Thirty adult subjects (16 male, age 18–31) participated in resting-state scans, and 20 subjects (9 male, age 19–33) participated in the disgusting/neutral image experiment, 4 of whom had also received resting runs. Subjects had no history of neurological or psychiatric impairment. All subjects gave written informed consent, and the protocol was approved by the Human Investigation Committee of the Yale University School of Medicine.
Brain images were acquired on a 3-T Siemens Magnetom Trio scanner. High-resolution T1-weighted structural images of the whole brain were acquired with a 3D MPRAGE sequence. Among the 30 subjects with resting-state scans, 21 subjects were given one pulse sequence (time repetition [TR] = 1900 ms; time echo [TE] = 2.96 ms; flip angle α = 9°; field of view [FOV] = 256 mm; matrix = 2562; slice thickness = 1 mm; 160 slices; number of excitations [NEX] = 1), while the remaining 9 subjects received a slightly different sequence (TR = 1230 ms; TE = 1.73 ms; flip angle α = 9°; FOV = 250 mm; matrix = 2502; slice thickness = 1 mm; 176 slices; NEX = 3). All 20 subjects performing the emotional task received the former pulse sequence. Because anatomical images were used only for coregistration and normalization, this difference did not influence our analyses of the functional data. For all subjects, functional images were acquired using a gradient-recalled echo planar pulse sequence (TR = 2000 ms; TE = 25 ms; flip angle α = 60°; FOV = 220 mm; matrix = 642; slice thickness = 4 mm; 34 slices).
During resting runs, subjects were presented a screen with a gray background and a black crosshair in the center; they were instructed to keep their eyes open, to remain awake, and to stay as still as possible. Resting scans lasted 6:40 or 200 volumes, within a single run.
To elicit responses to disgusting stimuli, we used an emotion regulation task, analogous to the design of Ochsner et al. (2002)
but with negative images chosen specifically to evoke disgust. In this task, subjects viewed disgusting or neutral images, taken from the International Affective Picture System (Lang et al. 2008
); disgusting pictures included moldy food, humans with unsightly skin conditions, rotten teeth, and so on. The relative number of faces and objects in neutral and disgusting conditions were approximately balanced. While viewing disgusting images, subjects were asked either to passively view the images or to increase or decrease their emotional response to the image using a specified strategy. In the present study, we only investigated evoked responses to passively viewed disgusting and neutral images; the influence of emotion regulation will be presented elsewhere. On each trial, subjects were first presented with instructions for 6 s (either “Look,” “Increase,” or “Decrease,” followed by a specific strategy for increase/decrease trials); the image for 4 s; an affect ranking screen for 6 s, in which subjects were asked to report their emotional response on a Likert scale from neutral (1) to disgusted (5) using a trackball mouse; and a screen reading “Relax” for 2 s. White text and images were presented on a black background. Between trials, subjects viewed a fixation cross for 2–6 s. Each subject received one 407-volume run, with 9 trials of each of 4 conditions (look-neutral, look-disgusting, increase, and decrease), as well as 12 s of fixation at the beginning of the experiment and 10 s at the end. Trials were presented in pseudorandom order, counterbalanced across subjects.
Data were processed using BrainVoyager QX 2.0 (Brain Innovation), along with in-house MATLAB scripts. The first 3 volumes from each resting data set and first 6 volumes from each task data set were removed to allow longitudinal magnetization to reach steady state. Functional data sets were then preprocessed; steps included rigid-body motion correction, slice scan timing correction, linear trend removal, and high pass filtering (3 cycles per series cutoff). Task data were also spatially smoothed using 4 mm–full-width at half-maximum (FWHM) Gaussian kernel. We did not spatially smooth resting-state data to avoid blurring differences in connectivity patterns between neighboring voxels. Instead, resting data were temporally smoothed with a 2.8 s-FWHM Gaussian kernel.
To reduce the influence of variation unrelated to neural activity on connectivity analyses, 9 nuisance variables were removed from resting-state data via linear regression. These included the global mean signal, a time course from white matter, a time course from the left lateral ventricle, and 6 motion parameters. White matter signal was taken from a 3 mm cubic ROI around the Talairach coordinate (−26, −13, 31), and ventricular signal was taken from a 3 mm cube around (−19, −35, 15); we verified that these coordinates fell in white matter and ventricle, respectively, in each subject's normalized anatomical image. We also applied a custom script to remove any pairs of consecutive volumes with an estimated 0.5 mm of translation in any direction or 0.5° of rotation about any axis between them to diminish potential effects of rapid motion occurring within the TR, which in some cases cannot be corrected with rigid-body motion correction. However, no volumes met these criteria.
