Blind source separation of fMRI data by means of factor analytic transformations

Abstract

In this study, the application of factor analytic (FA) rotation methods in the context of neuroimaging data analysis was explored. Three FA algorithms (ProMax, QuartiMax, and VariMax) were employed to carry out blind source separation in a functional magnetic resonance imaging (fMRI) experiment that involved a basic audiovisual stimulus paradigm. The outcomes were compared with those from three common independent component analysis (ICA) methods (FastICA, InfoMax, and Jade). When applied in the spatial domain (sFA), all three FA methods performed satisfactorily and comparably to the ICA methods. The QuartiMax and VariMax methods resulted in highly similar outcomes, while the ProMax results more closely resembled those from the FastICA and InfoMax ICA analyses. All methods were able to identify multiple distinct contributing factors of neural origin, including e.g. the central auditory system, the mediotemporal limbic lobe, the basal ganglia, and the motor system. In addition, various contributions from artifacts could be observed, but these constituted different factors that were well separated from those with neural effects. When applied in the temporal domain (tFA), the factor analytic methods performed drastically worse, in the sense that the spatial activation maps revealed activation much more diffusely throughout the brain and the corresponding time courses were less pronouncedly related to the employed stimulus paradigm. Temporal ICA performed better than tFA, with the possible exception of the Jade method, but still did worse than any of the spatial FA or ICA methods. In conclusion, the present findings suggest that sFA forms a viable and useful alternative to ICA in the context of fMRI data analyses, and indicate that sFA methods complement the range of blind source separation methods that are currently in use in fMRI already.