This repository contains the code for predicting microsatellite status in colorectal cancer from H&E-stained FFPE histopathology slides.

This code was developed and tested in the following settings.

• Ubuntu 18.04

• Nvidia GeForce RTX 2080 Ti

• python (3.6.10)
• numpy: (1.18.1)
• pandas (0.25.3)
• pillow (7.0.0)
• matplotlib (3.1.0)
• scikit-learn (0.21.3)
• scikit-image (0.15.0)
• opencv-python (4.1.2.30)
• openslide-python (1.1.1)
• staintools (2.1.2)
• h5py (2.9.0)
• pytables (3.5.1)
• pytorch (1.4.0)
• torchvision (0.5.0)
• fastai (1.0.55)

• Install Miniconda on your machine (download the distribution that comes with python3).

• After setting up Miniconda, install OpenSlide (3.4.1):

apt-get install openslide-tools

• Create a conda environment with environment.yml:

conda env create -f environment.yml

• Activate the environment:

conda activate msinet

• Download diagnostic whole-slide images from TCGA-COAD project and TCGA-READ project using GDC Data Transfer Tool Client.

gdc-client download -m gdc_manifest_tcga_coadread.txt

• Download NCT-CRC-HE-100K and CRC-VAL-HE-7K datasets (generated by Kather et al.) from here to train and test tissue type classifier.

python tissue_type_classifier.py

python svs2tile.py  
python svs_tile2hdf.py  

python tissue_type_inference.py  

python tmap2tile.py  
python tmap_tile2hdf.py  

python msi_predictor.py  

python msi_inference.py

Note: please edit paths in each .py file.

Lancet Oncology 2021;22(1):132–41

@ARTICLE{Yamashita2021deep,
  title = "Deep learning model for the prediction of microsatellite instability in colorectal cancer: a diagnostic study",
  author = "Yamashita, Rikiya and Long, Jin and Longacre, Teri and Peng, Lan and Berry, Gerald and Martin, Brock and Higgins, John and Rubin, Daniel L and Shen, Jeanne",
  journal  = "Lancet Oncol.",
  volume   =  22,
  number   =  1,
  pages    = "132--141",
  month    =  jan,
  year     =  2021,
  language = "en"
  }