Resting state connectivity of the bed nucleus of the stria terminalis at ultra‐high field

Abstract

The bed nucleus of the stria terminalis (BNST), a portion of the “extended amygdala,” is implicated in the pathophysiology of anxiety and addiction disorders. Its small size and connection to other small regions prevents standard imaging techniques from easily capturing it and its connectivity with confidence. Seed‐based resting state functional connectivity is an established method for mapping functional connections across the brain from a region of interest. We, therefore, mapped the BNST resting state network with high spatial resolution using 7 Tesla fMRI, demonstrating the in vivo reproduction of many human BNST connections previously described only in animal research. We identify strong BNST functional connectivity in amygdala, hippocampus and thalamic subregions, caudate, periaqueductal gray, hypothalamus, and cortical areas such as the medial PFC and precuneus. This work, which demonstrates the power of ultra‐high field for mapping functional connections in the human, is an important step toward elucidating cortical and subcortical regions and subregions of the BNST network. Hum Brain Mapp 36:4076–4088, 2015. © 2015 Wiley Periodicals, Inc.

INTRODUCTION

The bed nucleus of the stria terminalis (BNST) is a small structure in the basal forebrain whose lateral division is highly implicated in the autonomic arousal and sustained nature of anxiety [Davis et al., 2010; Johansen, 2013; Walker et al., 2003] and complex motivational responses, including the withdrawal and craving stages of addiction disorders [Jennings et al., 2014; Koob and Volkow, 2009; Silberman and Winder, 2013]. It is a key component of the extended amygdala [Heimer et al., 2007], sharing embryological origin [Johnston, 1923], chemoarchitecture [Lesur et al., 1989; Walter et al., 1991], and connections with the central nucleus of the amygdala (CeA) [Fudge et al., 2012; Sakamoto et al., 1999]. Although the BNST and many of its anatomical connections are evolutionarily conserved, its exact homology with humans remains under investigation [Fox et al., 2015], prompting recent efforts at delineating pathways in nonhuman primates [deCampo and Fudge, 2013; Fudge and Haber, 2001]. Figure 1 presents a summary of major findings of BNST anatomical connectivity in rodents and primates.

Figure 1.

BNST anatomical connectivity known from rodent and primate tract tracing studies. Box sizes and positioning are arbitrary. Nuc Acc, nucleus accumbens; PAG, periaqueductal gray; IL/CM, intralaminar/centromedial; VTA, ventral tegmental area; SNC/RRF, substantia nigra compacta/retrorubal field; N. Tractus Solitarius, nucleus of tractus solitarius; CeA, central amygdala nucleus; BLA, basolateral amygdala nucleus; SLEA, sublenticular extended amygdala; BNST, bed nucleus of the stria terminalis.

Given the BNST's role as a key node in networks relevant to anxiety and addiction, there is great need to understand the regions with which it communicates in humans. Connectivity estimates of endogenous, low‐frequency fluctuations during rest, using functional magnetic resonance imaging (fMRI), are an established method to map functional networks in the brain [Fox and Raichle, 2007]. Although researchers are still clarifying the mapping between functional and structural connectivity, there are nonetheless strong relationships between the two. The aim of the current study was to map human BNST functional connectivity with ultra‐high field (UHF) fMRI.

A prior study of BNST resting state functional connectivity in humans was published with promising results [Avery et al., 2013]. It identified structural and functional connectivity to a number of cortical and subcortical regions. In general, the functional findings aligned quite well with known anatomical connectivity from animal studies [Dong and Swanson, 2005; Dong et al., 2001; Heimer et al., 2007; Hsu and Price, 2009; Weller and Smith, 1982]. However, the findings did not have the spatial resolution to identify small but very critical regions implicated in anxiety and addiction such as the hypothalamus, periaqueductal gray, or subnuclei of thalamic structures also known to connect to the BNST [Dong et al., 2001; Hsu and Price, 2009; Weller and Smith, 1982]. Given the importance of mapping BNST connectivity to these deeper regions in the brain, and the importance of replication particularly in the affective sciences [Barch and Yarkoni, 2013], the present work aimed to replicate and extend this study by additionally investigating diencephalic and mesencelphalic structures using the high spatial resolution, reduced partial voluming [Newton et al., 2012] and increased Blood Oxygenation Level Dependent (BOLD) signal changes [van der Zwaag et al., 2009] that UHF fMRI offers.

A voxel‐wise, seed‐based resting state fMRI experiment was performed to map BNST functional connectivity at a level of spatial specificity that would be sensitive to many of the subnuclei of regions previously reported in tract tracing studies, as well as demonstrate connectivity with additional small regions not typically reported in fMRI research. More specifically, we predicted that the human BNST would be functionally connected with basal ganglia nodes of addiction circuitry, including nucleus accumbens, caudate, and ventral tegmental area (VTA), and to subcortical nodes of anxiety circuitry, including central amygdala and anterior hippocampal subregions, periaqueductal gray, and hypothalamus.

