Computational fluid dynamics of fetal left ventricles and aortic stenosis: impacts of fetal aortic valvuloplasty and application of physics-informed neural networks
File(s)
Author(s)
Wong, Hong Shen (Sean)
Type
Thesis
Abstract
Critical aortic stenosis with evolving Hypoplastic Left Heart Syndrome (CAS-eHLHS) is a severe congenital condition marked by the narrowing of the aortic valve, which restricts blood flow and leads to the underdevelopment of the left side of the heart. Fetal aortic valvuloplasty (FAV) has emerged as an important in-utero intervention aimed at preventing CAS-eHLHS from progressing to the univentricular morphology seen in Hypoplastic Left Heart Syndrome (HLHS). FAV seeks to relieve aortic valve stenosis, enhance blood flow and promote left heart growth based on the "no flow, no grow" theory, which highlights the importance of blood flow for heart development. Despite some success, FAV’s efficacy is inconsistent and carries significant procedural risks.
This thesis addresses gaps in understanding FAV's impact by investigating flow dynamics in the fetal left ventricle (LV) using computational fluid dynamics (CFD). Pre-FAV, flow is dominated by a high-velocity mitral inflow jet vortex, while post-FAV introduces further complexity with aortic valve regurgitation, creating chaotic secondary flow patterns and increasing wall shear stress. FAV also resulted in a highly inefficient flow field within the LV.
Traditional CFD models are time-consuming and technically demanding, limiting their clinical use. In this thesis, Physics-Informed Neural Networks (PINNs), a mesh-free method, are presented as a promising alternative by integrating clinical data and bypassing the need for precise flow boundary conditions. Tested within the thesis using synthetic Doppler data, PINNs successfully reconstructed detailed 3D flow fields, even when the data was sparse or noisy, outperforming conventional methods.
In conclusion, this thesis highlights significant differences in CAS-eHLHS LV flow dynamics compared to healthy LVs, even after FAV. While traditional CFD models are challenging to apply clinically, PINNs present a viable solution for reconstructing haemodynamic flows, addressing key limitations in current methods.
This thesis addresses gaps in understanding FAV's impact by investigating flow dynamics in the fetal left ventricle (LV) using computational fluid dynamics (CFD). Pre-FAV, flow is dominated by a high-velocity mitral inflow jet vortex, while post-FAV introduces further complexity with aortic valve regurgitation, creating chaotic secondary flow patterns and increasing wall shear stress. FAV also resulted in a highly inefficient flow field within the LV.
Traditional CFD models are time-consuming and technically demanding, limiting their clinical use. In this thesis, Physics-Informed Neural Networks (PINNs), a mesh-free method, are presented as a promising alternative by integrating clinical data and bypassing the need for precise flow boundary conditions. Tested within the thesis using synthetic Doppler data, PINNs successfully reconstructed detailed 3D flow fields, even when the data was sparse or noisy, outperforming conventional methods.
In conclusion, this thesis highlights significant differences in CAS-eHLHS LV flow dynamics compared to healthy LVs, even after FAV. While traditional CFD models are challenging to apply clinically, PINNs present a viable solution for reconstructing haemodynamic flows, addressing key limitations in current methods.
Version
Open Access
Date Issued
2024-09-18
Date Awarded
01/05/2025
License URL
Advisor
Yap, Choon Hwai
Sponsor
Imperial College London
Publisher Department
Department of Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)