Deep Learning Utilities for Pathology: PyTorch-lightning implementation of multiple instance learning models

March 18, 2022 ยท View on GitHub

==== WORK IN PROGRESS ====

  • Add bash scripts to run compilation of features from VISSL
  • Add bash scripts to run training of all classifiers
  • Add config files
  • Improve documentation

==========================

dlup_lightning_mil offers a set MIL classification models built with PyTorch-lightning using utilities from DLUP to ease the process of running Deep Learning classification/regression models on Whole Slide Images.

Tile supervision methods can be run with just this repository.

MIL methods require feature extraction using hissl.

Reproduce DeepSMILE (to be added)

The scripts in ~/dlup-lightning-mil/tools/reproduce_deepsmile will reproduce the models as described in DeepSMILE. It is expected you run ~/hissl/tools/reproduce_deepsmile/reproduce_deepsmile.sh from hissl on the same machine and hosting the repository in the same parent directory.

Running ~/dlup-lightning-mil/tools/reproduce_deepsmile$ bash reproduce_deepsmile.sh will run the following scripts:

  • 0_check_singularity/check_singularity.sh
    • Checks if a singularity image is available to run all following scripts
  • 1_create_splits_for_tcga_crck
    • Creates 5-fold train-val splits for TCGA-CRCk from the defined train set
    • Creates subsets of the training splits for low labelled data regime experiments
  • 2_compile_features/compile_h5_features.sh
    • Reorganizes the saved features from hissl to allow for easier and quicker loading
  • 3_train_models/*
    • Runs
      • Every combination of extractor and classifier
      • On both datasets
      • For all available labels
      • For all defined fractions of training data
    • Saves to ~/dlup-lightning-mil/logs
      • model checkpoints
      • training loss
      • validation and test AUCs
      • validation and test predictions and labels
      • MIL attention or predicted score for each tile

Upcoming features

  • Analysis of results of runs
  • Visualization of high- and low-attention tiles
  • Plotting of graphs
  • More MIL models & end-to-end MIL training