Abstract
BACKGROUND: Accurate segmentation of cartilage from magnetic resonance imaging (MRI) is crucial for the diagnosis and surgical planning of knee osteoarthritis. However, manual segmentation is time-consuming, and conventional computed tomography-based surgical systems are limited by their inability to visualize cartilage. OBJECTIVE: This study aimed to develop a clinically targeted deep learning framework, the Swin-UNet conditional generative adversarial network (cGAN), for the automatic segmentation of femoral and tibial cartilage in MRI. We then evaluated its performance against conventional UNet, UNet cGAN, and Swin-UNet baseline models. METHODS: Our dataset comprised 232 knee MRI scans. We conducted quantitative experiments on the proposed Swin-UNet cGAN model and compared the results with those of widely used UNet, UNet cGAN, and Swin-UNet models for femoral and tibial cartilage segmentation, using the Dice similarity coefficient, mean intersection over union, 95th percentile Hausdorff distance, and average symmetric surface distance. All performance metrics were statistically analyzed. In addition, the performance of the Swin-UNet cGAN model was evaluated on an external validation dataset. RESULTS: The proposed Swin-UNet cGAN achieved the highest mean Dice similarity coefficient and intersection over union scores for both femoral and tibial cartilage segmentation, demonstrating performance statistically comparable to the best-performing baseline (UNet) in the tibia. Regarding distance metrics (average symmetric surface distance and 95th percentile Hausdorff distance), the proposed model significantly outperformed all baselines in the tibia while achieving results comparable to the UNet cGAN in the femur. It also maintained consistently high segmentation performance on both the internal test set and an external validation dataset. CONCLUSIONS: These findings indicate that the proposed Swin-UNet cGAN achieves more accurate knee cartilage segmentation than UNet, UNet cGAN, and Swin-UNet, particularly in terms of boundary accuracy, while maintaining promising generalizability performance across both internal testing and external validation cohorts. This MRI-based deep learning approach addresses critical limitations of computed tomography-based patient-specific instrumentation systems by providing cartilage visualization, potentially improving surgical precision and outcomes in total knee arthroplasty.