MATERIALS AND METHODS

Subjects

Twenty‐nine right‐handed, healthy volunteers were recruited from a mixed urban and suburban population through internet listservs, flyers, and print advertisements and were compensated for their time. Exclusion criteria included: (a) current or past Axis I psychiatric disorder as assessed by SCID‐I/NP [First et al., 2007], (b) first degree relative with a known psychotic disorder, (c) a medical condition conflicting with safety or design of the study, (d) brain abnormality on MRI as assessed by a radiologist, (e) positive toxicology screen, (f) MRI contraindication, or (g) excessive head motion during the functional scans. Excessive head motion was defined as more than 15% of a subject's TRs censored, where the criterion for censoring was a Euclidean norm motion derivative greater than 0.3mm for temporally adjacent time points. Two subjects were removed for this reason. Written informed consent was obtained from subjects, approved by the National Institute of Mental Health (NIMH) Combined Neuroscience Institutional Review Board.

Functional Image Acquisition

Scanning was performed on a 7T Siemens Magnetom MRI with a 32‐channel head coil. Immediately prior to both structural and functional image acquisition, third order shimming was implemented to correct for magnetic inhomogeneities [Pan et al., 2011]. The high‐resolution, 0.7 mm isotropic, T1‐weighted MPRAGE had the following parameters: TR = 2,200 ms, TE = 3.01, flip angle 7°, acquisition matrix = 320 × 320, acquisition time = 10 min. Parallel imaging was turned off to increase signal to noise ratio (SNR). The functional EPI had an interleaved TR of 2.5 s, TE = 27 ms, flip angle 70°, 49 slices, 1.3 mm isotropic voxels, acquisition matrix = 154 × 154, with 240 images collected over the 10 minute scan. A 10‐min session was selected because resting state scans longer than the standard 5–8 min have higher reliability of network identification [Birn et al., 2013; Smith et al., 2011]. To maximize spatial resolution a non‐whole‐brain field of view (FOV), which covered approximately 2/3 of the brain (Supporting Information Fig. 2A, inset), was collected. From superior to inferior, the FOV was manually angled to capture dorsal medial prefrontal cortex, BNST, amygdala, and hippocampus and to avoid eyeballs, whose movements can cause ventral artifacts. Subjects were provided earplugs and head pads to further suppress sound and movement, and a pillow was placed underneath their knees for stability. Participants were instructed to keep their eyes open and look at a white fixation cross on a black background.

Figure 2.

BNST masking procedure. A: Figure from the drawing protocol showing BNST context at one coronal slice: anterior commissure (AC), internal capsule (IC), caudate (C), lateral ventricle (LV) and septal nuclei (s). B: Hypointensity of thalamostriate vein used as caudate/BNST border for all subjects. C: Mask average of three raters for one subject; sagittal and axial views; color bar corresponds to degree of rater agreement at each voxel (red = all three classified that voxel as BNST). D: Same subject's mask was thresholded at 2 and binarized to serve as correlation seed. This was performed for each subject. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Physiological Measures

To correct the fMRI data for physiological signals of non‐interest, respiration was measured with a pneumatic belt placed around the diaphragm and cardiac rhythm with a pulse oximeter clipped around the right index finger. Such correction has been shown to greatly improve temporal signal to noise and BOLD sensitivity at high spatial resolution [Hutton et al., 2011]. All physiological data were sampled at 500 Hz using AcqKnowledge software connected via a BioPac MP150 system (http://www.biopac.com).

BNST Mask

The boundaries of the BNST mask were defined by authors ST and JF, with reference to nonhuman primate specimens examined at the microscope, and previous studies on human BNST anatomy [Heimer et al., 2007]. Subsequently, three authors (ST, KO, and AD) separately drew left and right BNSTs on the subjects' structural images in native space, bias‐corrected using the tool 3dUnifize. These same structural scans were also used in EPI processing. Each rater first practiced on three pilot images and then, blind to subject identity, drew BNSTs in a randomized order to spread residual learning effects across the subject population. The inter‐rater reliability metric Dice Similarity Index [Dice, 1945] and intraclass correlation coefficient (ICC) were calculated using mask volumes. In addition, correlations were calculated between average time series extracted between all raters' masks for each subject.

