Purchase this article with an account.
Sergey V. Fogelson, Peter J. Kohler, Michael Hanke, Yaroslav O. Halchenko, James V. Haxby, Richard H. Granger, Peter U. Tse; STMVPA: Spatiotemporal multivariate pattern analysis permits fine-grained visual categorization. Journal of Vision 2011;11(11):814. doi: 10.1167/11.11.814.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Multi-voxel pattern analyses of neuroimaging data are traditionally performed on individual, temporally distinct observations (i.e. single brain volume acquisitions). However, as we have previously shown, the temporal dynamics of multi-voxel pattern classification of basic visual categories (faces and houses) vary systematically as a function of the visual area studied (Kohler et al., VSS 2010). We performed fast acquisition sequence fMRI experiments with slow event-related designs, using faces and houses as stimuli, and explored the spatiotemporal activity patterns within functionally defined regions of interest (ROIs) in occipital and ventral temporal cortex. We probed the temporal variability of the multi-voxel patterns by training and testing a classifier on multiple temporally contiguous acquisitions per observation, effectively “spatializing” time. This spatiotemporal multi-voxel pattern analysis (STMVPA) revealed higher classification accuracies within our ROIs as compared with single timepoint, spatial multi-voxel pattern analysis. In order to investigate patterns of activation outside our ROIs, we also performed a whole-volume analysis using a spatiotemporal searchlight. This analysis revealed regions not found using a standard, purely spatial searchlight (Kriegeskorte et al., 2006). Furthermore, STMVPA and the spatiotemporal searchlight allowed us to probe within-category distinctions, including the representation of gender and the identity of a face, that have thus far mostly resisted inquiry (except see Kriegeskorte et al., 2007). Our results suggest that reliable, temporally variable information is found within the hemodynamic response, and that this information can be usefully exploited using STMVPA. We propose this temporal extension to multi-voxel pattern analyses as a methodological advance that increases the sensitivity and scope of traditional multi-voxel analyses of neuroimaging data, at least in cases where the temporal sampling of the hemodynamic response is relatively fine.
This PDF is available to Subscribers Only