Scalable Bayesian methods for the analysis of neuroimaging data
The recent surge in large-scale population health datasets, such as the UK Biobank or the Adolescent Brain Cognitive Development (ABCD) study, requires the development of scalable statistical methods that are capable of analysing the rich multitude of data sources. This thesis focuses on the scalabl...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2024
|
| Subjects: |
| _version_ | 1826313267569491968 |
|---|---|
| author | Menacher, AK |
| author2 | Nichols, T |
| author_facet | Nichols, T Menacher, AK |
| author_sort | Menacher, AK |
| collection | OXFORD |
| description | The recent surge in large-scale population health datasets, such as the UK Biobank or the Adolescent Brain Cognitive Development (ABCD) study, requires the development of scalable statistical methods that are capable of analysing the rich multitude of data sources. This thesis focuses on the scalable analysis of Magnetic Resonance Imaging (MRI) neuroimaging data, such as binary lesion masks and task-based functional Magnetic Resonance Imaging (fMRI). In particular, we introduce two Bayesian spatial models with sparsity priors on the spatially varying coefficients and extend our work to suit the large sample sizes found in population health studies.
Firstly, we propose a scalable hierarchical Bayesian image-on-scalar regression model, called BLESS, capable of handling binary responses and of placing continuous spike-and-slab mixture priors on spatially varying parameters. Thereby, enforcing spatial dependency on the parameter dictating the amount of sparsity within the probability of inclusion. The use of mean-field variational inference with dynamic posterior exploration, which is an annealing-like strategy that improves optimisation, allows our method to scale to large sample sizes. We validate our results via simulation studies and an application to binary lesion masks from the UK Biobank.
Secondly, we extend our method to account for underestimation of posterior variance due to variational inference by providing an approximate posterior sampling approach inspired by Bayesian bootstrap ideas and spike-and-slab priors with random shrinkage targets. Besides accurate uncertainty quantification, this approach is capable of producing novel cluster size-based imaging statistics, such as credible intervals of cluster size, and measures of reliability of cluster occurrence.
Thirdly, we develop a Bayesian nonparametric scalar-on-image regression model with a relaxed-thresholded Gaussian process prior on the spatially varying coefficients in order to introduce sparsity and smoothness into the model. Our main contribution is the improved scalability, allowing for larger sample sizes and bigger image dimensions, which is made possible by replacing posterior sampling with a variational approximation. We validate our results via simulation studies and an application to cortical surface task-based fMRI data from the ABCD study. |
| first_indexed | 2024-09-25T04:10:24Z |
| format | Thesis |
| id | oxford-uuid:56e3f65e-3db2-4224-8481-5463c3b8e047 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-09-25T04:10:24Z |
| publishDate | 2024 |
| record_format | dspace |
| spelling | oxford-uuid:56e3f65e-3db2-4224-8481-5463c3b8e0472024-06-24T12:49:21ZScalable Bayesian methods for the analysis of neuroimaging dataThesishttp://purl.org/coar/resource_type/c_db06uuid:56e3f65e-3db2-4224-8481-5463c3b8e047Bayesian statisticsStatisticsNeuroimagingEnglishHyrax Deposit2024Menacher, AKNichols, THolmes, CThe recent surge in large-scale population health datasets, such as the UK Biobank or the Adolescent Brain Cognitive Development (ABCD) study, requires the development of scalable statistical methods that are capable of analysing the rich multitude of data sources. This thesis focuses on the scalable analysis of Magnetic Resonance Imaging (MRI) neuroimaging data, such as binary lesion masks and task-based functional Magnetic Resonance Imaging (fMRI). In particular, we introduce two Bayesian spatial models with sparsity priors on the spatially varying coefficients and extend our work to suit the large sample sizes found in population health studies. Firstly, we propose a scalable hierarchical Bayesian image-on-scalar regression model, called BLESS, capable of handling binary responses and of placing continuous spike-and-slab mixture priors on spatially varying parameters. Thereby, enforcing spatial dependency on the parameter dictating the amount of sparsity within the probability of inclusion. The use of mean-field variational inference with dynamic posterior exploration, which is an annealing-like strategy that improves optimisation, allows our method to scale to large sample sizes. We validate our results via simulation studies and an application to binary lesion masks from the UK Biobank. Secondly, we extend our method to account for underestimation of posterior variance due to variational inference by providing an approximate posterior sampling approach inspired by Bayesian bootstrap ideas and spike-and-slab priors with random shrinkage targets. Besides accurate uncertainty quantification, this approach is capable of producing novel cluster size-based imaging statistics, such as credible intervals of cluster size, and measures of reliability of cluster occurrence. Thirdly, we develop a Bayesian nonparametric scalar-on-image regression model with a relaxed-thresholded Gaussian process prior on the spatially varying coefficients in order to introduce sparsity and smoothness into the model. Our main contribution is the improved scalability, allowing for larger sample sizes and bigger image dimensions, which is made possible by replacing posterior sampling with a variational approximation. We validate our results via simulation studies and an application to cortical surface task-based fMRI data from the ABCD study. |
| spellingShingle | Bayesian statistics Statistics Neuroimaging Menacher, AK Scalable Bayesian methods for the analysis of neuroimaging data |
| title | Scalable Bayesian methods for the analysis of neuroimaging data |
| title_full | Scalable Bayesian methods for the analysis of neuroimaging data |
| title_fullStr | Scalable Bayesian methods for the analysis of neuroimaging data |
| title_full_unstemmed | Scalable Bayesian methods for the analysis of neuroimaging data |
| title_short | Scalable Bayesian methods for the analysis of neuroimaging data |
| title_sort | scalable bayesian methods for the analysis of neuroimaging data |
| topic | Bayesian statistics Statistics Neuroimaging |
| work_keys_str_mv | AT menacherak scalablebayesianmethodsfortheanalysisofneuroimagingdata |