Introduction

During resting-state functional magnetic resonance imaging (rs-fMRI), subjects remain relaxed with the eyes either closed (EC) or open (EO) without engaging in any task (Fox et al. 2009; Northoff et al. 2010). Several recent studies examining resting states have reported that operational conditions (i.e., eyes-closed, eyes-open, and eyes-open with fixation on an image) greatly affect the generalizability of scientific or clinical findings pertaining to connectivity networks, particularly in the sensory cortex (Andrews-Hanna et al. 2010; Feige et al. 2005; Fox et al. 2005; Kollndorfer et al. 2013; McAvoy et al. 2008; Patriat et al. 2013; Van Dijk et al. 2010; Yeo et al. 2011). In those studies, EC and EO conditions were employed in separate experimental runs, and the first 4 − 10 image volumes of the scanning session were discarded during the data preprocessing phase to allow for T1 equilibrium. Because the auditory instruction was normally brief enough, those discarded image volumes could have also minimized potential confounding effects from brain reactions to closing or opening the eyes. Nevertheless, abrupt auditory instructions could evoke functional states that deviate from previous accounts of resting states and have possible associations with motivation, anxiety and/or neuroendocrine activity (Henning et al. 2006; Yoshida et al. 2019). For example, an abrupt EO instruction could be aversive or novel and invade the consciousness of a subject when being relaxed quietly with eyes closed. Although research into EC/EO has uncovered interesting insights into rs-fMRI connectivity, there is a paucity of empirical data on temporal responses that occur immediately after closing/opening the eyes. There is also a lack of information pertaining to the roles played by neuroendocrine and psychological factors in the regulation of these responses. This study bridges this research gap by empirically assessing the strength and duration of short-term effects induced by brain reactions to closing/opening the eyes on a few well-known resting-state networks (RSNs) and the associations between these reactions and neuroendocrine activity.

Several RSNs have been identified corresponding to basic functions, such as vision, auditory perception, language, episodic memory, executive control and salience detection in the human brain (Damoiseaux et al. 2006; De Luca et al. 2006; Habas et al. 2009; Smith et al. 2009; Yeo et al. 2011). Six highly cited RSNs include the DMN (medial prefrontal cortex, posterior cingulate cortex, precuneus, lateral parietal cortex/angular gyrus, inferior temporal gyrus), fronto-parietal network (FPN; lateral prefrontal and inferior parietal cortices), salience network (SN; the anterior insula and anterior cingulate cortex), meso-paralimbic network (MPN; the amygdala, hippocampal formation, and temporal poles), visual network (VN; the primary and high level visual cortices), and sensorimotor network (SMN; the supplementary motor area, sensorimotor cortex, and secondary somatosensory cortex) (Syan et al. 2017). These networks have been identified primarily based on macroscopically defined regions of interest (ROIs). Some studies using the seed-based approach examined the basal ganglia, thalamus, and hypothalamus when investigating associations between cortisol levels and resting-state connectivity (Peters et al. 2019; Veer et al. 2012). Other studies examined the direct and indirect effects of neuroendocrine concentrations on different brain systems, such as the interaction between limbic structures (hippocampus and amygdala) and the prefrontal cortex (Albert and Newhouse 2019; Noack et al. 2019; Yoshida et al. 2019). The success of those studies has made it clear that any method used to investigate connectivity should be subjected to cross-validation by examining the interplay among connectivity networks, neuroendocrine systems, and psychological factors.

A number of recent studies have shown that movies or other naturalistic stimuli can synchronize brain responses among subjects, resulting in stable spatiotemporal activity patterns with a high degree of similarity (Hasson et al. 2010; Moraczewski et al. 2018; Nastase et al. 2019). In many of those studies, inter-subject correlation (ISC) analysis was used to assess the fMRI time courses that are consistently observed among subjects (Chen et al. 2016; Imhof et al. 2017). In ISC analysis, Pearson correlation is computed between two time-courses in a given ROI, the results of which are then applied to all pairs of subjects. The output values are Fisher Z-transformed, averaged, and finally re-transformed back to a correlation value indicating the average degree of similarity among subjects in terms of time courses pertaining to the given ROI (Schmälzle et al. 2017). Intraclass correlation (ICC), on the other hand, is applicable to cases involving more than a single pair of time-courses, and can be used to obtain a single ISC estimate without having to average multiple pairs of correlation values. Unlike Pearson correlation being scale-free, ICC is sensitive to mean- and variance-differences between time courses. Note that this is not necessarily a flaw, as these differences may reflect clinical severity or developmental changes. Theoretically, the asymptotic properties (e.g., standard error of estimation) of ICC are tractable. This makes the ICC approach a practical alternative to the averaging of correlation values across all pairs of subjects. Note also that ICC analysis is easily implemented within cytoarchitectonically (JuBrain) or macroscopically (AAL) defined ROIs, and remains computationally tractable even when dealing with a fairly large number of subjects.

