Registration and machine learning methods for brain imaging of Alzheimer's disease

• Main Author: Cattell, Liam
• Other Authors: Schnabel, Julia A. ; Hutton, Chloe
• Published: University of Oxford 2016
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.728895

Alzheimer's disease (AD) is an irreversible neurodegenerative disorder, characterised by memory loss and reduced cognitive function. One of the pathologic biomarkers of AD is the presence of neuritic plaques composed of beta amyloid peptides. Positron emission tomography (PET) imaging is increasingly used to assess the amyloid burden in patients with dementia, and therefore, tools for analysing amyloid PET scans could aid clinicians when diagnosing and treating AD. For this reason, this thesis presents novel methods for quantifying and classifying amyloid in brain PET images. The most important contributions made in this thesis are: 1) The formulation of a dual-modality deformable registration method, based on the demons framework, for simultaneously aligning PET and magnetic resonance (MR) images to a template space, in order to calculate standardised uptake value ratios (SUVRs). 2) The proposal of a novel machine learning method for classifying brain amyloid status, based on histograms of oriented image gradients. 3) An assessment of convolutional deep belief networks for amyloid status classification. 4) An initial investigation towards a complete classification framework for AD using multi-graph learning. The results of the registration experiments suggest that the proposed joint PET-MR registration method can perform as well as similar single-modality methods in terms of registration accuracy, and could provide an improved separation between populations of AD patients and healthy controls when used in the calculation of SUVRs. Although the investigation into convolutional deep belief networks indicated that, at present, they are impractical for amyloid status classification, our novel amyloid status classification method achieved a higher classification accuracy than two other established methods. Moreover, unlike SUVR, our proposed machine learning method does not require specific knowledge of neuroanatomy and can be applied to multiple amyloid PET tracers without substantial recalibration.