A deep learning-based rapid strategy for computing wall shear stress in intracranial aneurysms

The rupture of intracranial aneurysms (IAs) is a life-threatening emergency, often resulting in severe complications such as subarachnoid hemorrhage and cognitive impairment with high rates of mortality and morbidity. Wall shear stress (WSS) has been identified as an independent risk factor for IA rupture. However, the clinical applicability of existing computational fluid dynamics (CFD) methods has been limited due to their high computational cost and prolonged processing time. We propose a deep learning-based neural network for WSS prediction in IAs, aiming to provide an efficient and accurate approach for hemodynamic analysis. The proposed network, trained on 2000 idealized IA models with varying geometries, utilizes a dual-path PointNet architecture. It takes as input the three-dimensional point cloud of the IAs geometry and the boundary conditions to rapidly predict WSS. Compared to CFD methods, our approach significantly reduces computation time, enabling WSS prediction within 4 s. Unlike existing neural network-based methods, our model explicitly accounts for the influence of boundary conditions on WSS estimation. Validation on real IA models from 50 clinical patients demonstrates a high degree of agreement between predictions of the network and CFD results with a mean squared error of 0.0173 and a strong correlation (r = 0.89, p < 0.001) with statistical significance (0.07 ± 0.12 Pa, p < 0.001). The proposed WSS prediction network enhances computational efficiency while maintaining high accuracy, providing valuable technical support for patient-specific IA treatment and rupture risk assessment in the future.