Functional neuroimaging: a sparse modelling approach

Developments in technology have enabled scientists to study brain function in an unprecedented way. Functional neuroimaging is the use of neuroimaging technologies to capture information about the state of a brain, with the goal of studying the relationship between mental functions and brain activity. One such technology is functional magnetic resonance imaging (fMRI), which produces a signal that can be used to create cross-sectional images of the brain. These images can be used to measure brain activity in different sections of the brain. In fMRI recording there is a tradeoff between spatial and temporal resolution. My first contribution in this thesis is to present a novel algorithm for generating cross-sectional images. This is a signal processing problem with high dimensionality, but few measurements. My algorithm uses ideas from sparse modelling because variations in functional MR images are sparse over time in the wavelet domain. It will enable high resolution images to be generated using fewer measurements. Sequences of functional MR images are recorded while subjects perform different tasks. The second contribution of this thesis is a machine learning technique to predict different tasks from the captured fMRI sequences. Existing methods perform poorly at this prediction task due to the curse of dimensionality. I overcome this problem by designing a novel sparse modelling method based on the assumption that the active brain region in response to a target task is sparse in the whole brain area. The final contribution to this thesis is the design of different assessment criteria for selecting the most relevant voxels to interpret the neural activity. The conventional selection method uses the assessment of predictive performance, resulting in many false positive selections due to the small number of samples. To overcome this problem, I introduce the concept of stability. My method selects the relevant voxels using the assessments of both predictive performance and stability, which significantly reduces the selection error while maintaining the predictive performance.Open Acces