This work is published in IEEE Transactions on Medical Imaging (https://doi.org/10.1109/TMI.2022.3176598).

Introduction

This repository contains a PyTorch implementation of a deep learning based graph-transformer for whole slide image (WSI) classification. We propose a Graph-Transformer (GT) network that fuses a graph representation of a WSI and a transformer that can generate WSI-level predictions in a computationally efficient fashion.

To demonstrate the applicability of our approach, we selected 3,024 hematoxylin and eosin WSIs of lung tumors and the oneswith normal histology from the Clinical Proteomic TumorAnalysis Consortium (CPTAC), the National Lung ScreeningTrial (NLST) and The Cancer Genome Atlas (TCGA) and developed a model to distinguish adenocarcinoma (LUAD) and squamous cell carcinoma (LSCC) from those that havenormal histology. To understand how our model processes WSI data and visualize regions that are highly associated with the class label, we proposed a novel class activation mapping technique called GraphCAM on graphs. see below:

Usage

1. Graph Construction

(a) Tiling Patch

python src/tile_WSI.py -s 512 -e 0 -j 32 -B 50 -M 20 -o <full_patch_to_output_folder> "full_path_to_input_slides/*/*.svs"

Mandatory parameters:

-s is tile_size: 512 (512x512 pixel tiles)

-e is overlap, 0 (no overlap between adjacent tiles). Important: the overlap is defined as "the number of extra pixels to add to each interior edge of a tile". Which means that the final tile size is s + 2.e. So to get a 512px tile with a 50% overlap, you need to set s to 256 and e to 128. Also, tile from the edges of the slide will be smaller (since up to two sides have no "interior" edge)

-j is number of threads: 32

-B is Max Percentage of Background allowed: 50% (tiles removed if background percentage above this value)

-o is the path were the output images must be saved

-M set to -1 by default to tile the image at all magnifications. Set it to the value of the desired magnification to tile only at that magnification and save space

(b) Training Patch Feature Extractor

Go to './feature_extractor' and config 'config.yaml' before training. The trained feature extractor based on contrastive learning is saved in folder './feature_extractor/runs'. We train the model with patches cropped in single magnification (20X). Before training, put paths to all pathces in 'all_patches.csv' file.

python run.py

You could use pretrained feature extractor: feature_extractor/model.pth. The pre-trained models can be downloaded.

(c) Constructing Graph

Go to './feature_extractor' and build graphs from patches:

python build_graphs.py --weights "path_to_pretrained_feature_extractor" --dataset "path_to_patches" --output "../graphs"

2. Training Graph-Transformer

Run the following script to train and store the model and logging files under "graph_transformer/saved_models" and "graph_transformer/runs".

bash scripts/train.sh

To evaluate the model. run bash scripts/test.sh

Split training, validation, and testing dataset and store them in text files as:

sample1 \t label1
sample2 \t label2
LUAD/C3N-00293-23 \t luad
...

3. GraphCAM

To generate GraphCAM of the model on the WSI:

1. bash scripts/get_graphcam.sh

To visualize the GraphCAM:

2. bash scripts/vis_graphcam.sh

Note: Currently we only support generating GraphCAM for one WSI at each time.

More GraphCAM examples:

GraphCAMs generated on WSIs across the runs performed via 5-fold cross validation are shown above. The same set of WSI regions are highlighted by our method across the various cross-validation folds, thus indicating consistency of our technique in highlighting salient regions of interest.

Requirements

WSI software: PixelView (deepPath, Inc.)

Major dependencies are:

python

pytorch

openslide-python

Weights & Biases