To determine the BNST mask location and dimensions the raters followed a protocol developed by ST and JF that demarcated the BNST within surrounding anatomy (Fig. 2A). The drawn BNST masks were consistently located in native space, encompassing the area between the following structures: the center of the anterior commissure ventrally, medial edge of the internal capsule laterally, nadir of lateral ventricle or septal nuclei or fornix medially (depending on rostral/caudal level), and caudate nucleus and thalamostriate vein dorsolaterally. The thalamostriate vein appeared as a hypointensity on the T1 scans (figure 2B), and the nadir of the lateral ventricle on coronal slices was used as a boundary line to exclude the medial BNST. This decision to exclude medial BNST was based on the notion that this BNST subregion is not involved in the same functions or networks as the lateral portion that is implicated in anxiety through its connection with the CeA. The drawing process, performed in AFNI [Cox, 1996], was also informed by verifying anatomical landmarks using a detailed atlas [Mai et al., 2007] and was reviewed by a trained anatomist (JF). On coronal slices, the caudal BNST boundary was defined at the point posterior to the AC where it disappeared between the internal capsule and lateral ventricle, and the rostral BNST boundary was defined at the point anterior to the anterior commissure (AC) where the globus pallidus emerged through successive anterior slices. Because the BNST, by definition, tapers off and merges with the sublenticular extended amygdala sweeping under the AC, the decussation of the AC was chosen as the ventral boundary. The three authors' drawn masks were averaged and then thresholded at 2/3 (0.67; majority wins for every voxel) to determine the masks that were used as seeds for each subject's connectivity estimates (Fig. 2C,D). This averaging and thresholding procedure was chosen to lend greater confidence to demarcations of the few perceptually ambiguous (i.e., caudal and rostral) boundaries. Finally, and before reslicing to EPI resolution, the volumes of the left and right BNSTs for each subject were normalized to their estimated total intracranial volume (eTIV), i.e., BNSTvol divided by eTIV, and these values were compared between sexes with two‐sample t‐tests.

Preprocessing

Tissue segmentation and eTIV were first determined for each individual in FreeSurfer version 5.3 [Fischl et al., 2002]. Resting state preprocessing and analyses were performed within AFNI using the ANATICOR processing stream [Jo et al., 2010], which avoids problems resulting from regressing out the global signal [Murphy et al., 2009; Saad et al., 2012]. Subjects' first three functional volumes were removed to allow for scanner equilibrium, and the remaining functional volumes were slice‐time and motion corrected, and coregistered to their MPRAGE structural scan using a Local Pearson Correlation algorithm optimized for EPI‐T1 alignment [Saad et al., 2009]. Subjects' anatomy was then non‐linearly normalized with 3dQwarp [Cox and Glen, 2013] to a skull‐stripped “ICBM 2009b Nonlinear Asymmetric” template in Montreal Neurological Institute (MNI) space. The resulting affine plus nonlinear parameters were then applied to the hand‐drawn BNSTs, cortical and subcortical FreeSurfer parcellations and segmentations, and the functional data. To check the performance of this normalization on the hand‐drawn masks, a group average of the normalized BNSTs was created (Supporting Information Fig. 1). Functional data was smoothed with a 2.6 mm FWHM Gaussian kernel within a mask of FOV voxels that had valid data at every TR.

A number of time series were then modeled as regressors of non‐interest, in a single regression, projecting them out of the data to leave cleaned, “residual” time series with which voxel‐wise correlations were performed. Regressors of no interest included: six head motion parameters and their derivatives; 0.01–0.1 Hz bandpass filter regressors; 13 slice‐based cardiac (RETROICOR), and respiration volume per unit time measures [Birn et al., 2008; Glover et al., 2000] calculated with the RetroTS.m script; and time series from one lateral and one third+fourth ventricle masks that both had BNSTs voxels removed and were then eroded by 1 voxel to avoid partial voluming with gray matter. Finally, the regressions were performed at each voxel using 3dLocalStat within a 13 mm radius sphere that took into account local white matter, which controls for white matter signal heterogeneity, and hardware‐ (i.e., coil‐) related artifacts [Jo et al., 2010]. These white matter masks were also constructed by first removing miscategorization of tissue‐type by FreeSurfer (i.e., subtracting BNSTs) and were then eroded. Note that incorporation of the resulting physiological calculations occurred between the despiking and slice‐timing correction steps, so that physiological effects at each interleaved slice would be accounted for [Jo et al., 2013].

The residual maps after this regression contained the BOLD signals of interest, and were used to extract a mean time series from the BNST masks. This time series was correlated across the rest of the brain using 3dTcorr1D within a binarized mask that represented the EPI coverage of 95% or more of the subjects. Two separate analyses for the left and right BNSTs were originally performed; however, the resulting connectivity maps were so similar (not shown) that the map of correlations to the average of both left and right BNSTs together is reported here. The resulting Pearson correlations were converted to Z scores and entered into 3dttest++. Finally, using 3dClustSim's Monte Carlo simulation, incorporating the data's intrinsic plus applied smoothing, the resulting maps were thresholded to P < 0.00001, k = 10 (P = 1 × 10⁻⁵) . This represents a global correction at P < 0.01.

