ZoomMIL

ZoomMIL is a multiple instance learning (MIL) method that learns to perform multi-level zooming for efficient Whole-Slide Image (WSI) classification. This repository contains the PyTorch code to reproduce the results of our corresponding paper Differentiable Zooming for Multiple Instance Learning on Whole-Slide Images.

Overview

Installation

histocartography

This code relies on functionality from the histocartography library. To install it, simply clone the github repository and add the path to your PYTHONPATH:

git clone https://github.com/histocartography/histocartography.git
export PYTHONPATH="<PATH>/histocartography:$PYTHONPATH"

conda environment

Create a conda environment and install the required packages from the provided environment.yml:

git clone https://github.com/histocartography/zoommil.git && cd zoommil
conda env create -f environment.yml
conda activate zoommil

Install pytorch:

conda install -n zoommil pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=11.1 -c pytorch -c conda-forge

Preprocessing also requires the OpenSlide library. On Linux, you can install it with

conda install -n zoommil -c conda-forge conda-forge/linux-64::openslide-python

Please check the documentation for more information.

Getting started

After cloning the repository and creating the conda environment, you can follow the steps below to get started.

Datasets

We evaluated ZoomMIL on three publicly available datasets:

CRC:
- 1133 colorectal biopsy and polypectomy slides from non-neoplastic, low-grade, and high-grade lesions
- The dataset can directly be requested from the authors of CAD systems for colorectal cancer from WSI are still not ready for clinical acceptance
BRIGHT:
- 703 Breast WSIs from non-cancerous, precancerous, and cancerous subtypes
- Data: https://brightchallenge.na.icar.cnr.it/BRIGHT_Challenge/
- We only used data from the subfolders "WSIs" (not ROIs) under Train/Validation/Test
- Note that the test labels of BRIGHT are currently not public, but the BRIGHT challenge may reopen for submissions
CAMELYON16:
- 399 Breast WSIs with normal and metastatic cases
- Data: https://camelyon17.grand-challenge.org/Data/ (provides images from both CAMELYON16 and CAMELYON17)

The train/val/test splits for all datasets can be found here.

Preprocessing

Whole-slide images (e.g., from the BRIGHT dataset) can be preprocessed (tissue masking + patch feature extraction) as .h5 files:

python bin/preprocess.py --out_path <PATH_TO_PREPROCESSED_DATA> --in_path <PATH_TO_DOWNLOADED_DATA> --mode features --dataset BRIGHT

To only extract patches (without features), select the mode patches:

python bin/preprocess.py --out_path <PATH_TO_PREPROCESSED_DATA> --in_path <PATH_TO_DOWNLOADED_DATA> --mode patches --dataset BRIGHT

Training & testing

Adapt the paths in your config file, then run train.py to run the training and testing for ZoomMIL. This script expects WSIs that have been preprocessed as patch features.

python bin/train.py --config_path zoommil/config/sample_config.json

Citation

If you use this code, please consider citing our work:

@inproceedings{thandiackal2022zoommil,
  title={Differentiable Zooming for Multiple Instance Learning on Whole-Slide Images},
  author={Thandiackal, Kevin and Chen, Boqi and Pati, Pushpak and Jaume, Guillaume and Williamson, Drew FK and Gabrani, Maria and Goksel, Orcun},
  booktitle = {The European Conference on Computer Vision (ECCV)},
  year={2022}
}

cshnai / zoommil