Identification of discrete functional subregions
of the human periaqueductal gray
Ajay B. Satputea,1, Tor D. Wagerb, Julien Cohen-Adadc, Marta Bianciardid, Ji-Kyung Choid, Jason T. Buhlee,
Lawrence L. Waldd, and Lisa Feldman Barretta,f
aDepartment of Psychology, Northeastern University, Boston, MA, 02115;bDepartment of Psychology and Neuroscience, University of Colorado at Boulder,
Boulder, CO 80309;cDepartment of Electrical Engineering, École Polytechnique de Montréal, Montreal, QC, Canada H3T 1J4; Departments ofdRadiology
andfPsychiatry, Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, Charlestown, MA 02129;
andeDepartment of Psychology, Columbia University, New York, NY, 10027
Edited by Marcus E. Raichle, Washington University in St. Louis, St. Louis, MO, and approved September 9, 2013 (received for review March 30, 2013)
The midbrain periaqueductal gray (PAG) region is organized into
distinct subregions that coordinate survival-related responses
during threat and stress [Bandler R, Keay KA, Floyd N, Price J
(2000) Brain Res 53 (1):95–104]. To examine PAG function in humans,
researchers have relied primarily on functional MRI (fMRI), but tech-
nological and methodological limitations have prevented research-
ers from localizing responses to different PAG subregions. We used
high-field strength (7-T) fMRI techniques to image the PAG at high
resolution (0.75 mm isotropic), which was critical for dissociating the
PAG from the greater signal variability in the aqueduct. Activation
while participants were exposed to emotionally aversive images seg-
regated into subregions of the PAG along both dorsal/ventral and
rostral/caudal axes. In the rostral PAG, activity was localized to
lateral and dorsomedial subregions. In caudal PAG, activity was
localized to the ventrolateral region. This shifting pattern of activ-
ity from dorsal to ventral PAG along the rostrocaudal axis mirrors
structural and functional neurobiological observations in nonhu-
man animals. Activity in lateral and ventrolateral subregions also
grouped with distinct emotional experiences (e.g., anger and sad-
ness) in a factor analysis, suggesting that each subregion partic-
ipates in distinct functional circuitry. This study establishes the
use of high-field strength fMRI as a promising technique for re-
vealing the functional architecture of the PAG. The techniques de-
veloped here also may be extended to investigate the functional
roles of other brainstem nuclei.
homeostatic regulation important for affective responses and
stress (1–3). Subregions of the PAG underlie distinct, coordinated
behavioral responses to threat. For example, stimulation in the
lateral/dorsolateral portion produces active-coping responses
(e.g., “fight” or “flight”) that involve increasing heart rate and
arterial pressure, redistribution of the blood to the limbs, and
a fast-acting, nonopioid-mediated analgesia. Stimulation in the
ventrolateral portion produces passive-coping responses (i.e.,
disengagement, freezing) that involve reduced heart rate, decreased
reactivity to the environment, and a longer-term, opioid-mediated
analgesic response. These responses occur even when inputs to
PAG from the cortex are severed (1, 4).
The considerable animal literature on the critical role of the
PAG in coordinating emotional responses has led to a surge of
interest in studying the PAG in humans. The PAG plays a central
role in neurobiologically inspired theories of human emotion (5),
the neural circuitry underlying depression and anxiety (3, 6),
autonomic regulation (7), and pain (8–11). To examine PAG
function in humans, researchers have relied primarily on func-
tional MRI (fMRI). To date, dozens of human neuroimaging
studies have observed increased activation in the vicinity of the
PAG during administration of painful and aversive stimuli (8,
12–16) and across a variety of emotional states (17).
he periaqueductal gray (PAG) is a small tube-shaped region
of the midbrain involved in survival-related responses and
Unfortunately however, standard fMRI is fundamentally lim-
ited in its resolution, making it uncertain which fMRI results lie
in the PAG and which lie in other nearby nuclei. The overarching
issue is size and shape. The PAG is small and is shaped like a
hollow cylinder with an external diameter of ∼6 mm, a height of
∼10 mm, and an internal diameter of ∼2–3 mm. The cerebral
aqueduct, which runs through the middle, can prevent detecting
activations within the PAG [type II errors (18)] and also can
create artificial activations that appear to be in the PAG but are
not [type I errors (19)], making the PAG particularly challenging
to image among the subcortical nuclei. Standard smoothing and
normalization procedures, even with high-resolution scanning,
incorporate signal from the aqueduct (Fig. 1). This signal can be
overpowering. The variability of signal in the aqueduct can be
an order of magnitude greater than that of the surrounding
PAG. (Figs. S1 and S2).
Standard neuroimaging techniques also are fundamentally
limited in capturing the remarkable functional organization
that is internal to the PAG. In addition to being differentiated
into columns (1, 4), the PAG also is organized rostrocaudally.
In caudal PAG, neurons that contain endogenous opioids and
neuropeptides involved in nonopioid analgesia are concentrated
in the ventrolateral columns, whereas in rostral PAG this con-
centration is greater in the lateral and dorsomedial columns
(20, 21). Mirroring this distribution, administration of anxiogenic
drugs produces greater neural activity in caudal, ventrolateral
PAG and rostral, dorsolateral PAG [as measured from c-Fos
The periaqueductal gray is a brainstem region that is critical
for autonomic regulation and for defensive responses (e.g.,
“fight,” “flight,” “freeze”). It has been studied extensively in
rodents and cats, but less is known about the human peri-
aqueductal gray. The small size and shape of the periaqueductal
gray makes it challenging to study using standard noninvasive
MRI techniques. We used a high-field strength magnet to ex-
amine this region at high resolution while participants viewed
emotionally aversive or neutral images. Emotion-related func-
tional activity was concentrated in particular subregions and in
ways that are consistent with neurobiological observations in
nonhuman animals. This study establishes a technique to un-
cover the functional architecture of the periaqueductal gray
Author contributions: A.B.S., T.D.W., J.T.B., L.L.W., and L.F.B. designed research; A.B.S.,
J.C.-A., M.B., and J.-K.C. performed research; J.C.-A., and L.L.W. contributed new reagents/
analytic tools; A.B.S. analyzed data; and A.B.S., T.D.W., J.C.-A., M.B., J.-K.C., J.T.B., L.L.W., and
L.F.B. wrote the paper.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission.
