Detecting Colorectal Cancer MSI/MSS Status From High Resolution WSI with Deep Learning

This repo contains code used in the paper "Patient-level Microsatellite Stability Assessment from Whole Slide Images By Combining Momentum Contrast Learning and Group Patch Embeddings" by Shats et al. This paper will appear in the proceedings of the ECCV workshop on Medical Computer Vision (ECCV-MCV 2022). You can also read it on Arxiv here. BibTeX citation below:

@article{shats2022patient,
  title={Patient-level Microsatellite Stability Assessment from Whole Slide Images By Combining Momentum Contrast Learning and Group Patch Embeddings},
  author={Shats, Daniel and Hezi, Hadar and Shani, Guy and Maruvka, Yosef E and Freiman, Moti},
  journal={arXiv preprint arXiv:2208.10429},
  year={2022}
}

Dependencies

rich
wandb
numpy
pandas
PIL
pytorch
torchvision
torchmetrics
pytorch-lightning
lightly

Relevant Files

Data

src/data_stuff.sh download the subset of TCGA used in our experiments.

src/data_stuff/convert_tcga_to_folders.py converts the downloaded tcga data to folders in a hierarchy as follows:

train/
--- MSS/
--- --- patient1/
--- --- --- patch1.png
--- --- --- patch2.png
--- --- --- ...
--- --- patient2/
--- --- --- patch1.png
--- --- --- patch2.png
--- --- --- ...
--- --- ...
--- MSIMUT/
--- --- patient1/
--- --- --- patch1.png
--- --- --- patch2.png
--- --- --- ...
--- --- patient2/
--- --- --- patch1.png
--- --- --- patch2.png
--- --- --- ...
--- --- ...

Training

resnet_experiment.py trains the baseline approach that was used for comparison against our method. It is our implemenation of the work in Kather et al.

moco_experiment.py is used to train a patch feature extractor using momentum contrast learning.

downstream_experiment.py uses the trained feature extractor from moco_experiment.py to now train a group-patch level MLP which can classify WSI's with higher accuracy.