Abstract
The marriage of cutting edge technologies for 3D visualisation (including 3D printing and virtual reality) with volumetric medical imaging admits the construction and representation of high-fidelity models of patient-specific anatomy. These capture the structural insights of 3D scan data - those critical to the care of patients with congenital heart disease (CHD) - in a form that is accessible not only to the imaging or radiological specialist, but also to the remainder of the multidisciplinary team. In relatively small-numbered studies, this type of enhanced communication has fostered improved consensus decision-making and personalised treatment planning, amongst a host of clinically related applications. Despite their promise, we argue that the wider application of patient-specific models has been limited by the technical burden of manual image segmentation, an unavoidable step in their determination from medical images. In response, this thesis investigates methods from the burgeoning field of deep learning, in pursuit of automated solutions to the segmentation of CHD anatomy from 3D cardiac magnetic resonance (CMR) data. More specifically, we make a clinically focused appraisal of state of the art convolutional neural networks (CNNs), a family of non-linear models of high statistical capacity.Dependent on an underlying set of parameterised functions, CNNs can be tuned to the task of discriminative classification through data-driven optimisation. Observing the paucity of training examples appropriate to our task, we curate the Evelina London Children’s Hospital (ELCH) dataset, including: isotropic CMR volumes and 4D contrast enhanced scans of 150 patients with CHD; each labelled according to a clinically meaningful manual segmentation protocol expressing the haemodynamic continuity of up to eighteen cardiovascular structures (including the congenital defects therein) by pixel adjacency. In a comprehensive clinical characterisation and comparative analysis, we confirm the ELCH dataset as a quantitatively and qualitatively unique resource for both CNN training, and, more generally, for advancing our collective 3D understanding of the heart.
Leveraging this dataset within an assessment of CNN-based segmentation, we investigate different modes for combining 3D and 4D scan data within the U-Net architecture, observing inclusion of the latter to be associated with marginal gains in spatial overlap performance. More significantly, we extend our analyses beyond those encountered in the bulk of the technical literature. Presenting novel, clinically focused metrics sensitive to the presence of defects, we highlight limitations in conventional CNN optimisation: that the application of pixelwise loss functions, ignorant of extended spatial context, can result in predictions that lack coherence, and which fail to describe image data in a clinically meaningful fashion.
Interpreting these metrics through the lens of topology, we extend existing persistent homology (PH)-based loss functions for binary segmentation to the multiclass setting. Within a combinatorial framework sensitive to the topology of both individual and combined multi-class labels, these expose the differences between a predicted segmentation and a prior specification of topology according to abstract Betti numbers. We demonstrate the capacity of such losses to reliably make statistically significant improvements in multi-class segmentation topology across a range of 2D and (thanks to our highly efficient implementation based on cubical complexes and parallel execution) 3D cardiac image segmentation tasks, for the first time. Critically we show that our compact, multi-class description of topology informs patient-specific CHD diagnosis. Accordingly, by optimising our PH-based loss functions, CNNs learn a clinically meaningful representation of cardiac defects, overcoming the shortcomings of conventional pixelwise losses.
Though we cannot claim our work heralds an automated solution to the segmentation of patient-specific CHD anatomy from volumetric CMR, we believe that we have made valuable contributions in pursuit of this goal. Whether through our unique training dataset, keen clinical assessments or highly generalisable topological loss functions, we anticipate many applications and extensions of our work. I am incredibly grateful to have had the opportunity to make these contributions, and hope that they drive innovation in the personalised care of all members of the CHD population in the future.
| Date of Award | 1 Mar 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Andrew King (Supervisor) & Giovanni Montana (Supervisor) |