Data-driven fMRI data analysis based on parcellation

Functional Magnetic Resonance Imaging (fMRI) is one of the most popular neuroimaging methods for investigating the activity of the human brain during cognitive tasks. As with many other neuroiroaging tools, the group analysis of fMRI data often requires a transformation of the individual datasets to a common stereotaxic space, where the different brains have a similar global shape and size. However, the local inaccuracy of this procedure gives rise to a series of issues including a lack of true anatomical correspondence and a loss of subject specific activations. Inter-subject parcellation of fMRI data has been proposed as a means to alleviate these problems. Within this frame, the inter-subject correspondence is achieved by isolating homologous functional parcels across individuals, rather than by matching voxels coordinates within a stereotaxic space. However, the large majority of parcellation methods still suffer from a number of shortcomings owing to their dependence on a general linear model. Indeed, for all its appeal, a GLM-based parcellation approach introduces its own biases in the form of a priori knowledge about such matters as the shape of the Hemodynamic Response Function (HRF) and taskrelated signal changes. In this thesis, we propose a model-free data-driven parcellation approach to singleand multi-subject parcellation. By modelling brain activation without an relying on an a priori model, parcellation is optimized for each individual subject. In order to establish correspondences of parcels across different subjects, we cast this problem as a multipartite graph partitioning task. Parcels are considered as the vertices of a weighted complete multipartite graph. Cross subject parcel matching becomes equivalent to partitioning this graph into disjoint cliques with one and only one parcel from each subject in each clique. In order to solve this NP-hard problem, we present three methods: the OBSA algorithm, a method with quadratic programming and an intuitive approach. We also introduce two quantitative measures of the quality of parcellation results. We apply our framework to two fMRI data sets and show that both our single- and multi-subject parcellation techniques rival or outperform model-based methods in terms of parcellation accuracy