Domain Adaptation for Breast MRI: Deploying AI to New Clinical Sites Without Manual Labels

Artificial intelligence (AI) models for breast MRI often perform well on their training datasets but face challenges when deployed at different hospitals. This performance drop occurs because each clinical site utilizes different imaging scanners and scanning protocols. To overcome this domain shift, AI models typically require retraining with new expert-annotated datasets, a process that is both time-consuming and expensive. We developed a self-supervised framework that adapts an existing breast lesion segmentation model to new hospital data without requiring any new manual annotations.