| Summary: | <p>Cardiovascular diseases account for the highest number of annual deaths worldwide, a burden exacerbated by current limitations in disease understanding. Accurate clinical outcome prediction is key for reducing these fatalities. However, high phenotype heterogeneity of diseases like hypertrophic cardiomyopathy and myocardial infarction (MI) complicates this task. Current clinical methods used for assessing outcome risk, such as ejection fraction, are often seen as too coarse and non-specific, failing to take complex variations of myocardial shape and function into account. Thus there is a growing need for more holistic approaches. The rise of deep learning has revolutionised medical image analysis and boosted interest in generative methods for cardiac modelling, due to the networks’ inherent aptitude for capturing complex data distributions. The aim of this thesis was therefore to investigate their use in learning structure-function interactions in the heart to facilitate cardiac function modelling and aid with disease understanding and clinical outcome prediction.</p>
<p>This thesis shows that Generative Adversarial Networks (GANs) and Variational
Autoencoders (VAEs) can be used for synthetic 2D Cardiac Magnetic Resonance
(CMR) data generation, reaching high levels of image realness that can convince
trained observers. It further explores modes of variation within CMR data sets
and uses latent space manipulation to guide the image generation process towards creating sub-population specific cardiac data. Conditional GANs (cGANs) are then used for 2D and 3D cardiac function modelling whereby high levels of predictive accuracy are achieved. Aided by the inclusion of patient metadata, subpopulation specific cardiac contraction patterns are extracted for both healthy and diseased patient groups from which understanding of (patho)physiological differences between groups is derived. This thesis presents a state-of-the-art segmentation network for 2D long-axis UK Biobank CMR data using cGANs to achieve an epicardial Dice score of 0.971, thereby aiding in the development of 3D CMR reconstruction pipelines based on 2D slice segmentations. Additionally, a new state-of-the-art classification network for assessing the risk of developing Major Adverse Cardiac Events (MACE) in MI survivors is presented which achieves an accuracy of 74.8%. By using class activation maps, it thus enables the derivation of MACE-related pathophysiology understanding.</p>
<p>Overall the results presented in this thesis demonstrate the power of generative
methods to capture structure-function interactions from CMR data and provide initial validation of their applicability in clinical problems.</p>
|
|---|