Abstract
Category information represented in activation patterns can be decoded using multivariate pattern analysis of functional MRI (fMRI) measures. Cross-subject registration of brain anatomy aligns only coarse structure leaving fine-scale variability in representational patterns. This is a major challenge in building a functional brain atlas that can store common representational activation patterns. We have developed a new functional alignment method, ‘hyperalignment’, that aligns each individual's multi-voxel representational space to a common space that generalizes across experiments. We studied 14 subjects in two different imaging centers using different MRI scanners. We derived our alignment parameters using fMRI data that we obtained from ventral temporal cortex while subjects watched a movie. We then used these parameters to align fMRI data from two different experiments, in which ten subjects were shown images of seven categories of objects and faces in a block design and the other four were shown images from the same seven categories in a slow event-related design in a different scanner, into the same common space. A classifier trained on the hyperaligned face/object block design experiment data to classify these seven categories predicted the categories in the hyperaligned slow event-related data from the other four subjects with a mean accuracy of 52.7%. This between-subject classification (BSC) performance was equivalent to mean within-subject classification accuracy (WSC) for those four subjects (55.4%) and was significantly higher than BSC after anatomical alignment (BSCA) (35.8%). Mean BSC accuracy of hyperaligned data for subjects scanned in block design study was 61.3% (WSC=60.3%, BSCA=47.8%). These results demonstrate that hyperalignment provides a better way of deriving common representational patterns than does anatomical registration. Moreover, these common representational patterns can be mapped back into the brain of any reference subject opening doors for a new type of functional brain atlas that can store the high-dimensional patterns that are specific to an unlimited variety of neural representations.