Few-Shot Anomaly Detection for Medical Ultrasound Images Using Metric Learning and Multimodal BiomedCLIP Embeddings

Medical ultrasound imaging is extensively utilized in clinical practice due to its advantages of safety, cost-effectiveness, and real-time imaging capability. Nevertheless, inherent issues such as low signal-to-noise ratios, operator dependency, and speckle noise introduce significant challenges in automated anomaly detection. To overcome these limitations, we propose a novel few-shot anomaly detection framework specifically designed for medical ultrasound imaging. Our method employs BiomedCLIP, a multimodal model tailored for biomedical applications, to jointly encode ultrasound images and clinically relevant textual descriptions into semantically rich embeddings. Subsequently, these embeddings are refined through a projection network to create compact, discriminative representations optimized for anomaly classification. A prototype-based metric learning approach further enhances the separability of these embeddings by explicitly clustering normal and abnormal cases. Extensive evaluations conducted on representative ultrasound datasets demonstrate that our proposed method achieves superior anomaly detection performance compared to existing contrastive and multimodal learning frameworks, particularly in severely limited data scenarios. Our findings underscore the efficacy and clinical potential of combining multimodal embeddings and metric learning for robust and interpretable anomaly detection in medical ultrasound images.