Prediction of post-embolization recurrence in internal carotid-posterior communicating aneurysms with Vel-PointNet

Abstract

Prediction models for post-embolization recurrence with hemodynamic parameters from computational fluid dynamics (CFD) were widely studied to manage cerebral aneurysms. However, only spatiotemporally averaged or maximal scalars were used to develop the models. The hemodynamic information was suppressed from 3-dimensional (3D) to 0D after averaging or maximizing, and its fidelity was strongly dependent upon the accuracy of geometry, boundary conditions, and model parameters in CFD. We designed a deep learning network, Velocity-PointNet (Vel-PointNet), to predict the recurrence probability by extracting 3D morphological and hemodynamic features in aneurysm. Geometries of 52 internal carotid-posterior communicating (ICPC) aneurysms (eight recanalized and 44 stable) were acquired from our clinical study. The blood flow was simulated using CFD. Vel-PointNet was trained with 3D morphological-hemodynamic data from CFD results. Our Vel-PointNet model was compared to existing machine learning (ML) models using significant morphological-hemodynamic scalars to verify our assumption regarding the advantage of 3D features over 0D features. Furthermore, the performance of ML models trained with morphological scalars and conventional PointNet that extracted 3D morphological features was evaluated to verify importance of hemodynamic parameters in predictive models. The area under receiver-operating characteristic curve (AUC) and precision recall curve (AUPRC) of Vel-PointNet (1.000/1.000) was higher than the other models. As a result of extracting both morphological and hemodynamic features in 3D, Vel-PointNet was found to be more accurate than traditional approaches at predicting post-embolization recurrence of ICPC aneurysms.

Keywords

• hemodynamics
• deep neural network
• computational fluid dynamics (cfd)
• cerebral aneurysm
• pointnet
• flow pattern recognition