Development and validation of a deep learning markerless system for lower-limb kinematics in hip and knee osteoarthritis population.

With the advancement of deep learning technology, markerless systems have emerged as a cost-effective and user-friendly alternative to marker-based systems. However, most existing markerless systems are developed using datasets from healthy individuals, which limits their generalizability to patient populations. Therefore, this study developed a four-camera markerless system using a dataset of patients with osteoarthritis and validated its measurement accuracy in lower-limb kinematics. A total of 150 patients with hip or knee osteoarthritis were allocated to a training set (n = 120) and a test set (n = 30). Kinematic data during gait were simultaneously collected using both markerless and marker-based systems. We developed a four-camera markerless system on the training set. In the test set, the kinematic differences between the markerless and marker-based systems over the gait cycle were assessed using root mean square error (RMSE) and intraclass correlation coefficient (ICC). The grand mean position difference and ICC for the keypoints predicted by the markerless system were 13.4 mm and 0.93, respectively. Additionally, the mean RMSE for all joint angles was 4.1°. The ICC for the joint angle waveforms between the markerless and marker-based systems in the sagittal, frontal, and transverse planes were 0.93, 0.50, and 0.34, respectively. Our four-camera markerless system, developed using data from patient populations, shows high accuracy in keypoints and sagittal plane joint angles prediction. This indicates that our markerless system is suitable for osteoarthritis populations and offers a cost-effective and convenient tool for disease-related biomechanical research.