1To whom correspondence should be addressed. E-mail: email@example.com.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
| October 15, 2013
| vol. 110
| no. 42
expression (22)]. Connections from the central nucleus of the
amygdala terminate more extensively in lateral and dorsal rostral
PAG and ventrolateral caudal PAG (23). The ability to resolve
which of these circuits is involved in a given behavior is crucial for
understanding the implications of PAG activity in a given situation
and for mapping homologies across species.
Resolving activity to subregions of the PAG requires greater
precision than provided by any study to date. Indeed, the over-
whelming majority of previous studies have used 1.5- or 3-T MRI
systems that typically cannot exceed an isotropic resolution of
1.5–2 mm without incurring significant losses in the signal-to-
noise ratio (SNR) and significant increases in image distortion.
This resolution already introduces substantial partial-volume
effects (Fig. 1) that merge signal from the PAG with the ex-
ceedingly variable signal in the aqueduct (Fig. S2). Standard
intersubject normalization and smoothing procedures that gen-
erally increase SNR but diminish localization accuracy further
aggravate these issues. Ultimately, standard fMRI techniques—
even those using high-resolution scanning—are incapable of
capturing the functional organization of the PAG in humans.
To overcome these obstacles, we used ultra-high field (7-T)
fMRI combined with 32-channel parallel imaging (24), which
can boost sensitivity by as much as an order of magnitude
compared with low-field strength magnets (1.5–3 T) and vol-
ume coils (25). Higher-field strength magnets also provide
greater sensitivity to the susceptibility effects that underlie the
blood-oxygenation level–dependent (BOLD) signal measured
in fMRI (26) and greater sensitivity to microvasculature (27,
28). At a nominal voxel resolution of 0.75 mm isotropic (Fig.
1A), we isolated the PAG directly from the functional scans and
separated activity in the PAG from activity in the aqueduct
We examined activity in the PAG while participants viewed
aversive images (29), which included images of burn victims,
gory injuries, and other content related to threat, harm, and loss,
or while they viewed neutral images. Although prior studies have
demonstrated activity in the vicinity of the PAG during the
viewing of aversive images (15), we first tested whether the PAG
definitively showed greater activity during to these images when
signal was separated from the aqueduct and surrounding brain-
stem nuclei. Next, we used two approaches to test whether ac-
tivity was localized to subregions of the PAG. In one approach,
we segmented the PAG along dorsal/ventral and rostral/caudal
divisions and examined whether activity was localized within
specific segments. In a second but related approach, we per-
formed analyses on a voxel-by-voxel basis to test whether high-
field strength imaging can map the functional architecture
of PAG at the voxel level. Given that discrete voxels identified
in this approach necessarily would reside within just a few mil-
limeters of each other, we performed a factor analysis to dis-
cover whether they carried similar or unique information during
Activation in PAG During Aversive-Image Viewing. In a first analysis,
we averaged signal within the PAG and found that activity was
greater during the viewing of highly aversive images than during
the viewing of neutral images [t (10) = 3.538, P < 0.005]. There
were no effects in the aqueduct [t (10) = 0.348, P < 0.75]. This
finding provides definitive evidence of PAG activation during the
presentation of aversive images.
Identifying Discrete Functional Subregions of PAG During Aversive-
Image Viewing: Mask-Based Segmentation. Using the known in-
ternal organization of the PAG in nonhuman animals, we tested
whether activity in subregions of human PAG was localized to
particular columns, within particular rostrocaudal segments, or
both. First, we implemented a customized normalization pro-
cedure to align voxels in the PAG into a common subject space
without including signal from the aqueduct or surrounding tissue
(Fig. S1). We then divided the PAG radially (six segments ra-
dially divided along dorsoventral and left/right axes) and along
the rostrocaudal axis (six segments: three rostral and three
caudal) and averaged activity across voxels within these segments
for each subject (Fig. 2B).
A repeated-measures ANOVA across picture type (aversive
vs. neutral), radial locations, and longitudinal segments pro-
duced a significant three-way interaction [F (25, 250) = 2.173,
P < 0.0015]. These results indicate that differences in activity
during the viewing of aversive vs. neutral pictures depended on
the interaction between radial and longitudinal segments. No
other effects reached significance (aside from the main effect of
picture type). We further tested for lateralization in left and right
lateral and ventrolateral PAG using a second repeated-measures
ANOVA and observed no significant lateralization effects (P >
0.35 for all cases in which lateralization was a factor).
Unpacking the three-way interaction showed that activations
followed a spiral-like pattern similar to that observed in non-
human animals for the distribution of analgesia-related peptides
(20, 21), c-Fos expression upon anxiogenic drug administration
(22), and connections to PAG from the central nucleus of the
amygdala (23). Although exactly where a transition from caudal
ventrolateral PAG to rostral lateral/dorsomedial PAG may occur
remains unclear (particularly in humans), an approximation us-
ing the findings in nonhuman animals suggests that each extends
about halfway along the rostrocaudal axis.
Averaging within the rostral and caudal segments of the PAG,
we found that the difference in activity during the viewing of gory
vs. neutral images was greater in the rostral half in dorsomedial
PAG [t (10) = 3.19, P = 0.0096] and lateral PAG [t (10) = 3.06,
shows the PAG from a functional scan at ultra-high field strength (7-T) and
high resolution (0.75 mm isotropic). Scanning the PAG at lower resolutions
prevents clear separation of the PAG from the aqueduct and surrounds. (A)
The mean functional image for a single run at the 0.75-mm isotropic reso-
lution used in this study shows the PAG crisply as indicated by the red arrow.
(B) Downsampling the image to a resolution of 1.5 mm isotropic begins to
blur the boundary between the PAG and its surrounds because of partial-
volume effects. (C) Further downsampling the image to a resolution of 3 mm
isotropic eliminates the ability to distinguish PAG from the aqueduct with
any degree of confidence. (D–F) Smoothing with a standard 4-mm kernel
further increases the partial-volume effects that blend signal from PAG with
the aqueduct and surrounds, as shown for 0.75 mm isotropic (D), 1.5 mm
isotropic (E), and 3 mm isotropic (F) resolutions. Most neuroimaging studies
use a 3-mm isotropic resolution with a 4-mm or higher smoothing kernel.
