Anatomy-Guided Foundation Model Adaptation with Within-Case Prototype Supervision for Standard Plane Detection in Fetal Ultrasound Blind Sweeps

TL;DR - AnatoProto adapts a frozen BiomedCLIP encoder to detect rare fetal abdominal standard-plane frames in ultrasound blind sweeps. It achieves 67.72 F1 on ACOUSLIC-AI, outperforming the strongest reported foundation-model baseline by 13.20 points.

• Anatomy-weighted pooling uses nnU-Net abdominal-region probabilities to focus BiomedCLIP patch features on clinically relevant anatomy.
• Within-case prototype supervision pulls frame embeddings toward the mean positive-frame embedding from the same sweep.
• A frame-to-segment-to-case cascade and hybrid stability/boundary head reduce structurally implausible and boundary false positives.
• Prototype supervision improves recall only when paired with anatomy-guided pooling, suggesting that accurate within-case prototypes are critical.