Abstract
We present a novel method for controlling the effects of group differences in motion on functional connectivity studies. Resting-state functional magnetic resonance imaging (rs-fMRI) is a powerful tool that allows for the assessment of whole-brain functional organization across a wide range of clinical populations. However, as highlighted by recent studies, many measures commonly used in rs-fMRI are highly correlated with subject head movement. A source of this problem is that motion itself, and motion correction algorithms, lead to spatial smoothing which is then variable across the brain and across subjects or groups dependent upon the amount of motion present during scanning. Studies aimed at elucidating differences between populations that have different head-motion characteristics (e.g., patients often move more in the scanner than healthy control subjects) are significantly confounded by these effects. In this work, we propose a solution to this problem, uniform smoothing, which ensures that all subject images in a study have equal effective spatial resolution. We establish that differences in the intrinsic smoothness of images across a group can confound connectivity results, and link these differences in smoothness to motion. We demonstrate that eliminating these smoothness differences via our uniform smoothing solution is successful in reducing confounds related to the differences in head motion between subjects.
Keywords: Head Movement, Motion, Image Smoothness, Resting-state fMRI, Connectivity
Corresponding author: Dustin Scheinost Magnetic Resonance Research Center 300 Cedar St PO Box 208043 New Haven, CT 06520-8043 Tel: (203) 785-6148 Fax: (203) 785-6534 dustin.scheinost@yale.edu
1) Introduction
Resting-state functional magnetic resonance imaging (rs-fMRI) is an emerging tool that allows for the analysis of whole-brain functional organization without a priori knowledge (Smith, 2012). By measuring the functional connectivity of brain regions via correlation of spontaneous fluctuations in the blood-oxygen-level dependent (BOLD) signal (Biswal et al., 1995; Biswal et al., 2010; Lowe et al., 1998), rs-fMRI can easily be applied clinically as it can be task- and performance-free. This technique has great clinical potential in a range of neurological diseases including those populations for whom the burden of complex cognitive tasks is greatest. While rs-fMRI is maturing as a modality, a recent set of papers have shown that most functional connectivity measures are highly correlated with subject movement (Power et al., 2012; Satterthwaite et al., 2013; Satterthwaite et al., 2012; Van Dijk et al., 2012; Yan et al., 2013). In many cases, comparisons between control groups and clinical populations, where rs-fMRI may have the most potential, are confounded by systematic differences in head movement between the groups. The interaction between study group, motion, and functional connectivity is currently a major obstacle in the development and clinical application of rs-fMRI.
Current approaches aimed at reducing the impact of motion on functional connectivity have focused generally on controlling for subject head motion. Controlling for motion is achieved by removing high-motion data (Power et al., 2012), by regressing motion at a group level (Satterthwaite et al., 2012), by matching datasets for motion (Tian et al., 2006), or by regressing higher motion terms (Satterthwaite et al., 2013). However, these approaches do not entirely eliminate motion confounds (Yan et al., 2013). One potential issue with removing time points or regressing several motion terms is that potentially real changes in connectivity associated with motion can be removed along with artifacts (Scheinost et al., 2013). Other approaches that do not rely explicitly on controlling for motion, such as removal of global signal and additional normalization, have been suggested as potential solutions to motion confounds (Power et al., 2014; Yan et al., 2013).
The primary contribution of this paper is to introduce the use of iterative smoothing as a method to reduce motion confounds of the form that arise when significant differences in motion are present between experimental groups. This approach works without needing to explicitly control for motion. First, we establish that an image’s intrinsic smoothness is correlated with both region-of-interest (ROI) based and voxel-based measures of connectivity and show that differences in smoothness across a sample can confound connectivity. Next, we show that subject head motion is correlated with this intrinsic smoothness suggesting that increased image smoothness is caused by head motion and motion correction. Finally, we demonstrate that eliminating these differences in image smoothness, by smoothing all images to a uniform level across the sample, is an effective way to reduce motion-related confounds in functional connectivity studies. We demonstrate that our method has at least equivalent performance compared to other current strategies focused on minimizing motion confounds, while not relying on excluding high motion frames from the data.
2) Methods
2.1 Subjects
We selected the Oulu dataset from the 1000 functional connectivity project (Biswal et al., 2010) (http://www.nitrc.org/plugins/mwiki/index.php/fcon_1000/). This dataset was chosen due to the large number of subjects (n=103) and due to the tight age range (range=20-23 years, mean=21.5 years, standard deviation=0.6 years) in order to minimize any age-related effects on motion (Van Dijk et al., 2012) or connectivity (Fair et al., 2008; Hampson et al., 2012). Full demographic information and imaging parameters can be found elsewhere (Biswal et al., 2010). Briefly, for each subject, the dataset included a high-resolution anatomical MPRAGE (Magnetization Prepared Rapid Gradient Echo) and a resting-state functional image. The functional images were acquired with a TR of 1.8 seconds, an imaging matrix of 64×64, 28 slices, voxel dimensions of 4×4×4.4 mm, and 245 frames. We also selected the Cambridge dataset from the 1000 functional connectivity project to use in a replication study that is presented in the Supplemental Materials section.
