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

Area of Science:

• Medical imaging analysis
• Machine learning in radiology
• Musculoskeletal imaging

Background:

• Deep learning models for musculoskeletal imaging demand extensive annotated datasets for each anatomical region.
• The MURA dataset lacks fracture-specific labels, necessitating alternative approaches for abnormality detection.

Purpose of the Study:

• To investigate the potential of zero-shot transfer learning for abnormality detection in musculoskeletal imaging.
• To determine if anatomical similarity between body parts facilitates effective cross-part transfer training without target data access.
• To quantify the impact of anatomical proximity on transfer learning performance.

Main Methods:

• Utilized deep learning models for abnormality detection (normal vs. abnormal) on the MURA dataset.
• Employed study-level aggregation and Wilson 95% confidence intervals to assess cross-part transfer accuracy.
• Evaluated transfer learning performance as a function of anatomical similarity between source and target body parts.
• Included a replication subset with a second backbone to ensure architecture-independent findings.

Main Results:

• Higher accuracy was observed when the source and target anatomical regions were more similar.
• Anatomical proximity significantly influences the effectiveness of zero-shot transfer learning in musculoskeletal imaging.
• The study demonstrates the feasibility of transferring knowledge between anatomically related body parts.

Conclusions:

• Anatomical similarity is a key factor enabling effective zero-shot transfer learning in musculoskeletal imaging.
• Findings provide bounds for performance without semantic information or target adaptation.
• Motivates anatomy-aware data selection strategies for scalable deployment in data-scarce environments.