Abstract
In most neuroimaging studies, responses are measured across the whole brain, though not all regions are necessarily informative. To consider relevant regions, current practices range from selecting voxels with high overall activity to restricting all analyses to a few independently-defined regions of interest. However, these methods tend to rely on arbitrary activity thresholds and voxel counts, and often presume particular regional boundaries. Here, we introduce an alternative method – reliability-based voxel selection. This method first computes split-half reliability for each voxel by correlating response profiles over conditions in odd and even halves of the dataset. Next, a voxel-reliability cutoff is derived that optimizes both coverage and multi-voxel pattern reliability across conditions. We employed this method on an example dataset consisting of whole-brain responses to 60 short action videos. A voxel-wise split-half reliability threshold of r>0.3 selected a set of voxels over which the multivoxel patterns for the 60 conditions reached an average reliability of r=0.88 in group data (sd=0.03). We next considered an alternative subset of voxels selected based on overall activity (all>rest t>2.0). The reliable voxel method yielded a higher condition pattern reliability (mean r across items and subjects =0.65 for reliable voxels, 0.47 for active voxels, t(12) = 15.8, p < 0.001), and this relationship held over a range of possible thresholds defining active voxels. These results replicated in a separate dataset of whole-brain responses to 72 objects (t(10) = 21.65, p < 0.001). Simulations indicate this method is suitable for designs with 15 or more conditions. The key advantages of this voxel-selection method are (1) that it leverages the structure of the data itself, without any a priori hypotheses about regions of interest or the relationships among the conditions, and (2) it emphasizes data reliability as the first step when analyzing condition-rich neuroimaging data.
Acknowledgement: Funding for this project was provided by NIH grant S10OD020039 to Harvard University Center for Brain Science, NSF grant DGE1144152 to L.T., and the Star Family Challenge Grant to T.K.