We addressed these issues by separating PAG voxels from the aqueduct
before additional image processing (i.e., using the image shown in A) so that
only voxels within the PAG are incorporated into later stages of analysis. The
top of the transaxial image corresponds to the anterior portion of the head;
the bottom corresponds to the posterior portion of the head.
The PAG imaged at high resolution. The transaxial slice on the left
| www.pnas.org/cgi/doi/10.1073/pnas.1306095110Satpute et al.
P = 0.012] but not in the caudal half [t (10) = 1.03, P = 0.33, and
t (10) = 1.44, P = 0.18, respectively]. In contrast, ventrolateral
PAG showed the opposite pattern, with greater activity in the
caudal half [t (10) = 3.08, P = 0.012] but not in the rostral half [t
(10) = 1.21, P = 0.25]. Activity in ventromedial PAG paralleled
that in ventrolateral PAG numerically, but differences in caudal
PAG did not reach significance [caudal: t (10) = 2.07, P = 0.065;
rostral: t (10) = -0.039, P = 0.97, respectively]. The pattern of
results is illustrated in Fig. 2C.
Identifying Discrete Functional Subregions of PAG During Aversive-
Image Viewing: Voxelwise Analyses. Voxelwise analyses across the
entire PAG volume complement the mask-based analyses in two
ways. First, in assessing the ability of high-field strength and high-
resolution neuroimaging to identify the functional properties of
the human PAG, it is important to determine whether activity in
the PAG can be identified on the voxel level. This approach
allows investigators to discover, rather than stipulate, what the
functional organization of the human PAG may be and supports
PAG on a canonical brain image (using MRIcroGL software). (B) A sample-specific PAG template was generated directly from the functional scans (see Fig.
S1 for details). The template was divided into six radial segments for the dorsomedial PAG (dmPAG; highlighted in mustard), the lateral PAG (lPAG; light
blue), the ventrolateral PAG (vlPAG; purple), and the ventromedial PAG (vmPAG; green). It was divided further into six rostrocaudal or longitudinal
segments that are not depicted but which consisted of slabs that were perpendicular to the radial segments. (C) The plot illustrates the pattern of t scores
across PAG segments for the difference in neural activity during the viewing of aversive relative to neutral images. Radial locations of segments are
plotted on the x-axis, collapsed across laterality (which showed no significant effects). Longitudinal locations of segments are plotted on the y-axis.
Segments were interpolated to illustrate the pattern in the plot. The plot shows that activity was organized both radially and rostrocaudally. From caudal
to rostral PAG, activity transitioned from ventrolateral PAG to lateral and dorsomedial PAG. This spiral-like pattern mirrors neuroanatomical observations
in nonhuman animals (see main text). (D) The figure illustrates results from voxelwise analyses comparing activity during the viewing of aversive relative
to neutral images. Although mask-based segmentation approaches as presented in B and C rely on assuming a particular structural organization before
identifying functional activity within it, voxelwise analyses are important for studies aiming to discover novel functional architectures of human PAG.
Significant clusters were observed in the rostral lateral PAG and in caudal ventrolateral PAG (P < 0.05, corrected for the entire PAG volume; Fig. S3 shows
mean time-course plots). The voxelwise analyses display the utility of high-field strength imaging for detecting activity in specific voxels within the PAG.
Distribution of functional activity in PAG subregions during aversive-image viewing. (A) The sagittal brain rendering illustrates the location of the
Satpute et al.PNAS
| October 15, 2013
| vol. 110
| no. 42
the use of data-driven voxelwise analyses to examine PAG function
in future studies. Second, although mask-based approaches typ-
ically are more powerful statistically, they also introduce partial-
volume effects that can merge distinct signals when the underlying
neuroanatomy can only be approximated. Here, voxelwise analy-
ses serve to mitigate these limitations (at least to the resolution
of the voxel) and are less likely to merge variability arising from
We performed voxelwise analyses by comparing activity during
the viewing of aversive images relative to neutral images across
the PAG volume (corrected for multiple comparisons using the
FMRIB Software Library threshold-free cluster estimation tool,
tfce). As shown in Fig. 2D, the results showed three distinct
clusters. The mesh coloring in the figure is provided for visual
comparison with the mask-based analyses using segments, but
the voxelwise analyses did not incorporate these segments in the
analysis. Subregions were located in the left rostral lateral region
[t (10) = 6.03, P < 0.001, Ptfce< 0.05] and in the caudal ven-
trolateral region (bilaterally, left: t (10) = 5.17, P < 0.001, Ptfce<
0.05; right: t (10) = 5.81, P < 0.001, Ptfce< 0.05). These findings
show that functional activity was localized to discrete clusters at
the voxel level. Mean time-course plots for the three peak voxels
are presented in Fig. S3.
The clusters we identified were only a few millimeters apart.
Hence, we examined whether they showed differential patterns
of variability, which would suggest that they participate in dis-
tinct functional circuits, or whether they showed similar patterns
of variability, which would suggest that they do not participate
in distinct functional circuits. Specifically, we examined how ac-
tivation in these regions varied from each other and with reports
of emotional experience in an exploratory factor analysis. For
each participant, we extracted the average parameter estimate
from 0.75-mm spheres centered on the peak voxel for each
cluster during aversive- vs. neutral-image viewing and submitted
these scores along with the differences in self-reported emo-
tional experience for aversive vs. neutral images to an explor-
atory factor analysis. We reasoned that if activation in the lateral
and ventrolateral PAG clusters participated in the same func-
tional circuit, they would load on a single factor or at least
similarly across factors. However, if activation in these clusters
contributed unique variability during affective responses, then
more than one factor would be required to explain their vari-
ability, and they would load on different factors.
Three factors with eigenvalues above 1 were obtained in the
solution (Table S1). These factors explained 35, 24, and 18% of
the variability. Intercorrelations among factors were low (maxi-
mal r = 0.17), indicating they were likely to be independent. An
examination of the factor loadings showed that instead of loading
similarly across factors, each PAG cluster loaded primarily on
a separate factor and with different emotional experiences.
