Deep learning in detecting bucket-handle meniscal tears on knee radiographs: Comparison with surgeon interpretations.

backgroundDetecting bucket-handle meniscal tears (BHMTs) on knee radiographs remains challenging. Advances in deep learning, particularly convolutional neural networks, have shown strong potential in medical image analysis. This study evaluated the feasibility and diagnostic accuracy of a deep learning model for detecting BHMTs on knee radiographs and compared its performance to that of orthopedic surgeons.

methodsKnee radiographs, including anteroposterior (AP) and lateral (Lat) views, were collected from our institution and external hospitals. Radiographs were screened and labeled based on arthroscopic confirmation of the presence or absence of BHMTs. All images were cropped and standardized before analysis. In addition to AP and Lat inputs, composite images combining both views were generated. Separate models were trained for each group, evaluated on an independent test set, and the best-performing model was compared with interpretations by orthopedic surgeons.

resultsIn total, radiographs from 406 patients at our institution and 90 patients from external hospitals were included. The composite radiographs input achieved the highest performance in distinguishing BHMTs from non-BHMTs, with an area under the receiver operating characteristic curve (AUC) of 0.844, a sensitivity of 74.4%, a specificity of 85.0%, a positive predictive value (PPV) of 82.9%, a negative predictive value (NPV) of 77.3%, a precision of 82.9%, and an F1 score of 78.4%. Overall, the model demonstrated higher diagnostic performance for BHMT detection compared with orthopedic surgeons.

conclusionThis study demonstrates the potential utility of deep learning for detecting bucket-handle meniscal tears on knee radiographs.