Functional data were coregistered to high-resolution anatomical images, which were in turn normalized to Talairach space (Talairach and Tournoux 1988
). Normalization was performed in 2 steps: images were first aligned with stereotactic axes and then transformed to the Talairach grid using a piecewise affine transformation based on manual identification of the anterior and posterior commissure and the edges of cortex along each axis. Subsequent analyses were performed on preprocessed data in a space of 3-mm resolution aligned with Talairach space.
The left and right insula were defined anatomically by drawing insular gray matter on the Montreal Neurological Institute (MNI) 152 standard brain. The limits of the insula were taken to be the anterior, superior, and inferior periinsular sulci (Türe et al. 1999
; Naidich et al. 2004
). These ROIs were converted to Talairach space by normalizing the MNI brain in the same way that individual subject anatomical images were normalized. The resulting left insula ROI fell within a box bounded by the planes x
= −23 and −43, y
= −17 and 24, and z
= −12 and 20; the right insula ROI was contained within the reflection of this box about the x
-axis. Each voxel in the insular ROIs (converted to 3-mm resolution) was used as a seed in a whole-brain functional connectivity analysis: Their resting-state time series were normalized and used as regressors in general linear model (GLM)–based analyses for each subject. Resulting beta maps were averaged across subjects and treated as cross-subject connectivity maps for a given seed region.
We then applied k
-means clustering to these subject-averaged beta maps, treated as vectors, using squared Euclidean distance as the distance measure. Clustering was performed separately for the left and right insula. The k
-means algorithm was repeated 100 times, and the solution that minimized within-cluster variance was chosen to avoid the influence of random initial cluster membership on our results. We chose to implement the basic k
-means algorithm rather than more sophisticated graph-theoretic or spectral techniques (e.g., Meilă and Shi 2000
; Shi and Malik 2000
; Ng et al. 2002
) because the latter are intended for sparse graphs. Insofar, as voxels in the insula have time courses and connectivity maps that are correlated with those of many other voxels within the ROI, this system has a dense distance matrix.
We specified that the analysis find k
= 3 clusters; solutions for k
= 2 and 4 are presented in the Supplementary Materials
, available online. The choice of 3 clusters was initially based on exploratory connectivity analyses with seeds placed in various positions around the insula, which suggested that there were 3 distinct patterns of large-scale connectivity. The primary justification for this choice of k
is that when cluster analyses with higher k
were performed on our data set, each of the connectivity maps associated with the resulting clusters typically corresponded closely with one of the 3 maps found using k
= 3. For an illustration of this, see Supplementary Figure S2
, with connectivity maps associated with the k
= 4 solution.
We then performed connectivity analyses using the clusters as seeds. Whole-brain voxel-wise regression analyses were performed for each subject, using BOLD signal averaged over each cluster as regressors and combined across subjects in a random-effects analysis. Resulting t-maps were thresholded by controlling the false discovery rate at q
< 0.05 using the Simes procedure to correct for multiple comparisons (Genovese et al. 2002
). For display, these maps were overlayed on an inflated cortical surface using CARET surface mapping software (http://brainmap.wustl.edu/caret
; Van Essen et al. 2001
) and the PALS cortical atlas (Van Essen 2005
ROI-based analyses were performed to closely examine connectivity between cingulate and insular cortices and to statistically assess regional differences in connectivity strengths. Cingulate ROIs were defined as 9 mm spheres surrounding 3 coordinates placed along the middle to anterior cingulate, identified anatomically using the MNI 152 brain. The coordinates were each separated by 28 mm along the y-axis; their position along the z-axis was chosen such that the spheres fell just below the cingulate sulcus. Coordinates were placed at middle cingulate cortex (MCC, 0, −10, 41), dACC (0, 14, 35), and pACC(0, 38, 17). We assessed connectivity between the 3 cingulate ROIs and 3 insula subregions identified with cluster analysis, by regressing signal from cingulate regions on signal from insula regions (with means removed from all time series). Betas from each regression were statistically assessed with one-sample 2-tailed t-tests. Functional connections of a given cingulate region to different insular regions were statistically compared, using paired 2-sample 2-tailed t-tests. Effects of laterality on connectivity strengths are not investigated in the present study; such differences, if present, could reflect either true differences in connectivity or subtle differences in the extent of insula clusters on the left and right.
Whole-brain GLM-based analyses were also performed on task data. The model contained 7 regressors, modeling responses to the instruction period, affect rating period, and relax period, as well as the image presentation period for each of 4 conditions. Regressors were defined as boxcars peaking during each period, convolved with a double gamma hemodynamic response function. Beta values for regressors from look-neutral and look-disgusting conditions were averaged across insula subregions defined by cluster analysis. Betas for each region and each of 2 conditions were statistically assessed with one-sample 2-tailed t-tests; betas within each region were also compared between conditions (disgusting vs. neutral) using paired 2-sample 2-tailed t-tests.