Movement correction of fMRI time-series using intrinsic statistical properties of images: an independent component analysis approach

A 3D image registration method for alignment of functional magnetic resonance imaging (fMRI) time-series, based on independent component analysis (ICA), is described. Movement during acquisition of an fMRI time-series corrupts the statistics of the acquired data, resulting in an increase in the joint entropy of the data and a decrease in the entropy of a nonlinear function of the independent components calculated by ICA. Motion effects can therefore be mitigated by spatially adjusting the data to maximize the entropy difference between the data and the nonlinear function of the calculated ICA components without explicitly estimating motion parameters. By determining which linear combination of spatially transformed images maximizes entropy difference, interpolation error incurred by resampling the misaligned image to bring it into alignment with a reference image is minimized. We applied this method to synthetic and real fMRI data. The proposed results were more accurate than cubic interpolation even when the displacement was known. We conclude that this initial approach is completely automatic, noniterative and provides nonrigid-body motion correction, without the need for an explicitly-defined reference volume.