Factor 1 involved the right caudal ventrolateral PAG (λ = −0.73)
and the emotions disgust, arousal, and fear (all λ > 0.72). Factor
2 involved the left caudal ventrolateral PAG (λ = −0.85) and
anger (λ = 0.89). Factor 3 involved the left rostral lateral PAG (λ =
0.82) and sadness (λ = 0.92). All other loadings between PAG
regions or emotions and factors were less than 0.4 (Table S2).
These results suggest that the PAG clusters contribute to different
functional circuits. Indeed, pairwise correlations between PAG
regions were low and revealed no significant relationships (all
Ps > 0.25; regardless of sign, maximal r = −0.18). Overall, these
results indicate that the PAG subregions identified in voxelwise
analyses carried functionally different sources of variability
during affective responses.
Importantly, this study was not designed to address precisely
which regions of the PAG are related to precisely which kinds of
emotional experiences. Indeed, making strong predictions about
the relationship between self-reported experience of emotions
and animal behavioral responses during threat may be premature.
Self-reports of emotional states are not diagnostic of nor do they
show one-to-one correspondence with specific behavioral actions
in humans or in nonhuman animals (studies reviewed in refs. 30–
32). As such, aligning behaviors in animal research with emo-
tional experiences in human research is not likely to be strongly
justified. The high-resolution imaging techniques developed in
this study may help reveal what relationships are present in be-
havior, experience, and activation of PAG subregions during
emotion in future experiments.
Nonetheless, the shared loadings we observed between PAG
activity and emotional experiences may be useful for formulating
hypotheses in future research in humans. For descriptive pur-
poses, we computed pairwise correlations between emotional
experiences and PAG subregions that loaded similarly on a fac-
tor (Table S2). For factors 1 and 2, the pairwise correlations
showed that activity in the left caudal ventrolateral region was
negatively correlated with self-reported anger (r = −0.62, P <
0.05), and activity in the right caudal ventrolateral region was
negatively correlated with arousal to the aversive images (r =
−0.53, P < 0.05). These findings suggest that feelings of anger
and arousal may emerge with diminishing activity in ventrolateral
“passive-coping” columns. For factor 3, greater activity in the
rostral lateral region was correlated with self-reported sadness in
response to the aversive images (r = 0.53, P < 0.05). Although
sadness is considered a prototypically low-arousal emotion (33),
studies also have identified high-arousal and approach-oriented
forms of sadness (34, 35), which may occur robustly with the
aversive-image stimuli used here (36). As such, factor 3 suggests
that feelings of sadness also may involve greater activity in a
lateral active-coping column. We reiterate, however, that these
specific interpretations are speculative.
More generally, and consistent with recent conceptual and
theoretical developments in emotion research (30, 32, 37), these
findings suggest that mappings between subjective emotional ex-
periences, active-/passive-coping responses that and activity in
PAG subregions may vary with the particular experimental
contexts involved and are unlikely to involve simple one-to-one
formulations of specific emotions with active and passive be-
havioral coping strategies in nonhuman animals. Precisely where
and how emotional experiences in humans map onto activity in
PAG would be an important question for future research using
the techniques we developed here.
Using high-field strength, high-resolution fMRI, we made four
observations related to the human PAG. First, we observed de-
finitive activation in the human PAG while participants passively
viewed aversive images. Prior studies have been unable to correct
for the partial-volume issues that blend signal from PAG with
the aqueduct and surrounding brainstem nuclei. Second, seg-
menting the PAG into both radial and longitudinal subregions
illustrated that activity during negative affect was not diffuse but
was concentrated along a spiral pattern from ventrolateral cau-
dal PAG to lateral and dorsomedial rostral PAG. This pattern
mirrors functional and structural observations in nonhuman
animals (20–23). Indeed, it bears a striking resemblance to the
distribution of c-Fos expression in rodents upon the adminis-
tration of anxiogenic drugs (22). Third, analyses on a voxel-by-
voxel basis showed peak activations in caudal ventrolateral PAG
and in rostral lateral PAG. These findings indicate that high-field
strength neuroimaging can be used to discover, on the voxel
level, the functional architecture of human PAG. Finally, we
found that activity in these subregions responded differentially;
functional activity in these subregions loaded on separate factors
with distinct emotional experiences. Thus, although they were
located only a few millimeters apart, variability in these voxels
was functionally distinct.
| www.pnas.org/cgi/doi/10.1073/pnas.1306095110Satpute et al.
The results of this study highlight several points for un-
derstanding functional activity in human PAG using neuro-
imaging. First, submillimeter high-resolution imaging (such as
the 0.75-mm resolution used here) is critical to separate activity
in the PAG from the high variability observed in the aqueduct.
Imaging studies that use lower resolutions, inexact normalization
procedures, or smoothing without extracting high-variability
signal from the aqueduct beforehand are prone to partial-volume
effects that blend these signals.
Second, this study stresses the importance of imaging func-
tional activity at the resolution of the underlying circuitry. We
observed that portions of the lateral and ventrolateral PAG
contributed distinct sources of information while processing
aversive stimuli. Thus, although the PAG showed greater activity
overall during the viewing of aversive images, the psychological
interpretation of this activity rests on which portions of the PAG
are being engaged. Notably, our approach may be extended to
isolate BOLD activity in nearby midbrain regions, such as the
substantia nigra, red nucleus, colliculi, potentially the dorsal
raphe nucleus, and others that also are known be involved in
emotional and cognitive functions (see Fig. S4 for an example).
Separating these sources of signal will help resolve whether ac-
tivity in the vicinity of the PAG reflects activity within the PAG
or these other, nearby structures.
Third, this method may be used to explore homologies be-
tween human and animal PAG and to extend PAG studies to the
more complex emotional reactions present in humans but diffi-
cult to elucidate in animals. Animal models have shown that
escapable and inescapable stressors elicit active- and passive-
coping responses, respectively, and these responses are related to
distinct subregions of the PAG. However, whether this finding
also holds for humans in similarly threatening contexts is un-
known. Equally important, however, is to examine how the hu-
man PAG functions across a variety of aversive and appetitive
situations. People vary widely in the complex array of cognitive,
behavioral, and affective responses they may have in response
to a given stimulus. We observed that variability in emotional
experiences also relates to activity in discrete subregions of the
PAG (e.g., sadness and anger). This finding indicates a role for
PAG in emotional experience in addition to its known role in
visceromotor responses to threat.
