Automatic correction of performance drift under acquisition shift in medical image classification

This repository contains all the necessary code to reproduce the experiments in our article

M. Roschewitz, G. Khara, J. Yearsley, N. Sharma, J.J. James, E. Ambrózay, A. Heroux, P. Kecskemethy, T. Rijken, B. Glocker. Automatic correction of performance drift under acquisition shift in medical image classification. Nature Communications (2023)

Preliminaries

This code base runs in Python. You can download a version of python via Anaconda, see instructions here. Additionally, you will need to install the following dependencies via pip.

pip install numpy seaborn matplotlib pandas scikit-learn jupyter tqdm

Code structure

This repository has a very simple structure:

• two notebooks necessary to reproduce all the plots for the breast screening task (see analysis_breast_screening.ipynb) and the Wilds-CameLyon histopathology task (see analysis_histopathology.ipynb). Just run all cells to re-create all the plots, it takes about 3 mins to run on the CameLyon data and 30 mins to run for the breast task (if you want to accelerate execution you can reduce the number of bootstrap samples).
• upa.py is a stand-alone file that contains all the code related to fitting and applying unsupervised prediction alignement. It is built as a class, similar to sklearn estimators with a fit and predict method, so that it can easily be applied on any model predictions. See code for in-depth documentation.
• the other files are simply utils files to compute metrics, run the experiments and prepare the plots for the manuscript.
• the necessary model outputs are provided in the breast_outputs and camelyon folders. They contain csvs with the raw model predictions and labels for each task and each dataset.