Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion

Abstract

Here, we demonstrate that subject motion produces substantial changes in the timecourses of resting state functional connectivity MRI (rs-fcMRI) data despite compensatory spatial registration and regression of motion estimates from the data. These changes cause systematic but spurious correlation structures throughout the brain. Specifically, many long-distance correlations are decreased by subject motion, whereas many short-distance correlations are increased. These changes in rs-fcMRI correlations do not arise from, nor are they adequately countered by, some common functional connectivity processing steps. Two indices of data quality are proposed, and a simple method to reduce motion-related effects in rs-fcMRI analyses is demonstrated that should be flexibly implementable across a variety of software platforms. We demonstrate how application of this technique impacts our own data, modifying previous conclusions about brain development. These results suggest the need for greater care in dealing with subject motion, and the need to critically revisit previous rs-fcMRI work that may not have adequately controlled for effects of transient subject movements.

Figure 1. Frame-by-frame changes in rs-fcMRI signal are related to frame-by-frame changes in head position even after motion regression

(A) rs-fcMRI timecourses from 3 ROIs in a single subject. Each ROI is a 10 mm diameter sphere centered on the coordinates listed at the right, which are all left occipital locations (see Figure S2 for further examples elsewhere in the brain). (B) The six parameters calculated for frame-by-frame realignment of the subject’s data, which indicate the translational (solid line) and rotational (dotted line) displacements of the head from a fixed position in space. Rotational displacements were calculated at a 50 mm radius. (C) The absolute values of the differential of each of the timecourses in (A). (D) The framewise displacement (FD) of head position, calculated as the sum of the absolute values of the differential of the realignment estimates in (B). Note the correspondence of high values in (C) and (D) indicating that large changes in BOLD signal co-occur with large changes in head position. RMS movement for this subject is 0.50 mm, which is well within traditionally acceptable limits of subject motion (e.g. 1.5 mm RMS movement with 3 mm isotropic voxels). Coordinates are in MNI space, and ROIs are shown on a PALS fiducial surface (Van Essen, 2005).

Figure 2. Frame-by-frame head displacement is related to frame-by-frame changes in rs-fcMRI signal throughout the brain and across subjects

(A) For each frame of data in the same subject used in Figure 1, the framewise displacement (FD) of a frame of data is plotted against the absolute values of the differentials of rs-fcMRI timecourses of 264 ROIs (locations listed and shown in Table S1 and Figure S3). These data are fitted with a loess curve (black line) sampling the nearest 5000 data points. (B) Identically produced loess curves from all 22 subjects in Cohort 1 are plotted against framewise displacement. There is a clear trend for larger frame-by-frame head displacement to co-occur with larger changes in rs-fcMRI signal. The inset magnifies the plot between framewise displacements of 0 and 1, demonstrating that this relationship exists even for very small movements.

Figure 3. Two framewise indices of data quality, and a method for flagging frames of suspect quality

(A) The framewise displacement (FD) of head position, calculated as the sum of the absolute values of the 6 translational and rotational realignment parameters (see Figure 1C,D). (B) The DVARS measure, calculated as the RMS of the differential of all timecourses of all voxels within a spatial brain mask at each frame. (C) Two temporal masks, indicating frames of suspect quality in red. Here, the upper row indicates frames that were flagged as having FD over 0.5 (see the dotted red line in (A)), and the bottom row indicates frames that were flagged as having DVARS over 5 (dotted red line in (B)). These temporal masks are quite similar, but not identical, just as the plots in (A) and (B) are quite similar, but not identical. (A-C) show data for the same subject examined in Figure 1 (RMS movement = 0.50 mm), and (D-F) show data from another subject, whose RMS movement is 0.82 mm. Here, thresholds are chosen to simply identify the most egregiously suspect frames.

Figure 4. Examples of how motion scrubbing impacts rs-fcMRI data

(A,B) Seed correlation maps in unscrubbed and scrubbed data in two subjects using a medial parietal seed (−7 −55 27). Scrubbing removed 35% of the data in (A) and 39% of the data in (B), leaving 265 and 151 frames for analysis, respectively. The red ovals indicate locations where correlations are clearly altered. Seed maps produced from scrubbed data demonstrate much greater resemblance to the canonical default mode network. (C) Changes in the strength of rs-fcMRI correlations between medial parietal (−7 −55 27, red sphere) and medial prefrontal cortex (−7 50 −1, green sphere) across all 22 subjects in Cohort 1. Δr is produced by subtracting unscrubbed correlation values from scrubbed correlation values. Note that scrubbing increases this long-distance correlation in most subjects, does not substantially alter it in others, and reduces it in a small number of subjects.