As with fMRI experiments involving naturalistic stimuli, it is possible to synchronize brain responses among subjects simply by providing the same auditory instructions to keep EC-then-EO (or vice versa) for a specific duration. Individual ROIs in RSNs may exhibit high degrees of ISC when temporal responses induced by brain reactions to auditory instructions are incorporated in data analysis. Because a high degree of synchronicity is likely observed immediately after EC/EO onset and at no other times during resting-state scans, the time courses masked out by the ICC method may still convey information specific to individual subjects, varying as a function of psychological traits and cortisol levels. Within the context of well-known RSNs, our primary objective in the current study was to clarify the strength and duration of short-term effects induced by brain reactions to closing and opening eyes on the DMN, FPN, SN, and MPN, rather than to detail the issue of active versus activated processes under different resting states (Joel et al. 2011; Morcom and Fletcher 2007). We also considered EC/EO effects on the stress-response network (SRN; Lucassen et al. 2014) which is distributed in limbic structures as well as the orbitofrontal cortex. The SRN has been widely discussed in studies on the synthesis of rs-fMRI connectivity and neuroendocrine concentrations. We finally examined the neuroendocrine (cortisol) correlates in these networks to further validate our empirical findings.

Methods

fMRI Data Acquisition and Analysis

Participants and Tasks

A total of 55 right-handed young and healthy adults (27 males; average age 22.93 ± 3.08) were recruited for the rs-fMRI experiment and 36 of them had previous experiences performing tasks in an MRI scanner. The experiment was approved by the Human Participant Research Ethics Committee/Institutional Review Board at Academia Sinica (Taiwan) in accordance with the Declaration of Helsinki, and all subjects gave informed written consent before their enrollment into the study. During the fMRI experiment, each subject received exclusively auditory instructions to close or open the eyes through high-quality MR-compatible insert earphones (Sensimetrics Model S14) worn under hearing protection ear muffs. The 8-min experimental task in this study involved one 4-min cycle under the EC condition followed by one 4-min cycle under the EO condition (herein referred to as EC or EO onset). Under the EC condition, the subject was instructed to rest quietly with eyes closed, and the visual presentation was turned off. Under the EO condition, the subject was instructed to rest quietly with their eyes fixated on a central crosshair, and a red crosshair on a dark background image was presented using an MR-compatible visual stimulation system. Note that the order (EC-EO) was fixed based on the results of a pilot study involving 6 subjects.Footnote 1 The pilot study revealed that this configuration was more relaxing for the subjects. We adopted the 4-min duration with the aim of minimizing motion artifacts, and stabilizing the strength of RSNs (Van Dijk et al. 2010).

Image Acquisition and Preprocessing

The fMRI scan was performed using a 3T MAGNETOM Skyra scanner (Siemens Healthcare, Erlangen, Germany) and a standard 20-channel head-neck coil. The echo planar imaging (EPI) scans were performed with parameters TR/TE = 2000-ms/30-ms, flip angle = 84°, 35 slices, slice thickness = 3.4 mm, FOV = 192 mm, and resolution 3 × 3 × 3.74 mm to cover the whole brain including the cerebellum. There were 120 image volumes collected per condition with a total of two conditions (240 volumes) in each session. In addition, four “dummy” volumes were acquired before the EC/EO paradigm to allow the MR signal to reach steady state, and those volumes were discarded by the scanner. T1-weighted anatomical images were also acquired with the following parameters: TR/TE = 2530-ms/3.30-ms, flip angle = 7°, 192 slices, FOV = 256 mm, and resolution 1 × 1 × 1 mm. The data were preprocessed using SPM12 (http://www.fil.ion.ucl.ac.uk/spm). In line with other rs-fMRI studies, empirical time courses were preprocessed by including motion correction, slice timing, linear detrending, and spatial transformation to the MNI-152 template (Ciric et al. 2017; Elliott et al. 2019; Satterthwaite et al. 2013). Nuisance regressors (e.g., motion parameters) were not included in the preprocessing phase to avoid inflated connectivity values (Caballero-Gaudes and Reynolds 2017; Nalci et al. 2019). Also, bandpass filtering (0.01–0.1 Hz) was not performed in the preprocessing phase for reasons detailed in the Discussion section.

Head motion was corrected by the SPM routine which estimated 6 motion parameters consisting of three translations and three rotations at each time point. The 6 parameters time series were represented by the framewise displacement (FD) and root-mean squared (RMS) displacement values relative to a single reference volume (Satterthwaite et al. 2013). The mean FD relative displacement (MRD) value was computed for each subject. The whole brain signal change was summarized by the derivative of RMS displacement over voxels (DVARS) (Power et al. 2011; Power et al. 2012). All the displacement values were computed by the BRAMILA tools available at (https://users.aalto.fi/~eglerean/bramila.html). There were three subjects whose parameters exceeded 2-mm and \({2}^{\circ }\). Their image volumes were manually divided into two segments. Realignment was performed in accordance with a new reference volume formulated for each segment. The aligned image volumes within individual segments were co-registered individually to the standard MNI template. Table 1 provides a summary of mean and standard deviation of MRD and DVAR values of the 55 subjects along with the subgroup (ten subjects) with the highest motion parameters and the subgroup (ten subjects) with the lowest motion parameters.