Finally, a two‐sample t‐test (3dttest++) was used to examine sex differences in connectivity.

RESULTS

Subject Demographics, Subjective High‐Field Sensory Effects, and BNST Mask Reliability

Table 1 provides summary statistics for the subjects' demographics, subjective sensory effects at high field and inter‐rater and volume measurements for the drawn BNST masks. The final analysis consisted of 27 subjects (13 male, 14 female) aged 27.3 ± 6 years. The racial distribution of our pool was 37% Caucasian, 30% Asian, 15% African American, 11% mixed and 7% unknown. Most subjects (78%) experienced mild but acceptable dizziness upon entering or exiting the scanner bore, but this sensation faded within a couple minutes at isocenter, consistent with a previous report at UHF [Theysohn et al., 2007]. Although the final drawn BNST masks were satisfactory by visual inspection, the inter‐rater reliability based on separate rater volumes was only acceptable: Dice similarity indices, using author ST as reference, were on average 0.61 for both left and right BNSTs. However, the averaged correlations (with SD) between all averaged time series extracted from the raters' individual masks, before combination with the two‐thirds procedure, were r = 0.92 (0.07) for the right BNST and r = 0.90 (0.08) for the left BNST (all significant), showing a high degree of inter‐rater reliability on the BNST fMRI time series (Table 1).

Table 1.

Subject demographics, anxiety, dizziness, and mask drawing

Subject number and age (μ ±SD)	Sex (M,F)	STAI trait, STAI state	Average dizziness 1 (imperceptible) to 10 (severe)	Inter‐rater drawing reliability (based on mask volumes)	Correlations (SD) between time series from all raters' masks	average female vs. male BNST volumes, normalized to eTIV
N = 27 27.3 ± 6 years	13, 14	29.8 ± 8, 27.3 ± 6	Entering bore: 4.0 Static: 1.7 Exiting: 2.5	lBNST: DSIa = 0.61 ICC = 0.452rBNST: DSI = 0.61 ICC = 0.409	rBNST: r = 0.92 (0.07) P < 0.001rBNST: r = 0.90 (0.08) P < 0.001	T‐test results: L: p=0.90 R: p=0.97 L vs. R: p = 0.26

^aDice Similarity Index (DSI) calculated with one rater (ST) as reference; the value is the average of the other two. ICC = Intraclass correlation coefficient. L, left; R, right; eTIV, estimated total intracranial volume.

Finally, sex had no effect on left or right BNST volumes after head size (eTIV) normalization.

Neuroimaging

The pattern of BNST functional connectivity was very similar, but not identical, to the anatomical connectivity reported in lower mammals and primates. The findings are summarized in Table 2 and Figures 3 and 4A–G, and include significant subcortical and cortical clusters.

Table 2.

Spatial coordinates of highly significant clusters of BNST functional connectivity within a group mask of 95% probability

