Abstract
Numerous deep learning-based solutions have been developed for the automaticrecognition of breast cancer using mammography images. However, theirperformance often declines when applied to data from different domains,primarily due to domain shift - the variation in data distributions betweensource and target domains. This performance drop limits the safe and equitabledeployment of AI in real-world clinical settings. In this study, we presentDoSReMC (Domain Shift Resilient Mammography Classification), a batchnormalization (BN) adaptation framework designed to enhance cross-domaingeneralization without retraining the entire model. Using three large-scalefull-field digital mammography (FFDM) datasets - including HCTP, a newlyintroduced, pathologically confirmed in-house dataset - we conduct a systematiccross-domain evaluation with convolutional neural networks (CNNs). Our resultsdemonstrate that BN layers are a primary source of domain dependence: theyperform effectively when training and testing occur within the same domain, andthey significantly impair model generalization under domain shift. DoSReMCaddresses this limitation by fine-tuning only the BN and fully connected (FC)layers, while preserving pretrained convolutional filters. We further integratethis targeted adaptation with an adversarial training scheme, yieldingadditional improvements in cross-domain generalizability. DoSReMC can bereadily incorporated into existing AI pipelines and applied across diverseclinical environments, providing a practical pathway toward more robust andgeneralizable mammography classification systems.