Table 1 Motion parameters in the complete group and subgroups
Full size table

The T1-weighted high-resolution image volume was co-registered to the mean of the realigned EPI images and was spatially normalized with voxel size 2 × 2 × 3 mm to the standard MNI-152 space (Note: The anisotropic voxel size was chosen to reflect the resolution of raw data). The signal-to-noise ratio of functional images was enhanced by spatial smoothing using a 4-mm FWHM (full width at half maximum) Gaussian kernel.

Data Analysis

ICC Index

In fMRI applications, two ICC indices have been widely used to assess test–retest reliability between brain activation maps and connectivity networks: ICC(C, M) and ICC(A, M), where M denotes the number of experimental replicates, C refers to consistency, and A refers to agreement (Brandt et al. 2013; Caceres et al. 2009; Fiecas et al. 2013; Kristo et al. 2014; Termenon et al. 2016; Wang et al. 2017; Zanto et al. 2014; Zhu et al. 2014). ICC(C, M) is insensitive to mean differences in intensity among subjects, and therefore gives a slightly higher ISC value when MR images are acquired using different scanners. The empirical values of the indices range from − ∞ to 1 (Lahey et al. 1983). The magnitude of ICC indices is determined by many factors, including the size of M. When computed using M of different sizes, the indices cannot be directly compared with one another unless they are normalized according to standard errors. The standard error of ICC(C, M) for temporally dependent time courses was derived under a stationarity assumption (Kuo et al. 2019). In the sequel, the asymptotic standard error of ICC(A, M) is derived by an analogous approach. In a single voxel, the data forms an n(row)-by-M(column) matrix of image intensities, where the M columns indicate subject-level time courses with n time points in each column. Let S denote an M-by-M column-wise cross-covariance matrix with zero lagged time points among the preprocessed fMRI time courses, and let V denote an n-by-n row-wise covariance matrix. An off-diagonal element in S is computed using a pair of time courses from two subjects as a reflection of their synchronization strength (or similarity). An off-diagonal element in V is computed using a pair of intensity vectors between two points in time as an indication of whether between-subject changes in intensity remain the same at the two time points. The ICC(A, M) or agreement index introduces a mean-sensitive correlation measure by considering the variability in mean intensity values among M subjects and among n time points (McGraw and Wong 1996).

The index has been widely applied in clinical research when investigating test–retest reliability of imaging techniques and connectivity networks (Fiecas et al. 2013; Kristo et al. 2014; Zanto et al. 2014). The index can be expressed as follows:

$$\widehat{\text{A}}={ (\frac{\text{M}}{\text{M}-1})[ {\underline{1}}^{{^{\prime}}}}\mathbf{S}\underline{1}-\text{tr}\left(\mathbf{S}\right) ]/\left\{{\underline{1}}^{{^{\prime}}}\mathbf{S}\underline{1} \left[ 1+\left(\frac{\left(\text{n}-1\right){\varvec{\Gamma}}}{\text{n}\left(\text{M}-1\right)}\right)-\left(\frac{\text{M}}{\text{n}\left(\text{M}-1\right)}\right)\frac{\text{tr}\left(\mathbf{S}\right)}{{\underline{1}}^{{^{\prime}}}\mathbf{S}\underline{1}}+ \frac{1}{\text{n}\left(\text{M}-1\right)}\right]\right\},$$
(1)

which is also identical to the ICC(2, M) index in Shrout and Fleiss (1979). A large and positive \(\widehat{\text{A}}\) value indicates that the subject-to-subject variation in their time courses is small; the size of \(\widehat{\text{A}}\) is decreased, on the other hand, if MR images are acquired from different scanners or brain reactivity is unique to individual subjects. In the denominator of (1), Γ denotes the ratio between \(\left({\underline{1}}^{{^{\prime}}}\mathbf{V}\underline{1}/{\text{n}}^{2}\right)\) and \(\left({\underline{1}}^{{^{\prime}}}\mathbf{S}\underline{1}/{\text{M}}^{2}\right)\) where \(\underline{1}\) is the summing vector of order M with the transpose \({\underline{1}}^{^{\prime}}\), and tr(S) denotes the trace of S.

