simongraham / Masterclass-Hackathon

A deep learning model of mutation prediction using a form of a transfer learning from H&E images. It is provided as a baseline to the Hackathons participants in PathLAKE masterclass, Royal College of Pathologists, London 2020.

Masterclass-Hackathon

Introduction

A simple baseline provide training and validation script to predict mutation from the H&E stained whole slide images. This pipeline reproduces the results reported in [1, 2].

Dataset requirements

1. Tumor region of H&E stained whole slide images tiled into blocks of size 512 by 512, arranged in a separate directory per whole slide image.
2. Train and test set csv files are provided in the “data” folder of this repository.
3. In data/train.csv each row contains:

• “SLIDES”: are the names of directories where whole slides image are stored as tiles of size 512 by 512
• “Mutation”: is the label information for each slide (1 for mutation present, 0 otherwise)

This implementation is also inspired from and reuses few routines from another github repository [3] which contains implementation of Campanella et al. 2019.

References:

1. Kather, J.N., Pearson, A.T., Halama, N. et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat Med 25, 1054–1056 (2019) doi:10.1038/s41591-019-0462-y
2. Kather, J.N., Lara, R. Heij, et al. Pan-cancer image-based detection of clinically actionable genetic alterations. bioRxiv 833756; doi: https://doi.org/10.1101/833756
3. MIL-nature-medicine-2019.https://github.com/MSKCC-Computational-Pathology/MIL-nature-medicine-2019