# Multi-Head Attention Pooling

Zero-shot anomaly detection has leaned on multi-scale sliding-window inference over CLIP, or on prompt learning that needs segmentation masks from auxiliary datasets. RADIOlogy instead exploits the patch–text alignment that multi-head attention pooling encoders preserve out of the box, scoring the average similarity gap between normal and abnormal prompts so that a single forward pass yields both an image-level anomaly score and a pixel-level anomaly map. Across seven industrial and seven medical benchmarks it reaches mean image AUROC of 92.6 industrial and 97.3 medical, with pixel AUROC of 95.2 and 87.3.

Zero-shot anomaly detection has leaned on multi-scale sliding-window inference over CLIP, or on prompt learning that needs segmentation masks from auxiliary datasets. RADIOlogy instead exploits the patch–text alignment that multi-head attention pooling encoders preserve out of the box, scoring the average similarity gap between normal and abnormal prompts so that a single forward pass yields both an image-level anomaly score and a pixel-level anomaly map. Across seven industrial and seven medical benchmarks it reaches mean image AUROC of 92.6 industrial and 97.3 medical, with pixel AUROC of 95.2 and 87.3.