Figure 1.
Voxel-wise characteristics of head motion during an fMRI scan.
Examples of different patterns of voxel-wise displacement within the time frames of one illustrative subject (A) and temporally averaged voxel-wise displacement of six illustrative subjects (B).
Figure 2.
Short names of atlas regions in the glass brain plot used to demonstrate results. Full names and additional information about regions can be seen in Table 1. Red spheres imply the axial projection of the center of mass of brain regions. Note that this plot does not indicate the axial depth of the regions.
Table 1.
Brain atlas regions, left hemisphere.
Figure 3.
Histogram of group averaged voxel-wise displacement correlations based on 5000 random permutations.
Temporally averaged standard-space voxel-wise displacement maps were averaged across the subjects of two randomly assigned groups. Spatial Pearson correlation was calculated between these group-mean voxel-wise displacement maps ). The histogram of this inter-group voxel-wise displacement correlation was computed based on 5000 random group formulation. Group-pairs with extrem inter-group differences in voxel-wise displacement were chosen for further analysis.
Table 2.
Voxel-vise displacement-dependent groups.
Figure 4.
Regional motion-BOLD relationship scales with the global motion-BOLD relationship.
Within-subject correlation of and
(horizontal axis) plotted against the correlation of FD and DVARS (vertical axis) for all brain regions and time frames of N = 183 subjects. The regional effect of motion on BOLD seems to scale with the global relationship which is correlated (
(p<0.000001) with the mean FD.
Figure 5.
Filled contour plots visualizing the Regional Displacement Interaction (RDI) effect: how the predicted connectivity strength (color-coded) changes depending on the simultaneously varying values of and
, in case of no nuisance regression (A∶NOREG) and all investigated first-level nuisance regression methods, i.e., NOREG+M6 (E), WMSCF (B), COMPCORR (C), GSREG (D), WMSCF+M6 (F), COMPCORR+M6 (G), GSREG+M6 (H), and SAT36 (I).
Vertical and horizontal axes of plots B-I are the same as those of plot A. Gray bars next to the legends indicate the (−1,1) interval to ease interpretation of color-coded Z-score values.
Figure 6.
RDI effect in case of a demonstrative connection: occipital fusiform gyrus (A = 20) - prefrontal gyrus (B = 12).
On the left, the partial residuals for from the model defined by Eq. (8) (dependent variable: connectivity data of the healthy control population (N = 105), without nuisance signal regression (NOREG)) are plotted against
. Although the model reveals no significant relationship (t = −1.46,p = 0.15) between connectivity strength and
, the
interaction effect is significant (t = 3.31, p = 0.0013), implying that the effect of
on connectivity strength differs depending on the value of
. This is demonstrated by dividing the data into four groups based on the value of
(color-coded on the left plot) and visualizing the corresponding cross-sectional CCPR (component and component-plus-residual) plots (on the left). Mean value of
corresponding to the cross-section is indicated. Partial residuals are plotted with colored dots corresponding to the cross-section group. The corresponding regression line estimated from the full model fit and the corresponding 95% confidence interval is displayed in black and gray, respectively. The horizontal and vertical axes of the cross-sectional CCPR plots are the same as those of the partial residual plot on the left. Cross-sectional CCPR plots imply that in each group the relationship between
and partial residual connectivity strength is significant but the regression lines have different slopes, which makes this latent relationship not observable without accounting for the interaction of the regional displacement covariates.
Figure 7.
Network pattern of connections where utilizing RDI significantly improves second-level modeling.
Statistical parametric networks presenting the model comparison performed by F-tests between the STD and STD+RDI models (Eq. (9) and (8)). Connections are only visualized, when the STD+RDI model explains significantly more variance than the STD model (the null hypothesis of the F-test can be rejected) with a false discovery rate of q<0.05. The proposed STD+RDI method proves to be most efficient with the nuisance signal regression methods NOREG, WMCSF, WMCSF+M6, GSREG, and COMPCOR+M6, and seems to demonstrate no significant improvement in case of SAT36, after correction for multiple comparison.
Figure 8.
The effect of RDI on motion-related group differences.
The number of significantly (p<0.01) differing connections is plotted against the spatial correlation coefficient of group-mean voxel-wise displacement maps for RD-based group-pairs for each nuisance signal regression method. The number of significant group differences is plotted on logarithmic axis. Improvement in the reduction of motion-related group differences was most pronounced for the NOREG, WMCSF and COMPCOR methods. Nuisance regression methods incorporating GSREG seem not to be sensible for differences in group-mean voxel-wise displacement patterns however, they show relatively high number of group differences by all group comparisons.
Table 3.
Motion-related group differences.
Figure 9.
Autism related group comparisons.
Autism related group differences for the investigated correction strategies. Colors denote significance levels as detailed in the legend.
Table 4.
Autism-related group differences (without nuisance signal regression of six motion parameters).
Table 5.
Autism-related group differences (with nuisance signal regression of 6 motion parameters).