Validity of Markerless Three-dimensional Wrist and Finger Motion Analysis Using Deep Learning: A Comparison with Goniometric Measurement

BackgroundMarkerless motion analysis using deep learning is attracting attention in the field of rehabilitation; however, the three-dimensional measurement accuracy in finger joints, which are prone to self-occlusion, has not been sufficiently validated. This study aimed to validate the accuracy of finger joint angle measurements obtained using a marker-less system based on DeepLabCut (DLC) and Anipose by comparing it with the clinical standard of goniometric measurements. MethodsForty-one healthy adults were recruited. Videos from ten participants were used for DLC training, whereas the remaining 31 served as the analysis subjects. Eight flexion movements (wrist, thumb, index, and middle finger) were recorded using five synchronized cameras. The DLC tracked 2D keypoints, and Anipose performed 3D reconstruction to calculate the angles. Goniometric measurements were performed simultaneously for comparison. The agreement between methods was evaluated using Bland-Altman analysis to identify fixed and proportional biases. FindingsHigh agreement was observed between the angles estimated by marker-less analysis and goniometer measurements, and most data points were within the 95% limits of agreement. However, a significant proportional bias, where the error increased with an increase in flexion angle, was observed in distal joints, such as the thumb interphalangeal joint and index/middle proximal interphalangeal joints. InterpretationThis system demonstrated clinically acceptable validity for measuring finger range of motion. However, underestimation is likely to occur in the distal finger joints and at maximal flexion owing to the influence of occlusion, necessitating consideration of proportional bias. This method represents a noninvasive, low-cost tool for assessing hand function.