Motion Artifact in Magnetic Resonance Imaging: Implications for Automated Analysis Jonathan D. Blumenthal,* Alex Zijdenbos,† Elizabeth Molloy,* and Jay N. Giedd* *Child Psychiatry Branch, NIMH, NIH, Bethesda, Maryland 20892-1367; and †Montreal Neurological Institute, McGill University Received September 20, 2001 Automated measures of cerebral magnetic reso- nance images (MRI) often provide greater speed and reliability compared to manual techniques but can be particularly sensitive to motion artifact. This study employed an automatic MRI analysis program that quantified regional gray matter volume and created images for verification and quality control. Motion ar- tifact was assessed on each image and assigned a rat- ing of none, mild, moderate, or severe. Greater motion artifact was associated with smaller gray matter vol- umes. Severity of motion artifact is an important, but often overlooked, consideration in the interpretation of automated MRI measures. © 2002 Elsevier Science (USA) Key Words: brain; magnetic resonance imaging; mo- tion artifact; segmentation. INTRODUCTION As the boundaries of most cerebral structures are the interface of gray matter, white matter, and/or cerebro- spinal fluid, morphometry of cerebral magnetic reso- nance images relies on accurate classification of image voxels into these types. Hand tracing by human raters is time-consuming and may be vulnerable to poor intra and inter-rater reliability. Therefore, there is a strong impetus to apply automated techniques to the quanti- fication of cerebral MR images, especially when dealing with large sample sizes. Advances in image analysis technology have made great strides in improving the reliability and validity of automated techniques. The technique discussed here uses a probabilistic atlas that classifies tissue according to its location in standardized space combined with an artificial neural network based method that classifies tissue according to voxel intensity (Collins et al., 1999). Face-validity for this technique is high and age, sex, and psychiatric diagnosis effects have been reported based on its out- put (Castellanos et al., 2001; Giedd et al., 1999, 2001; Paus et al., 1999; Rapoport et al., 1999). However, the effects of motion artifact, frequently present in MR images, on this technique have not previously been addressed. To examine this issue we analyzed gray matter volume from 180 healthy volunteers and, in addition, one healthy volunteer scanned multiple times with varying degrees of motion artifact. MOTION ARTIFACT Pharmacologic sedation can minimize movement during scan acquisition but is not appropriate in all situations (e.g., patients with medical contraindica- tions to sedation and, because of the risks associated with sedation, healthy volunteers participating in re- search protocols). Mechanical attempts to limit motion, such as placing foam padding around the subject’s head, are often in- adequate to eliminate mild movement that can result in concentric bands of high intensity on the MR im- ages. An alternative approach is to measure gross patient motion and correct for its effects during image analy- sis. Investigators have been working on ways to sup- press the effect of motion in an MRI scan, using post- processing techniques that involve the minimization of measures of motion artifact (Atkinson et al., 1999; Hed- ley and Yan, 1992; Manduca et al., 2000; Su et al., 2001). These methods typically require the acquisition of the frequency domain (k-space) data from the MRI scanner, which is uncommon or even unfeasible in routine acquisition protocols. As such, many MRI re- searchers are not able to make use of these techniques because they are unable to alter the scanning protocol (e.g., ongoing longitudinal studies) or because these autofocus or autocorrection methods may not be able to correct for all degrees of freedom (e.g., through-plane rotation) of the head motion expected from a child. Automated measures of gray matter (GM) volumes are especially susceptible to motion artifact because motion blurs image intensities. A narrow, highly con- voluted region, such as cortical gray matter, neighbor- ing a comparatively large homogeneous white matter region will tend to “disappear” when blurred by motion. Algorithms designed to quantify GM volume, including the ANIMAL + INSECT (Collins, 1999) method used in this study, typically rely, at least in part, on the NeuroImage 16, 89 –92 (2002) doi:10.1006/nimg.2002.1076, available online at http://www.idealibrary.com on 89 1053-8119/02 $35.00 © 2002 Elsevier Science (USA) All rights reserved.