Deep learning for cardiac MRI segmentation and event prediction with extracted features
Publication Date
December 31, 2025
Creator
Abstract
Background: Ventricular Tachycardia (VT) is associated with significant mortality and morbidity. Ischaemic Heart Disease (IHD) can predispose individuals to VT through the presence of anatomical substrate (i.e. scar). Management of VT remains challenging with progression to ablation not uncommon and high risk of recurrence. Cardiac magnetic resonance imaging(CMR), particularly late-gadolinium enhanced (LGE) MRI, is beneficial for procedures through the identification of scar, thereby facilitating scar localisation, navigation of ablation as well as, through quantification, VT risk prediction. Unfortunately, the task of segmentation for either purpose is resource-intensive with suboptimal reproducibility and scar quantification does not appreciate scar complexity or identify high risk scar. Artificial intelligence (AI) can be used to rapidly and reproducibly undertake routine tasks utilising unseen features and associations in data. Its application to LGE CMR for reproducible, rapid scar segmentation has significant potential. Furthermore, subsequent extraction of data from that scar for risk prediction may be beneficial.
Aims: This thesis aims to explore applications of AI to LGE CMR for patients at risk of VT with IHD. A particular focus is placed on cardiac structure & scar delineation and prediction of clinical events.
Methods: (1) A systematic review of current scar segmentation was undertaken to identify key limitations in current models. (2) A deep learning, multilabel model was developed utilising a new, large, disease specific CMR dataset – collected and labelled with clinical data. Inter- and intra-rater assessments were undertaken to reinforce confidence in underlying labels. Testing utilised multi-model evaluation. (3) Validation of the model was undertaken utilising an external cohort. (4) Features were extracted from original labels using a bespoke developed program and assessed for reproducibility. Exploratory univariate analysis with arrhythmia and all-cause mortality were undertaken after correction for co-linearity and dimension reduction.
Results & Conclusions: (1) Feasibility of LGE CMR segmentation is shown with significant limitations in sample size, disease aetiology and application to VT. (2) Performance of the multilabel segmentation model was in keeping with previously published models with the advantage of additional structure labelling for VT ablation and building on previous limitations. (3) External validation was suboptimal, potentially due to distribution shift. The need for large, varied databases for model development including multiple scanners and sequences was highlighted. (4) Several features were associated with clinical endpoints resulting in hypothesis generation and requiring future confirmatory studies. Based on findings and limitations of data availability throughout the study, a prospective, multicentre paired imaging and clinical database has been designed and recruitment started to expand available datasets for further segmentation and prediction model refinement and testing. In addition, a retrospective study between LGE CMR findings and invasive electrophysiological testing has started recruitment for early assessments of clinical correlation.
Item Type
ethesis
Thesis Type
PhD
Supervisors
Subjects (LC)
Associated Schools / Departments
School of Medicine
eprints ID
82306
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
Jathanna Nikesh, 20123633, Corrections.pdf
Type
Full-text
Description
Examined. Corrections
Size
3.96 MB
Format
Adobe PDF
Checksum (MD5)
5b4a8f16de0f340754392dc2314342f9