	# voxels	X	Y	Z	T statistic
Left hemisphere
BNST/caudate/nucleus accumbens	1681	−6.2	3.1	−1.2	25.6
Amygdala SF	216	−16.7	−6.1	−16.8	9.1
Anterior hippocampus CA	74	−20.6	−13.8	−18	9.7
Anterior hippocampus CA	27	−31	−10	−16.8	8
Posterior hippocampus CA	22	−23.2	−35.9	−6.3	7.8
Posterior hippocampus FD	17	−27.1	−32.1	−9	7.4
Posterior hippocampus CA	25	−32.2	−28.2	−12.9	7.2
Parahippocampus	25	−25.8	−42.4	−3.8	10.4
Putamen	12	−31	8.2	−2.5	8
Dorsomedial thalamus	53	−5	−15.2	7.9	7.7
Centromedian/parafascicular thalamus	83	−1.1	−12.6	2.8	8.1
Para‐periaqueductal gray°	84	−5	−34.7	−11.5	9.5
Suborbital sulcus (BA 10/11)	678	−2.4	47.2	−12.9	10.2
Lateral orbitofrontal (BA 47)	15	−38.8	29.1	−19.3	7.5
Subparietal sulcus/precuneus (BA 31)	132	−10.2	−54.2	26.2	11.7
Subparietal sulcus/precuneus (BA 31)	13	−11.5	−38.6	33.9	7.4
Middle occipital gyrus (BA 19)	22	−51.8	−77.6	0.2	7.5
Parahippocampal part of medial occipito‐temporal gyrus (BA 36)	20	−19.2	−38.6	−14.2	7.5
Anterior cingulate gyrus & sulcus (BA 32)	60	−1.1	45.9	14.4	7.9
Superior temporal sulcus (BA 21)	12	−57	−10	−10.2	6.4
Calcarine sulcus (BA 30/19)	12	−12.8	−47.7	1.4	7.1
Precuneus (BA 7/23)	12	−7.6	−60.7	11.8	6.6
Parieto‐occipital sulcus (BA 31)	12	−11.5	−59.3	19.7	6.5
Cingulate posterior ventral gyrus (BA 29)	11	−7.6	−47.7	10.5	6.4
Posterior insula (BA 13)	10	−37.5	−23	5.3	6.6
Right hemisphere
BNST/caudate/nucleus accumbens	1793	6.8	4.3	1.4	26.4
Amygdala SF	59	22.3	0.4	−14.2	10.44
Amygdala CeA	33	18.4	−3.4	−11.5	7.4
Amygdala BLA	20	23.6	−2.2	−19.3	9.2
Amygdalostriatal area°	28	31.4	−7.3	−15.5	7.2
Hippocampus CA	65	19.7	−15.2	−14.2	8.1
Hippocampus CA	12	30.1	−37.2	−6.3	7.8
Hippocampus CA	10	34	−17.8	−12.9	7.5
Hippocampus FD	14	30.1	−25.6	−12.9	7.6
Hippocampus SUB	12	24.9	−29.4	−18	7.1
Periaqueductal gray	16	1.5	−32.1	−6.3	7.2
Periaqueductal gray	17	2.8	−33.3	−11.5	7.4
Ventroanterior thalamus	19	0.2	−7.3	−1.2	6.8
Dorsomedial thalamus	14	2.8	−16.5	9.2	6.1
Dorsomedial thalamus	10	6.8	−23	7.9	6.5
Dorsal nucleus of raphe	17	0.2	−28.2	−20.7	6.6
Habenula	15	5.4	−26.8	−1.2	6.6
Precuneus (BA 7)	1011	1.5	−63.2	18.3	9.8
Orbital gyrus (BA 47)	18	36.6	34.2	−18	6.5
Gyrus rectus (BA 11)	15	0.2	43.3	−23.2	6.3
Superior frontal gyrus (BA 10)	24	2.8	55.1	15.8	6.9
Superior frontal gyrus (BA 10)	33	2.8	61.6	27.4	7.2
Superior frontal sulcus/dlPFC (BA 9)	21	23.6	38.1	39.1	7.3
Superior temporal sulcus (BA 21)	12	56.1	−10	−9	7.6
Calcerine sulcus (BA 30)	25	14.5	−47.7	5.3	6.9

Peak clusters are organized first by hemisphere and then whether they are subcortical or cortical (separated by horizontal line). Maximum probability labeling for hippocampal subregions from [Amunts et al., 2005], thalamus subnuclei and cortical gyri and sulci from additional maps [Destrieux et al., 2010; Metzger et al., 2010]. Brodmann's area (BA) labels (≤ 3 mm from peak) are from the Talairach Daemon for general reference. Amygdala subnuclei and other subcortical labels without standard atlas reference were visually estimated with the assistance of JF, CM and JO (see acknowledgements). ° reflect uncertainty in labeling. Clusters are corrected at P < 0.00001, k = 10. Coordinates are in MNI space.

Figure 3.

“Whole brain” results in axial, overlaid on averaged group anatomy nonlinearly normalized to MNI space. P < 0.00001, k = 10. Left is left. This thresholding and color scale was used for Figures 4 and 5 as well. Note in the structural image the hyper‐intense arteries seen on the average of our T1 scans. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Figure 4.

A–G: BNST functional connectivity with subcortical regions, magnified views. A: Crosshairs at region consistent with anterior hippocampal CA1' (sagittal slice). B: Posterior hippocampal CA (sagittal). C: Habenula (coronal and axial). D: centromedian/parafascicular thalamus. Also visible are dorsomedial thalamus clusters (coronal). E. Periaqueductal gray (axial and sagittal). F: Lateral hypothalamus with crosshairs at [−5 −8 −9] (sagittal and axial). G: Nucleus accumbens and sublenticular extended amygdala (SLEA) on equally‐spaced anterior‐to‐posterior (left‐to‐right) coronal slices, right panels show connectivity to CeA amygdala. Thresholding and color scale the same as in Figure 3, although cluster size was changed to five voxels at this magnification. Bottom right corner coordinates are rounded. Anatomical underlay is a high resolution MNI template (0.5 mm) downsampled to the resolution of our acquired structural scans. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

The caudate nucleus showed the highest concentration of functional connectivity with the BNST (Fig. 3), as it was the most prominent portion of the cluster of connectivity that spread from the BNST in anterior and ventrolateral directions. Anteriorly, this cluster covered the head of the caudate and extended into the body, with particular specificity along the medial edges. Ventrally, it extended into the hypothalamus (Fig. 4F). The tail of the caudate showed no BNST functional connectivity.

