Robust deep learning methods for accountable contrast-agent-free CMR imaging in clinical applications
<p>Cardiovascular magnetic resonance (CMR) T1 mapping is a crucial non-invasive imaging technique for quantifying myocardial tissue characteristics, providing valuable insights into a range of ischaemic and non-ischaemic cardiac conditions, including myocardial infarction, inflammation, infilt...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2024
|
| Subjects: |
| _version_ | 1824459334108577792 |
|---|---|
| author | Gonzales, RA |
| author2 | Piechnik, SK |
| author_facet | Piechnik, SK Gonzales, RA |
| author_sort | Gonzales, RA |
| collection | OXFORD |
| description | <p>Cardiovascular magnetic resonance (CMR) T1 mapping is a crucial non-invasive imaging technique for quantifying myocardial tissue characteristics, providing valuable insights into a range of ischaemic and non-ischaemic cardiac conditions, including myocardial infarction, inflammation, infiltration, and fibrosis. Despite its clinical utility, the application of T1 mapping in routine practice faces several technical challenges, especially: motion artefacts that compromise image quality, limited visualisation tools for ease of interpretation, the need for generalisable segmentation of myocardial tissues, and the absence of reliable automated methods for annotating anatomical landmarks. This thesis addresses these challenges by proposing and validating novel deep learning approaches to enhance the clinical applicability of CMR T1 mapping.</p>
<br>
<p>The first goal of this thesis is to provide motion-corrected T1 maps by developing a deep learning-based method for correcting motion artefacts. A convolutional neural network architecture is employed to correct motion artefacts, to improve the quality of the T1 maps to be used for diagnosis. The proposed method is rigorously evaluated against traditional motion correction techniques, demonstrating superior performance in both simulated and real-world datasets.</p>
<br>
<p>The second goal is to provide an advanced visualisation technique for assessing T1 maps, including during pharmacologic stress through stress T1 reactivity maps. Leveraging vision transformers, this thesis presents a pathway for generating pixel-wise visualisations of stress-induced changes in left ventricular (LV) myocardial tissue properties, namely the delta T1 (dT1) map. This approach enhances the clinician's ability to interpret the effects of stress on the LV myocardium, providing a better visual display of the spatial extent of myocardial tissue changes during stress conditions.</p>
<br>
<p>The third goal is to provide a solution for automated segmentation of myocardial tissue in T1 mapping-based virtual native enhancement images. This is based on a quality control-driven deep ensemble method that combines the strengths of multiple deep learning models, to achieve robust and accurate segmentation, with an integrated quality control mechanism that flags potentially erroneous segmentations. This approach reduced inter-observer variability and improved the consistency of myocardial segmentation.</p>
<br>
<p>The fourth goal is to assist the automated segmentation of CMR short-axis slices into the American Heart Association 16-segment model for clinical applications across various CMR modalities. A generalist deep learning model was developed for the automated annotation of anatomical landmarks, specifically the anterior right ventricular insertion point and the LV centre point, using a dual-stage residual neural network framework, which accurately tracks these key landmarks. This work demonstrates many benefits for generalist training, validated against specialist models across diverse CMR datasets.</p>
<br>
<p>In conclusion, this thesis presents a comprehensive set of deep learning solutions aimed to enhance the clinical utility of CMR T1 mapping. By addressing the key challenges of motion correction, advanced visualisation, automated segmentation, and landmark annotation, the methods developed in this work pave the way towards more reliable, efficient, and standardised CMR imaging, which may ultimately translate into better clinical decision-making and patient outcomes.</p> |
| first_indexed | 2025-02-19T04:40:08Z |
| format | Thesis |
| id | oxford-uuid:cd1bc104-527c-40d5-a461-6f01a030cf5b |
| institution | University of Oxford |
| language | English |
| last_indexed | 2025-02-19T04:40:08Z |
| publishDate | 2024 |
| record_format | dspace |