Based on the law of large numbers, the Γ ratio consistently estimates the ratio γ = \({\sigma }_{r}^{2}\)/\({\upsigma }_{\text{c}}^{2}\) as n and M become large. The parameters \({\sigma }_{r}^{2}\) and \({\sigma }_{c}^{2}\) are the mean values of elements in V and S, respectively. The \(\widehat{\text{A}}\) index can be approximated by

$$\widehat{\text{A}}\simeq { \left(\frac{\text{M}}{\text{M}-1}\right)[ {\underline{1}}^{^{\prime}}}{\varvec{S}}\underline{1}-\text{tr}\left({\varvec{S}}\right) ]/\left\{{\underline{1}}^{^{\prime}}{\varvec{S}}\underline{1} \left[1+\left(\frac{1}{\text{M}-1}\right){\varvec{\Gamma}}+O\left(\frac{1}{\text{nM}}\right)\right]\right\} \simeq \widehat{\text{C}}\left[\frac{\text{M}-1}{\left(\text{M}-1\right)+{\varvec{\Gamma}}}\right]+\text{O}\left(\frac{1}{\text{nM}}\right)$$
(2)

In (2), O \(\left(\frac{1}{\text{nM}}\right)\) denotes a remainder term converging to zero with order n times M (Note: The proof can be found in Table A in the Supplementary Materials). The \(\widehat{\text{C}}\) index in (2) denotes the estimate of ICC(C, M) index, which is identical to ICC(3, M) proposed by Shrout and Fleiss (1979). Because \({\varvec{\Gamma}}\) is a positive value, the range of \(\widehat{\text{A}}\) is smaller than that of \(\widehat{\text{C}}\); the \({\varvec{\Gamma}}\) value is affected by the mean differences in intensity among subjects. The variance of \(\widehat{\text{C}}\) has been given in the literature (Browne 1974; Kuo et al. 2019; van Zyl et al. 2000); the asymptotic result in (2) suggests that the variance of \(\widehat{\text{A}}\) can be approximated by

$${\text{Var}}(\widehat{\text{A}}) \backsimeq \text{Var}(\widehat{\text{C}}){\left(\frac{\text{M}-1}{\left(\text{M}-1\right)+\upgamma }\right)}^{2}{|}_{\gamma =\Gamma }$$
(3)

In applications, γ can be evaluated at Γ. The matlab program for computing Var(\(\widehat{\text{C}}\)) is available at (https://github.com/PoChihKuo/icc-neuroimage), which has the following form:

$${\text{Var}}\left( {\widehat{{\text{C}}}} \right) = 2{\text{n}}^{{ - 1}} \left. {\eta ^{\prime}{\text{K}}_{M} \left( {\Sigma \otimes \Sigma } \right){{\text{K}^{\prime}}_{M}} {\eta} } \right|_{{\sum { = S} }} ,$$
(4)

where \({\eta }^{^{\prime}}\) is the derivative of ICC(C, M) with respect to \({\text{vech}}^{\prime}{\varvec{\Sigma}}\) = \(({\text{vech}}{\varvec{\Sigma}})^{\prime},\) and \(\otimes\) is the Kronecker product. \({\varvec{\Sigma}}\) is the population counterpart of S; \({\text{vech}}{\varvec{\Sigma}}\) is the half-vectorization of the matrix Σ; \({\text{K}}_{M}\) is the transformation matrix of order M(M + 1)/2-by-M2, with the identity \({\text{vech}}{\varvec{\Sigma}}={\text{K}}_{M}({\text{vec}}{\varvec{\Sigma}})\), where \({\text{vec}}{\varvec{\Sigma}}\) is the vectorization of the matrix Σ and is of order M2-by-1 and \({\varvec{\Sigma}}\) can be evaluated at S in application. The derivative can be expressed as

$$\eta ^{\prime} = \frac{{\text{M}}}{{({\text{M}} - 1)({\mathbf{\underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle-}$}}{1}^{\prime} }}\Sigma {\mathbf{\underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle-}$}}{1} }})^{2} }}\left. {\left[ { - ({\mathbf{\underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle-}$}}{1}^{\prime} \Sigma \underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle-}$}}{1} }}){\text{vec}^{\prime}}({\text{I}}_{M} ) + {\text{tr}}({\mathbf{\Sigma }})({\mathbf{\underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle-}$}}{1}^{\prime} }} \otimes {\mathbf{\underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle-}$}}{1}^{\prime} }})} \right]{\text{G}}_{M} } \right|_{{\sum { = S} }} ,$$
(5)

where IM is the identity matrix of order M, and \({\text{G}}_{M}\) is the transformation matrix of order M2-by-M(M + 1)/2 with the identity \({\text{vec}}{\varvec{\Sigma}}={\text{G}}_{M}\text{(vech}{\varvec{\Sigma}}\)). The t-value or standardized value for evaluating the empirical agreement index for significance is defined as

$${t}_{\widehat{\text{A}}}=\frac{\widehat{\text{A}}}{\sqrt{\text{Var}\left(\widehat{\text{A}}\right)}}$$
(6)

The data acquisition provided n = 240 and M = 49 in ISC analysis. A schematic illustration of computing the agreement index, error variance, and \({t}_{\widehat{\text{A}}}\) value in the ICC method can be found in Fig. A-1 in the Supplementary Materials.

