Deep learning tools for structural and functional analysis of cardiac MRI
File(s)
Author(s)
Galazis, Christoforos
Type
Thesis
Abstract
Cardiovascular diseases are the leading cause of death worldwide, underscoring the need for earlier detection and improved risk stratification. Cardiac magnetic resonance imaging (MRI) is a key modality for assessing myocardial structure and function. Regional functional analysis has the potential to enhance diagnostic and prognostic accuracy by revealing subtle early-stage abnormalities. However, cardiac MRI is constrained by trade-offs in acquisition time, spatial resolution, and coverage. Standard protocols typically rely on thick-sliced short-axis stacks and complementary long-axis views, which only partially compensate for limited through-plane detail.
This thesis investigates deep learning approaches to overcome these limitations and enable more comprehensive, patient-specific assessments of cardiac function. First, we propose a multi-view super-resolution framework that combines standard 2D and 3D cardiac MRI views to reconstruct high-resolution volumetric representations. While promising in design, this method did not demonstrate substantial improvements in practice.
We then extend the multi-view concept to segmentation of the right ventricle, a structurally complex and clinically significant chamber. By leveraging spatial relationships across views, the proposed model improves segmentation accuracy over single-view approaches.
Next, we present a regional motion analysis framework for the left atrium, incorporating segmentation, motion estimation, strain computation, and atlas-based analysis to detect functional impairments. To address data scarcity, we integrate test-time optimisation, data augmentation, and regularisation.
Finally, we develop spatial-aware physics-informed neural networks aimed at eventual integration as a motion regulariser in cardiac analysis. By embedding physical constraints into the learning process, we have shown in an example application that the model can yield more consistent and physiologically plausible estimations.
Together, these contributions aim at advancing the capabilities of cardiac MRI analysis by enabling a more detailed and personalised assessment of cardiac function. The methods developed in this thesis lay the groundwork for integrating regional functional analysis into clinical workflows, ultimately supporting earlier diagnosis.
This thesis investigates deep learning approaches to overcome these limitations and enable more comprehensive, patient-specific assessments of cardiac function. First, we propose a multi-view super-resolution framework that combines standard 2D and 3D cardiac MRI views to reconstruct high-resolution volumetric representations. While promising in design, this method did not demonstrate substantial improvements in practice.
We then extend the multi-view concept to segmentation of the right ventricle, a structurally complex and clinically significant chamber. By leveraging spatial relationships across views, the proposed model improves segmentation accuracy over single-view approaches.
Next, we present a regional motion analysis framework for the left atrium, incorporating segmentation, motion estimation, strain computation, and atlas-based analysis to detect functional impairments. To address data scarcity, we integrate test-time optimisation, data augmentation, and regularisation.
Finally, we develop spatial-aware physics-informed neural networks aimed at eventual integration as a motion regulariser in cardiac analysis. By embedding physical constraints into the learning process, we have shown in an example application that the model can yield more consistent and physiologically plausible estimations.
Together, these contributions aim at advancing the capabilities of cardiac MRI analysis by enabling a more detailed and personalised assessment of cardiac function. The methods developed in this thesis lay the groundwork for integrating regional functional analysis into clinical workflows, ultimately supporting earlier diagnosis.
Version
Open Access
Date Issued
2025-04-10
Date Awarded
2026-02-01
Copyright Statement
Attribution 4.0 International Licence (CC BY)
License URL
Advisor
Varela, Marta
Bharath, Anil A.
Sponsor
UK Research and Innovation
Grant Number
EP/S023283/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)