Clustering the Brain With "CluB": A New Toolbox for Quantitative Meta-Analysis of Neuroimaging Data.

In this paper we describe and validate a new coordinate-based method for meta-analysis of neuroimaging data based on an optimized hierarchical clustering algorithm: CluB (Clustering the Brain). The CluB toolbox permits both to extract a set of spatially coherent clusters of activations from a database of stereotactic coordinates, and to explore each single cluster of activation for its composition according to the cognitive dimensions of interest. This last step, called "cluster composition analysis," permits to explore neurocognitive effects by adopting a factorial-design logic and by testing the working hypotheses using either asymptotic tests, or exact tests either in a classic inference, or in a Bayesian-like context. To perform our validation study, we selected the fMRI data from 24 normal controls involved in a reading task. We run a standard random-effects second level group analysis to obtain a "Gold Standard" of reference. In a second step, the subject-specific reading effects (i.e., the linear t-contrast "reading > baseline") were extracted to obtain a coordinates-based database that was used to run a meta-analysis using both CluB and the popular Activation Likelihood Estimation method implemented in the software GingerALE. The results of the two meta-analyses were compared against the "Gold Standard" to compute performance measures, i.e., sensitivity, specificity, and accuracy. The GingerALE method obtained a high level of accuracy (0.967) associated with a high sensitivity (0.728) and specificity (0.971). The CluB method obtained a similar level of accuracy (0.956) and specificity (0.969), notwithstanding a lower level of sensitivity (0.14) due to the lack of prior Gaussian transformation of the data. Finally, the two methods obtained a good-level of concordance (AC = 0.93). These results suggested that methods based on hierarchical clustering (and statistics) and methods requiring prior Gaussian transformation of the data can be used as complementary tools, with the GingerALE method being optimal for neurofunctional mapping of pooled data according to simpler designs, and the CluB method being preferable to test more specific, and localized, neurocognitive hypotheses according to factorial designs.