SAM2-based Bayesian Prompt Adaptation for Cross-Modality Medical Segmentation

Abstract

Segmentation foundation models such as SAM2 generalize well in natural images, but adapting them to medical imaging remains challenging because of cross-modality gaps and limited annotations. We propose BayesPrompt, a SAM2-based Bayesian Prompt Adaptation framework for few-shot medical segmentation. BayesPrompt combines Bayesian meta-prior adaptation, which regularizes lightweight head updates through source-target posterior alignment, with a probabilistic prompt module that encodes support-set feature statistics and uncertainty into decoder prompts. Experiments on ultrasound and MRI show improvements in few-shot generalization, calibration, and efficiency over existing adaptation baselines.