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

Area of Science:

• Orthopedics
• Radiology
• Sports Medicine

Background:

• Traumatic anterior glenohumeral dislocation often involves bony Bankart lesions.
• Plain radiography is the initial imaging modality, but its diagnostic accuracy for these lesions is debated.
• Computed tomography (CT) is considered a more sensitive tool for bony Bankart lesion detection.

Purpose of the Study:

• To compare the diagnostic accuracy of plain radiography versus CT for bony Bankart lesions.
• To identify clinical factors predicting the need for CT using machine learning.

Main Methods:

• Retrospective analysis of 69 traumatic anterior glenohumeral dislocations.
• Blinded radiographic evaluation by two observers, with inter-observer reliability assessed (Cohen's kappa).
• Diagnostic metrics calculated using CT as the gold standard; machine learning models (decision tree) used for predictor identification.

Main Results:

• Plain radiography showed fair inter-observer agreement (κ = 0.272) and low sensitivity (27.27%) but high specificity (100%).
• Time from first dislocation was a significant predictor for CT-detected lesions (p = 0.019).
• Machine learning identified a critical window of 36-42 weeks post-dislocation with a 55.6-70% probability of 'hidden' bony Bankart lesions.

Conclusions:

• Conventional radiographs have limited sensitivity and fair reliability for bony Bankart lesions compared to CT.
• Negative radiographs do not exclude bony Bankart lesions in traumatic anterior shoulder instability.
• CT is recommended within 42 weeks of initial dislocation and for recurrent instability (≥3 episodes) to identify occult osseous pathology and guide surgical planning.