Exploring anatomical similarity in zero-shot learning for bone abnormality detection.

Deep learning models for musculoskeletal imaging often require large, annotated datasets per anatomical region. In this study we explicitly treat abnormality detection (normal vs. abnormal) on MURA (which does not provide fracture-only labels) and ask whether anatomical similarity enables zero-shot (out-of-domain) transfer-training on one body part and evaluating on unseen parts without target data access. Using study-level aggregation and Wilson 95% confidence intervals, we quantify cross-part transfer as a function of anatomical proximity. We also include a small, labeled replication subset with a second backbone to verify that observed trends are not architecture-specific. Results show higher accuracy when source and target are anatomically similar. These findings bound what can be achieved without semantic side information or target adaptation and motivate anatomy-aware data selection for scalable deployment in data-limited settings.