List of potential DPhil/Postdoc projects in the FMRIB Analysis Group

For all Analysis Group projects, students will need good mathematical/engineering and computing skills, and through the
projects will acquire a strong set of skills in all or most of the following areas: medical image and signal processing, Bayesian modelling, machine learning, multivariate model-free techniques (e.g. independent component analysis), biophysical modelling.

Contact primary supervisors: Steve Smith, Mark Jenkinson, Mark Woolrich, Tim Behrens

Project Titles:

Alignment of Histological and Post-Mortem MR Images
Supervisors: Mark Jenkinson & Jesper Andersson
Combining information from histology at microscopic scales (microns or less) with the information from MRI (millimetre resolution) is crucial in order to (i) understand the relationship between the MRI signal and histology, to develop biomarkers for disease and aid diagnoses, and (ii) explore relationships between distant, but connected, brain regions for probing disease mechanisms. This project aims to register images from cutting-edge post-mortem MRI to histological sections taken from the same brain. Current registration methods fail in this task as there are significantly different tissue distortions and structures highlighted in the different modalities. The proposed work will develop a novel registration technique by exploring an integration of feature-based and standard intensity-based similarity measures, and apply pattern recognition methods for the detection of tears and holes within the histologically-stained tissue slices. Development and optimisation of the overall imaging/histology protocol and analysis of the registered images will also form part of the project.

Bayesian Decoding of Mental States in functional MRI and MEG data
Supervisor: Mark Woolrich
The aim of this work is to develop multivariate Bayesian methods for decoding functional MRI and MEG data, for example, in order to predict "mental states" in the brain. Crucially, these mental states are not just experimentally controlled variables (e.g. when a reward gets received), but internal mental states derived from a hypothesized mechanism of how the brain is performing a particular task (e.g. the brain's prediction of whether a reward will be received). This work will use techniques from machine learning and pattern recognition, and combine methods such as Bayesian PCA and ICA with multivariate linear models, Gaussian processes and relevance vector machines. This will lead to powerful and general model-based approaches to FMRI and MEG data that will make a substantial impact on the field of neuroimaging.

Brain connections: Finding the routes of information flow in the brain
Supervisors: Tim Behrens, Saad Jbabdi & Mark Woolrich
The understanding of connections and functional interactions in the brain is a fundamental goal in neuroscience. This project will use two types of neuroimaging data and develop computational models of how information flows from region-to-region in the brain. Using diffusion MRI we can examine the structure of brain pathways ("tractography"), the anatomical basis of functional interactions. Using functional MRI, we can examine how these pathways are used to allow the passage of information through successive brain regions. By combining these sources of information, this project will make major advances in our ability to measure functioning brain pathways. The project will involve several Bayesian learning techniques, hierarchical models and dynamic models.

Integrating multi-subject image analysis across Diffusion, Structural, Functional and Resting-Functional MRI
Supervisors: Steve Smith, Mark Woolrich, Mark Jenkinson & Christian Beckmann
This project aims to model how spatial patterns across the brain vary across subjects, using several MRI modalities, separately and together. The different MRI modalities image brain structures, activation networks, resting "activity", white matter global connectivity and blood flow; the spatial patterns in the different image types covary across different subjects in complicated ways, allowing, for example, suitably advanced analysis techniques to attempt to classify subjects into those having schizophrenia, Alzheimer's disease and multiple sclerosis. This project will be linked to the "Bayesian decoding" project described above, and will also develop multivariate machine learning methods, integrated with Independent Component Analysis data-driven multivariate analysis. An important component will be developing optimal ways of combining (and weighting relative to each other) the different modalities. The project will have access to several large, complex multi-modal imaging datasets such as a recent study of 200 Alzheimer's and cognitively impaired subjects; we wish to learn more about the effects of genetics on the disease, and develop methods whereby MR imaging can be reliably used to predict disease early enough for new treatments to be of value.

