Deep learning-based automated knee joint segmentation integrated with wireless electromagnetic navigation for total knee arthroplasty: development and clinical validation

Total knee arthroplasty (TKA) requires accurate preoperative planning and intraoperative guidance to achieve optimal implant alignment and functional outcomes. Manual segmentation of computed tomography (CT) images is time-intensive and prone to interobserver variability, limiting the scalability of computer-assisted orthopedic surgery (CAOS). This study aimed to develop and clinically validate a deep learning-based automated CT segmentation system integrated with a wireless electromagnetic navigation platform. A hybrid deep learning model combining convolutional neural networks and transformer-based attention mechanisms was developed to automatically segment the femur, tibia, patella, and fibula from CT images. The dataset included 568 lower extremity CT scans (osteoarthritis: n = 248; normal: n = 98; periprosthetic fractures: n = 222). The segmentation system was integrated into a wireless electromagnetic navigation platform and prospectively evaluated in 42 patients undergoing TKA (navigation group: n = 22; conventional group: n = 20). Outcomes included segmentation accuracy (Dice similarity coefficient), processing time, bone resection accuracy, component sizing, and postoperative alignment. The model achieved mean Dice coefficients of 0.984 ± 0.008 (femur), 0.981 ± 0.011 (tibia), 0.972 ± 0.015 (patella), and 0.976 ± 0.012 (fibula). Mean processing time was 54.3 ± 6.2 s per case. Clinically, the navigation group demonstrated significantly improved bone resection accuracy (medial distal femur: 0.38 ± 0.21 mm vs. 0.82 ± 0.47 mm; p < 0.01) and postoperative alignment accuracy, assessed by mean absolute HKA deviation, was improved in the navigation group compared with the conventional group (0.64 ± 0.35° vs. 1.28 ± 0.54°, p < 0.01). Component sizing accuracy reached 100% for both femoral and tibial implants in the navigation group. The proposed workflow demonstrated high segmentation accuracy and was associated with improved bone-resection accuracy and postoperative alignment. These findings support the early clinical feasibility of integrating automated CT segmentation with wireless electromagnetic navigation for TKA and warrant further multicenter prospective validation.