Automatic landmark detection and angle measurement in radiographs based on deep learning in application to coronal plane alignment of the knee classification

The Coronal Plane Alignment of the Knee (CPAK) based on radiographs is essential both for both preoperative planning and postoperative evaluation of knee arthroplasty. However, challenges in accurately detecting knee joint landmarks lead to imprecise knee-related angles measurement results, which in turn result in unreliable CPAK classification. In this paper, we propose an automated method for detecting landmarks, calculating knee-related angles, and performing CPAK classification using bilateral lower limb radiographs. Specifically, we construct a dataset and retrain a YOLO-based network to identify candidate regions for landmark detection. Landmark detection in radiographs is challenging due to the complexity of image details. To enhance the network’s ability to leverage important features, we introduce a Dual-path Fusion Attention Module, which uses Spatial Transformer Networks to focus on the skeletal region, and employs an Efficient Channel Attention Module to enhance edge features. A Coordinate Correction Module is proposed to facilitate multi-scale feature interaction, enabling accurate landmark localization. With precise landmark detection, our model achieves reliable angle measurement and CPAK classification. Extensive experiments demonstrate the superior performance of our network. The mean absolute errors for hip-knee-ankle angle, mechanical lateral distal femoral angle, mechanical medial proximal tibia angle and joint line convergence angle were 0.18°, 0.33°, 0.75° and 0.80° respectively. The intraclass correlation coefficients for all four angles were above 0.9.