Unsupervised adaptive sampling graph autoencoder for 3D surface encoding and mesh representation transfer.

Standardizing anatomical surface representations across heterogeneous datasets is critical for population-level modeling and downstream medical image analysis tasks. Although existing mesh autoencoders can achieve standardized latent surface representations, most produce uninterpretable encodings and rely on spatial correspondences, pose assumptions, pre-processing or simple dimensionality reduction operations such as pooling that compromise accuracy, reproducibility and generalizability. We present the Adaptive Sampling Graph Autoencoder (ASGAE), an unsupervised geometric learning framework that decouples surface geometry from mesh topology to enable standardized surface representations. ASGAE integrates a novel Fuzzy Segmentation module that learns probabilistic mappings between input surfaces with variable pose, resolution, point distribution and connectivity, and a geometrically consistent latent subspace with uniform dimensionality identified in a fully unsupervised manner from heterogeneous training data, enabling interpretable and pose-standardized encodings. A novel Adaptive Sampling Decoder reconstructs surfaces from their latent encodings using either their original mesh topology or the topology from any other reference mesh, thus enabling mesh representation transfer. ASGAE achieved state-of-the-art encoding and reconstruction performance in both a public synthetic dataset and a 3D photogrammetry dataset of children with craniofacial pathology, providing point-to-surface coordinate and texture reconstruction errors of 0.43 ± 0.07 mm and 4.61 ± 0.87%, respectively. It also enabled mesh representation transfer between heterogeneous surfaces with a point-to-surface error of 0.44 ± 0.08 mm (p = 0.001). Finally, we demonstrated the use of ASGAE's capabilities in two common downstream analysis tasks: landmark identification and pathology classification. Unlike existing methods, ASGAE produces interpretable surface encodings and enables mesh standardization, which is key in many medical image analysis tasks.