Computing network centrality measures on fMRI data using fully weighted adjacency matrices

Abstract: A lot of interesting research is currently being done in the field of neuroscience, a recent subject being the effort to analyse the the human brain connectome and its functional connectivity. One way this is done is by applying graph-theory based network analysis, such as centrality, on data from fMRI measurements. This involves creating a graph representation from a correlation matrix containing the correlations over time between all measured voxels. Since the input data can be very big, this results in computations that are too memory and time consuming for an ordinary computer. Researchers have used different techniques to work around this problem, examples include thresholding correlations when creating the adjacency matrix and using a smaller input data with lower resolution.This thesis proposes three ways to compute two different centrality measures, degree centrality and eigenvector centrality, on fully weighted adjacency matrices that are built from complete correlation matrices computed from high resolution input data. The first is reducing the problem by doing the calculations in optimal order and avoiding the construction of the large correlation matrix. The second solution is to distribute and do the computations in parallel on a large computer cluster using MPI. The third solution is to calculate as large sets as possible on an ordinary laptop using shared-memory parallelism with OpenMP. Algorithms are presented for the different solutions, and the effectiveness of the implementations of them is tested.