YOLO11m-Based Artificial Intelligence Versus Orthopaedic Physician Panels for Arthroscopy-Referenced Three-Category MRI Classification of the Posterosuperior Rotator Cuff: A Multi-Reader Diagnostic Accuracy Study.
Authors
Affiliations (3)
Affiliations (3)
- Department of Orthopaedics and Traumatology, Gaziosmanpaşa Hospital, Faculty of Medicine, Istinye University, Istanbul 34245, Türkiye.
- Department of Orthopaedics and Traumatology, Kanuni Sultan Süleyman Training and Research Hospital, University of Health Sciences, Istanbul 34303, Türkiye.
- Department of Orthopaedics and Traumatology, Faculty of Medicine, Istanbul Yeni Yüzyıl University, Istanbul 34010, Türkiye.
Abstract
<b>Background/Objectives:</b> Artificial intelligence (AI) may support rotator cuff MRI interpretation; however, its performance relative to clinicians at different experience levels remains uncertain. We compared a locked YOLO11m-based model with prespecified orthopaedic physician panels using arthroscopy as the reference standard. <b>Methods:</b> This retrospective single-centre study included a separate 300-patient model-development dataset and an arthroscopy-referenced reader cohort of 258 patients (60 intact cuffs, 86 partial-thickness tears, and 112 full-thickness tears). Twelve blinded physicians were assigned to four three-reader experience panels, and panel diagnoses were defined by a two-of-three majority. Paired accuracy comparisons used exact McNemar tests with Holm adjustment. <b>Results:</b> AI achieved 0.841 overall accuracy (95% CI, 0.792-0.881). The senior-specialist panel had numerically higher accuracy (0.911), but the difference from AI was not statistically significant after adjustment (Holm <i>p</i> = 0.060); the junior-specialist panel also did not differ from AI (0.822; Holm <i>p</i> = 0.649). AI showed higher paired accuracy than the senior-resident (0.729; Holm <i>p</i> = 0.005) and junior-resident panels (0.671; Holm <i>p</i> < 0.001). Partial-thickness tears had the lowest AI sensitivity (0.709). <b>Conclusions:</b> In this restricted-cine, three-category benchmark, AI outperformed both resident panels, whereas neither specialist panel comparison reached the adjusted significance threshold. These findings do not establish equivalence to specialists or benefit from AI-assisted reading.