Deep learning-enabled segmentation of knee cartilage in conventional magnetic resonance images: Internal and external validation of different models.

Abstract

Accurate evaluation of the cartilage anatomy of the knee is helpful for clinical evaluation of the source of knee pain and the classification and treatment of knee osteoarthritis (OA). This study proposes a deep learning model for segmentation of knee articular cartilage in conventional proton density fat-saturated MRI sequences to assess cartilage morphology for subsequent injury grading. This retrospective study was conducted at two radiology centers, involving 254 knees from 254 patients who had previously undergone MRI scans. The training-internal validation cohort included 219 knees from Center 1. The external validation cohort comprised 35 knees from Center 2. Two musculoskeletal radiology experts manually annotated the cartilage regions. A 3D Res U-net model was employed for segmentation, and its performance was compared with 3D U-net and 3D V-net models. Segmentation results were evaluated using the Dice coefficient and Jaccard index. The 3D Res U-net model demonstrated superior segmentation performance compared to the other deep learning methods. For cartilage in the lateral femorotibial joint, medial femorotibial joint, and patellofemoral joint, the average Dice coefficients with 3D Res U-net were 0.871, 0.860, and 0.858 in internal validation and 0.846, 0.837, and 0.819 in external validation, respectively. The Jaccard index followed a similar trend. The 3D Res U-net model improves knee cartilage segmentation in conventional MR imaging, contributing to the understanding of cartilage morphology and the improvement of clinically relevant decisions.