COMPARATIVE EVALUATION OF ARTIFICIAL INTELLIGENCE AND RADIOLOGIST INTERPRETATION OF SHOULDER MRI

Abstract

Background: Shoulder MRI provides detailed assessment of the rotator cuff, labrum, and other periarticular structures, but interpretation can be challenging in examinations with multiple or subtle abnormalities. Artificial intelligence (AI) has emerged as a potential adjunct to conventional radiological interpretation. This study evaluated the diagnostic performance of AI in shoulder MRI using the recorded final diagnosis as the reference standard. Methods: A comparative observational study was conducted on 60 patients who underwent shoulder MRI at SGT Hospital, Gurugram. We recorded MRI acquisition parameters, including repetition time (TR), echo time (TE), flip angle, and sequence information. The same MRI examinations were analyzed using a designated AI system, and AI findings were compared with the documented final/reference diagnosis. Diagnostic performance was assessed using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and Cohen's kappa. Results: The study included 38 (63.3%) men and 22 (36.7%) women, with a mean age of 44.18 ± 15.15 years. 54 examinations (90.0%) were abnormal and six (10.0%) were normal. AI correctly identified the main diagnostic category in 53/60 examinations (88.3%). For binary classification of normal versus abnormal examinations, AI demonstrated 100.0% sensitivity, 83.3% specificity, 98.2% PPV, 100.0% NPV, and 98.3% accuracy, with substantial agreement (κ ≈ 0.90). Sensitivity was less for full-thickness supraspinatus tears (66.7%). Conclusion: AI demonstrated high sensitivity and substantial agreement with the reference diagnosis for detecting abnormalities on shoulder MRI. However, limitations in lesion-specific characterization, particularly full-thickness supraspinatus tears, indicate that AI should currently be considered an adjunct to radiologist interpretation rather than a replacement.