Mathematical modelling of resting-state functional networks in the brain
Supervisors: Steve Smith (FMRIB), Mark Woolrich (OHBA MEG Centre) & Christian Beckmann (Donders, Netherlands)
In recent years the study of functional networks in the "resting" brain, as imaged by Functional MRI, has become an exciting area of brain imaging research. For example, the $30m NIH-funded "Human Connectome Project", in which we are a major partner, is using resting-state networks as one of the primary approaches for creating the most detailed mapping of brain connectivity to date. Resting-state networks (RSNs) have been the subject of many studies into their true nature ("Are RSNs really neural functional networks?") and their applications ("Are RSNs sensitive early markers for diseases such as Alzheimer's and schizophrenia?"). There are, however, many fundamental questions that still need thorough research, a good number of which relate to the mathematical techniques (e.g., independent component analysis) used to analyse resting FMRI data. In this project we will address issues such as: developing optimal analysis techniques for comparing and contrasting the spatial and temporal characteristics of RSNs across different subjects and different pathology groups; investigating temporal relationships between different resting networks; characterising the hierarchy of different resting networks and investigating the consistency of this across different subjects, in part to produce an "RSN atlas"; optimal discrimination of the resting FMRI signal into that truly caused by resting functional networks and that part caused by "uninteresting" non-neural physiological changes; investigating how the networks' spatial patterns are also present as structured covariance in other MRI modalities such as structural MRI and functional activation databases. We will also utilise recent exciting advances in accelerating the speed with which resting-FMRI data can be acquired (at least 10x MRI acceleration), in order to study the temporal dynamics of RSNs, and to discover new functional brain networks.

Modelling Brain Connectivity in functional MRI and MEG data
Supervisors: Mark Woolrich, Tim Behrens & Saad Jbabdi
Understanding the interactions between networks of brain regions, and how these relate to underlying connectional anatomy, is of central importance for a mechanistic understanding of brain function. Dynamic Causal Models (DCMs) are a unique way of testing hypotheses about the way in which different areas of the brain interact with each other, and the external environment, using FMRI and MEG data. These models have the potential to advance our understanding of the mechanisms of drugs and diseases linked to abnormalities of connectivity and synaptic plasticity. The aim of this project is to develop innovative approaches to DCM that can overcome existing challenges such as the influence of hidden sources of brain activity, and how to search for the location of brain regions in the network adaptively. This will also allow for the integration of functional and anatomical connectivity information, by combining functional network models with the global tractography methods recently developed for diffusion MRI data. The project will involve Bayesian learning techniques and dynamic models.

Multiple-Sclerosis Lesion Segmentation
Supervisor: Mark Jenkinson
Multiple-Sclerosis (MS) is a disease that shows patterns of inflammation (lesions) in the white matter of the brain, which are detectable with MRI. Automatic, accurate segmentation of lesions is very important for studying the progress of the disease, monitoring treatments, diagnosing patients and assessing clinical trials for new treatments. Current approaches, however, fail to be sufficiently accurate and robust, leaving most segmentations to be done manually with associated problems of speed and repeatability. Machine learning techniques show promising initial results but optimising their performance in conjunction with the image acquisition has not been explored. This project aims to apply existing and novel machine learning techniques to the problem of white-matter lesion segmentation as well as exploring how the performance of these methods is affected by the characteristics of the input images and features derived from them. Optimisation of the image acquisition methods, within clinically-relevant constraints, will also be investigated.

Physiological Signal Modelling and Compensation
Supervisor: Mark Jenkinson
Measured signal fluctuations due to the respiratory and cardiac cycles are a major problem in neuroimaging, especially at high field strengths (e.g. 7 Tesla) and for spinal cord imaging. RETROICOR is a traditional method of modelling and compensating for the signal but is not highly accurate and can cause a loss of statistical sensitivity, making it difficult to find the signals of interest (e.g. brain activation). This project aims to extend some of our existing work with Bayesian model comparison techniques and more sophisticated signal modelling, to improve the ability to separate signals of interest from confounding physiological signals. Applications of this work will be explored in high-field FMRI and spinal cord imaging.