Three-dimensional markerless pose estimation for anatomical landmarks of the shoulder and upper limb.

This study aimed to develop and validate a 3D markerless pose estimation algorithm for anatomical landmarks of the shoulder and upper limb. Twenty-six healthy participants were asked to hold 21 static positions, while both markerless and marker-based (applied after palpation) images were recorded using eight video cameras. A pre-trained convolutional neural network based on ResNet-50 was fine-tuned on 2612 markerless images to estimate the poses of 20anatomical landmarks. The model was tested on 1680 images by calculating the 3D Euclidean distances between predicted coordinates and those labeled from marker-based images. Across all positions, median and 90th percentile Euclidean distances were below 15 mm and 30 mm, respectively for all anatomical landmarks, except for the 8th thoracic vertebra, inferior angle of the scapula and medial epicondyle, which presented the highest Euclidean distances. For most of the anatomical landmarks, loss rates inferior to 6 % were observed for predicted coordinates. The neural network accuracy was similar between movements tested and not influenced by the degree of arm elevation. To conclude, a neural network was developed and validated for estimating shoulder and upper-limb anatomical landmarks poses, demonstrating promising accuracy for future clinical applications.