Source-free domain adaptation for multi-institutional chest X-ray images

Machine learning models for chest X-ray (CXR)-based tuberculosis (TB) detection often suffer from performance drops when deployed across institutions due to domain shifts caused by differences in patient populations, imaging protocols, and equipment. Domain adaptation has been widely studied to address the problem, but privacy regulations restrict data sharing, making conventional approaches infeasible.

We aim to enhance TB detection performance when a detector model is shared across multiple clinical sites without accessing any source data or statistics. This problem is related to source-free domain adaptation (SFDA) but requires special designs for imbalanced binary classification.

We propose two methods for TB detection, which are based on Source HypOthesis Transfer (SHOT) and SFDA via source Distribution Estimation (SFDA-DE), respectively. As the main ideas of our methods, we propose ranking-based pseudo-labeling and selective fine-tuning of network layers to address severe class imbalance and prevent overfitting.

On six unseen target domains with high class imbalance, our proposed SFDA methods significantly improved the average F1 score and AUROC compared to non-adapted models, all without accessing source data during adaptation.

Our methods offer a privacy-preserving and effective strategy to improve the generalization of TB detection models across institutions, addressing domain shifts while complying with data-sharing restrictions.