Mesh deep support vector data description: Unsupervised anomaly detection in 3D human leg morphologies for adaptive compression stocking design

Ready-to-wear compression stockings are not adapted to the morphological changes of the human legs. At the same time, they have to provide a predicted effect, which is only possible if the fit of these products adapts to the underlying morphology of the wearer. The reevaluation of the sizing system used in the design of such products has been evoked several times in the state-of-the-art; yet, no real shape-related morphological classes have been defined. Moreover, the 3D body scans that are used in geometric learning anthropometric studies often contain outliers and surface anomalies that can hinder the outcome of data partitioning. In this paper, we propose Mesh Deep Support Vector Data Description (MeshDeepSVDD), a novel unsupervised anomaly detection framework that integrates the Fully Convolutional Mesh Autoencoder (FCMA) with Deep SVDD for 3D mesh processing. Unlike existing methods that rely on pre-trained feature extractors (e.g., PointMAE or PointBERT), our model is pre-training-free and trained from scratch. It is also dataset-agnostic and generalizable across various 3D anomaly detection scenarios without task-specific adaptation. The method outperformed current state-of-the-art 3D anomaly detection models on the MulSen-AD dataset.When applied to our dataset of 131 human legs, it ranked morphologically distinctive subjects as top anomalies, providing qualitative and limited quantitative evidence of its usefulness for exploring leg-shape variability. This framework can support the identification of morphologically distinctive legs and geometric patterns that may be relevant for compression stocking design. However, because the human leg dataset contains only limited clinical labels, the detected outliers should be interpreted as morphology-based candidates rather than clinically validated pathological cases. The results therefore represent an initial step toward defining leg morphotypes for adaptive compression stocking design.