ReCAP: Input-Conditioned Prototypes for Language-Free Medical Anomaly Detection
A new framework called ReCAP (Bounded Prototype Conditioning) is introduced for medical anomaly detection, addressing limitations in existing CLIP-based methods. The approach replaces static text prompts or learned visual tokens with input-conditioned visual prototypes, enabling query-adaptive anomaly scoring. ReCAP uses a bounded gated modulation to re-center normal and abnormal prototypes for each image, constraining context-induced drift. For few-shot settings, a non-parametric normal-reference memory preserves instance-level target-domain variation. The method is language-free, eliminating the need for text annotations. The paper is available on arXiv (2608.00442) and is categorized as a cross-domain medical imaging study. This development could improve generalization across organs and modalities in medical imaging, reducing annotation requirements.
Key facts
- ReCAP is a language-free framework for medical anomaly detection.
- It replaces static anchors with input-conditioned visual prototypes.
- Uses bounded gated modulation to re-center prototypes per image.
- Introduces a non-parametric normal-reference memory for few-shot learning.
- Aims to improve cross-domain generalization across organs and modalities.
- Paper available on arXiv with ID 2608.00442.
- Addresses limitations of CLIP-based methods with static references.
- Method is query-adaptive and constrains prototype drift.
Entities
Institutions
- arXiv