Deep learning surrogates for finite element modelling of cardiac biomechanics
File(s)
Author(s)
Mu, Siyu
Type
Thesis
Abstract
Cardiac biomechanics provides a quantitative framework for understanding cardiac function and enables patient-specific assessment beyond conventional imaging biomarkers. However, the clinical adoption of image-driven cardiac biomechanical modelling remains limited by the difficulty of reconstructing patient-specific cardiac motion from medical images and the high computational cost of finite element analysis (FEA) for simulating cardiac mechanics throughout the cardiac cycle.
To address these challenges, this thesis develops a series of machine learning-based methods that accelerate key stages of the left ventricular (LV) biomechanical modelling pipeline.
First, Elastix Fourier Deep Learning (EFDL) is proposed as an unsupervised motion-tracking framework that combines image registration with deep-learning-based temporal regularisation to recover cycle-consistent cardiac motion trajectories. The method enables reliable 2D+t and 3D+t motion tracking across multiple imaging modalities.
Building upon these motion priors, the thesis introduces IMC-PINN-FE, a motion-consistent physics-informed neural network (PINN) framework for volume-driven ventricular mechanics. By integrating governing-equation constraints with tracked motion information, the method enables efficient estimation of myocardial stiffness and active tension while reconstructing physiologically coherent pressure-volume (P-V) behaviour.
To overcome the scalability limitations of patient-specific optimisation, the thesis further develops graph-based surrogate models. HeartUnloadNet predicts unloaded LV geometries directly from end-diastolic configurations on heterogeneous tetrahedral meshes, substantially reducing the computational burden of inverse cardiac mechanics. Extending this concept, CardioGraphFENet (CGFENet) is proposed as a unified surrogate model for full-cycle LV mechanics. Using a volume- and time-driven mapping, the framework predicts displacement fields and chamber pressure while supporting both forward loading and inverse unloading.
Overall, this thesis establishes an integrated computational pipeline for motion tracking, physics-informed parameter estimation, unloaded-state recovery, and fast surrogate simulation, improving the efficiency and scalability of image-driven cardiac biomechanics.
To address these challenges, this thesis develops a series of machine learning-based methods that accelerate key stages of the left ventricular (LV) biomechanical modelling pipeline.
First, Elastix Fourier Deep Learning (EFDL) is proposed as an unsupervised motion-tracking framework that combines image registration with deep-learning-based temporal regularisation to recover cycle-consistent cardiac motion trajectories. The method enables reliable 2D+t and 3D+t motion tracking across multiple imaging modalities.
Building upon these motion priors, the thesis introduces IMC-PINN-FE, a motion-consistent physics-informed neural network (PINN) framework for volume-driven ventricular mechanics. By integrating governing-equation constraints with tracked motion information, the method enables efficient estimation of myocardial stiffness and active tension while reconstructing physiologically coherent pressure-volume (P-V) behaviour.
To overcome the scalability limitations of patient-specific optimisation, the thesis further develops graph-based surrogate models. HeartUnloadNet predicts unloaded LV geometries directly from end-diastolic configurations on heterogeneous tetrahedral meshes, substantially reducing the computational burden of inverse cardiac mechanics. Extending this concept, CardioGraphFENet (CGFENet) is proposed as a unified surrogate model for full-cycle LV mechanics. Using a volume- and time-driven mapping, the framework predicts displacement fields and chamber pressure while supporting both forward loading and inverse unloading.
Overall, this thesis establishes an integrated computational pipeline for motion tracking, physics-informed parameter estimation, unloaded-state recovery, and fast surrogate simulation, improving the efficiency and scalability of image-driven cardiac biomechanics.
Version
Open Access
Date Issued
2026-03-27
Date Awarded
2026-06-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Yap, Choon Hwai
Sponsor
Imperial College London
Grant Number
#1019496
Publisher Department
Department of Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)