This repository is a software system containing an end-to-end Whole Slide Imaging pre-processing pipeline from The Cancer Genome Atlas download documents, as well as a complete implementation of deep learning tumor segmentation from WSI binary labels as detailed in "Weakly supervised multiple instance learning histopathological tumor segmentation".
Example of Whole Slide Image tumor segmentation (black background; blue: normal tissue; pink: neoplastic tissue).
This software is entirely written in Python3 and contains two major parts:
- a tool to automatically download data from TCGA GDC Data Portal, which also handles tiles extraction, background removal, and tumor label extraction.
- an end-to-end pytorch software that can train many types of common image classifier architectures for the task of tumor segmentation on WSI based on weak binary WSI labels indicating the presence of tumor in each WSI.
- a collection of 6481 semi-automatically generated tumor maps for the entire snap-frozen WSI of TCGA repository for breast, kidney, and bronchus and lung locations.
Use python3 and install mandatory libraries:
virtualenv -p python3 --system-site-packages venv
source venv/bin/activate
pip install -r requirements.txt
- Download the GDC Data Transfer Tool executable (not included here for license issues)
- Constitute any cohort on the TCGA GDC Data Portal, then download
the associated manifest file, and place it in a
source_folder - Launch the download and pre-processing pipeline with:
python -m code.data_processing.main --gdc gdc_executable_path source_folder
This script first downloads all files in the manifest file, then tiles WSI, extracts tiles of a given magnification, removes background tiles, and finally seeks to extract per-slide binary labels from their name. More information here (in construction).
After data download and pre-processing has been performed, launch the training pipeline using:
python -m code.training --preprocessed-data-folder ./data/preprocessed --alpha 0.1 --beta 0.
Many parameters are tunable, see python -m code.training --help
More informations about the training pipeline, including available imaging models here (in construction).
This software is released under the GNU Affero General Public License v3.0 license.
If you use this software or any part of this software in your research, please use the following BibTeX entry.
@misc{lerousseau2020weakly,
title={Weakly supervised multiple instance learning histopathological tumor segmentation},
author={Marvin Lerousseau and Maria Vakalopoulou and Marion Classe and Julien Adam and Enzo Battistella and Alexandre Carré and Théo Estienne and Théophraste Henry and Eric Deutsch and Nikos Paragios},
year={2020},
eprint={2004.05024},
archivePrefix={arXiv},
primaryClass={eess.IV}
}