Automatic identification of shoulder landmarks with a single camera in healthy participants

Abstract Objective . To develop and validate a markerless algorithm for automatic detection of scapular and thoracic anatomical landmarks in healthy individuals using a single camera. Methods . Ninety healthy adults (55 males and 35 females) held twelve static upper-limb postures. For each posture, two images were acquired: one markerless and one with anatomical landmarks indicated by a mark placed by palpation, serving as reference. A ResNet-50–based convolutional neural network was fine-tuned to predict 2D landmark coordinates from the markerless images. Accuracy was quantified using the Euclidean distance to palpation-based reference points, as well as differences in humerothoracic, scapulothoracic, and scapulohumeral joint angles between methods. Results . Median errors across all landmarks ranged from 5 mm to 14 mm, with 90th percentiles below 38 mm. Joint angle estimation yielded mean absolute errors of 2.0° to 6.6°. Performance was consistent across elevation planes, arm elevation percentages, and sexes. However, a small but statistically significant reduction in accuracy was observed in the sagittal plane at the highest elevation level. Scapulothoracic angle estimates remained stable, while scapulohumeral and humerothoracic errors increased moderately (≈+6–7°) under this condition. Conclusion . This marker-free, single-camera workflow offers a low-cost, rapid-setup alternative to conventional motion capture systems, enabling objective assessment of shoulder kinematics in routine clinical and field applications with minimal equipment and preparation time.