Integrating Physics-Informed Neural Networks and 3D Vascular Geometry Learning for Cerebral Aneurysm Detection and Multimodal Rupture-Risk Prediction

Abstract

Cerebral aneurysms are localized dilations of intracranial arteries that may rupture and cause subarachnoid hemorrhage. Current assessment relies on human interpretation of imaging and clinical risk factors, but integrating vascular shape, flow-related information, and patient-level variables into a unified quantitative model remains challenging. This study develops a modular framework for cerebral aneurysm detection and rupture-risk prediction using 3D vascular geometry learning, physics-informed hemodynamic descriptors, and clinical variables. A PointNeXt-based detector first identified aneurysm presence from vascular point clouds. For aneurysm-positive cases, an unsteady physics-informed neural network then generated geometry-conditioned pressure, velocity, wall shear stress (WSS), time averaged WSS, oscillatory shear index (OSI), and relative residence time descriptors under prescribed Navier-Stokes residual and boundary-condition constraints. Multimodal models then integrated vascular morphology, physics-informed hemodynamic descriptors, and clinical variables to produce rupture-risk scores. The aneurysm detector achieved pooled out-of-fold area under the receiver operating characteristic curve (AUROC) of 0.959 and area under the precision-recall curve (AUPRC) of 0.859. For rupture-risk prediction, fixed 70/30 late fusion achieved the highest performance among evaluated models, with pooled AUROC of 0.827 and AUPRC of 0.732, exceeding all comparison models after Holm-corrected paired DeLong testing (all adjusted p < 0.05). Feature analysis identified OSI distribution, aneurysm location, radial geometry, and TAWSS descriptors as important contributors to cross-sectional rupture-risk discrimination. Together, these results provide a quantitative, multimodal strategy for case-specific aneurysm assessment.