

1. Run download_data.ipynb to download the dataset and to create the graphs.
2. Run train_model_mlp.ipynb to train the baselines.
3. Run train_ae.ipynb to train the Autoencoder to embed the node features.
4. Run train_model_gat_ae.ipynb to train the Hybrid GAT.
5. Run train_model_simple.ipynb to train the End-to-End GAT (best model performance).
6. Run visualize_graphs.ipynb to visualize the attention scores on a fixed spatial grid.

Results Summary

Model	Accuracy	AUC
GAT (End-to-End)	0.775	0.962
GAT (Hybrid)	0.689	0.942
MLP Baseline	0.638	0.925