Statistical Evaluation of the Stationarity Assumption

The derivation of \(\widehat{A}\) in (2) follows from the law of large numbers (i.e., the sample size n or M is reasonably large). The distribution approximations in (3) and (4) are valid when voxel responses \(\left\{{X}_{k,i}: 1 \le k \le n, 1 \le i \le M\right\}\) across the n image scans have bounded fourth moments and satisfy two basic conditions: (i) strictly stationary along the sequence of image scans, and (ii) ρ-mixing which satisfies the condition that the maximal modulus of lagged correlation values between \({X}_{k,i}\) and \({X}_{k+\upsilon , i}\) converge toward 0, that is, for 1 ≤ k ≤ n–υ and 1 ≤ i ≤ M, or max k, i\(\left|\rho \left( {X}_{k,i},{X}_{k+\upsilon , i}\right)\right|\) → 0 as υ → ∞. Fig. A-2 in the Supplementary Materials provides a graphical proof for the ρ-mixing condition, which shows the bivariate plots of the maximal size of lagged-correlation coefficients against the increasing sequence 10 ≤ υ ≤ 230 given the maximal length n = 240. In the plots, it is clear that the lagged sample correlation coefficients decrease from 0.30 to 0.13 as υ is increasing, suggesting that it can decrease toward 0 when υ increases toward infinity.

It follows from Theorem 1.3 in Bradley (2012) that the stationary elements \({s}_{i,j}\) satisfy asymptotic normality of S, which is the basis for deriving the asymptotic variance of \(\widehat{\text{C}}\) in (4). Several empirical studies have suggested that rs-fMRI time courses tend to be non-stationary (Cole et al. 2010; Lee et al. 2013), which may restrict the use of Var(\(\widehat{\text{A}})\) in (3). When temporal responses induced by brain reactivity to closing and opening the eyes are incorporated into data analysis, however, the problem of non-stationarity might become less relevant. To validate the application of (3) and (4), stationarity condition (i) must be examined. The Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test was applied to individual time courses in all in-brain voxels for evaluating the null hypothesis that time courses are trend-stationary against the alternative of a unit root (Kwiatkowski et al. 1992). In addition, the Augmented Dickey–Fuller (ADF) test was also applied to evaluate the null hypothesis of non-stationarity against the alternative hypothesis of stationarity. A time course can be assumed to be stationary if the KPSS test fails to reject the null hypothesis, and the ADF test rejects the null hypothesis. Among in-brain voxels in the current study, a small portion of time courses (2.02%) were consistently rejected by the KPSS test in at least 70% subjects (4.82% in 60% and 11.79% in 50% subjects). After the weak FDR control, the proportion was reduced to 1.96% in at least 70% subjects (4.64% in 60% and 11.14% in 50% subjects). The rejected time courses were mainly distributed in the subgenual anterior cingulate (Areas s24 and s32), orbitofrontal cortex (Areas Fo1, Fo2, and Fo3), and frontal pole (Fp1, Fp2). None of the time courses was consistently accepted by the ADF test across subjects (i.e., 0.15% time courses was consistently accepted in five subjects). Based on the results, the stationarity condition (i) was found valid in most brain voxels.

ICC Maps

As demonstrated by the stationarity tests, the time courses of in-brain voxels generally satisfied the assumption of stationarity, which partially validated the use of Fourier phase randomization to generate surrogate data for thresholding the empirical \({t}_{\widehat{\text{A}}}\) values when determining statistical significance of these values (Lerner et al. 2011). Familywise Type-I error rate was controlled at α = 0.05 using the false discovery rate (FDR) control procedure (Benjamini and Hochberg 1995; Benjamini and Yekutieli 2001; Genovese et al. 2002; Langers et al. 2007). In the FDR procedure, a sequence of ordered p-values (i.e., the probability of randomly simulated t values in the surrogate distributions greater than or equal to the observed \({t}_{\widehat{\text{A}}}\) values) were compared with the critical value (i/K)[α/C(K)] for the i-th p-value in the ordered sequence to control the FDR at α, where K is the total number of voxels considered and C(K) is a predetermined constant. The choice of constant depends on the joint distribution of p-values in the sequence. It has previously been established that FDR control may increase the false-positive or false-negative rate depending on the overall proportion of voxels that are truly synchronized among subjects (Chumbley et al. 2010). In the current study, C(K) = \({\sum }_{i=1}^{K}{i}^{-1}\) was selected for a weak control of the false positive rate (Genovese et al. 2002). The ICC maps were constructed by identifying the supra-threshold voxels with the \({t}_{\widehat{\text{A}}}\) values exceeding the FDR critical values. The supra-threshold maps were co-registered to the JuBrain Cytoarchitectonic Atlas available at (http://www.fz-juelich.de/inm/inm-7/tools) for the topographical definition of area-specific effects (Eickhoff et al. 2005).