Figure 4 magnifies sections of the brain‐level map to show a number of other connected regions and subregions that have not yet been reported in humans. Strong connectivity is observed in Figure 4A in an anterior hippocampal CA region [Amunts et al., 2005], (consistent with the uncal CA1') [Fudge et al., 2012; Saunders et al., 1988], and in clusters along the hippocampal rostrocaudal axis (figure 4B). In figure 4C we see connectivity with the right habenula. In the thalamus, activity over a region consistent with the centromedian nucleus is highlighted at the crosshairs in figure 4D, with adjacent clusters in the dorsomedial nucleus of the thalamus [Metzger et al., 2010]. The left side of Figure 4E reveals a cluster in the ventral periaqueductal gray (PAG) and the right side a cluster in the dorsal PAG. Figure 4F shows sagittal and axial slices of the BNST‐centered cluster extending into the hypothalamus. Additional sagittal slices of some of these findings can be seen in Supporting Information Figure 3C–D.

Figure 4G illustrates BNST functional connectivity with the sublenticular extended amygdala (SLEA), which can be seen sweeping ventrolaterally below the internal capsule from the BNST seeds, through the nucleus accumbens and ultimately connecting with the dorsal amygdala subnuclei along the ventroamygdalofugal path (Figure 4G, far right) including the CeA and the superficial nucleus [Amunts et al., 2005].

Anterior and posterior cortical regions also showed significant functional connectivity with the BNST. Connectivity is shown anteriorly in the anterior cingulate cortex (BA32; Fig. 5A), posteriorly with the precuneus (consistent with BA7 and BA31; Fig. 5B), and also anteriorly with the orbital frontal cortex (BA47 and BA11), superior and anterior prefrontal cortices (BA9, BA10), and right superior frontal sulcus (Supporting Information Fig. 3A,B; Table 2).

Figure 5.

A: Medial prefrontal cortical BNST functional connectivity. B: Connectivity with posterior cingulate (precuneus) cortex. Anatomical underlay is also the high resolution MNI template. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

Finally, a between‐group, two‐sample t‐test showed no significant sex difference on BNST functional connectivity across the brain (results not shown).

DISCUSSION

This is the first ultra‐high field, high‐resolution fMRI study to examine the functional connectivity of the BNST, a structure that plays a key role in anxiety and addiction. Overall, the findings replicate much of the results of the recently published study in humans using resting state fMRI [Avery et al., 2013], showing BNST functional connectivity with basal ganglia structures (accumbens, caudate, putamen and pallidum), amygdala, thalamus, hippocampus and medial PFC. Importantly, however, the present findings also provide much greater specificity with regard to subregions of these structures, and extend the findings to very small regions such as the SLEA, PAG, hippocampal subregions, the habenula, and hypothalamus, which cannot be reliably identified with conventional fMRI. The following discussion focuses on six new findings: the first five subcortical and the sixth cortical.

First, the BNST functional connectivity to the head and body, but not the tail, of the caudate nucleus replicates findings in Avery et al. (2013). Such connectivity is interesting in light of the role of the caudate head and body in human cognition and emotion [Robinson et al., 2012]. However, we extend these findings by showing that this connectivity is highest in the caudate's medial portion, which corroborates a recent histological and DTI‐based report of human BNST structural connectivity [Kruger et al., 2015], as well as primate work demonstrating that the medial division preferentially processes emotional information [Haber, 2003].

Observing functional connectivity throughout the sublenticular extended amygdala is the second important finding. Heimer et al. [2007], who advanced the ‘extended amygdala’ concept, have shown that the SLEA in humans consists of interdigitated cells in a path along the route shown in Figure 4G [deCampo and Fudge, 2013; Sakamoto et al., 1999]. This connecting path has also been shown, albeit with less spatial specificity, by Kalin and colleagues, who tested non‐human primate and human resting functional connectivity by seeding the central nucleus of the amygdala [Birn et al., 2014; Oler et al., 2012]. The CeA (and SLEA), as well as the corticoamygdala transition region and basal nucleus (parvicellular subdivision) provide major inputs to the BNST [deCampo and Fudge, 2013], consistent with the present results.

Third, while strong connectivity with the hippocampus [Avery et al., 2013] is replicated, it is localized more specifically to the anterior CA region rather than its posterior segment (Fig. 4A,B) [Amunts et al., 2005]. This pattern is consistent with the involvement of the anterior region in stress responses and motivation processing, in contrast to the posterior hippocampus which is preferentially involved in the learning and memory of spatial navigation [Fanselow and Dong, 2010]. This finding is also consistent with rodent work reporting that the ventral hippocampus (the homologue of the human anterior hippocampus) has, along with the dorsal amygdala, strong efferent projections to the BNST [Lee and Davis, 1997].

