Lower Extremity CT Segmentation for Patients with Knee Osteoarthritis Based on 3D U-Net

Total Knee Arthroplasty (TKA) is the most effective treatment for end-stage Knee Osteoarthritis (KOA). Preoperative planning using the patient's CT images can significantly improve the performance of the surgery. During the preoperative planning process, segmenting the femur and tibia, especially knee joint regions, and extracting 3D models are crucial steps. Deep learning methods such as U-Net provide effective tools for image segmentation, but the training requires high-quality image datasets which takes experienced operators a lot of time to label CT images. It is crucial to reduce the cost of manual labeling when establishing a 3D CT dataset and improve robustness against knees with severe KOA and metal artifacts. Therefore, we propose a two-stage labeling and modeling method involving low-quality and high-quality stages. Initially, a small amount of CT data was labeled to train a low-quality segmentation U-Net model. The operators were then only required to correct the segmentation of the CT data based on the results predicted by the low-quality model, which saved them a lot of workload for labeling. The results indicate that the model has a high-precision segmentation performance of lower limb CT, demonstrating robustness against knees with severe KOA and metal artifacts.