2.2 Preprocessing
A standard preprocessing pipeline was used. All images were slice time and motion corrected with 4th order B-spline interpolation using SPM (http://www.fil.ion.ucl.ac.uk/spm/). Unless otherwise specified, all further analysis was performed using BioImage Suite (www.bioimagesuite.org; (Joshi et al., 2011)). The functional images were then smoothed with a Gaussian kernel with full width half max (FWHM) of 6 mm or the uniform smoothing algorithm (see Section 2.3). Several covariates of no interest were regressed from the data including linear and quadratic drift, six rigid-body motion parameters, mean cerebrospinal fluid (CSF) signal, mean white-matter signal, and mean global signal. The white matter and CSF areas were defined on a template brain (Holmes et al., 1998), eroded to ensure only white matter or CSF signal would be included, and warped to the subjects’ space using a series of transformations described below. Finally, the data were low-pass filtered via temporal smoothing with a zero mean unit variance Gaussian filter (approximate cutoff frequency=0.12Hz).
2.3 Uniform smoothing
In order to create a uniform level of smoothness across the dataset (thus minimizing group differences associated with image smoothing), each subject’s functional run was smoothed with AFNI’s 3dBlurToFWHM (http://afni.nimh.nih.gov/afni ). This program iteratively smoothes a functional series using a diffusion-based smoothing scheme until the images are smoothed to approximately the desired level. Specifically, the -detrend, -automask, and -temper options were used. These options mask the data so only voxels within the brain are used in the smoothing, remove high-order polynomials trends from the data so that the estimated smoothness minimizes the impact of spatial structure, and increases the tolerance used for matching the estimated image smoothness to the desired smoothness. Both global and local smoothness are smoothed to approximately the desired level of smoothness. The input was the slice-timed and motion corrected data and was smoothed to a FWHM of 6mm. Additional information about this program can be found in the Supplemental Materials section.
2.4 ROI-based connectivity metrics
To evaluate the effect of image smoothness and motion on connectivity, we performed a standard ROI-based (“seed”) analysis using an ROI centered in the Posterior Cingulate Cortex (PCC, MNI coordinates: 0,-55,26 ). The PCC ROI was defined on the MNI reference brain as a 9mm cube and transformed back (via the inverse of the transforms described below) into individual subject space. The time course of the PCC in a given subject was then computed as the average time course across all voxels in the PCC ROI. This time course was correlated with the time course for every other voxel in the gray matter to create a map of r-values, reflecting ROI-to-whole-brain connectivity. These r-values were transformed to z-values using Fisher’s transform yielding one map for each subject representing the strength of correlation to the PCC ROI. This PCC connectivity was chosen to be consistent with other studies (Satterthwaite et al., 2013; Van Dijk et al., 2012; Yan et al., 2013).
2.5 Voxel-based connectivity metrics
In addition to ROI-based connectivity analysis, we examined the relationship between image smoothness/motion and a voxel-wise connectivity metric based on the network theory metric degree. Degree is simply the sum of all connection weights to a particular node in a network. For our purposes, each voxel is treated as a separate node and all connections are functions of the correlation between the timecourses for any two voxels. We examined degree based on a binary network (Buckner et al., 2009; Martuzzi et al., 2011). In this case, a connection (correlation) threshold of r=0.25 was used to determine if two voxels were connected and degree was simply the count of all such connections above this threshold. After preprocessing, degree was calculated for each functional run. A gray matter mask was first applied to the data so only voxels in the gray matter were used in the calculation. The gray matter mask was defined on a template brain (Holmes et al., 1998), dilated to ensure full coverage of the gray matter, and warped to each individual subjects’ space using a series of transformations described below. The timecourse for each voxel was correlated with every other timecourse in the gray matter and the voxel-based metrics described above were calculated. As global signal regression is known to create ambiguity in the sign of the correlation, we only considered positive correlations in calculating the voxel-base connectivity metrics (Buckner et al., 2009; Cole et al., 2012). To account for differences in brain size across participants, individual degree maps were normalized by one of two methods. For the first method, the degree maps were divided the total number of voxels in the individual subject’s gray matter mask. For the second method, the degree maps were converted to z-scores by subtracting the mean across all voxels and dividing by the standard deviation across all voxels. In the following sections, we refer to the output of the first normalization method simply as degree or degree maps; while, we refer to the output of the second normalization method as normalized degree or normalized degree maps. While degree can be sensitive to the choice of connection threshold (Scheinost et al., 2012),the presented motion and smoothing confounds were robust over a large range of thresholds (0.10<r<0.75). We chose this metric instead of ROI-to-ROI connectivity based on a high resolution parcellation of the brain (Finn et al., 2013; Shen et al., 2013) because degree represents a generalization of this approach as each voxel is treated as an ROI and has potential clinical utility (Constable et al., 2013).
2.6 Common space registration
To make inferences at the group level, all single-subject results were warped to a common template space through the concatenation of a series of linear and non-linear registrations. The functional series were linearly registered to the MPRAGE image. The MPRAGE anatomical image was non-linearly registered to the template brain. All transformation pairs were calculated independently and combined into a single transform warping the single-subject results into common space. This single transformation allows the single-subject images to be registered to common space with only one transformation, reducing interpolation error. All transformations were visually inspected for accuracy and were estimated using the intensity-only component of the method implemented in BioImage Suite (Papademetris et al., 2004).
