Hemodynamic Characterization of Intracranial Aneurysms

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

enIntracranial aneurysms affect approximately 3.2% of the general population. Rupture causes subarachnoid hemorrhage, a life-threatening emergency with a case fatality rate approaching 40%. Existing risk scores aggregate morphological and clinical variables but do not capture patient-specific blood flow patterns inside the aneurysm sac. This work pursues two complementary objectives: (i) identifying which hemodynamic indices separate growing from stable aneurysms in a matched cohort, and (ii) assessing the ability of a physics-informed neural network (PINN) to reproduce the hemodynamic fields obtained from a reference finite element (FEM) simulation on the same geometries, together with the associated computational cost. Both objectives are addressed on six patient-specific intracranial aneurysm geometries drawn from the Vascular Model Repository, arranged into three anatomically matched growing–stable pairs. FEM simulations with svMultiPhysics serve as the ground truth. PINNs trained with the NVIDIA PhysicsNeMo framework are evaluated against this reference on identical geometries and boundary conditions. Absolute wall-averaged time-averaged wall shear stress (TAWSS) does not discriminate the growing from the stable group at this sample size, because the two anatomical sites carry markedly different baseline shear levels. Normalising dome shear stress by the paired parent-vessel value removes this confound. The dome-to-parent TAWSS ratio is consistently lower in the growing group across all three pairs (0.26, 0.40, 0.39 versus 0.79, 0.58, 0.58), with a large effect size (Cohen’s d = −2.99). This provides cross-sectional matched-pair evidence that relative dome shielding is associated with growth status. The pure pulsatile PINN achieves a cycle-averaged relative ℓ2 velocity error between 0.81 and 0.90, with TAWSS underestimated by 65 to 89%. Training requires approximately 24 hours on a single A100 GPU per case, which is 8 to 37 times more expensive than the FEM solve for five of the six cases. Adding sparse FEM supervision reduces the velocity error to 0.10–0.30 and closes the TAWSS gap to within ±32% of the FEM reference. In the best case, on the smallest mesh, about 2% of the FEM mesh nodes are enough to drive most of this gain. The hybrid approach is a practical strategy when even a small number of reference measurements is available.