hide
Free keywords:
-
Abstract:
Introduction
In recent years, several quantitative meta-analysis techniques have emerged aiming at (1) the identification and modelling of individual brain regions that show consistent responses across experiments, and (2) the search for functional networks that capture multivariate co-activation patterns across several brain regions. Building on these techniques, we present a meta-analysis processing chain that facilitates the discovery of partially directed functional networks from meta-analysis fMRI data. The processing chain consists of Activation Likelihood Estimation (ALE) (Turkeltaub et al., 2002), model-based clustering (Neumann et al., 2008) and replicator dynamics (Neumann et al., 2005), three well-established meta-analysis techniques, and structure learning of Bayesian networks, which so far has not been used in the context of fMRI meta-analysis.
Methods
The consecutive application of ALE, model-based clustering, and replicator dynamics facilitates the transformation of lists of activation coordinates from independently performed imaging studies into a set of frequently activated functional regions. The co-activation of these regions across studies then serves as input to a structure-learning algorithm for Bayesian networks, where network nodes represent functional regions, and arcs between nodes represent statistical interdependencies between these regions. Specifically, if the activation of region A statistically depends on the activation of region B, structure learning will result in a directed connection from B to A.
Results
21136 activation coordinates from 2505 individual contrasts obtained in over 500 fMRI experiments were extracted from the Brainmap database (Laird et al., 2005). Experiments contained a large variety of motor, visual, and cognitive control paradigms and were the result of searching the database for the key word “fMRI”. Meta-analyses as search results were excluded from further processing. ALE resulted in 13 functional regions (Fig 1) which were subsequently refined into 49 sub-regions by model-based clustering. Replicator dynamics determined the 14 most frequently co-occurring functional regions (Fig 2) which were subjected to structure learning of three networks with 5, 7, and 10 nodes, the first containing motor-related regions, the second containing regions typically found in cognitive control tasks, and the third containing all cortical regions of the first two networks. Results are presented in Fig 3 and 4. Reliability of the results was determined by 100 repetitions of the learning process using a randomly sampled subset of all available data.
Conclusions
The proposed processing chain provides a tool for the detection of partially directed functional networks between brain regions on a meta-level. Structure learning of Bayesian networks thereby represents a useful exploratory data analysis technique complementing existing methods for the coordinate-based meta-analyses of functional imaging data.