2.7 Smoothness metric
The smoothness for each participant’s data was estimated using the AFNI program, 3dFWHMx. This program estimates the underlying smoothness in the x, y and z directions for a 4d fMRI dataset as a function of the ratio of the variance for each spatial derivative (x,y,z) to the total variance. Specifically, the -automask and -detrend options were used. The input was the slice time corrected, motion corrected, and smoothed images for each subject. The output of this program was estimates of smoothness in the x, y, and z directions which were combined into a single number using the geometric mean, (FWHMx*FWHMy*FWHMz)1/3, resulting in a single metric of smoothness for any given participant. Additional information about and evaluations of this program can be found in the Supplemental Materials section.
2.8 Motion metric
The motion for each participant’s data was estimated using a motion metric (Jenkinson et al., 2002) as calculated using the REST toolbox (Song et al., 2011). This metric summarizes the motion between any two frames in the fMRI timecourse into a single summary metric. For each adjacent pair of frames in the fMRI timecourse, this metric was calculated resulting in a timecourse of motion. These displacements were then averaged over the motion timecourse to attain a single motion metric for any given subject.
2.9 Analysis of image smoothness, motion, and connectivity
We primarily investigated the relation between image smoothness, motion, and connectivity using three methods. First, we determined how image smoothness and motion is related to connectivity. In this procedure, the smoothness metric or the motion metric was correlated with the connectivity measures in a voxel-wise manner. Second, we determined how motion influenced the BOLD timecourse by correlating the motion timecourse and the BOLD timecourse. Third, we tested the hypothesis that controlling for differences in image smoothness via a uniform smoothing algorithm (see section 2.3) minimizes group differences in connectivity between high motion subjects and low motion subjects. The data was split into two groups with significantly different motion (p<0.05): a high motion group (n=52, average motion=0.042) and a low motion group (n=51, average motion=0.023). There were no differences in age or handedness between the groups. Significance was assessed at p<0.05 and AFNI AlphaSim was use to correct for multiple statistical comparisons. The smoothness parameters used in the Monte Carlo simulations were calculated as described in Section 2.7. The data with a standard Gaussian filter required a cluster size of 125 voxels in 3mm MNI space and the uniformly smoothed data required a cluster size of 90 voxels in 3mm MNI space. Anatomical locations were identified using BioImage Suite’s digital Brodmann atlas.
2.10 Comparison between uniform smoothing and other strategies to reduce motion confounds
We compared uniform smoothing to several strategies recently proposed in the literature to minimize motion confounds due to group differences in motion. First, we censored high motion time points (Power et al., 2012). If the motion for any frame was greater than 0.11 mm, that frame as well as the frame before and two frames after were censored and removed from further analysis after preprocessing. The threshold used to censor frames was chosen, first, by selecting a common value in the literature (threshold=0.2mm) (Power et al., 2012; Yan et al., 2013) and, then, correcting this threshold for the different motion metric used in this work. It has been shown that the motion metric in this work is approximately a factor of 1.72 smaller than other metrics (Yan et al., 2013). For censored data, we also censored the same time points from the motion timecourse and the calculation of our motion metric. Second, we regressed higher order motion terms in addition to the six motion parameters from the individual timecourse (Satterthwaite et al., 2013). Specifically, the six motion parameters, their temporal derivative (calculated via backwards difference), and the quadratic term for these twelve parameters were included in the regression model. Last, when comparing the high and low motion group in Supplemental Material, we added motion as an additional covariate at the group level (Hampson et al., 2012; Satterthwaite et al., 2012).
We used several different procedures to compare the different strategies to minimize motion confounds. First, for comparing the correlation between the motion timecourse and the BOLD timecourse, we performed a paired t-test comparing the fisher transformed correlations produced with each strategy with the fisher transformed correlations produced with standard preprocessing. Second, for comparing the correlation between motion and connectivity, we show these confound maps and the distribution of these correlations. As these distributions are generally centered around zero, a narrower distribution would indicate that there are less voxels exhibiting a high correlation with motion. We used entropy to measure the spread of these distributions without making an assumption of the actual shape of the distribution. A lower entropy would indicate a narrower distribution and less overall correlation with motion. These correlation maps were generally not significantly different at the voxel level. Third, for comparing the differences between the high and low motion group, we show the group differences thresholded at p<0.05 corrected for multiple comparisons.
3) Results
The Results section is organized as follows. First, in Section 3.1, we present results demonstrating that image smoothness is significantly correlated with functional connectivity measures. Next, in Section 3.2, we relate image smoothness to motion. Then, in Section 3.3, we present results showing that the uniform smoothness approach can reduce potential confounds due to motion. In Section 3.4, we compare our uniform smoothing approach to others solutions to minimize motion related confounds.
