An alternative approach towards assessing and accounting for individual motion in fMRI timeseries

Abstract

Motion is a significant problem for the analysis of functional MRI data. This manuscript addresses the question of whether an individualized assessment of motion may be informative, and whether it may be beneficial with regard to explaining motion-related variance. Two independent datasets are used to explore and test this hypothesis, from a total of 21 healthy children, performing either no externally-cued task (resting state) or an active listening paradigm (beep story). Translations and rotations are combined into one single, individual measure of total displacement, which is demonstrated to be substantially different between brain regions as a function of their distance from the individual origin. An increasing number of covariates leads to a loss of detection power, but more so on the first than on the second level, and more so in less-powerful designs. Synthetic timeseries are calculated from which the direct effects of motion as well as motion*B0 effects can be isolated, allowing to extract individual timecourses which reflect both direct and indirect motion effects. Including three timecourses from such an individually-derived “motion fingerprint” into first-level statistical analyses explains variance to a similar degree as the commonly-used approach of including the realignment parameters, and performance is statistically equivalent to including the realignment parameters on the second level. A more individualized approach to explaining motion-related variance may therefore be beneficial, depending on the scenario.

Introduction

Functional magnetic resonance imaging (fMRI) has become one of the most widely-used research tools in the basic and clinical neurosciences (Detre, 2006, Logothetis, 2008). While offering many advantages, it is very vulnerable to subject motion (Lemieux et al., 2007, Lund et al., 2005, Wu et al., 1997). Therefore, motion is commonly “corrected for” by aligning the successive images in an fMRI timeseries, usually to the first volume (Friston et al., 1996); however, this does not fully remove all motion effects from the data (Andersson et al., 2001, Wu et al., 1997). Consequently, it was recommended to include these realignment parameters into ensuing statistical analyses (Friston et al., 1996, Lemieux et al., 2007, Salek-Haddadi et al., 2003), potentially as a function of the experimental design (Johnstone et al., 2006). The inclusion of derivatives of these parameters (for example, the Voltera expansion; Friston et al., 1996, Lemieux et al., 2007, Salek-Haddadi et al., 2003) has been shown to additionally explain variance in the data that may be attributed to motion.

However, there seem to be two points that are as yet not fully taken into account, namely that, first, the effects of motion depend on the individual geometry of the brain. As rotations around any given point will lead to a larger displacement further away from the center of rotation than closer to it, identical motion will lead to different effects in different brain regions, or across different subjects. Hence, an individualized assessment of these motion effects may be beneficial. A second concern is that including more covariates will lead to a loss of power in the final statistical model, impacting the individual as well as the group level. Hence, using fewer parameters to explain motion-related variance may be helpful, particularly in less-efficient designs with fewer degrees of freedom (Friston et al., 1999, Josephs et al., 1997, Liu et al., 2001). This technical note is aimed at exploring these aspects.