CFD-Trained Machine Learning Algorithm to Predict Hemodynamic Features in Patient-Specific Vascular Geometries

Abstract

study presents a novel approach by linking computational fluid dynamics (CFD) and machine learning algorithms (ML) to identify growing cerebrovascular aneurysms from stable ones. The growth of cerebral aneurysms has been linked to local hemodynamic conditions; thus, the main objective of this thesis is to apply our in-house developed approach to predict hemodynamic parameters such as pressure, velocity, wall shear stress within patient-specific vascular geometries, with emphasize on accuracy and shortening the computational time. Our ultimate goal is to predict patient-specific hemodynamic features which will help guide neurosurgeons by making a rapid assessment is to identify the growing aneurysms based on predicted hemodynamic parameters and decide on treatments that are most likely to work to minimize risk of aneurysm rupture. Our predictive approach has been developed by A) pre-processing of patient-specific computed tomography angiography (CTA) images to reconstruct 3D geometry of an artery with aneurysm, B) simulating the blood flow within 3D vascular geometries to compute hemodynamic features via CFD method, C) training different machine learning algorithms such as regression models with CFD-produced results, D) reproducing hemodynamic features via ML algorithms, E) testing accuracy of ML algorithms in predicting hemodynamics features.