Types of Short-Term Effects

Table 2 presents a list of cytoarchitectonic areas involved in the DMN, MPN, FPN and SN and their corresponding Brodmann nomenclature. We annotated brain areas linked to the DMN in a meta-analysis of nine Positron Emission Tomography (PET) studies (132 subjects) by Shulman et al. (1997). After re-analyzing the PET data, Buckner et al. (2005) identified DMN areas that were more active during passive tasks than during active tasks (See Figure 2 in Buckner et al. (2008)). Zilles (2017) included anatomical correspondence between the DMN in Buckner et al. (2008) and cytoarchitectonic areas in JuBrain. In the current study, we also considered cortisol levels, and therefore examined brain structures associated with the SRN (Cerqueira et al. 2008; Lucassen et al. 2014), including the orbitofrontal cortex, hippocampus, entorhinal cortex, amygdala, nucleus accumbens, basal forebrain, and subgenual anterior cingulate cortex. Brain areas unavailable in JuBrain were defined using the Neuromorphometrics Inc. Atlas (http://neuromorphometrics.com) and MRIcro software (http://www.mricro.com) in support of network analysis.

Table 2 Cytoarchitectonic brain areas referred to in the connectivity networks
Full size table

For each ROI (e.g., Area 25 in the sACC) in Table 2, we identified supra-threshold voxels with significant \({t}_{\widehat{A}}\) values. We initially classified the corresponding mean time courses across subjects within each target ROI using iterative K-means clustering and the Silhouette statistics. This was meant to ensure that time courses were homogeneous within each ROI. In cases where an ROI contained two types of time courses (e.g., anterior insula), the ROI was further partitioned into two sub-ROIs. Following an initial examination, the time courses were averaged within each ROI or sub-ROI at the subject level for use in deriving connectivity values. Mean time courses at the subject-level were pooled and averaged at the group level (i.e., one mean time course in each ROI or sub-ROI) for classification into homogeneous networks using iterative K-means clustering (i.e., This procedure was to identify various types of short-term effects). The optimal size for K (number of networks) was determined using the Silhouette statistic. Interested readers may refer to p.88 in Kaufman and Rousseeuw (2009) for a qualitative interpretation of the Silhouette statistic.

Correlation values were computed using subject-level mean time courses between each pair of ROIs within a given network or cluster (Betti et al. 2013; O'Neil et al. 2014; Wang et al. 2007) and between two networks. Fisher's Z transformation was used to enhance the Gaussianity of correlation values. We identified 55 Fisher Z-transformed correlation values between each pair of ROIs (each connectivity value), the sample mean of which was evaluated for statistical significance using the Student’s t-test (Betti et al. 2013; O'Neil et al. 2014; Wang et al. 2007) with the weak FDR control of family-wise error rates. To facilitate visualization, connectivity networks were reconstructed using the BrainNet Viewer (** regions in human posterior parietal cortex. Neuroimage 191:234–242" href="/article/10.1007/s10548-022-00897-x#ref-CR92" id="ref-link-section-d20627697e6477">2019). The SPC is heavily interconnected with the dorsal premotor cortex (Wise et al. 1997), which is important in response selection, stimulus–response association, and response preparation (Passingham 1993). Function in the left AIns is specific to the perception of visual emotional stimuli (Caria et al. 2010). As mentioned previously, Area vHC initiated the orienting response to novel (aversive) stimuli, thereby engaging motor immobility as a parasympathetic brake on the otherwise active motor system to sustain goal-directed behavior. To summarize, it appears that auditory instructions could decrease activity in brain areas that were functionally unrelated to controlling eye-movement. Moreover, decreased activity in Area vHC and the motor system might be an indication of a suppressed fight-or-flight response with the aim of sustaining the on-going resting-state.

Stress-Response Network

One recent review in which animal models were combined with diffusion imaging tractography proposed a revision to the conventional limbic model, which treats the anterior temporal cortex (AM and TP) and OFC as a network functionally associated with behavioral inhibition (Catani et al. 2013). In the theory of behavioral inhibition proposed by Gary and Naughton, the septal nucleus (Note: Nucleus Ch1 in the BF is the medial septal nucleus) and hippocampus play critical roles in monitoring approach-avoidance conflicts. In that schema, activation of the HC constitutes non-anxious rumination, activation of the AM constitutes pure fear, and activation of both constitutes anxiety (McNaughton and Gray 2000). Thus, the AM is strongly engaged in increasing arousal when subjects encounter fear potentiated startle. Neuroimaging and post-mortem histological studies on patients suffering from chronic anxiety and depression have also revealed stress-induced changes in brain structures, including the OFC, sACC, HT, insular cortex, and EnC (Cerqueira et al. 2008; Lucassen et al. 2014). The insular cortex is commonly included as part of the stress-response system (especially the anterior insula) (Bogdan et al. 2016; Hofman and Falk 2012; Nieuwenhuys 2011); however, time courses obtained from the anterior insula in this study are suggestive of two processes: One associated with the TPN and the other associated with the TNN. Generally, brain areas in the SRN overlap considerably with the MPN, which deals with information related to emotion processing and introspection.

