Abstract
Manual segmentation of regions of interest (ROIs) has been a source of inconsistency across the literature, often with individual labs varying in their specific procedures. Recently, Julian et al. (Neuroimage 2012) developed a method of algorithmically segmenting functionally defined ROIs in the ventral visual pathway and validated their method with linear activity contrasts. Multi-voxel patterns, however, might be more sensitive to the specific localization method than linear contrasts. We explored how manually and automatically generated ROIs affect multi-voxel pattern analysis (MVPA) and repetition suppression (RS), two methods frequently used to probe high-level visual representations. We systematically explored different methods of localizing ROIs along the ventral visual stream and compared the morphology of the resulting ROIs. We then measured the effects of the localization methods on MVPA and RS using data from a recent study of natural scene representation from our lab (O'Connell et al., VSS 2013). We found that different localization methods gave consistent results with few minor exceptions. This finding imbues confidence in using an objective and fully automated method for identifying ROIs, like the one proposed by Julian et al., for both univariate and multivariate analyses. Mass adoption of an automated ROI segmentation procedure should lead to improved reproducibility and comparability of results between different labs, while drastically reducing experimenter bias in the localization process by removing the necessity of manually adjusting parameters on a subject-by-subject basis.
Meeting abstract presented at VSS 2014