Numerical methods for fMRI data analysis

Brain imaging data are increasingly analyzed via a range of machine-learning methods. In this thesis, we discuss three specific contributions to the field of neuroimaging analysis methods: 1. To apply a recently-developed technique for identifying and viewing similarity structure in neuroimaging data, in which candidate representational structures are ranked (Kemp et al. [34]); 2. Provide side-by-side analyses of neuroimaging data by a typical non-hierarchical (SVM) versus hierarchical (Decision Tree) machine-learning classification methods; and 3. To develop a novel programming environment for PyMVPA, a current popular analysis toolbox, such that users will be able to type a small number of packaged commands to carry out a range of standard analyses. We carried out our analysis with an fMRI data set generated using auditory stimuli. “Tree” and “Ring” were the best-voted structural representations we obtained by applying the Kemp et al. algorithm to our data. Machine-learning classification resulted in accuracy values that were similar for both decision tree and SVM algorithms. Coding for different sound categories primarily occurred in the temporal lobes of the brain. We discovered a few non-temporal regions of the brain coding for these auditory sounds as well.