In the current study, the SRN exhibited a prolonged decrease in activity following the eyes-closed instruction, and a short-term decrease succeeded by a long-term increase in activity following the eyes-open instruction. Limbic deactivation has been interpreted as the suppression of unwanted emotional reactions that are aroused by tasks (Dagher et al. 2009; Dedovic et al. 2009; Lederbogen et al. 2011; Pruessner et al. 2008; Soliman et al. 2011), or conversely as full relaxation responses under resting states (Kalyani et al. 2011; Shetkar et al. 2019). As mentioned previously, a short-term decrease in the vHC was an indication that the subject was under parasympathetic control and free from approach-avoidance conflicts. Thus, it would be reasonable to hypothesize that while the subjects were relaxed under the EC condition, the SRN played a role in monitoring the external world using sensory information. A prolonged decrease in activity when the subject became accustomed to the external environment might be an indication of fMRI repetition suppression (Barron et al. 2016; Bunzeck and Thiel 2016; Grill-Spector et al. 2006). In other words, the prolonged decrease in activity might be an indication of reduced connectivity between the network and cortical structures (Barron et al. 2016; Kuo et al. 2019). Researchers have previously demonstrated that goal-directed eye movements can activate a dorsal fronto-parietal network (i.e., TPN) and transiently deactivate the amygdala (i.e., SRN) via a ventromedial prefrontal pathway to enable the cognitive regulation of emotions (de Voogd et al. 2018). These findings are consistent with our observation of time courses within the SRN under the EO condition. To summarize, the SRN presented reduced connectivity to cortical structures under the EC condition. It is possible that when the novel (aversive) stimulus was non-threatening, the observed short-term decrease succeeded by a long-term increase in activity under the EO condition was mediated by either the ventromedial prefrontal pathway or Area vHC (Patrick et al. 2019).

Pre-scan and Post-scan Cortisol Levels

Cortisol is a catabolic hormone in normal life and an adaptation hormone in response to stressful situations (Coenen and Flik 2017). Cortisol acts on low-affinity glucocorticoid receptors (GRs) as well as high-affinity mineralocorticoid receptors (MRs). GRs are widely distributed throughout the brain with high concentrations in the hypothalamic corticotrophin-releasing-hormone neurons and pituitary corticotropes and, by contrast, the distribution of MRs is highly restricted (Andersen et al. 2013; de Kloet and Joëls 2020; Jelić et al. 2005). Both GRs and MRs are highly expressed in limbic structures, such as the hippocampus, amygdala, and brain areas associated with fear and anxiety (De Kloet et al. 2018; Deuter et al. 2019). In the current study, average cortisol levels in saliva collected prior to the fMRI scanning sessions were significantly lower than in saliva collected after scanning. Connectivity values in the three networks were also more strongly associated with pre-cortisol levels than with post-cortisol levels. As mentioned, subjects were passively exposed to 20-min of natural sound stimulation (human voices and animal vocalizations) and speech (same sex-marriage controversy) following the completion of the resting-state scan. The STAI-Trait and SQ-ISMA scores were more strongly correlated with post-cortisol levels than with pre-cortisol levels. In other words, trait-anxiety or stress scores could be predictive of post-cortisol levels.

The results in Tables 4 and 5 show that after controlling for demographic factors, connectivity within the SRN was significantly positively correlated with pre-cortisol levels. Inter-hemispheric connectivity within the SRN and TNN was also significantly correlated with pre-cortisol levels. These results are in good agreement with the high concentrations of GRs and MRs in limbic structures. As mentioned previously, ROIs in the SRN were stable loci for probing cortisol activity, despite their sensitivity to the EO instruction while the subjects were resting quietly. The strong connection between the TNN and SRN under the EC condition might account for the significant correlation between connectivity in the TNN and pre-cortisol levels. It is also possible that under the EC condition, connectivity within and between the SRN and TNN could serve as a predictor of post-cortisol levels. The EC protocol is generally well-suited to research linking resting-state connectivity and GR/MR concentration in limbic structures, considering the weakness of brain reactions to the anticipated EC instruction. Our empirical data collected from one cycle of EC-EO resting state scanning did not support the hypothesis that decreased activity in the ventral hippocampus is regulated by cortisol activity, as indicated by the fact that the TNN was unaffected by pre-cortisol levels under the EO condition. Moreover, the correlation values listed in Tables 4, 5, C and D do not necessarily provide evidence supporting the role played by pre-cortisol levels in regulating EC/EO effects. As mentioned, subject-level mean time courses showed a high degree of synchronicity after EC/EO onset and at no other times. This would be a reason that we found a significant correlation between connectivity values and pre-cortisol levels within the SRN. Therefore, the ICC method provided ample information specific to individual subjects when searching for synchronized brain reactions to EC/EO instructions.

