Beyond Static Anchors: Bounded Prototype Conditioning for Language-Free Medical Anomaly Detection

2026-08-01

AI Summary

This work addresses the limited generalization of existing medical anomaly detection methods that rely on static references—such as fixed textual prompts or visual tokens—across diverse domains. To overcome this, the authors propose ReCAP, a novel framework that discards static anchors and instead dynamically generates conditioned normal and abnormal visual prototypes from the input image via a bounded gated modulation mechanism. ReCAP further incorporates a non-parametric memory bank to store instance-level normal variations from the target domain, enabling language-free, query-adaptive anomaly scoring. Built upon the CLIP vision encoder, the method operates without textual prompts or test-time gradient updates. Evaluated on six medical benchmarks, ReCAP achieves state-of-the-art image-level AUROC in zero-shot and 23 out of 24 few-shot settings, leads in pixel-level AUROC across all three segmentation datasets, and reduces inference latency by over 70%.

Abstract

Medical anomaly detection identifies abnormal images and localizes lesions under scarce supervision while generalizing across organs and modalities. Existing CLIP-based methods reduce annotation requirements through vision--language alignment, but their normal and abnormal references, whether text prompts or learned visual tokens, remain fixed across test images. Such static references may not transfer reliably to unseen targets in a cross-domain medical imaging scenario. To address this, we propose ReCAP, a language-free framework that replaces static anchors with input-conditioned visual prototypes. ReCAP re-centers separated normal and abnormal prototypes for each image through a bounded gated modulation, enabling query-adaptive anomaly scoring while constraining context-induced prototype drift. For the few-shot setting, we introduce a non-parametric normal-reference memory to preserve instance-level target-domain variation and complement the conditional prototype branch. Across six medical benchmarks, ReCAP achieves the best image-level AUROC on all zero-shot and 23 of 24 few-shot settings, and the best zero-shot pixel-level AUROC on all three segmentation datasets. Particularly, it reduces inference latency by over 70% compared to the fastest baseline, without text prompts or test-time gradient updates.

Problem

• medical anomaly detection
• cross-domain generalization
• static reference
• vision-language alignment
• prototype drift

Innovation

• prototype conditioning
• language-free anomaly detection
• bounded gated modulation
• few-shot medical imaging
• zero-shot generalization