| spelling | oxford-uuid:cd1bc104-527c-40d5-a461-6f01a030cf5b2025-02-18T09:46:37ZRobust deep learning methods for accountable contrast-agent-free CMR imaging in clinical applicationsThesishttp://purl.org/coar/resource_type/c_db06uuid:cd1bc104-527c-40d5-a461-6f01a030cf5bMagnetic resonance imagingDiagnostic imagingBiomedical engineeringDeep learning (Machine learning)Artificial intelligenceEnglishHyrax Deposit2024Gonzales, RAPiechnik, SKFerreira, VMZhang, QLewis, AJMFrangi, AF<p>Cardiovascular magnetic resonance (CMR) T1 mapping is a crucial non-invasive imaging technique for quantifying myocardial tissue characteristics, providing valuable insights into a range of ischaemic and non-ischaemic cardiac conditions, including myocardial infarction, inflammation, infiltration, and fibrosis. Despite its clinical utility, the application of T1 mapping in routine practice faces several technical challenges, especially: motion artefacts that compromise image quality, limited visualisation tools for ease of interpretation, the need for generalisable segmentation of myocardial tissues, and the absence of reliable automated methods for annotating anatomical landmarks. This thesis addresses these challenges by proposing and validating novel deep learning approaches to enhance the clinical applicability of CMR T1 mapping.</p> <br> <p>The first goal of this thesis is to provide motion-corrected T1 maps by developing a deep learning-based method for correcting motion artefacts. A convolutional neural network architecture is employed to correct motion artefacts, to improve the quality of the T1 maps to be used for diagnosis. The proposed method is rigorously evaluated against traditional motion correction techniques, demonstrating superior performance in both simulated and real-world datasets.</p> <br> <p>The second goal is to provide an advanced visualisation technique for assessing T1 maps, including during pharmacologic stress through stress T1 reactivity maps. Leveraging vision transformers, this thesis presents a pathway for generating pixel-wise visualisations of stress-induced changes in left ventricular (LV) myocardial tissue properties, namely the delta T1 (dT1) map. This approach enhances the clinician's ability to interpret the effects of stress on the LV myocardium, providing a better visual display of the spatial extent of myocardial tissue changes during stress conditions.</p> <br> <p>The third goal is to provide a solution for automated segmentation of myocardial tissue in T1 mapping-based virtual native enhancement images. This is based on a quality control-driven deep ensemble method that combines the strengths of multiple deep learning models, to achieve robust and accurate segmentation, with an integrated quality control mechanism that flags potentially erroneous segmentations. This approach reduced inter-observer variability and improved the consistency of myocardial segmentation.</p> <br> <p>The fourth goal is to assist the automated segmentation of CMR short-axis slices into the American Heart Association 16-segment model for clinical applications across various CMR modalities. A generalist deep learning model was developed for the automated annotation of anatomical landmarks, specifically the anterior right ventricular insertion point and the LV centre point, using a dual-stage residual neural network framework, which accurately tracks these key landmarks. This work demonstrates many benefits for generalist training, validated against specialist models across diverse CMR datasets.</p> <br> <p>In conclusion, this thesis presents a comprehensive set of deep learning solutions aimed to enhance the clinical utility of CMR T1 mapping. By addressing the key challenges of motion correction, advanced visualisation, automated segmentation, and landmark annotation, the methods developed in this work pave the way towards more reliable, efficient, and standardised CMR imaging, which may ultimately translate into better clinical decision-making and patient outcomes.</p> |
| spellingShingle | Magnetic resonance imaging Diagnostic imaging Biomedical engineering Deep learning (Machine learning) Artificial intelligence Gonzales, RA Robust deep learning methods for accountable contrast-agent-free CMR imaging in clinical applications |
| title | Robust deep learning methods for accountable contrast-agent-free CMR imaging in clinical applications |
| title_full | Robust deep learning methods for accountable contrast-agent-free CMR imaging in clinical applications |
| title_fullStr | Robust deep learning methods for accountable contrast-agent-free CMR imaging in clinical applications |
| title_full_unstemmed | Robust deep learning methods for accountable contrast-agent-free CMR imaging in clinical applications |
| title_short | Robust deep learning methods for accountable contrast-agent-free CMR imaging in clinical applications |
| title_sort | robust deep learning methods for accountable contrast agent free cmr imaging in clinical applications |
| topic | Magnetic resonance imaging Diagnostic imaging Biomedical engineering Deep learning (Machine learning) Artificial intelligence |
| work_keys_str_mv | AT gonzalesra robustdeeplearningmethodsforaccountablecontrastagentfreecmrimaginginclinicalapplications |