Domain-continual learning for multi-center anatomical detection via prompt-enhanced and densely-fused MedSAM

Anatomical structure detection is crucial for automated quality assessment and abnormality identification in medical imaging. Unlike natural images, medical images are often noisy and complex. Variations in imaging devices, scan protocols, and patient populations introduce domain gaps that hinder the generalization of conventional detectors in clinical settings. Thus, we propose DCA-Det, a novel multi-center detection framework tailored for fetal anatomical structure recognition in ultrasound images. The proposed framework enhances domain continual learning across multi-center scenarios by integrating the pretrained medical foundation model (MedSAM) and architectural innovations. More specifically, built on MedSAM’s strong generalization, DCA-Det improves domain adaptation ability through a two-stage training scheme that combines hierarchical freezing/unfreezing, random weight restoration, and efficient model tuning. Moreover, we introduce a lightweight pyramid architecture and a self-prompted salient region decoder to suppress noise and guide feature hierarchies toward specific anatomical targets. Benefiting from the rich semantics and global context provided by the large model backbone, a dense granularity-aware fusion module is further proposed to enhance cross-scale interaction and local fine-grained detail modeling. Extensive experiments on a multi-center fetal ultrasound dataset confirm the effectiveness of our approach, which outperforms several state-of-the-art CNN- and ViT-based detectors in generalization and domain adaptation for fetal brain and abdominal anatomical structure detection.