Development of a deep learning system for comprehensive classification of common knee abnormalities from MRI: a large-scale, multi-center study

Motivation: Due to the complex anatomy of the knee, MRI interpretation is prone to oversight and misdiagnosis, demanding meticulous and time-intensive efforts. Goal(s): To develop a deep learning system (DLS) to effectively classify nine knee abnormalities and improve radiologists' performance and workload efficiency. Approach: Trained and validated on 14,847 patients, the DLS was assessed with radiologists across multiple diverse test sets. Results: The DLS effectively classified nine knee abnormalities and demonstrated strong generalization on external test sets, improving radiologists' accuracy by 1.0% to 7.2%. Impact: Our DLS markedly enhanced radiologists' diagnostic accuracy in knee MRI interpretation, streamlining workflows and reducing reliance on radiologist experience, thereby ensuring more consistent management and showcasing DLS's transformative potential in clinical radiology practice.