Repo for the "Re-identification of patients from imaging features extracted by foundation models" paper (currently under revision).

Installation

All needed dependencies are stored in file requirements.txt

Quickstart

We make available RETFound features extracted from the GRAPE dataset.

To reproduce our re-identification results on this dataset, we provide

• data/GRAPE/grape_img_info_w_file.csv: file listing all images and their corresponding patient (we used visit number as a proxy for timestamp)
• features/GRAPE: folder with pickle files with RETFound features extracted from the images
• results/GRAPE: folder including results from the re-identification experiments (will be overwritten when re-runnning experiments)

To run experiments, move to the re_identification_scripts folder and run the provided shell script.

cd re_identification_scripts
sh run_grape.sh

Further information

We make available the code used to:

• extract features from the RETFound and CXR Foundation models: here
• run re-identification experiments: here
• train demographic feature prediction logistic/linear regression models: here