Evaluation of data-driven network analysis approaches for functional connectivity MRI

Correlated low frequency fluctuations in the blood oxygenation level dependent signal have been widely observed in highly connected brain regions and are considered to be indicative of coordinated activity within those regions. A typical functional connectivity MRI study consists of hundreds of time points acquired from thousands of image voxels, and thus exploratory data analysis is a significant challenge. This paper investigates the utilization of analytical methods based upon graph theory that can potentially provide a data-driven approach to examining the relationships between and within groups of voxels. Three algorithms, based on reachable groups, path-length analysis, and hierarchical clustering, are described and evaluated in the relatively simple context of the rodent brain. Analysis indicates that (based on the cross-correlation coefficient) cortical voxels are the most strongly connected network nodes. These voxels exhibit stronger clustering than would be expected in a randomly connected graph, and the amount of clustering is dependent on the cross-correlation threshold chosen. The analysis algorithms identify core groups in somatosensory areas and indicate that left and right somatosensory regions are more strongly connected to each other than to midline cortical areas. The results show that algorithms based on graph theory are well-suited for the data-driven analysis of functional connectivity studies.