The global signal and observed anticorrelated resting state brain networks

Abstract

Resting state studies of spontaneous fluctuations in the functional MRI (fMRI) blood oxygen level dependent (BOLD) signal have shown great promise in mapping the brain's intrinsic, large-scale functional architecture. An important data preprocessing step used to enhance the quality of these observations has been removal of spontaneous BOLD fluctuations common to the whole brain (the so-called global signal). One reproducible consequence of global signal removal has been the finding that spontaneous BOLD fluctuations in the default mode network and an extended dorsal attention system are consistently anticorrelated, a relationship that these two systems exhibit during task performance. The dependence of these resting-state anticorrelations on global signal removal has raised important questions regarding the nature of the global signal, the validity of global signal removal, and the appropriate interpretation of observed anticorrelated brain networks. In this study, we investigate several properties of the global signal and find that it is, indeed, global, not residing preferentially in systems exhibiting anticorrelations. We detail the influence of global signal removal on resting state correlation maps both mathematically and empirically, showing an enhancement in detection of system-specific correlations and improvement in the correspondence between resting-state correlations and anatomy. Finally, we show that several characteristics of anticorrelated networks including their spatial distribution, cross-subject consistency, presence with modified whole brain masks, and existence before global regression are not attributable to global signal removal and therefore suggest a biological basis.

FIG. 1.

Anticorrelated brain networks are replicable across datasets and statistical technique. A: anticorrelated brain networks reproduced from the dataset of Fox et al. (2005) using fixed effects analysis showed correlations within a system and negative correlations between systems. B: Z-score map from the current independent dataset shows voxels significantly correlated with a seed in the task-positive network (area MT+) using random effects analysis. C: Z-score map from the current dataset shows voxels significantly correlated with a seed in the task negative network (posterior cingulate/precuneus) using random effects analysis.

FIG. 2.

The impact of preprocessing and global regression on seed-based correlation maps. Z-score maps show voxels significantly correlated with various seed regions at 3 processing stages: no regression (left), movement, ventricle, and white matter regression (middle), and global regression (right). Histograms of voxel intensities for the 3 processing stages are shown to the right using blue (no regression), green (movement, vent and white matter), and red (global regression) lines. The location of each seed region is shown on the far left and include the posterior cingulate cortex/precuneus (Pcc), area MT+ (MT), the somatomotor cortex (MC), and primary visual cortex (V1). Talairach slice coordinates for Z-score maps: z = 45 (Pcc); z = 36 (MT); z = 54 (MC); z = −6 (V1).

FIG. 3.

The spatial distribution of the global signal. A: Z-score map showing all voxels significantly correlated with the whole brain (WB) or global signal. B: Z-score maps showing voxels significantly more correlated with the global signal than the average voxel.

FIG. 4.

Global signal regression shows fine neuroanatomical specificity not seen without global signal correction regardless of thresholding. Z-score maps (random effects across 17 subjects) showing voxels significantly correlated with seeds in (A) the primary visual cortex (top), (B) prefrontal cortex (middle), and (C) temporal cortex (bottom) at 3 stages of processing: no regression (left), movement, ventricle, and white matter regressed (middle), and global signal regressed (right). Raising the threshold of the Z-score map after movement, ventricle, and white matter regression does not show the neuroanatomical specificity achievable with global regression such as the correlation between the visual cortex (V1) and the lateral geniculate nucleus (LGN). For the bottom 2 rows, images were masked to focus on the thalamus and cortical seed regions were generated per (Zhang et al. 2008). Transverse slices: z = −6, Coronal slices: y = −27.

FIG. 5.

Global signal regression shows a unique distribution of negative correlations compared with post hoc distribution centering. Z-score maps showing voxels significantly correlated with seeds in the posterior cingulate (top) and area MT+ (bottom) after global signal regression (left) and post hoc distribution centering (right). Histograms show the distribution of voxel values across the entire brain (blue = whole brain regression; green = post hoc distribution centering images). Although both techniques center the distribution of correlations around 0, only global regression shows neuroanatomically specific negative correlations. Seed region locations are as shown in Fig. 1.

FIG. 6.

Global signal regression mandates negative correlations at the single subject level but not at the population level. A: single-subject regression coefficients (beta maps) for seeds in the posterior cingulate (left, blue line) and in the white matter (right, green line) for a representative subject. B: random effects Z-score maps show voxels significantly correlated with seeds in the posterior cingulate (left) and white matter (right) across the population of 17 subjects. The sum of voxel values across the entire brain is shown below each image and voxel histograms are shown to the right. Although the voxelwise sum of beta maps must be 0 for each subject and histograms similar, these measures can vary greatly in the population level Z-score maps depending on the consistency across subjects.

FIG. 7.

Anticorrelated networks persist despite modified whole brain masks that eliminate the mathematical constraints imposed by global regression. Modified whole brain masks (left) were created by removing 0, 30, 70, 80, or 95% of the voxels most correlated or anticorrelated with area MT. The intensity of the shown whole brain mask values reflects the absolute value of correlations with MT+. These masks were used to regress out a modified global signal, and Z-score correlation maps for a seed in MT+ were generated using the residual (right). The systems of interest remain anticorrelated even after being excluded from the whole brain mask, suggesting that the relationship is not an artifact of global regression. The MT+ seed region is as shown in Fig. 1.