This repository contains the code and results included in the paper (currently under review, link coming later).
Provided are the implementations of:
- Confounder generation,
- fine-tuning, and
- performance analysis as described in the paper.
For access to the data used in the paper, please refer to:
- RadImageNet for pre-trained weights,
- NIH CXR14, and
- LIDC-IDRI.
Our dataset splits are provided in the data folder.
The main files to run are:
include_tag.py
, include_lowpass.py
, include_noise.py
, gender.py
(for confounder generation as described in the paper)
fine-tuning.py
(for fine-tuning the models on confounded targets and logging results)
results_plots.ipynb
and statistical_tests.ipynb
(for pulling results together and plotting)
Feel free to contact us for help with the reproduction of our experiments.