Fourth, the BNST is functionally connected with the PAG, an area that has multiple roles in emotion, pain and stress‐ and defense‐related behaviors [Bandler and Shipley, 1994; Panksepp et al., 2011]. More specifically, connectivity is observed to both the ventral and dorsal regions of the PAG. It is important to note that both in vitro animal and in vivo human UHF work has shown that the PAG does not operate as a unitary functional node [De Oca et al., 1998; Satpute et al., 2013]. In rodents, for example, the ventral PAG is thought to be involved in freezing responses, while the dorsal PAG is preferentially involved in pain‐induced increases in vocalization [McLemore et al., 1999]. Therefore our demonstration that the PAG is differentially connected to the BNST could be useful for future, human in vivo studies.

Fifth, a region consistent with the lateral hypothalamus, a major BNST output structure [Kim et al., 2014] and one that is rarely reported in human neuroimaging [Baroncini et al., 2012], is shown as part of the BNST network. Our spatial resolution, however, precludes further identification of other hypothalamic subregions, such as the periventricular or dorsomedial divisions, which are also known to be anatomically connected to the BNST [Dong and Swanson, 2005] and to play a crucial role in downstream activation of the HPA axis through the release of CRH and AVP hormones [Ulrich‐Lai and Herman, 2009].

The PAG, hypothalamus and habenula (Table 2) are all small structures for which no standardized masks are available. To palliate this problem, we catalogued standard space coordinates for these structures that have been reported from a number of studies. We then averaged these coordinates and calculated their Euclidean distance to the coordinates of the present study (Supporting Information Table 1). We expect that subcortical masks derived from UHF imaging may soon eliminate the necessity for this type of check.

The cortical findings in the medial prefrontal cortex across the BA 10/BA32 border ([Bludau et al., 2014]; Figure 5A) mirror clusters functionally connected to the BNST in reports by others [Avery et al., 2013; McMenamin et al., 2014; Motzkin et al., 2015] as well as DTI‐based structural findings in humans [Kruger et al., 2015]. Elsewhere in the prefrontal cortex (see Table 2), additional clusters are observed in the medial orbitofrontal cortex, which modulates the BNST threat response [Fox et al., 2010], and in the left lateral orbitofrontal cortex and the right superior frontal sulcus, which are novel human findings reported here (Supporting Information Fig. 3A,B). These prefrontal connections may support the interface between higher cognitive functions and subcortical physiological effectors.

Finally, figure 5B reveals a strongly connected cluster in the precuneus, which appears to replicate the one reported by Avery and colleagues in their voxel‐level analysis [Avery et al., 2013], cf. Fig. 5, row 2, column 5). Similar to prefrontal cortical activity patterns, this finding suggests that the BNST might be connected to the default mode network (DMN), of which the precuneus is a key node [Fox and Raichle, 2007]. Alternatively, the rostral BA 32/precuneal region has also been shown to serve as a general hub node for broader connectivity across the brain [Tomasi and Volkow, 2010 ; van den Heuvel and Sporns, 2013].

It is interesting to note that despite strong evidence of the involvement of the anterior insula in anxiety [Alvarez et al., 2011; Etkin and Wager, 2007; Paulus and Stein, 2006] or addiction [Koob and Volkow, 2009], we did not find the insula to be functionally connected to the BNST. Several explanations of this negative result can be proposed. First, the (anterior) insula may be involved in a different network than that of the BNST. Second, the functional connectivity of the BNST with the anterior insula may be too weak or nonlinearly coupled to be detected in a correlation‐based resting state study. Third, the relationship of insula with the BNST may only come online during the anxious state itself.

A diagram summarizing our BNST functional connectivity results is presented in Figure 6. The background lines in solid black represent the major anatomical BNST afferents and efferents identified by primate and rodent tract tracing studies [deCampo and Fudge, 2013; Dong and Swanson, 2005; Dong et al., 2001; Fudge et al., 2012; Hsu and Price, 2009], and the foregrounded colored lines summarize the present findings. This diagram suggests that a study adapted to the benefits of UHF has the strength to detect a large percent of known connections with the BNST, supporting a high level of evolutionary conservation of this connectivity across species. The diagram also shows (dashed lines) that the BNST is functionally connected to cortical regions that have not been described in animal research, suggesting recent phylogenetically developed functional connections.

Figure 6.

