Ensemble deep learning model for predicting anterior cruciate ligament tear from lateral knee radiograph

Objective: To develop an ensemble deep learning model (DLM) predicting anterior cruciate ligament (ACL) tears from lateral knee radiographs and to evaluate its diagnostic performance. Materials and methods: In this study, 1433 lateral knee radiographs (661 with ACL tear confirmed on MRI, 772 normal) from two medical centers were split into training (n = 1146) and test sets (n = 287). Three single DLMs respectively classifying radiographs with ACL tears, abnormal lateral femoral notches, and joint effusion were developed. An ensemble DLM predicting ACL tears was developed by combining the three DLMs via stacking method. The sensitivities, specificities, and area under the receiver operating characteristic curves (AUCs) of the DLMs and three radiologists were compared using McNemar test and Delong test. Subgroup analysis was performed to identify the radiologic features associated with the sensitivity. Results: The sensitivity, specificity, and AUC of the ensemble DLM were 86.8% (95% confidence interval [CI], 79.9–92.0%), 89.4% (95% CI, 83.4–93.8%), and 0.927 (95% CI, 0.891–0.954), achieving diagnostic performance comparable with that of a musculoskeletal radiologist (P = 0.193, McNemar test; P = 0.131, Delong test). The AUC of the ensemble DLM was significantly higher than those of non-musculoskeletal radiologists (P = 0.043, P < 0.001). The sensitivity of the DLM was higher than that of the radiologists in the absence of an abnormal lateral femoral notch or joint effusion. Conclusion: The diagnostic performance of the ensemble DLM in predicting lateral knee radiographs with ACL tears was comparable to that of a musculoskeletal radiologist.