Remarks

Rs-fMRI studies have previously used 0.01 Hz (or 0.008 Hz) as a standard cutoff to remove low frequency effects such as slow drifts in the fMRI signal (Wang 2021). Physiological noise from respiratory and cardiac functions is concentrated at higher frequencies (> 0.1 Hz) (Davey et al. 2013; Mascali et al. 2021). Thus, RSNs of interest are generally associated with frequencies in the range of 0.01–0.1 Hz. Fig. H in the Supplementary Materials plots supra-threshold time courses of the 49 subjects (receiving EC-EO instructions) after either high-pass (i.e., > 0.01 Hz) or bandpass (0.01–0.1 Hz) filtering. Note that the supra-threshold time courses were extracted using the ICC method based on pre-processed data that had not undergone filtering. In a comparison of Figs. 3, 4, 5 and H, it appears that the TPN was less affected by data filtering than were the TNN and SRN. Figures H-(b), H-(c), H-(bʹ) and H-(cʹ) further suggest that ROIs in the TNN and SRN after data filtering could be grouped into one cluster, such that activity in the ventral hippocampus and other motor areas would be indifferentiable from that in the SRN. The main hypothesis generated from our findings was the approach-avoidance conflicts which distinguished activity in the hippocampus (the transition negative network; TNN) from that in limbic structures (the stress-response network; SRN). The two networks were highly associated with each other under the EC condition, but disconnected under the EO condition. As mentioned, the SRN is significantly associated with pre-cortisol levels under both EC and EO conditions, but the TNN is significantly associated with pre-cortisol levels only under the EC condition. The filtering procedure (especially high-pass filtering) in the pre-processing stage could affect the scientific interpretation of TNN and SRN.

Changes in heart-rate and respiration could have confounding effects on rs-fMRI time courses. Because electrocardiograms (ECGs) and breathing patterns were not collected during resting-state scans, it is difficult to provide a concise assessment of these effects. However, high-pass and bandpass filtering provided similar results as indicated in Fig. H. It appears that heart-rate and respiration might not necessarily have a serious confounding effect on the analysis of brain reactions to closing/opening the eyes. Head motion after EC/EO onset could also affect the assessment of synchronicity in the ICC method. Values of FD and DVARS across the time scale are presented in Fig. F in the Supplementary Materials for the 49 subjects receiving EC-EO instructions. These values were computed using realigned image volumes within individual subjects before the stages of removing trends and normalization to the MNI template. There were two subjects who did show larger FD values after EC/EO onset compared with other subjects. In Fig. F, the large mean values after EC/EO onset (especially after EC onset) mainly reflect head motion of these two highlighted subjects (e.g., See blue and red curves). Because the ICC method considers synchronicity across all subjects, we assume that head motion of a few subjects after EC/EO onset had minor effects on interpretation of the three networks.

Limitations

The order of EC and EO instructions had minimum effects on the TPN, but had significant effects on the TNN and SRN according to brain reactions to auditory instructions in the 6 pilot subjects. However, the order of instructions did not show significant effects on the assessment of correlation between connectivity and cortisol levels. Our experiment was limited to one cycle of the EC-EO transition, and findings in the Results section could partially reflect a variety of diverse processes such as listening to instructions, opening the eyes, and novelty responses. The experimental design did not allow to disentangle the specific contributions of these processes to the observed data. A further study could adopt multiple cycles of the EC-EO transitions to investigate effects from different processes (Feige et al. 2005). Budget considerations limited us to 25 assays for cortisol levels (including cortisol levels in the 6 pilot subjects). Given the small sample size in the current study, we were unable to provide a reliable test on the hypothesis of approach-avoidance conflicts as regulated by Area vHC (or alternatively by a ventromedial prefrontal pathway).

Conclusion

The study is the first attempt to connect cytoarchitectonically defined brain structures to brain reactivity patterns associated with closing and opening the eyes under resting-state conditions. We recommend that research based on rs-fMRI adopts the EC condition when probing resting-state connectivity and its neuroendocrine correlates. We also suggest that abrupt instructions to open the eyes (while subjects are resting quietly with EC) could be used as a probe to characterize brain reactivity to aversive stimuli in the ventral hippocampus and other limbic structures. Finally, we encourage further work in this direction, particularly from a clinical perspective.