A correlation-based method for extracting subject-specific components and artifacts from group-fMRI data

We present a simple but effective correlation-based method (maxCorr) for extracting subject-specific components from group-fMRI data. The method finds signal components that correlate maximally with the data set of one subject and minimally with the data sets of the other subjects. We show that such subject-specific components are often related to movement and physiological noise (e.g. cardiac cycle, respiration). We further demonstrate that removing the most subject-specific components for each subject reduces the overall data variance and improves the statistical identification of true fMRI activations. We compare the performance of maxCorr with CompCor, a commonly used artifact-finding method in fMRI analysis. We show that maxCorr is less likely than CompCor to remove actual stimulus-related activity, especially when no information about the stimulus is available. MaxCorr operates without stimulus information and is therefore well suitable for analyses of fMRI experiments employing naturalistic stimuli, such as movies, where stimulus regressors are difficult to construct, and for brain decoding techniques benefiting from reduced subject-specific variance in each subject's data.