More broadly, activity in the vicinity of the PAG has also been
observed across a tremendous variety of human functions. For
example, responses in the vicinity of the PAG have also been
observed during cognitive and social cognitive functions that are
not traditionally associated with the PAG, including perception,
attention, memory, and language (Fig. S5 and ref. 12). However,
whether—and if so, how—the functionality of PAG in fact
contributes to these abilities in humans remains unclear. The
methodology developed in this study allows for pinpointing ac-
tivity to subregions of the PAG. Future research may discover,
with precision, what role the PAG plays in humans across these
Participants. Thirteen healthy, right-handed participants provided informed
consent in accordance with guidelines set forth by the Partner’s Health In-
stitutional Review Board. They received US$40/h in compensation. Two
participants were excluded because functional data could not be collected as
the result of a hard drive failure (one subject) or poor image quality caused
by ghosting (one subject). For one participant, data from only two of three
runs were included because of a programming failure in the third run. The
analyzed sample included 11 participants (five male, age range, 20–35 y).
Task and Stimuli. To elicit affective responses, participants were shown 30
highly aversive images and 30 neutral images sampled from a database of
normed photographs [the International Affective Picture System (29)].
Images were organized into 17.5-s blocks consisting of five images of each
type, each randomly sampled once during the block and presented for 2 s
followed by variable interstimulus intervals of 0.5, 1, 1.5, 2, or 2.5 s. After
each block of images, participants reported their emotional response to the
set of images using a five-button response box with the emotion labels “Ac-
tivated” (for arousal), “Angry,” “Disgusted,” “Sad,” and “Scared” in a random
order and a numbered scale from 0–4 in which “0” indicated none and “4”
indicated very much of the affect or emotion. After the 16-s reporting period,
the next image block commenced. In other blocks, participants viewed images
from another stimulus set (for details: K. Kveraga http://nmr.mgh.harvard.edu/
∼kestas/affcon) that were not the focus of this analysis. Confirming the aver-
sive intensity of the images, self-report measures indicated that subjects expe-
rienced greater anger [trobust(9) = 8.86, P < 0.00001], disgust (trobust[10) = 5.74,
P < 0.001], sadness [trobust(10) = 3.96, P < 0.01], and fear [scared; trobust(10) =
4.60, P < 0.001], but did not increase self-reported arousal [activated; trobust
(10) = 0.22, P < 0.84]. Self-reports of emotional experiences were not signifi-
cantly correlated with each other (all Ps > 0.08), suggesting that the measures
contributed different sources of variability and were unlikely to be reduced to
valence and arousal dimensions (see Table S2).
Neuroimaging Parameters. Gradient-echo echo-planar imaging BOLD-fMRI
built 32-channel RF loop coil head array was used for reception. Transmit was
provided by a custom-built detunable band-pass birdcage coil. The functional
imaging used single-shot gradient-echo Echo Planar Imaging: echo time = 26
ms, repetition time = 3s, flip angle = 90°, number of slices = 40, slice ori-
entation = oblique axial/coronal (slices were oriented approximately per-
pendicular to the aqueduct), nominal voxel size = 0.75 mm isotropic, gap
between slices = 0 mm, field of view = 192 × 192 mm2, number of repetitions =
90, GRAPPA acceleration factor = 4; echo spacing = 1.04 ms, effective echo
spacing = 1.04 ms/4 = 0.26 ms, bandwidth = 1,148 Hz per pixel, partial
Fourier in the phase encode direction: 6/8.
Neuroimaging Analysis. For initial preprocessing, neuroimaging data were
realigned using the linear registration tool in the FMRIB Software Library
(FSL) (38–40) software package, and the data were filtered through a high-
pass temporal filter (100 s); data were not smoothed or normalized at this
stage (similar to procedures in ref. 41). We then manually segmented the
PAG and the aqueduct for each functional run to normalize data across sub-
jects without introducing partial volume between the PAG and the aqueduct.
Taking advantage of the high variability in signal from the aqueduct, we
computed a variance map to guide the generation of an aqueduct mask (Fig.
S1), which then was inspected manually. A 2.25-mm (three-voxel) dilation
around the aqueduct defined the PAG, which then was inspected manually.
Resulting PAG masks ranged from 8.25–12 mm longitudinally across partic-
ipants. PAG masks were applied to the functional data to include signal from
the PAG and to exclude signal from the aqueduct and surrounding tissue.
General linear models consisting of regressors of the onsets for each image
categoryconvolved with thedouble gamma hemodynamic response function
along with their first-order temporal derivatives were applied to the aver-
aged PAG signal (for examining the undivided PAG) or to each voxel (for
examining subregions) using robust regressions (42) to minimize the in-
fluence of outliers. Across scans, the maximal relative displacement in mo-
tion had a median of 0.39 mm, an interquartile range of 0.29–0.61 mm, and
minimum/maximum range of 0.15–1.45 mm. To account for fluctuations in
response caused by motion, motion parameters were included in the model,
and robust regressions were used in calculations of paired t tests and cor-
relations to limit the influence of outliers upon the statistical analyses.
To examine PAG subregions, a custom PAG template was generated to
align data across participants (Fig. S1). First, each slice within each individual
PAG mask was centered to reshape it into a cylinder. Second, a template
mask was generated by summing across individual maps and including
voxels that showed overlaps by at least two participants’ masks. Registration
of individual masks to the template mask was performed using FSL’s FLIRT
with trilinear interpolation. The parameters from these transformations
then were applied to the functional data (for which aqueduct and sur-
rounding tissue signal already had been masked out). The resulting maps
were smoothed minimally by applying a 1 × 1 × 1 mm Gaussian kernel
[Statistical Parametric Mapping 8 software (43)].
For the mask-based analyses, the PAG template was divided radially (six
segments) and rostrocaudally (six segments), and parameter estimates for
aversive and neutral image viewing conditions were averaged across voxels
within these segments for each participant. A repeated-measures ANOVA
was performed to test whether functional activity was segregated to specific
subregions of the PAG. Follow-up robust paired t tests were performed
to test whether activations were concentrated in rostrocaudal and radial
organizations of PAG. For voxelwise analyses, robust t tests were calculated for
Satpute et al.PNAS
| October 15, 2013
| vol. 110
| no. 42
each voxel within the PAG template, and familywise error correction was de-
termined by using FSL’s tfce. Resulting Ptfcevalues and uncorrected t and P
values for comparison are reported only for voxels that met significance.
