Deep learning-based analysis on X-ray images of full lower limbs

This study proposes an automated deep learning-based method for segmenting full lower limb X-ray images and calculating critical angles required in preoperative planning for High Tibial Osteotomy (HTO). The method utilizes an improved U-Net architecture, enabling accurate segmentation of the femur, tibia, fibula, and joint space. After segmentation, affine registration is applied to the resulting point cloud data, and key reference points are manually integrated to calculate the mechanical angles needed for surgery. Experimental results indicate that the proposed model achieves an average Intersection Over Union (mIOU) of 89.23% and pixel accuracy (mPA) of 94.63%, outperforming widely used models such as the standard U-Net (mIOU: 86.36%) and DeepLabv3-Net (mIOU: 81.21%) in segmentation accuracy. Additionally, the proposed method reaches a precision of 93.13%, significantly surpassing other baseline models. Compared to conventional manual methods, this automated preoperative planning process reduces the average processing time per image to 1 second, resulting in an approximate 98.33% improvement in efficiency. This approach demonstrates high accuracy, speed, and reliability in High Tibial Osteotomy preoperative planning. Future work will focus on enhancing the model’s robustness and adaptability to accommodate a wider range of imaging types and clinical scenarios.