A probabilistic approach to full-brain functional connectivity analyses

Recent work suggests that our brains’ sub-structures change how they communicate with one another depending on the particular functions or computations that our ongoing cognitive processes demand (for review see [1]). The standard approach to estimating these so called functional connectivity patterns in functional magnetic resonance imaging (fMRI) data is to compute the correlation between the time series of (pairs of) voxel activations during a particular experimental condition. However, this voxel-based approach carries a substantial computational burden (of computing time and memory), which has led most researchers to focus their connectivity analyses on a small number of pre-selected regions of interest (ROIs). Here we present a technique, termed Hierarchical Topographic Factor Analysis (HTFA), for efficiently discovering full-brain networks (without pre-selecting ROIs) in large multi-subject neuroimaging datasets. HTFA approximates each subject’s full-brain functional connectivity network through a smaller number of network nodes. The number of nodes, along with their locations, sizes, and activations (over time) are determined in an unsupervised manner from the dataset. Because the number of nodes is typically substantially smaller than the number of voxels in an fMRI dataset, HTFA can be orders of magnitude more efficient than voxel-based functional connectivity approaches. Among other benefits, this enables researchers to apply polynomial time algorithms (which includes many pattern classification algorithms) to full-brain functional connectivity networks without paying the typical huge increase in computational time and memory that voxel-based methods demand. We show that HTFA recovers the connectivity patterns underlying a synthetic dataset, and provide a case study illustrating how HTFA may be used to discover full-brain connectivity patterns in real fMRI data.