Detecting Shortcuts in Medical Images - A Case Study in Chest X-rays

by Amelia Jiménez-Sánchez, Dovile Juodelyte, Bethany Chamberlain, Veronika Cheplygina

This repository provides a PyTorch implementation of our ISBI 2023 submission -> [arXiv]

Overview

Data-centric approaches, bias assessment, and validation are increasingly important as datasets get larger, but are still understudied in medical imaging. We review the literature and present a validation study on detecting shortcuts in chest X-rays. Our systematic experiments on two large benchmarks generalize earlier findings which show overoptimistic and biased performance. We share our code and a set of non-expert drain labels for CheXpert dataset under the preprocess folder.

Usage

1. Cloning the repository

$ git clone https://github.com/ameliajimenez/shortcuts-chest-xray.git
$ cd shortcuts-chest-x-ray/

2. Preprocessing: create development subsets

Detailed steps under preprocess folder.

3. Training, testing & visualizations

Detailed steps under bin folder.

Citation

If this work is useful for your research, please cite our paper:

@misc{https://doi.org/10.48550/arxiv.2211.04279,
  doi = {10.48550/ARXIV.2211.04279},
  url = {https://arxiv.org/abs/2211.04279},
  author = {Jiménez-Sánchez, Amelia and Juodelye, Dovile and Chamberlain, Bethany and Cheplygina, Veronika},
  keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Detecting Shortcuts in Medical Images - A Case Study in Chest X-rays},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}

Acknowledgments

Our repository is based on jhealthcare/CheXpert and purrlab/hiddenfeatures-chestxray. We thank Kasper Thorhauge Grønbek and Andreas Skovdal for early discussions and providing the labels used in our experiments.