Advances in statistical methods for large-scale binary-valued neuroimaging data

Thesis

White matter lesions are common in the ageing brain and their size, location, and evolution have been shown to be informative of diagnosis, treatment, or prevention of neurological conditions such as multiple sclerosis. The work presented in this thesis explores the use of voxel-based approaches for modelling binary lesion data obtained from brain Magnetic Resonance Imaging scans. We seek to develop methods that are computationally efficient and stable for massive datasets with very low lesion incidence, and demonstrate the value of these methods with a real dataset to disentangle contributing risk factors of lesion incidence.

Our contributions are spread across three main chapters including two published articles and one preprint, and they could be summarised as

Chapter 2

Kindalova et al. (2021a) explore whether the potential gains in estimator accuracy justify the use of a more computationally intensive spatial modelling approach as opposed to a mass-univariate approach to modelling voxel-wise binary lesion data. A method comparison of three crosssectional lesion mapping approaches is facilitated through the development of a novel simulation framework of artificial lesion masks, which mimics features of real lesion masks.

Chapter 3

Veldsman et al. (2020) use data on 13,680 healthy ageing UK Biobank participants at one time point to explore the effects of the individual cerebrovascular risk factors (e.g. waist-to-hip ratio and smoking) on lesion load and on lesion probability, which has been an obstacle in the literature so far due to the dominating effect of hypertension and the presence of comorbidities.

Chapter 4

Kindalova et al. (2021b) adopt a generalized estimating equations approach to modelling longitudinal binary-valued outcomes. By adding a Jeffreys-prior penalty to log-link generalized estimating equations for relative risk regression, finiteness of the estimates along with superior convergence rate are demonstrated in an extensive simulation study as well as in a UK Biobank application on 1,578 participants with data from 2 visits.

Authors: Kindalova, P

• DOI: 10.5287/ora-ker9g8y29
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English