Library for "Sparse Convolutional Context-Aware Multiple Instance Learning for Whole Slide Image Classification" (Best paper at the MICCAI workshop on Computational Pathology 2021).
This repository contains the full kit for training SparseConvMIL with any type of ResNet architecture as tile embedder. The provided data is from The Cancer Genome Atlas.
SparseConvMIL is a powerful and generic multiple instance learning architecture specifically designed to leverage spatial information in whole slide images. This is done by building a sparse map that contains the embeddings of sampled tiles, which are placed at the locations of the associated tiles within the source whole slide image.
This framework has shown state-of-the-art performance for subtype classification compared to conventional multiple instance learning approaches.
To run the demo version:
python -m training
More info about the hyper-parameters with:
python -m training --help
All models, including the tile embedder, the sparse-input pooling, the WSI embedding classifier and SparseConvMIL are located within the model.py file. If you want to change one of several of these, check this file!
To accomodate for your data, either copy the architecture as displayed in the sample_data, or check the dataset.py file for expected data architecture or to change data loading.
Clone this repo, create a virtual environment and download necessary packages:
git clone git@github.com:MarvinLer/SparseConvMIL
cd SparseConvMIL
virtualenv -p python3 venv; source venv/bin/activate
pip install -r requirements.txt
This library also relies on SparseConvNet:
# Get and install sparseconvnet
git clone git@github.com:facebookresearch/SparseConvNet.git
cd SparseConvNet/
bash develop.sh
If you find this code useful in your research then please cite:
@inproceedings{lerousseau2021sparseconvmil,
title={SparseConvMIL: Sparse Convolutional Context-Aware Multiple Instance Learning for Whole Slide Image Classification},
author={Lerousseau, Marvin and Vakalopoulou, Maria and Deutsch, Eric and Paragios, Nikos},
booktitle={MICCAI Workshop on Computational Pathology},
pages={129--139},
year={2021},
organization={PMLR}
}
Lerousseau, M., Vakalopoulou, M., Deutsch, E. and Paragios, N., 2021, September.
SparseConvMIL: Sparse Convolutional Context-Aware Multiple Instance Learning for Whole Slide Image Classification.
In MICCAI Workshop on Computational Pathology (pp. 129-139). PMLR.
SparseConvMIL is GNU AGPLv3 licensed, as found in the LICENSE file.