There was no uncorrected threshold for voxelwise analyses.
Factor Analysis. The central purpose of the exploratory factor analysis was to
examine whether PAG subregions identified in the voxelwise analyses shared
variability and loaded similarly across factors or whether they contained
meaningfully different variability andloaded on separate factors. Our central
purpose was not to identify precise relationships between emotional expe-
riences and activity in PAG subregions, although these relationships are of
interest for future work using the technique we developed here. Our goals
are generally consistent with suggested guidelines for exploratory factor
analyses with small samples (n < 20) as described recently (44).
Spherical masks (0.7-mm radius) were created centered on peak voxels for
the three activations that reached significance in the voxelwise analyses. The
difference in parameter estimates for the comparison of aversive and neutral
stimulus conditions was extracted for each participant. Self-reported mea-
sures of emotional responses to the aversive vs. neutral blocks of images were
averaged for each participant. These measures were submitted together to
a principle components analysis with direct oblimin rotation, which does not
force orthogonal factor solutions. Using an orthogonal rotation did not alter
the pattern of results. Communalities were in the moderate to high range
(minimum = 0.62, maximum = 0.88) and indicated no Heywood cases. The
analysis revealed three factors with eigenvalues above 1.
ACKNOWLEDGMENTS. We thank Jochen Weber for assistance with figures.
This research was carried out in part at the Athinoula A. Martinos Center
for Biomedical Imaging at the Massachusetts General Hospital using resour-
ces provided by the Center for Functional Neuroimaging Technologies,
P41EB015896, a P41 Regional Resource supported by the National Institute
of Biomedical Imaging and Bioengineering, National Institutes of Health
(NIH). This work involved the use of instrumentation supported by the NIH
Shared Instrumentation Grant Program Grants S10RR023043, S10RR019307,
and S10RR023401. This work also was funded by NIH Director’s Pioneer Award
DP1OD003312 (to L.F.B.), Army Research Institute Contract W5J9CQ-11-C-0046
(to L.F.B.), and by National Institute of Mental Health Grants 2R01MH076136
and R21MH082308 (both to T.D.W.). The views, opinion, and/or findings con-
tained in this article are solely those of the authors and should not be construed
as an official Department of the Army or Department of Defense posi-
tion, policy, or decision.
1. Bandler R, Keay KA, Floyd N, Price J (2000) Central circuits mediating patterned au-
tonomic activity during active vs. passive emotional coping. Brain Res Bull 53(1):
2. Behbehani MM (1995) Functional characteristics of the midbrain periaqueductal gray.
Prog Neurobiol 46(6):575–605.
3. Fanselow MS (1991) The midbrain periaqueductal gray as a coordinator of action
in response to fear and anxiety. The Midbrain Periaqueductal Gray Matter, eds
Depaulis A, Bandler R (Plenum, New York), Vol 105, pp 151–173.
4. Bandler R, Shipley MT (1994) Columnar organization in the midbrain periaqueductal
gray: Modules for emotional expression? Trends Neurosci 17(9):379–389.
5. Panksepp J (2011) The basic emotional circuits of mammalian brains: Do animals have
affective lives? Neurosci Biobehav Rev 35(9):1791–1804.
6. Drevets WC (2001) Neuroimaging and neuropathological studies of depression: Im-
plications for the cognitive-emotional features of mood disorders. Curr Opin Neuro-
7. Cechetto DF, Shoemaker JK (2009) Functional neuroanatomy of autonomic regula-
tion. Neuroimage 47(3):795–803.
8. Tracey I (2005) Nociceptive processing in the human brain. Curr Opin Neurobiol 15(4):
9. Fields H (2004) State-dependent opioid control of pain. Nat Rev Neurosci 5(7):
10. Bittar RG, et al. (2005) Deep brain stimulation for pain relief: A meta-analysis. J Clin
11. Borsook D (2012) Neurological diseases and pain. Brain 135(Pt 2):320–344.
12. Linnman C, Moulton EA, Barmettler G, Becerra L, Borsook D (2012) Neuroimaging
of the periaqueductal gray: State of the field. Neuroimage 60(1):505–522.
13. Mobbs D, et al. (2007) When fear is near: Threat imminence elicits prefrontal-peri-
aqueductal gray shifts in humans. Science 317(5841):1079–1083.
14. Tracey I, et al. (2002) Imaging attentional modulation of pain in the periaqueductal
gray in humans. J Neurosci 22(7):2748–2752.
15. Buhle JT, et al. (2013) Common representation of pain and negative emotion in the
midbrain periaqueductal gray. Soc Cogn Affect Neurosci 8(6):609–616.
16. Kober H, et al. (2008) Functional grouping and cortical-subcortical interactions in
emotion: A meta-analysis of neuroimaging studies. Neuroimage 42(2):998–1031.
17. Lindquist KA, Wager TD, Kober H, Bliss-Moreau E, Barrett LF (2012) The brain basis of
emotion: A meta-analytic review. Behav Brain Sci 35(3):121–143.
18. Lund TE, Madsen KH, Sidaros K, Luo WL, Nichols TE (2006) Non-white noise in fMRI:
Does modelling have an impact? Neuroimage 29(1):54–66.
19. Dagli MS, Ingeholm JE, Haxby JV (1999) Localization of cardiac-induced signal change
in fMRI. Neuroimage 9(4):407–415.
20. Moss MS, Basbaum AI (1983) The peptidergic organization of the cat periaqueductal
gray. II. The distribution of immunoreactive substance P and vasoactive intestinal
polypeptide. J Neurosci 3(7):1437–1449.
21. Moss MS, Glazer EJ, Basbaum AI (1983) The peptidergic organization of the cat
periaqueductal gray. I. The distribution of immunoreactive enkephalin-containing
neurons and terminals. J Neurosci 3(3):603–616.
22. Singewald N, Sharp T (2000) Neuroanatomical targets of anxiogenic drugs in the
hindbrain as revealed by Fos immunocytochemistry. Neuroscience 98(4):759–770.
