Deep learning for interpretable pathology detection from cardiac magnetic resonance imaging
File(s)
Author(s)
On, Yu Hon
Type
Thesis or dissertation
Abstract
Cardiovascular disease (CVD) remains the leading cause of death and disability worldwide, creating an urgent need for early, accurate, and scalable diagnostic tools. Cardiac magnetic resonance imaging (MRI) provides comprehensive, non-invasive assessment of cardiac structure, function, and tissue characteristics, but the growing volume and complexity of imaging data challenge manual interpretation and clinical workflow efficiency. This thesis addresses these challenges by developing automated, interpretable machine learning frameworks—spanning traditional algorithms, deep learning, and self-supervised techniques—for the detection and characterisation of cardiac abnormalities from MRI.
The work first assesses traditional feature engineering and classical machine learning classifiers for disease identification using hand-crafted morphological features, clarifying both the value and limitations of explicitly interpretable models when data are limited. Building on this, deep convolutional neural network (CNN) architectures are developed for key clinical tasks, including segmentation of atrial from LGE MRI and left ventricular structures and automated classification of valvular heart disease from cine MRI. To narrow the gap between model accuracy and clinical trust, post-hoc interpretability methods (such as saliency mapping) are used to visualise and explain model predictions. Finally, the thesis investigates self-supervised learning to exploit large volumes of unlabelled cardiac MRI, improving data efficiency in settings with scarce expert annotation. Multi-centre datasets and public challenge benchmarks are used to rigorously evaluate the proposed methods, demonstrating improved robustness, generalisability, and clinical relevance across heterogeneous imaging conditions.
Collectively, this thesis advances automated cardiac MRI analysis by coupling methodological innovation with a sustained focus on clinical interpretability and practical deployment. The results highlight the potential for machine learning to deliver earlier, explainable, and more personalised diagnostic support in cardiovascular medicine, paving the way for optimised imaging workflows and more efficient, patient-centred care.
The work first assesses traditional feature engineering and classical machine learning classifiers for disease identification using hand-crafted morphological features, clarifying both the value and limitations of explicitly interpretable models when data are limited. Building on this, deep convolutional neural network (CNN) architectures are developed for key clinical tasks, including segmentation of atrial from LGE MRI and left ventricular structures and automated classification of valvular heart disease from cine MRI. To narrow the gap between model accuracy and clinical trust, post-hoc interpretability methods (such as saliency mapping) are used to visualise and explain model predictions. Finally, the thesis investigates self-supervised learning to exploit large volumes of unlabelled cardiac MRI, improving data efficiency in settings with scarce expert annotation. Multi-centre datasets and public challenge benchmarks are used to rigorously evaluate the proposed methods, demonstrating improved robustness, generalisability, and clinical relevance across heterogeneous imaging conditions.
Collectively, this thesis advances automated cardiac MRI analysis by coupling methodological innovation with a sustained focus on clinical interpretability and practical deployment. The results highlight the potential for machine learning to deliver earlier, explainable, and more personalised diagnostic support in cardiovascular medicine, paving the way for optimised imaging workflows and more efficient, patient-centred care.
Version
Open Access
Date Issued
2025-11-30
Date Awarded
2026-07-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Varela, Marta
Bharath, Anil
Cole, Graham
Publisher Department
National Heart & Lung Institute
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)