Anatomically informed Bayesian inference for physiological imaging

Thesis

Anatomy interacts with all aspects of physiological imaging. Notably, volumetric techniques of limited spatial resolution struggle to image complex anatomies such as the cerebral cortex. One consequence of this is the partial volume effect, which introduces bias or confound into parameter measurements derived from such images. Fundamentally, the effect arises because the manner of image acquisition and analysis is not well-suited to the anatomy in question. Correction for partial volume effects is generally regarded as an optional post-processing step without a commonly agreed-upon strategy and it does not address the root cause of the problem.

Recently, surface-based methods have been shown to offer substantial advantages for the study of the cortex. The benefits follow from the fact that a surface representation is more appropriate for the anatomy of the cortex, namely, a thin and highly folded sheet that forms a topological sphere. Thus far, little work has investigated perfusion measurement in the cortex via arterial spin labelling in an explicitly surface-based manner, though such measurements could improve the understanding of brain function and disease.

This work represents the convergence of these two themes of partial volume effects and surface-based analysis. The major contribution is a framework for performing parameter estimation in a simultaneous surface-aware and volumetric manner, i.e., in the spaces that are most appropriate for the different anatomies present within the brain. The motivation for doing so is to realise the benefits of surface-based analysis for perfusion measurement in the cortex via arterial spin labelling, without negatively impacting measurement in the subcortex. As a consequence of the combined surface and volumetric approach, however, a new treatment of partial volume effects is obtained. Rather than considering these effects as a secondary problem that can be mitigated via post-processing, the entire approach is in fact built around them. By incorporating the anatomy that causes PVE directly into the generative model that is fit to the data, correction becomes an intrinsic feature of the estimation framework and not an afterthought.

Authors: Kirk, T

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: arterial spin labelling; perfusion; surface; physiological imaging; variational bayes
• Subjects: Magnetic resonance imaging; Perfusion (Physiology); Inference