Artificial Intelligence Segmentation In Musculoskeletal Ultrasound: A Novel Study Across Five Joints Using Expert Ground Truth
In brief
AI segmentation matched experts with over 90% agreement across five joints
In 79 adults, AI software segmented most of 120 anatomical structures in ultrasound clips with positive and negative agreement above 90% compared with expert annotations; agreement reached at least 97% for tendons, muscles, and nerves. Results were similar in people with and without sonographic disorders, but whether the tool improves real-time scanning or clinical decisions remains untested.
- Journal
- American journal of physical medicine & rehabilitation (Q1)
- Published
- 14 October 2026
- Study design
- Cross-sectional study
- Evidence level
- Level 3, Low (CEBM 3b)
- Authors
- Fatih Tok, Michael Dakkak, Gary S Schwartz, Mark Kronenfeld, Levent Özçakar
- PMID
- 42870329
- DOI
- 10.1097/PHM.0000000000003181
Why clinicians should know about it
- Picked for Anatomy (paper of the day, 18 October 2026): AI segmentation, musculoskeletal US anatomy
Abstract
OBJECTIVE: To evaluate the performance of an artificial intelligence (AI) technology in recognizing and segmenting anatomical structures across five joints (shoulder, elbow, wrist/hand, knee, and ankle/foot) in musculoskeletal ultrasound (US). DESIGN: This cross-sectional study included 79 adults. Software segmentations were compared to ground truth annotations provided by four experts. Performance was assessed using positive predictive agreement (PPA) and negative predictive agreement (NPA), with a true/false threshold defined as a dice similarity coefficient (DSC) ≥0.50. Subjects with and without sonographic disorders were evaluated to assess robustness. RESULTS: 3017 cine clips (15600 frames) covering 120 anatomical structures were analyzed. PPA and NPA exceeded 0.90 for most structures; tendons, muscles, and nerves showed near-perfect agreement (≥0.97). Similar performance was observed in healthy subjects (PPA: 0.93-0.96; NPA: 0.94-0.99) and subjects with sonographic disorders (PPA: 0.91-0.96; NPA: 0.91-0.99) (P<0.001). CONCLUSION: This first large-scale study demonstrates the high accuracy and reliability of AI-based segmentation. By improving anatomical identification consistency during image acquisition, the software may help reduce technical operator dependency. The results further indicate strong potential for real-time assistance, supporting clinicians and trainees, and underscore the broader applicability of AI-based software in musculoskeletal US.
Abstract as published, via PubMed.
For healthcare professionals. The summary is generated by AI from the published abstract, and the evidence level is assigned automatically from the study design on the Oxford CEBM hierarchy. Neither is medical advice. Read the full paper before changing practice.