3.1 Image smoothness significantly confounds connectivity
ROI connectivity from the PCC showed significant correlation (p<0.05 corrected) with image smoothness in several regions on the edges of the default mode network as shown in Figure 1A. These included positive correlations between image smoothness and PCC connectivity in the inferior medial prefrontal/orbitofrontal cortex and negative correlations in portions of the precuneus and right inferior parietal lobe. The voxel-based measure degree showed significant correlation with image smoothness throughout the gray matter (Figure 1B). Unlike, the relationship between ROI connectivity and smoothness, smoother images only showed a positive correlation with the number of high strength connections. Notably, degree for the PCC/precuneus was not significantly correlated with image smoothness. While most of the brain is highly correlated with smoothness, the strength of the correlation varies throughout the brain. Normalizing degree with the z-score normalization will “shift” which regions of the brain are correlated with smoothness (Figure 1C). For example, performing this normalization will make the PCC significantly inversely correlated with smoothness.
Figure 1.
Correlation of connectivity and image smoothness. An image’s intrinsic smoothness is significantly correlated (p<0.05 corrected) with both A) ROI-based and B-C) voxel-based metrics. A) Connectivity to the PCC showed both increased and decreased connectivity as image smoothness increased. B) The voxel based network measure degree showed only increased connectivity as image smoothness increased. C) When the degree maps were normalized to z-scores, areas (such as the PCC) that were not originally correlated with smoothness became negatively correlated with smoothness. Areas of the strongest correlation with smoothness (such as the temporal lobe) in the original map were still significantly correlated with smoothness. Such differences in image smoothness across a sample could lead to artifactual group differences and confound group comparisons.
When the data was preprocessed with the iterative smoothing algorithm, image smoothness and the two measures of functional connectivity were no longer significantly correlated in a voxel-wise manner. The lack of correlation between image smoothness and connectivity was primarily due to the reduction of variance in image smoothness across the sample. The smoothness for the uniformly smoothed images had a standard deviation of 0.025mm with minimum value of 5.79mm and a maximum of 5.93mm. The smoothness for the standard preprocessed images had a standard deviation of 0.135mm with minimum value of 6.94mm and a maximum of 7.59mm. Overall, uniformly smoothed images (mean FWHM=5.87mm) were significantly (p<0.05) less smoothed than the standard preprocessed images (mean FWHM=7.18). Further evaluation of how uniform smoothing minimizes differential smoothness across the group at the voxel level is presented as Supplemental Material.
3.2 Image smoothness is correlated with motion
Image smoothness showed a significant positive correlation with motion (r=0.28, p<0.004, Figure 2A). After the data was preprocessed with the uniform smoothing algorithm, image smoothness was no longer significantly correlated with motion (r=0.004, p=0.97, Figure 2B). Supplemental Material S2 further explores how motion and motion correction modulates image smoothness most notably via the interpolation step.
Figure 2.
Correlation of motion and image smoothness. A) Image smoothness showed a significant positive correlation with motion (r=0.28, p<0.004). B) After the data was preprocessed with the uniform smoothing algorithm, image smoothness was no longer significantly correlated with motion (r=0.004, p=0.97).
3.3 Uniform smoothing reduces confounds related to motion
When the motion timecourse was correlated with the timecourses for each voxel, motion was significantly correlated (p<0.05 corrected) with the BOLD timecourse for several regions in the brain (Figure 3A). These regions included parts of the sensory/motor network, the visual network, and attentional areas of the frontal and parietal cortices. When this correlation analysis was repeated with data that was uniformly smoothed, many of these regions were significantly less correlated (p<0.05 corrected) with motion as shown in Figure 3B. Importantly, no regions were significantly more correlated with motion after uniform smoothing. Figure S3 shows the correlation maps for the motion timecourse for each of the preprocessing strategies outlined in Section 2.10.
Figure 3.
Uniform smoothing reduces the correlation of the motion timecourse and the BOLD timecourses. A) When the motion timecourse was correlated with the BOLD timecourses for each voxel, motion was significantly correlated (p<0.05 corrected) with the BOLD timecourse for several regions in the brain. B) After uniform smoothing, the magnitude of these correlations were significantly reduced (p<0.05). Figure S3 in Supplemental Material shows the correlation maps for the motion timecourse for each of the preprocessing strategies outlined in Section 2.10.
ROI connectivity from the PCC and the voxel-based measure degree showed significant correlation (p<0.05 corrected) with motion as shown in Figure 4. For each, uniform smoothing reduced the correlation with motion (See Section 3.4). While the uniformly smoothed maps were less correlated with motion than standard preprocessing, these decreases in correlation were not significant at the voxel level.
Figure 4.
Correlation between connectivity and motion with and without uniform smoothing. ROI connectivity from the PCC and the voxel-based measure degree showed significant correlation (p<0.05 corrected) with motion. For each, uniform smoothing reduced the correlation with motion. Distributions of these correlations are shown in Figure 8.
To explore how uniform smoothing modulates individual maps, we compared the ROI-based and voxel-based connectivity maps from the uniformly smoothed data to the maps from regularly smoothed data. Comparisons of the ROI-based maps for all subjects revealed that uniform smoothing primarily reduces local connectivity in regions neighboring the PCC ROI (Figure 5A). Comparisons of the voxel-based maps for all subjects revealed that uniform smoothing reduces degree (Figure 5B) and normalized degree (not shown) throughout the grey matter.
Figure 5.