Main resting state findings (colored glow) overlaid on select BNST anatomical connectivity known from rodent and primate tract tracing studies. Dotted lines represent connections not reported in primate studies. dlPFC, dorsolateral prefrontal cortex; mPFC, medial prefrontal cortex; Nuc Acc, nucleus accumbens; PAG, periaqueductal gray; IL/CM, intralaminar/centromedial; VTA/SNC/RRF, ventral tegmental area/substantia nigra compacta/retrorubal field; N. Tractus Solitarius, nucleus of tractus solitarius; CeA, central amygdala nucleus; BLA, basolateral amygdala nucleus; SLEA, sublenticular extended amygdala; BNST, bed nucleus of the stria terminalis. Note that this diagram is meant to illustrate compelling parallels but does not equate functional with structural connectivity. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]

LIMITATIONS

One implicit assumption throughout the present work is that a strongly thresholded functional connectivity map reflects underlying structural connectivity. This is not necessarily the case, as anatomical structure is thought to constrain but not determine function. Some of the fMRI results presented here may, therefore, be functionally mediated by other regions and not directly, anatomically connected with the BNST. The relation between structural and functional connectivity is still under active investigation [Honey et al., 2010; Ng et al., 2013; Skudlarski et al., 2008] and further studies, possibly with animal models or with UHF and a range of statistical thresholds, are awaited.

The resolution of the structural scans we used to create the masks could also not resolve BNST subdivisions. Some borders of the drawn masks (i.e., the anterior and medial), which were not visible under T1 contrast, were instead defined in relation to more distal landmarks to ensure a conservative estimate of the BNST and to promote consistency among raters (the averaging and thresholding technique was also designed to provide such consistency). Thus, the resulting drawn masks are likely to contain more of the lateral, central, juxtacapsular, and oval BNST subdivisions, and less of the medial, ventral, and posterior subdivisions. The results should be considered in this light. Regarding the reliability of the drawing of the BNST masks, although our DICE and ICC metrics reflecting raters' drawn masks were relatively low, (a) the correlations between all time series extracted from these masks were statistically significant and quite high (Table 1), and (b) those traditional metrics do not reflect the “consensus correction” that the two‐thirds method was designed for.

Another limitation is that in postmortem human studies and especially animal studies of the BNST structure, sexual dimorphisms have been observed in some subdivisions [Chung et al., 2002]. Our failure to detect sexual dimorphisms might be related to our masks that average across BNST subdivisions, or to the fact that sexual dimorphisms are more detectable with structural than functional measures.

Related to the limits of UHF fMRI, very small regions, such as the parabrachial nuclei, or very distributed regions such as the reticular formation, which are known to work in concert with the BNST, are still beyond the spatial discriminability of UHF fMRI.

Of note is the restricted field of view employed in this study (Supporting Information Fig. 2A, inset). The selection of a non‐whole brain FOV was driven by a common trade‐off of brain coverage with fMRI spatial resolution. For this study, the field of view was focused on the most important regions implicated in the BNST network. However, some of the most dorsal and posterior brain regions, which have not been reported in animal studies to be connected with the BNST, but could theoretically be functionally connected in humans, were missed. Future fMRI studies could retain the same general spatial and temporal resolution with full brain coverage using specialized pulse sequences [Moeller et al., 2010].

Finally, for this study we selected a strong statistical threshold based on several factors, including the general size of the subcortical regions known to connect with the BNST, the effects of the preprocessing on spatial specificity (smoothing and registration), and the desire to report a reasonably‐sized table of results for future meta‐analyses and coordinate‐based ROI definitions. Such an approach increases the probability of false negatives (type II error) in the final statistical maps. Therefore, axial slices across the FOV at one uncorrected (P = 1.0) and two other corrected thresholds (P = 0.0001 and P = 0.001) are offered to the reader as Supporting Information Figures S2A‐C.

Taken together, however, the results demonstrate precise functional mapping of a BNST network, known to be relevant to both anxiety and addiction. Although the regions involved in anxiety and addiction are not mutually exclusive, the BNST network identified here includes areas especially relevant to addiction (caudate and nucleus accumbens) and to anxiety (dorsal amygdala, anterior hippocampus, PAG, and hypothalamus). Future work should test how the BNST intrinsic functional connectivity network responds to psychological tasks, and maps to abnormalities in psychopathology, particularly in anxiety and addiction, using large samples with high phenotypic variability.

To conclude more generally, small and important regions have largely been neglected from in vivo human neuroimaging research in the psychiatric field. However, recent advances in UHF imaging and data processing greatly facilitate their structural and functional identification. The current study is one of the first forays into the mapping of the functional connectivity of one small but very important region, the BNST, and is an approach that has implications for better understanding the neural mechanisms underlying normative and deviant behaviors, particularly with regards to anxiety and addiction.

Supporting information

Supporting Information