Consistency Margin Adapter for Few-Shot Medical Anomaly Detection

Vision-language models such as CLIP have become a versatile foundation for zero and few-shot medical anomaly detection, yet most adapter designs still rely on shal-low projections between visual tokens and handcrafted textual prompts. We introduce a Consistency Margin Adapter that aligns multi-layer visual representations with normal prompts via learnable directional shifts, semantic margin objectives, layer-wise consistency, and entropy regularization. The adapter produces reliable heatmaps from only k-shot support samples and integrates seamlessly with existing CLIP+CoOp pipelines. Extensive experiments on BUSI, BrainMRI, and CheXpert demonstrate superior image and pixel-level AUROC over prior adapters, particularly in the low-shot regime.