Bias-Resilient Feature Learning for Robust Domain Adaptation in Mammography

Distribution shifts are common in medical imaging datasets due to variations in scanner types, imaging protocols, and patient populations. Such shifts between source and target datasets/domains can significantly degrade the performance of deep learning models trained solely on source data. There is a strong need on developing methods that can effectively account for data shifts and heterogeneity for data harmonization. In this study, we propose Bias-Resilient Feature Learning (BRFL), a supervised domain adaptation approach designed to address this challenge. BRFL adopts a "trust but verify" strategy: the model is first pretrained on source data, then identifies potentially domain-intrinsic features by evaluating feature-label correlations on the target dataset, and is subsequently fine-tuned by swapping those dataset-dependent features. We evaluate BRFL on breast cancer diagnosis using two real-world mammography datasets. Extensive experiments show that BRFL outperforms conventional fine-tuning and dataset-merging baselines, achieving superior performance on target datasets with distribution shifts. Code is available at https://github.com/usernameoliver/BRFL.