Unsupervised Learning and Statistical Shape Modeling of the Morphometry and Hemodynamics of Coarctation of the Aorta

Abstract

Image-based patient-specific modeling of blood flow is a current state of the art approach in cardiovascular research proposed to support diagnosis and treatment decision. However, the approach is time-consuming, and the absence of large data sets limits the applicability of Machine Learning (ML) technology. This study employs Statistical Shape Models (SSM) and unsupervised ML to interconnect the morphometry and hemodynamics for the congenital heart disease coarctation of the aorta (CoA). Based on magnetic resonance imaging (MRI) data of 154 subjects, an SSM of the stenosed aorta was developed using principal component analysis, and three clusters were identified using agglomerative hierarchical clustering. An additional statistical model describing inlet boundary velocity fields was developed based on 4D-flow MRI measurements. A synthetic database with shape and flow parameters based on statistic characteristics of the patient population was generated and pressure gradients (dP), wall shear stress (WSS), kinetic energy (KE) and secondary flow degree (SFD) were simulated using Computational Fluid Dynamics (STAR CCM+). The synthetic population with 2652 cases had similar shape and hemodynamic properties compared to a real patient cohort. Using Kruskal Wallis test we found significant differences between clusters in real and synthetic data for morphologic measures (H/W-ratio and stenosis degree) and for hemodynamic parameters of mean WSS, dP, KE sum and mean SFD. Synthetic data for anatomy and hemodynamics based on statistical shape analysis is a powerful methodology in cardiovascular research allowing to close a gap of data availability for ML.