Bi, Ning ORCID: https://orcid.org/0000-0002-7505-3997
(2024)
Bayesian Deep Learning for Cardiac Motion Modelling and Analysis.
PhD thesis, University of Leeds.
Abstract
Cardiovascular diseases (CVDs) remain a primary cause of mortality globally, with an estimated 17.9 million deaths in 2019, accounting for 32%
of all global fatalities. In recent decades, non-invasive imaging, particularly Magnetic Resonance Imaging (MRI), has become pivotal in diagnosing CVDs, offering high-resolution, multidimensional, and sequential cardiac data. However, the interpretation of cardiac MRI data is challenging,
due to the complexities of cardiac motion and anatomical variations. Traditional manual methods are time-consuming and subject to variability. Deep
learning (DL) methods, notably generative models, have recently advanced
medical image analysis, offering state-of-the-art solutions for segmentation,
registration, and motion modelling.
This thesis encapsulates the development and validation of deep-learning
frameworks in the field of cardiac motion modelling and analysis from sequential cardiac MRI scans. At its core, it introduces a probabilistic generative model for cardiac motion modelling, underpinned by temporal coherence, capable of synthesising new CMR sequences. Three models are
derived from this foundational probabilistic model, each contributing to
different aspects.
Firstly, through the innovative application of gradient surgery techniques, we address the dual objectives of attaining high registration accuracy and ensuring the diffeomorphic characteristics of the predicted motion
fields. Subsequently, we introduce the joint operation of ventricular segmentation and motion modelling. The proposed method combines anatomical precision with the dynamic temporal flow to enhance both the accuracy
of motion modelling and the stability of sequential segmentation. Furthermore, we introduce a conditional motion transfer framework that leverages
variational models for the generation of cardiac motion, enabling anomaly
detection and the augmentation of data, particularly for pathologies that
are less commonly represented in datasets. This capability to transfer and
transform cardiac motion across healthy and pathological domains is set
to revolutionize how clinicians and researchers understand and interpret
cardiac function and anomalies.
Collectively, these advancements present novelty and application potentials in cardiac image processing. The methodologies proposed herein have
the potential to transform routine clinical diagnostics and interventions,
allowing for more nuanced and detailed cardiac assessments. The probabilistic nature of these models promises to deliver not only more detailed
insights into cardiac health but also to foster the development of personalised medicine approaches in cardiology.
Metadata
Supervisors: | Zakeri, Arezoo and Taylor, Zeike and Frangi, Alejandro and Ravikumar, Nishant |
---|---|
Keywords: | Deep Learning, Medical Image Analysis, Cardiac Image Analysis, Bayesian Inference, Generative Model, Cardiac MRI. |
Awarding institution: | University of Leeds |
Academic Units: | The University of Leeds > Faculty of Engineering (Leeds) > School of Computing (Leeds) |
Depositing User: | Dr Ning Bi |
Date Deposited: | 08 May 2024 10:24 |
Last Modified: | 08 May 2024 10:24 |
Open Archives Initiative ID (OAI ID): | oai:etheses.whiterose.ac.uk:34762 |
Download
Final eThesis - complete (pdf)
Filename: Thesis_ningb_revised.pdf
Licence:
This work is licensed under a Creative Commons Attribution NonCommercial ShareAlike 4.0 International License
Export
Statistics
You do not need to contact us to get a copy of this thesis. Please use the 'Download' link(s) above to get a copy.
You can contact us about this thesis. If you need to make a general enquiry, please see the Contact us page.