Machine Learning - Powered Assessment of Plain Radiographs for Detecting Bony Bankart Lesions and Guiding CT Decision-Making.

Abstract

Background. To evaluate the diagnostic accuracy of plain radiography in detecting bony Bankart lesions compared to computed tomography (CT) and to identify clinical factors that delineate the necessity for CT investigation using machine learning-powered analysis. Material and methods. A retrospective analysis of 69 cases of traumatic anterior glenohumeral dislocation was performed. Two blinded observers evaluated plain radiographs for bony Bankart lesions. Inter-observer reliability was assessed using Cohen’s kappa (κ). Diagnostic metrics (sensitivity, specificity, PPV, NPV) were calculated using CT as the gold standard. Multivariate logistic regression and machine learning (decision tree) models were utilized to identify predictors for ‘hidden’ lesions (X-ray negative, CT positive). Results. Inter-observer agreement for X-ray was fair (κ = 0.272). Observer 1 demonstrated 100% specificity but low sensitivity (27.27%). Multivariate analysis identified ‘Time from first dislocation’ as a significant predictor for CT-detected lesions (p = 0.019). ROC and machine learning models identified a critical cut-off of 36–42 weeks; patients presenting within this window had a 55.6-70% probability of a hidden lesion, compared to 16.7% for chronic cases. Conclusions. 1. Conventional radiographs demonstrated high specificity but limited sensitivity and only fair in­terobserver reliability for detecting bony Bankart lesions compared with CT. 2. A negative radiographic result does not reliably exclude a bony Bankart lesion in patients with traumatic anterior shoulder instability. 3. CT examination should be performed in patients presenting within 42 weeks after their first anterior shoulder dislocation, particularly when radiographs are negative but clinical suspicion persists. 4. CT should also be strongly considered in patients with recurrent instability, especially those with ≥3 dislocation episodes, to facilitate early identification of occult osseous pathology and guide appropriate surgical planning.