23. Rizvi TA, Ennis M, Behbehani MM, Shipley MT (1991) Connections between the
central nucleus of the amygdala and the midbrain periaqueductal gray: Topography
and reciprocity. J Comp Neurol 303(1):121–131.
24. Keil B, Triantafyllou C, Hamm M, Wald L (2010) Design optimization of a 32-channel
head coil at 7 T. Proceedings of the International Society of Magnetic Resonance in
Medicine (Stockholm), p 1493.
25. Wiggins GC, et al. (2006) 32-channel 3 Tesla receive-only phased-array head coil with
soccer-ball element geometry. Magn Reson Med 56(1):216–223.
26. Duyn JH (2012) The future of ultra-high field MRI and fMRI for study of the human
brain. Neuroimage 62(2):1241–1248.
27. U? gurbil K, et al. (1999) Functional mapping in the human brain using high magnetic
fields. Philos Trans R Soc Lond B Biol Sci 354(1387):1195–1213.
28. Yacoub E, et al. (2001) Imaging brain function in humans at 7 Tesla. Magn Reson Med
29. Lang PJ, Bradley MM, Cuthbert BN (2008) International affective picture system
(IAPS): Affective ratings of pictures and instruction manual. Technical Report A-8.
(University of Florida, Gainesville, FL).
30. Barrett LF (2006) Solving the emotion paradox: Categorization and the experience
of emotion. Pers Soc Psychol Rev 10(1):20–46.
31. Barrett LF, et al. (2007) Of mice and men: Natural kinds of emotions in the mam-
malian brain? A response to Panksepp and Izard. Perspect Psychol Sci 2(3):297–311.
32. Barrett LF (2012) Emotions are real. Emotion 12(3):413–429.
33. Russell JA, Barrett LF (1999) Core affect, prototypical emotional episodes, and other
things called emotion: Dissecting the elephant. J Pers Soc Psychol 76(5):805–819.
34. Witvliet CV, Vrana SR (1995) Psychophysiological responses as indices of affective
dimensions. Psychophysiology 32(5):436–443.
35. Wilson-Mendenhall CD, Barrett LF, Barsalou LW (2013) Neural evidence that human
emotions share core affective properties. Psychol Sci 24(6):947–956.
36. Mikels JA, et al. (2005) Emotional category data on images from the International
Affective Picture System. Behav Res Methods 37(4):626–630.
37. Wilson-Mendenhall CD, Barrett LF, Simmons WK, Barsalou LW (2011) Grounding
emotion in situated conceptualization. Neuropsychologia 49(5):1105–1127.
38. Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM (2012) Fsl. Neuro-
39. Smith SM, et al. (2004) Advances in functional and structural MR image analysis and
implementation as FSL. Neuroimage 23(Suppl 1):S208–S219.
40. Woolrich MW, et al. (2009) Bayesian analysis of neuroimaging data in FSL. Neuro-
image 45(1, Suppl):S173–S186.
41. Satpute AB, Mumford JA, Naliboff BD, Poldrack RA (2012) Human anterior and
posterior hippocampus respond distinctly to state and trait anxiety. Emotion 12(1):
42. Wager TD, Keller MC, Lacey SC, Jonides J (2005) Increased sensitivity in neuroimaging
analyses using robust regression. Neuroimage 26(1):99–113.
43. Friston KJ, Ashburner JT, Kiebel SJ, Nichols TE, Penny WD (2011) Statistical Parametric
Mapping: The Analysis of Functional Brain Images (Academic, London).
44. de Winter J, Dodou D, Wieringa P (2009) Exploratory factor analysis with small sample
sizes. Multivariate Behav Res 44(2):147–181.
| www.pnas.org/cgi/doi/10.1073/pnas.1306095110 Satpute et al.
Satpute et al. 10.1073/pnas.1306095110
partial-volume effects that merge signal from the aqueduct with signal from the PAG because of large voxel sizes, imprecise normalization, and smoothing
Isolating the periaqueductal gray (PAG). PAG masks were generated directly from the functional data. Standard imaging procedures introduce
Legend continued on following page
Satpute et al. www.pnas.org/cgi/content/short/1306095110 1 of 6
procedures. The figure illustrates how ultra-high resolution 7-T neuroimaging and a procedure for generating targeted masks can be used to isolate signal
from the PAG. (A) The mean image of a single functional run illustrates the advantage of high-resolution scanning at 7 T by clearly showing the PAG and other
subcortical structures including the red nucleus and substantia nigra. (B) The variability image of the functional scan clearly distinguishes the aqueduct (in
white) from the surrounding PAG (in gray). (C) A mask is generated for aqueduct as shown in red using the variability image for each functional run. (D) The
PAG is identified as surrounding the aqueduct, shown in blue, using a dilation of three voxels (2.25 mm) for each functional run. Masks were inspected visually
and edited manually as needed. (E) A sagittal slice showing the aqueduct in red and PAG in blue (yellow regions extended beyond the PAG and were removed
manually). (F) An example of a PAG mask for one run depicted in 3D.
the aqueduct and the PAG. On average and across runs, the SD of signal in the aqueduct is 6.11 times greater than in the PAG (or 36 times greater in variance).
Signal in the aqueduct can mask findings because of added noise and also can create artifactual observations attributed to the PAG because of the influence of
heart rate on cerebrospinal fluid flow (1).
Comparison of signal variability in the PAG and the aqueduct. The SD of signal is plotted for 32 functional runs for signal averaged within masks for
1. Dagli MS, Ingeholm JE, Haxby JV (1999) Localization of cardiac-induced signal change in fMRI. Neuroimage 9(4):407–415.
Satpute et al. www.pnas.org/cgi/content/short/13060951102 of 6
three peak voxels: the caudal lateral PAG (Top), the left rostral ventrolateral PAG (Middle), and the right rostral ventrolateral PAG (Bottom). The first gray bar
at time point 0 marks the onset of the block of images (either aversive or neutral), which lasted for 17.5 s; the second bar marks the offset of the block. SE bars
for within subject effects show the variability of the differences between conditions over time points. The plots demonstrate that, on average, the increase in
signal intensity is greater during aversive-image viewing (red lines) than during neutral-image viewing (blue lines). Images were presented in separate blocks,
but activity is overlaid here for ease of comparison. Intriguingly, the averaged hemodynamic responses appear slightly different among the regions, a finding
that may be of interest for future investigations that optimize detecting the properties of the hemodynamic response to various kinds of affective stimuli.