Contrast of connectivity results from regular smoothing and uniform smoothing. When the connectivity maps from data smoothed with uniform smoothing were compared with the connectivity maps from data smoothed with standard Gaussian smoothing, connectivity was reduced in several areas for both A) ROI-based and B) voxel-based metrics. Uniform smoothing reduced ROI-based correlations surrounding the PCC ROI and reduced degree throughout the grey matter.
To ensure that uniform smoothing is not removing typical patterns of connectivity, we show group averaged maps of the ROI-based and voxel-based metrics in Figure 6. After uniform smoothing, the average PCC connectivity (Figure 6A) displayed typical patterns of large positive correlations to other areas of the default mode and large negative correlations to “task-positive” networks (Fox and Raichle, 2007). Similarly, the average degree map (Figure 6B) and average normalized degree map (not shown) revealed hub regions as previously shown (Buckner et al., 2009; Cole et al., 2010; Zuo et al., 2012).
Figure 6.
Group average connectivity maps from uniformly smoothed data. Typical patterns of connectivity are observed using uniformly smoothed data. While uniform smoothing reduces connectivity and controls for motion, uniform smoothing does not disrupt the observation of normal brain networks as shown A) for ROI based connectivity to the PCC node and B) for voxel based connectivity using the degree metric.
3.4 Uniform smoothing compares well with other strategies to minimize motion confounds
All three strategies to minimize motion confounds significantly reduced the correlation between the motion timecourse and the BOLD timecourse when compared to standard processing as shown in Figure 7. Regressing higher order motion terms showed the greatest reduction in correlation. Both uniform smoothing and regressing higher order motion terms only reduced these correlations. Censoring increased correlation between the motion timecourse and the BOLD timecourse in the PCC and the inferior caudate. The distributions of these correlations are shown in the bottom row of Figure 7. Uniform smoothing, censoring high motion time points, and regression of higher order motion terms all produced narrower correlation distributions as measured by entropy (H). In agreement with the voxel wise comparisons, regression of higher order motion terms showed the lowest entropy (H=4.83); followed by uniform smoothing (H=5.50), censoring (H=5.65), and standard preprocessing (H=5.8). Even with this reduction in correlation between motion and the BOLD signal, many of the regions shown in Figure 3 remained significantly correlated with motion for all strategies. Confound maps are shown for each strategy in Supplemental Material.
Figure 7.
Changes in correlation between the motion timecourse and the BOLD timecourses for different strategies to minimize motion confounds. (top) All three strategies to minimize motion confounds significantly (p<0.05 corrected) reduced the magnitude of the correlations between the motion timecourse and the BOLD timecourses when compared to standard processing as indicated by the blue regions. Censoring high motion frames increased this correlation in the PCC as indicated by the red/yellow regions. (bottom) In agreement with the voxel wise comparisons, the distributions of these correlations are narrower, as measured by entropy, when compared to standard processing.
All strategies reduced confounds related to motion for PCC connectivity (first row of Figure 8). Censoring produced the narrowest distribution of correlations (H=5.85). Uniform smoothing and regressing higher order motion terms produced similar confound maps and distribution of correlations (H=5.93 and 5.94, respectively). For both degree and normalized degree, censoring and uniform smoothing reduced the entropy of the distribution of correlation between motion and connectivity. However, for degree, the distribution for censoring was not centered around zero and was skewed towards positive correlations (second row of Figure 8). For the distributions present in Figure 8, a distribution centered around zero indicates that the data is less correlated with motion. Likewise, distributions with a larger skew have less mass on zero correlations and more mass on non-zero correlations. Thus, these skewed distributions represent data that is more correlated with motion. Normalizing degree helped to center these distributions around zero reducing the mean correlation with motion. Confound maps are shown for each strategy in Supplemental Material.
Figure 8.
Distributions of correlations between connectivity and motion for different strategies to minimize motion confounds. (first row) For PCC connectivity, all strategies produced narrower distributions of correlations with motion as measured by entropy. (second row) For degree, only uniform smoothing produced a narrower distribution of correlations with motion and still remain centered around zero. While the distributions for degree using censoring produced the lowest entropy, these distribution showed an increase in the mean correlation. This increased correlation with motion is also shown in Figure S5. (third row) For normalized degree, uniform smoothing and censoring produced a narrower distribution of correlations with motion. Normalizing helps to center the distributions of correlations.
In Supplemental Material S5, we show comparisons between the high and low motion groups for PCC connectivity and degree. Additionally, as the uniformly smoothed data was significantly less smooth than the regular smoothed data, we present results from uniformly smoothed data with a larger effective smoothing kernel in Supplemental Material.
4) Discussion
Using a dataset with a large number of healthy subjects within a narrow age range, these data demonstrate for the first time that controlling for differences in the intrinsic smoothness of rs-fMRI data is an effective method to reduce motion confounds. We show that motion and motion correction leads to differences in smoothing across brain areas both within and across subjects. The intrinsic smoothness of the data was shown to be correlated with both ROI-based and voxel-based measures of functional connectivity suggesting that systematic differences in smoothness that arise due to motion could confound studies of connectivity. Iteratively blurring each subject’s data until the smoothness was uniform across each group removed this confound. Further, this intrinsic smoothness was significantly correlated with motion with subjects who moved the most having the greatest image smoothness. This uniform smoothing solution reduced motion effects in the BOLD timecourse but did not destroy typical patterns of connectivity. Uniform smoothing also reduced the number voxels highly correlated with motion across the sample on par with other solutions proposed in the literature.
