The latent shape space of intracranial saccular aneurysms.

P. Eulzer, Henrik Voigt, Monique Meuschke, Kai Lawonn

TL;DR: A unified framework that learns a compact but expressive latent representation of aneurysm morphology for generative modeling and rupture-label classification is developed, providing a scalable and interpretable basis for quantitative aneurysm morphometry.

BACKGROUND AND OBJECTIVE Underlying biomechanical instability of the vessel wall is believed to drive the substantial morphological variability observed in saccular intracranial aneurysms. Existing approaches to quantify this shape variance rely largely on handcrafted descriptors, which capture only limited geometric complexity and show inconsistent findings across studies. Deep learning-based models have shown promise for rupture risk classification but do not provide explicit, analyzable representations of shape. We therefore develop a unified framework that learns a compact but expressive latent representation of aneurysm morphology for generative modeling and rupture-label classification. METHODS Dense point correspondences were computed for 958 patient-derived aneurysm surfaces (338 ruptured) from five public datasets using uniform parametric mapping. Autoencoder and variational autoencoder models were trained on corresponded meshes at three resolutions (700, 3k, and 12k points) to learn 2-dimensional latent spaces. Reconstruction was evaluated by mean squared error, volumetric error, and Hausdorff distance. We tested rupture-label classification from latent features using logistic regression, support vector machines, and k-nearest neighbors. Baselines included statistical shape models, diffusion maps, UMAP, and established morphological descriptors. RESULTS The 2-dimensional variational autoencoder achieved high-fidelity reconstruction (mean Hausdorff distance 0.27 ± 0.24 mm), comparable to a 50-dimensional principal component model. Latent features outperformed handcrafted descriptors and other dimensionality-reduction baselines for rupture-label classification, reaching AUC 0.78 and accuracy 0.76 across classifiers (only shape/size parameters). The learned latent spaces showed interpretable continuous transitions between morphological phenotypes, including elongation and multilobularity. CONCLUSIONS The proposed framework unifies correspondence mapping, generative modeling, and discriminative analysis in a single workflow. The 2-dimensional latent space preserves clinically relevant geometry, enables real-time synthesis and exploration of aneurysm variability, and improves rupture-label discrimination over existing techniques, providing a scalable and interpretable basis for quantitative aneurysm morphometry.