Sample mean time-course plots from single voxels in the PAG. Raw signal intensity was extracted from each functional scan for each subject for the
Satpute et al. www.pnas.org/cgi/content/short/1306095110 3 of 6
strength and high-resolution imaging. The figure is provided for illustrative purposes only and does not reflect the precise methods used to acquire signal
from the PAG specifically (which are illustrated in Fig. S1). The underlay is the mean functional image for a single participant. The overlay is functional activity
during the viewing of aversive relative to neutral images for that participant (uncorrected, P < 0.05). Blue arrows point to functional activations that track
closely with the particular morphology of sulci and gyri of the participant’s lateral occipital and occipitotemporal cortex. The enlarged view in the blue box
shows separation among brainstem nuclei directly in the functional scans. The colliculi, the PAG, the red nucleus, and substantia nigra are visible landmarks
that also may be used to triangulate activation in adjacent nuclei. The functional parameters were tuned to maximize signal in the PAG. Adjustments would be
required to maximize signal in the red nucleus and substantia nigra (which are higher in iron content and may improve with a reduced echo time) and possibly
in other brainstem nuclei.
Functional activity in brainstem nuclei as revealed by neuroimaging at 7 T. The figure illustrates the greater precision obtained when using high-field
Term-based meta-analysis of PAG function
Activation in or
PAG across a broad array of domains, suggesting that regulation of homeostasis may be important to a broad range of psychological events. We used a term-
based summary map from a database of more than 6,000 neuroimaging studies (NeuroSynth) (1) and identified 145 studies that showed activation in or near
the PAG. NeuroSynth indexes the psychological content domains on which these studies focused based on the frequency counts of words. The bar plot shows
the number of articles that focused on various kinds of psychological content and also show activity in or near PAG. Of the 145 total studies, studies examining
“emotion” were most frequent; however, studies investigating a variety of other psychological contents also frequently engage voxels in the vicinity of the
PAG. The standard neuroimaging methods used in these studies cannot pinpoint whether the PAG is in fact frequently engaged in these domains.
Term-based meta-analysis of PAG function from functional MRI and PET studies. Neuroimaging studies routinely show activity in the vicinity of the
1. Yarkoni T, Poldrack RA, Nichols TE, Van Essen DC, Wager TD (2011) Large-scale automated synthesis of human functional neuroimaging data. Nat Methods 8(8):665–670.
Satpute et al. www.pnas.org/cgi/content/short/13060951104 of 6
Table S1.Pattern matrix from factor analysis
MeasureFactor 1Factor 2 Factor 3
PAG, left rostrallateral
PAG, left caudal ventrolateral
PAG, right caudal ventrolateral
The table presents the pattern matrix from a factor analysis involving
participants’ self-reported emotional experiences and neural activity in
PAG clusters for aversive vs. neutral images. A principle components extrac-
tion with direct oblimin rotation was performed on these measures. Three
factors were extracted with eigenvalues above 1, accounting for 35, 24, and
18% of the variability, respectively, indicating that variability was spread
fairly evenly, given the principle components extraction method. Commu-
nalities were in the moderate to high range (0.62–0.88), indicating no Hey-
wood cases. Factor loadings at 0.72 and above are in bold. Instead of loading
similarly across factors or uncovering a single-factor solution, the analysis
suggests that the PAG clusters loaded on separate factors. Moreover, each
factor comprised activity in a PAG cluster and a self-report measure of emo-
tion. Factor 1 involved the left caudal ventrolateral PAG and the emotions
disgust, arousal, and fear. Factor 2 involved the right caudal ventrolateral
PAG and the emotion anger. Factors 1 and 2 suggest that feelings of disgust,
arousal, fear, and anger emerge as activity in ventrolateral passive coping
columns diminishes. Factor 3 involved the rostral lateral PAG and sadness.
Although sadness is considered a prototypically low-arousal emotion, studies
also have identified high-arousal and approach-oriented forms of sadness,
which may be more likely to occur with the aversive images used here (see
main text). As such, Factor 3 suggests that more feelings of sadness may
relate with greater activity in the ventrolateral active coping column. In
general, these findings suggest that mappings between subjective emo-
tional experiences, active-/passive-coping responses, and activity in PAG sub-
regions may vary with the particular experimental contexts involved and are
unlikely to involve simple one-to-one formulations of specific emotions with
active and passive behavioral coping strategies in nonhuman animals.
Satpute et al. www.pnas.org/cgi/content/short/1306095110 5 of 6
Table S2.Correlation between emotional experience and PAG subregions Download full-text
PAG, left rostral lateral
PAG, left caudal ventrolateral
PAG, right caudal ventrolateral
The table presents the correlation matrix (obtained using robust correlations for outlier correction) between
emotional experience reports and between emotional experience with PAG subregions identified from the
voxelwise analyses. Correlations in bold were significant at P < 0.05 (uncorrected). The factor analysis indicates
that variability in activity in PAG subregions loaded on three separate factors. Each factor also was associated
with different emotional experience reports. The results suggest that although subregions of PAG were only
millimeters apart, they appear to operate in different functional circuits based on showing low intercorrelations
among each other (as described in the main text) and differential patterns of factor loadings and correlations with
the emotional experience reports. Greater activity in the rostral lateral region was correlated with self-reported
sadness in response to the aversive images (r = 0.53, P < 0.05), whereas activity in the caudal ventrolateral regions
were not (left caudal ventrolateral: r = 0.10, P < 0.4; right caudal ventrolateral: r = −0.22, P < 0.3). Alternatively,
greater activity in the left caudal ventrolateral region was negatively correlated with self-reported anger (r = −0.62,
P < 0.05), and activity in the right caudal ventrolateral region was negatively correlated with arousal to the
aversive images (r = −0.53, P < 0.05). Activity in the right rostral lateral region was not associated with self-
reported anger (r = −0.28, P < 0.3) or with arousal (r = −0.10, P < 0.4; Table S1). Robust regressions were used for
Satpute et al. www.pnas.org/cgi/content/short/1306095110 6 of 6