P 33 Different spatial observation scales in the analysis of an fMRI language task lead to substantially differing functional interpretations

Abstract

In the last two decades, neuroimaging research has provided important insights into structure-function relationships in the human brain. In this period, more than 40,000 studies using functional magnetic resonance imaging (fMRI) have been published in peer-reviewed academic journals. Recently, however, the methods and software used for analyzing fMRI data have come under increased scrutiny. From these investigations several lines of enquiry emerge: the cross-software comparability ( Pauli, 2016 ), the validity of statistical inference ( Eklund et al., 2016 , Wager et al., 2009 ) and interpretation, and the influence of the spatial filter on the anatomical assignment of activation peaks ( Ball, 2012 , White, 2001 ). Here, we use a wide range of Gaussian filters from 1 to 20 mm at full-width half maximum (FWHM) in analyzing fMRI data from a speech repetition paradigm in 25 subjects to investigate the scale-dependent variability of the fMRI analysis. We analyzed the data both at the coordinate level and at the level of regions-of-interests defined by particular contrasts in individuals and at the group level. We show that depending on the filter size, the results from fMRI analyses and their functional interpretation vary substantially ( Figs. 1 A and B, Figs. 2 A and B). We demonstrate how small spatial filter kernels bias the results towards subcortical and cerebellar activation clusters ( Fig. 1 B and B. We also show how the atlas-based anatomical assignment of regions shifts in relation to the filter size and how this may lead to substantially different functional interpretations of networks for language processing ( Fig. 1 A and 1B). Furthermore, we found substantially different scale-dependent cluster size dynamics between cortical and cerebellar clusters ( Fig. 2 D). We conclude that multiple spatial single-scale analyses expose the inter-individual anatomical variability of fMRI group data and may inform the interpretation of structure-function relationships. We propose to develop true spatial multiscale analyses to fully explore the deep structure of brain activations across Gaussian scale space.