Figure 5. Scrubbing high-motion frames from rs-fcMRI data decreases short-distance correlations and augments long-distance correlations

(A) Within the 22 subjects of Cohort 1, 264 ROIs were applied to scrubbed and unscrubbed data to produce two 264×264×22 correlation matrices. The unscrubbed matrix was subtracted from the scrubbed matrix and then averaged over subjects to produce a difference matrix (Δr). The values of this matrix are plotted as a function of the Euclidean distance between ROIs. Scrubbing high-motion frames from the data substantially decreases short-distance correlations and increases medium- to long-distance correlations. (B) To demonstrate that these effects arise from the removal of high-motion frames and not frame removal in general, the number of frames and the size of contiguous chunks of removed data were calculated for each subject, and identical sized chunks of data and numbers of frames were removed at random from each subject’s data. Difference matrices were calculated as in (A), and data are presented after (A). This process was repeated 10 times with similar results. Motion scrubbing has a much greater impact upon Δr values than random scrubbing (paired two-tail t-test: t = 305; p = 0). Linear fits are plotted over each dataset, demonstrating a relationship between distance and Δr when motion-targeted scrubbing is performed (r² = 0.18), but not when random scrubbing is performed (r² = 0.03).

Figure 6. The spatial distribution of scrubbing effects upon correlations

The top 0.5% (A) and 1% (B), and 2% (C) of Δr changes, as indexed by the absolute value of the change. Blue vectors represent correlations that decrease with scrubbing, and red vectors are correlations that increase with scrubbing. Most blue vectors are short- to medium-range, while most red vectors are medium- to long-range. The locations of the 264 ROIs are shown as small black spheres. Data are shown on a transparent PALS fiducial surface (Van Essen, 2005), and cerebellar ROIs are shown without a cerebellar surface.

Figure 7. Scrubbing produces direction- and distance-dependent changes in correlations

For Cohort 1, the top 1%, 3%, and 10% of changes in correlation are plotted as a function of the projection of each pairwise correlation vector onto the X, Y, and Z axes of Tailarach space. Here, X is the lateral axis, Y is the anterior-posterior axis, and Z is the vertical axis. In addition to the dependence of Δr upon distance, it appears that strongly lateral correlations tend to be mainly weakened by scrubbing, whereas more vertical or anterior-posterior correlations tend to be strengthened by scrubbing.

Figure 8. Similar effects of scrubbing are found in 3 additional cohorts

(A-C): Cohort 2, (D-F): Cohort 3, (G-I): Cohort 4. (A) The Δr values plotted against the Euclidean distance between ROIs. (B,C) Dorsal and lateral views of the top 1% of Δr changes, indexed by the absolute value of the change. Blue vectors represent correlations that decrease with scrubbing, and red vectors are correlations that increase with scrubbing. (D-F) and (G-I) are after (A-C) in different cohorts. The trend for scrubbing to augment long-distance correlations and to decrease short-distance correlations is present in each cohort (A,D,G), but the precise distribution of maximal change in correlation varies between cohorts. The magnitude of scrubbing effects reflects the average RMS movement of the cohort, and by extension the amount of data that is removed in the scrubbing process (see Table 1). An ANOVA comparing the magnitude of Δr across all cohorts was significant for a main effect of cohort (p = 0), and post-hoc two-sample two-tail t-tests demonstrated that 55.3% of child-adult, 8.5% of child-adolescent, and 2.1% of adolescent-adult comparisons of Δr were significant beyond p < 0.05, FDR corrected (children are Cohort 1 for these t-tests).

Figure 9. Motion-induced colored noise does not arise from standard regressions in functional connectivity processing

Each row presents analyses performed on Cohort 1 using functional connectivity data processing streams that differed only in the nuisance variables included in the multiple regression performed as the final step of functional connectivity processing. Each row plots the changes in correlation produced by motion scrubbing as a function of distance between ROIs, and also shows a histogram of these correlations. Nuisance variables in the top row included movement estimates at each frame (the 6 head realignment parameters and their temporal derivatives), as well as whole-brain signal and its derivative, white matter signal and its derivative, and ventricular signal and its derivative. Progressive analyses removed regressors (including derivatives) as indicated, and the bottom panels use no regression at all. Motion regression does not produce (or completely eliminate) the artifact in question, though it does produce a modest reduction in the changes in correlation produced by motion scrubbing.

Figure 10. The modular organization of brain-wide networks is fundamentally altered by head motion

Top and bottom rows depict data from children and adults, respectively. Left and right columns depict unscrubbed and scrubbed data, respectively. Each panel shows the sub-network (community) organization within the appropriate dataset of a network composed of the 264 ROIs studied in this report. Colors indicate sub-networks, and are independent in each panel, though congruent colors have been chosen across panels for ease of visual comparison. Only right hemispheres are shown, but results are generally symmetric across hemispheres. The numbers between panels indicate the normalized mutual information of community assignments between panels, a standard measure of how similar two sets of community assignments are (values of 1 indicate identical assignments). Scrubbing adult data produced little change in community assignments (NMI = 0.94), whereas scrubbing child data produced substantial changes in community assignments (NMI = 0.69). Moreover, though network organization in children and adults was initially quite dissimilar (NMI = 0.56), it became more similar with scrubbing (NMI = 0.70). This increase in similarity is not observed with random scrubbing (NMI = 0.58 ± 0.01 over 10 repetitions of random scrubbing in children and adults). The reorganization of functional network in children can be seen in the large contiguous patches of color (e.g. orange) in unscrubbed data, which become parts of distributed communities in scrubbed data (see red circles for an example). All graphs are thresholded at 10% edge density (r > 0.16, 0.15, 0.15, and 0.15 clockwise from upper left). For ease of visualization, nodes in communities with fewer than 4 members are colored white, and thus white nodes are explicitly not a single community.