BiomedAP: A Vision-Informed Dual-Anchor Framework with Gated Cross-Modal Fusion for Robust Medical Vision-Language Adaptation

BiomedAP is a new adaptation framework for medical vision-language models that boosts robustness to imperfect prompts.

It uses gated interaction between image and text prompts, plus dual-anchor regularization from expert templates and visual prototypes.

The method improves cross-modal alignment and few-shot diagnosis performance under prompt noise.

Researchers report stronger results and better robustness across 11 benchmarks, suggesting more reliable clinical use.