Recursive extraction and modelling of coronary artery trees
File(s)
Author(s)
Batten, James
Type
Thesis
Abstract
Accurately extracting and modelling coronary artery trees from medical images is key for the effective diagnosis, treatment planning and simulation of cardiovascular disease. However, current approaches are faced with challenges due to the morphological variability of vasculature across the patient population. Conventional grid-based models often introduce discretisation artefacts, face resolution constraints and provide weak guarantees about the structural properties of their outputs. Post-processing techniques are commonly used to derive representations such as centerlines; however, these methods have many failure modes when dealing with projective imaging ambiguities. This thesis describes three novel algorithms to tackle the vessel tree modelling and extraction problems. These techniques incorporate strong inductive biases which emphasise the geometric and topological properties of the vasculature. First, we introduce a recursive 2D tree extraction model which decodes vessel connectivity from projective images. Second, we present a method for acquiring compact and grid-free representations of single arteries which uses fixed-length descriptors to capture continuous vessel shape. Third, we describe a 3D tree modelling technique that can represent both the morphology and branching structure of coronary arteries. By formulating both the tree extraction and modelling algorithms as recursive processes, we can ensure that the outputs of these models are topologically correct trees. These three methods are evaluated on real and synthetic datasets, and we demonstrate their advantages with respect to grid-based approaches in both 2D and 3D settings. Our proposed models are more effective at preserving tree connectivity and outperform the baselines on vascular geometry reconstruction tasks. We also describe three frameworks to generate synthetic data, which are used to evaluate these vessel tree modelling and extraction algorithms in conjunction with the clinical data.
Version
Open Access
Date Issued
2025-04-26
Date Awarded
2026-02-01
Copyright Statement
Attribution 4.0 International Licence (CC BY)
License URL
Advisor
Glocker, Ben
O'Regan, Declan
Wolterink, Jelmer
Sponsor
UK Research and Innovation
HeartFlow (Firm)
Grant Number
EP/S023283/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)