We showed that increased head motion is associated with increased image smoothness. This increased smoothness is caused, in part, by the interpolation step during motion correction (Oakes et al., 2005), intrascan subject motion, and spin excitation history effects, and is not necessarily dependent upon the spatial autocorrelation or local functional homogeneity of the brain (Zuo et al., 2013). Once the smoothness of an image is increased, it is difficult to remove this added smoothness from the data. However, extra blurring can be applied allowing regions and/or images not blurred by motion to be blurred to the same level as the images with the largest smoothness. Uniform smoothing operates in this manner by slightly blurring each subject’s data such that all subjects are studied with approximately the same spatial smoothness. Subjects with the least motion and image smoothness will be additionally blurred to match subjects with the most motion and greatest image smoothness. Overall, images are less smooth after preprocessing with uniform smoothing than with standard smoothing. Nevertheless, uniform smoothing cannot overcome differences in image smoothness when there is large interpolation error. For example, as shown in Supplemental Material S2, linear interpolation introduces larger interpolation errors into the data when compared to other higher-order interpolation methods. Uniform smoothing cannot compensate for these interpolation errors.
An important finding of this work is that motion confounds can be reduced by controlling an experimental factor other than explicit motion measurements. Uniform smoothing or other methods to reduce the across-subject variability in smoothness are potentially very attractive in studies where a particular population or group may be expected to have more head movement than another group. Controlling or matching for motion in these cases may introduce other confounds in studies where disease severity is correlated with motion by excluding the most severe disease cases. Another example is studies investigating normal aging: It is known that the very old and very young subjects move more (Hampson et al., 2012; Satterthwaite et al., 2012; Van Dijk et al., 2012) and trying to match these datasets for motion may lead to keeping the data with the least motion from the two extreme groups and the data with the most motion for the middle-aged group.
Additionally, preprocessing steps that do not explicitly remove the high motion time points or regress higher order motion, yet still control for motion, may reveal interesting changes in connectivity that reflect the neural correlates of head movement. Recent results have shown for example that head motion is correlated with connectivity in motor processing areas of the brain (Scheinost et al., 2013; Yan et al., 2013). While, in some ways, it seems obvious that motion would be correlated with connectivity in the motor strip, these findings demonstrate that it is difficult to differentiate which portions of the presented results represent changes in brain activity associated with motion and which represent artifacts. Explicitly removing all effects related to motion reduces the likelihood of observing real changes in brain activity associated with moving one’s head. While it is essential to remove spurious differences in connectivity due to motion artifacts, equal care should be used in order not to remove any real changes in brain activity.
While uniform smoothing is a promising technique to control for motion confounds, not all confounds were removed. All methods used for comparison with uniform smoothing were unable to eliminate all group differences. These data, echo recent reports in the literature (Yan et al., 2013) which suggest that even the most sophisticated strategies to remove motion confounds cannot completely eliminate these confounds when large differences in group motion exist. As always, careful inspection of any data and results is recommended as any of the methods currently proposed in the literature may not be applicable for every situation. Different strategies to account for motion confounds may have different advantages for different experiments. For preprocessing, we removed the global signal via regression. While this processing step can change patterns of correlations (Murphy et al., 2009; Saad et al., 2012), it has also been shown to help minimize motion related confounds (Yan et al., 2013). We have not tested for smoothing confounds or for the relationship between smoothing and motion without this step. Future work would include characterizing the impact of global signal regression on image smoothness and the relationship between smoothness and motion.
Resting-state functional connectivity is finding increasing application in basic science and clinical investigations of the functional organization of the brain and yet the problem of subject motion and motion-related confounds remains a challenge. We have presented a novel approach to correct for motion confounds without explicitly controlling for subject motion (i.e. removing high-motion timeframes). Our approach, uniform smoothing, minimizes the variance of spatial smoothness across subjects using an iterative smoothing scheme where each image is progressively smoothed until a desired level of smoothness is achieved. We demonstrated that controlling for image smoothness provides an effective way to control for motion confounds in functional connectivity. This approach has the advantage that by not explicitly controlling for motion, potentially real changes in functional connectivity that may lead to, or in some way be associated with, increased motion can be explored. Furthermore, this approach can be combined with existing methods. The uniform smoothing approach introduced here holds promise as a mechanism to control for group differences in both smoothness and motion and reduce the impact of these confounds in functional connectivity analyses.
Supplementary Material
5) Acknowledgements
This study was funded in part by R01 NS052344, and R01 EB009666.
