Improved Motion Correction of fMRI Time-Series Corrupted with Major Head Movement Using Extended Motion-Corrected Independent Component Analysis

Abstract

An extension of previously-described Motion-Corrected Independent Component Analysis (MCICA) for improved correction of significant patient head motion in fMRI data is proposed. For fMRI time-points corrupted with relatively large motion, i.e. on the order of half a voxel, only partial images subject to minimal interpolation artifact are initially used in MCICA, allowing for an accurate estimation of the activation weights of the underlying ICA components. The remaining voxels that are irretrievably corrupted with gross motion in the motion-corrupted time-points are treated as missing data, so the final component maps of the ICA components are estimated from an optimally motionless reference ensemble. Interpolation artifact therefore is minimized in the final registered image, which can be mathematically expressed as a weighted combination of the extended reference ensemble. Experiments demonstrate that the proposed method was robust to the presence of simulated activation and the number of reference images used. While the previous version of MCICA already achieved noticeably decreased registration error than SPM and AIR, the proposed method further reduced the error by thirty percent when correcting simulated gross movements applied on real fMRI time-points. With a real fMRI time-series acquired during a motor-task, further increased mutual information and more clustered activation in the primary and supplementary motor areas were observed.