A quantitative comparison of motion detection algorithms in fMRI

An important step in the analysis of fMRI time-series data is to detect, an d as much as possible, correct for subject motion during the course of the scanning session. Several public domain algorithms are currently available for motion detection in fMRI. This paper compares the performance of four c ommonly used programs: AIR 3.08, SPM99, AFNI98, and the pyramid method of T hevenaz, Ruttimann, and Unser (TRU). The comparison is based on the perform ance of the algorithms in correcting a range of simulated known motions in the presence of various degrees of noise. SPM99 provided the most accurate motion detection amongst the algorithms studied. AFNI98 provided only sligh tly less accurate results than SPM99, however, it was several times faster than the other programs. This algorithm represents a good compromise betwee n speed and accuracy. AFNI98 was also the most robust program in presence o f noise. It yielded reasonable results for very low signal to noise levels. For small initial misalignments, TRU's performance was similar to SPM99 an d AFNI98. However, its accuracy diminished rapidly for larger misalignments . AIR was found to be the least accurate program studied. (C) 2001 Elsevier Science Inc. All rights reserved.