Footnotes
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
References
- Biswal B, Zerrin Yetkin F, Haughton VM, Hyde JS. Functional connectivity in the motor cortex of resting human brain using echo-planar mri. Magnetic Resonance in Medicine. 1995;34:537–541. doi: 10.1002/mrm.1910340409. [DOI] [PubMed] [Google Scholar]
- Biswal BB, Mennes M, Zuo XN, Gohel S, Kelly C, Smith SM, Beckmann CF, Adelstein JS, Buckner RL, Colcombe S, Dogonowski AM, Ernst M, Fair D, Hampson M, Hoptman MJ, Hyde JS, Kiviniemi VJ, Kötter R, Li SJ, Lin CP, Lowe MJ, Mackay C, Madden DJ, Madsen KH, Margulies DS, Mayberg HS, McMahon K, Monk CS, Mostofsky SH, Nagel BJ, Pekar JJ, Peltier SJ, Petersen SE, Riedl V, Rombouts SA, Rypma B, Schlaggar BL, Schmidt S, Seidler RD, Siegle GJ, Sorg C, Teng GJ, Veijola J, Villringer A, Walter M, Wang L, Weng XC, Whitfield-Gabrieli S, Williamson P, Windischberger C, Zang YF, Zhang HY, Castellanos FX, Milham MP. Toward discovery science of human brain function. Proc Natl Acad Sci U S A. 2010;107:4734–4739. doi: 10.1073/pnas.0911855107. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Buckner RL, Sepulcre J, Talukdar T, Krienen FM, Liu H, Hedden T, Andrews-Hanna JR, Sperling RA, Johnson KA. Cortical Hubs Revealed by Intrinsic Functional Connectivity: Mapping, Assessment of Stability, and Relation to Alzheimer’s Disease. The Journal of Neuroscience. 2009;29:1860–1873. doi: 10.1523/JNEUROSCI.5062-08.2009. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Cole MW, Pathak S, Schneider W. Identifying the brain’s most globally connected regions. NeuroImage. 2010;49:3132–3148. doi: 10.1016/j.neuroimage.2009.11.001. [DOI] [PubMed] [Google Scholar]
- Cole MW, Yarkoni T, Repovs G, Anticevic A, Braver TS. Global connectivity of prefrontal cortex predicts cognitive control and intelligence. J Neurosci. 2012;32:8988–8999. doi: 10.1523/JNEUROSCI.0536-12.2012. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Constable RT, Scheinost D, Finn ES, Shen X, Hampson M, Winstanley FS, Spencer DD, Papademetris X. Potential use and challenges of functional connectivity mapping in intractable epilepsy. Front Neurol. 2013;4:39. doi: 10.3389/fneur.2013.00039. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Fair DA, Cohen AL, Dosenbach NU, Church JA, Miezin FM, Barch DM, Raichle ME, Petersen SE, Schlaggar BL. The maturing architecture of the brain’s default network. Proc Natl Acad Sci U S A. 2008;105:4028–4032. doi: 10.1073/pnas.0800376105. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Finn ES, Shen X, Holahan JM, Scheinost D, Lacadie C, Papademetris X, Shaywitz SE, Shaywitz BA, Constable RT. Disruption of Functional Networks in Dyslexia: A Whole-Brain, Data-Driven Analysis of Connectivity. Biol Psychiatry. 2013 doi: 10.1016/j.biopsych.2013.08.031. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Fox MD, Raichle ME. Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging. Nat Rev Neurosci. 2007;8:700–711. doi: 10.1038/nrn2201. [DOI] [PubMed] [Google Scholar]
- Hampson M, Tokoglu F, Shen X, Scheinost D, Papademetris X, Constable RT. Intrinsic brain connectivity related to age in young and middle aged adults. PLoS One. 2012;7:e44067. doi: 10.1371/journal.pone.0044067. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Holmes CJ, Hoge R, Collins L, Woods R, Toga AW, Evans AC. Enhancement of MR Images Using Registration for Signal Averaging. Journal of Computer Assisted Tomography. 1998;22:324–333. doi: 10.1097/00004728-199803000-00032. [DOI] [PubMed] [Google Scholar]
- Jenkinson M, Bannister P, Brady M, Smith S. Improved optimization for the robust and accurate linear registration and motion correction of brain images. Neuroimage. 2002;17:825–841. doi: 10.1016/s1053-8119(02)91132-8. [DOI] [PubMed] [Google Scholar]
- Joshi A, Scheinost D, Okuda H, Belhachemi D, Murphy I, Staib LH, Papademetris X. Unified framework for development, deployment and robust testing of neuroimaging algorithms. Neuroinformatics. 2011;9:69–84. doi: 10.1007/s12021-010-9092-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lowe MJ, Mock BJ, Sorenson JA. Functional connectivity in single and multislice echoplanar imaging using resting-state fluctuations. NeuroImage. 1998;7:119–132. doi: 10.1006/nimg.1997.0315. [DOI] [PubMed] [Google Scholar]
- Martuzzi R, Ramani R, Qiu M, Shen X, Papademetris X, Constable RT. A whole-brain voxel based measure of intrinsic connectivity contrast reveals local changes in tissue connectivity with anesthetic without a priori assumptions on thresholds or regions of interest. NeuroImage. 2011;58:1044–1050. doi: 10.1016/j.neuroimage.2011.06.075. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Murphy K, Birn RM, Handwerker DA, Jones TB, Bandettini PA. The impact of global signal regression on resting state correlations: are anti-correlated networks introduced? Neuroimage. 2009;44:893–905. doi: 10.1016/j.neuroimage.2008.09.036. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Oakes TR, Johnstone T, Ores Walsh KS, Greischar LL, Alexander AL, Fox AS, Davidson RJ. Comparison of fMRI motion correction software tools. Neuroimage. 2005;28:529–543. doi: 10.1016/j.neuroimage.2005.05.058. [DOI] [PubMed] [Google Scholar]
