UBC Ovarian Cancer Subtype Classification and Outlier Detection (UBC-OCEAN)

Introduction

This repository contains a PyTorch implementation of a graph-based vision transformer (GraphTrans) framework for whole slide image (WSI) classification. The data employed in this work can be found from the UBC-OCEAN Kaggle challenge https://www.kaggle.com/competitions/UBC-OCEAN/overview

How to use

Preprocessing

Download the UBC-OCEAN dataset and put it in ./data folder

Run the patch-tiling code using the following command

python src/tile_WSI.py -s 512 -e 0 -j 32 -B 50 -M 20 -o ./data/output_tiled ./data/UBC-OCEAN/train_images/*.png

Contruct the graph using the following command

python build_graphs.py --weights ./resnet18.pth --dataset ../data/output_tiled --output ../graphs

Training GraphTrans

Run the following command to train and store the model and log files

bash scripts/train.sh

To evaluate the model run bash script/test.sh

NOTE: You can change the training settings in the bash files.

GraphCAM

To generate the saliency maps of the model on the WSI, run the following command