Reliability and Dual Validation of a Deep Learning Model for Anterior Cruciate Ligament Tear Detection Using Magnetic Resonance Imaging

The anterior cruciate ligament (ACL) is essential for knee stability, and its tears are common, especially in athletes. Magnetic resonance imaging (MRI) is the gold standard for diagnosis. Recent advances in deep learning (DL) offer promising opportunities for automating ACL tear detection, potentially reducing workload and enhancing diagnostic accuracy. This study evaluates DL models for ACL tear detection using MRI. A model was trained on the MRNet dataset, which includes 1,370 knee MRIs from Stanford University Medical Center. The data were split into training (60%), tuning (20%), and validation (20%) sets using stratified random sampling. Convolutional neural networks were applied to axial, sagittal, and coronal planes, followed by a stacking ensemble method. External validation was performed using the KneeMRI dataset (917 scans) from Clinical Hospital Centre Rijeka, Croatia. Model performance was also compared with two experienced radiologists on a subset of 120 MRIs. Sensitivity, specificity, and area under the curve (AUC) were used for evaluation. The model achieved AUCs of 0.885 (95% CI 0.865–0.902) on training, 0.878 (95% CI 0.807–0.925) on validation, and 0.912 (95% CI 0.879–0.937) on external validation. It demonstrated 84% accuracy compared to radiologists, though McNemar’s test (p = 0.001) showed a statistically significant difference in classification. These findings confirm the model’s robustness across datasets and its potential as a diagnostic support tool. DL models demonstrate high reliability in identifying ACL tears on MRI and may assist radiologists, particularly in resource-constrained environments. Further integration of arthroscopic data and larger datasets may improve clinical applicability.