Meniscal lesion detection and characterization in adult knee MRI: A deep learning model approach with external validation

PurposeEvaluation of a deep learning approach for the detection meniscal tears and their characterization (presence/absence migrated fragment).MethodsA large annotated adult knee MRI database was built combining medical expertise radiologists data scientists’ tools. Coronal sagittal proton density fat suppressed-weighted images 11,353 examinations (10,401 individual patients) paired with standardized structured reports were retrospectively collected. After curation, models trained validated on subset 8058 examinations. Algorithm performance evaluated test set 299 reviewed by 5 musculoskeletal specialists compared to general radiologists’ reports. External validation performed using publicly available MRNet database. Receiver Operating Characteristic (ROC) curves results Area Under Curve (AUC) values obtained internal external databases.ResultsA combined architecture localization lesion classification 3D convolutional neural networks reached AUC 0.93 (95% CI 0.82, 0.95) medial 0.84 0.78, 0.89) lateral tear detection, 0.91 0.87, 0.94) 0.95 0.92, 0.97) migration detection. resulted in an 0.83 0.75, 0.90) without further training 0.89 fine tuning.ConclusionOur algorithm demonstrated high menisci characterization,