Unsupervised Clustering of 3D Hand Scans Using K-Means and Gaussian Mixture Models (GMM)

The precise segmentation and classification of anatomical structures from three-dimensional (3D) scan data are critical for advancing personalized rehabilitation protocols, custom prosthetic fabrication, and the development of ergonomically optimized assistive technologies. Accurate identification of anatomical subregions facilitates improved device fit, enhances biomechanical performance, and supports patient-specific interventions. This study investigates the applicability and comparative performance of two unsupervised clustering algorithms, K-Means and Gaussian Mixture Models (GMM), for the segmentation of high-resolution 3D hand scan datasets. The hand models were acquired using a structured light scanning system, enabling the capture of complex surface geometries with sub-millimeter accuracy. Geometric descriptors were extracted from the reconstructed models and served as the input feature space for clustering, eliminating the need for pre-labeled training data. The performance of K-Means and GMM was evaluated based on segmentation accuracy, adaptability to morphological variations, and computational efficiency. Results analysis demonstrate that while K-Means offers reduced computational complexity and suitability for real-time processing, GMM provides superior adaptability and segmentation fidelity, attributed to its probabilistic representation of data distributions. These findings suggest that GMM is a more robust approach for capturing the inherent variability of anatomical structures, thereby holding greater potential for applications requiring high precision in patient-specific modeling and biomechanical analysis.