Enhancing Medical Anomaly Detection via Text-Adapted Few-Shot Learning with Visual-Language Models

Medical image anomaly detection (AD) is crucial for early disease diagnosis, yet it faces challenges such as data heterogeneity and scarcity of annotated samples. This paper introduces a text-adapted few-shot training framework using CLIP, which extends the text encoder to incorporate fine-grained descriptions and introduces a text feature adapter for better alignment with image representations. A text-image feature alignment module and a contrastive learning mechanism are presented to enhance cross-modal integration and the distinction between normal and abnormal samples. Experimental evaluations on six medical imaging datasets demonstrate that our method significantly outperforms state-of-the-art techniques in both classification and segmentation tasks, achieving an average improvement of 1.13% in AUC. The implementation code is available at https://github.com/clownddd/TAFT.