- Papademetris X, Jackowski A, Schultz R, Staib L, Duncan J, Barillot C, Haynor D, Hellier P. Integrated Intensity and Point-Feature Nonrigid Registration. Medical Image Computing and Computer-Assisted Intervention-MICCAI 2004. Springer; Berlin / Heidelberg: 2004. pp. 763–770. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Power JD, Barnes KA, Snyder AZ, Schlaggar BL, Petersen SE. Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion. Neuroimage. 2012;59:2142–2154. doi: 10.1016/j.neuroimage.2011.10.018. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Power JD, Mitra A, Laumann TO, Snyder AZ, Schlaggar BL, Petersen SE. Methods to detect, characterize, and remove motion artifact in resting state fMRI. Neuroimage. 2014;84:320–341. doi: 10.1016/j.neuroimage.2013.08.048. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Saad ZS, Gotts SJ, Murphy K, Chen G, Jo HJ, Martin A, Cox RW. Trouble at rest: how correlation patterns and group differences become distorted after global signal regression. Brain Connect. 2012;2:25–32. doi: 10.1089/brain.2012.0080. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Satterthwaite TD, Elliott MA, Gerraty RT, Ruparel K, Loughead J, Calkins ME, Eickhoff SB, Hakonarson H, Gur RC, Gur RE, Wolf DH. An improved framework for confound regression and filtering for control of motion artifact in the preprocessing of resting-state functional connectivity data. Neuroimage. 2013;64:240–256. doi: 10.1016/j.neuroimage.2012.08.052. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Satterthwaite TD, Wolf DH, Loughead J, Ruparel K, Elliott MA, Hakonarson H, Gur RC, Gur RE. Impact of in-scanner head motion on multiple measures of functional connectivity: relevance for studies of neurodevelopment in youth. Neuroimage. 2012;60:623–632. doi: 10.1016/j.neuroimage.2011.12.063. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Scheinost D, Benjamin J, Lacadie CM, Vohr B, Schneider KC, Ment LR, Papademetris X, Constable RT. The Intrinsic Connectivity Distribution: A Novel Contrast Measure Reflecting Voxel Level Functional Connectivity. NeuroImage. 2012 doi: 10.1016/j.neuroimage.2012.05.073. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Scheinost D, Papademetris X, Constable RT. Correlation of head movement and fMRI temporal signal in motor processing areas: What is artifact?. 19th Annual Meeting of Organization for Human Brain Mapping; Seattle, WA. 16-20 June 2011.2013. [Google Scholar]
- Shen X, Tokoglu F, Papademetris X, Constable RT. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. Neuroimage. 2013;82:403–415. doi: 10.1016/j.neuroimage.2013.05.081. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Smith SM. The future of FMRI connectivity. Neuroimage. 2012 doi: 10.1016/j.neuroimage.2012.01.022. [DOI] [PubMed] [Google Scholar]
- Song XW, Dong ZY, Long XY, Li SF, Zuo XN, Zhu CZ, He Y, Yan CG, Zang YF. REST: a toolkit for resting-state functional magnetic resonance imaging data processing. PLoS One. 2011;6:e25031. doi: 10.1371/journal.pone.0025031. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Tian L, Jiang T, Wang Y, Zang Y, He Y, Liang M, Sui M, Cao Q, Hu S, Peng M, Zhuo Y. Altered resting-state functional connectivity patterns of anterior cingulate cortex in adolescents with attention deficit hyperactivity disorder. Neurosci Lett. 2006;400:39–43. doi: 10.1016/j.neulet.2006.02.022. [DOI] [PubMed] [Google Scholar]
- Van Dijk KR, Sabuncu MR, Buckner RL. The influence of head motion on intrinsic functional connectivity MRI. Neuroimage. 2012;59:431–438. doi: 10.1016/j.neuroimage.2011.07.044. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Yan CG, Cheung B, Kelly C, Colcombe S, Craddock RC, Di Martino A, Li Q, Zuo XN, Castellanos FX, Milham MP. A comprehensive assessment of regional variation in the impact of head micromovements on functional connectomics. Neuroimage. 2013;76:183–201. doi: 10.1016/j.neuroimage.2013.03.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Zuo XN, Ehmke R, Mennes M, Imperati D, Castellanos FX, Sporns O, Milham MP. Network centrality in the human functional connectome. Cereb Cortex. 2012;22:1862–1875. doi: 10.1093/cercor/bhr269. [DOI] [PubMed] [Google Scholar]
- Zuo XN, Xu T, Jiang L, Yang Z, Cao XY, He Y, Zang YF, Castellanos FX, Milham MP. Toward reliable characterization of functional homogeneity in the human brain: preprocessing, scan duration, imaging resolution and computational space. Neuroimage. 2013;65:374–386. doi: 10.1016/j.neuroimage.2012.10.017. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.