Graph-Convolutional-Beta-VAE for synthetic abdominal aortic aneurysm generation

Limited clinical datasets and privacy constraints make it difficult to assemble large collections of abdominal aortic aneurysm geometries. This work combines graph convolution with a β-variational autoencoder to learn nonlinear anatomical variation in a compact, disentangled latent space.

A low-impact augmentation procedure based on Procrustes analysis expands the training set while preserving anatomical structure. The learned representation supports deterministic and stochastic generation strategies designed to increase geometric diversity without sacrificing realism.

Evaluation on unseen data shows greater robustness than the compared principal-component-based methods. The resulting synthetic geometries can support statistical analysis, computational modelling, and device testing without directly exposing patient records.

Paper: https://doi.org/10.1007/s11517-025-03491-y