The retrosplenial cortex: A memory gateway between the cortical default mode network and the medial temporal lobe

Abstract

The default mode network (DMN) involves interacting cortical areas, including the posterior cingulate cortex (PCC) and the retrosplenial cortex (RSC), and subcortical areas, including the medial temporal lobe (MTL). The degree of functional connectivity (FC) within the DMN, particularly between MTL and medial‐parietal subsystems, relates to episodic memory (EM) processes. However, past resting‐state studies investigating the link between posterior DMN‐MTL FC and EM performance yielded inconsistent results, possibly reflecting heterogeneity in the degree of connectivity between MTL and specific cortical DMN regions. Animal work suggests that RSC has structural connections to both cortical DMN regions and MTL, and may thus serve as an intermediate layer that facilitates information transfer between cortical and subcortical DMNs. We studied 180 healthy old adults (aged 64–68 years), who underwent comprehensive assessment of EM, along with resting‐state fMRI. We found greater FC between MTL and RSC than between MTL and the other cortical DMN regions (e.g., PCC), with the only significant association with EM observed for MTL‐RSC FC. Mediational analysis showed that MTL‐cortical DMN connectivity increased with RSC as a mediator. Further analysis using a graph‐theoretical approach on DMN nodes revealed the highest betweenness centrality for RSC, confirming that a high proportion of short paths among DMN regions pass through RSC. Importantly, the degree of RSC mediation was associated with EM performance, suggesting that individuals with greater mediation have an EM advantage. These findings suggest that RSC forms a critical gateway between MTL and cortical DMN to support EM in older adults.

1. INTRODUCTION

A network of dynamically coupled brain regions referred to as the default‐mode network (DMN) (Andrews‐Hanna et al., 2010; Buckner et al., 2008; Campbell et al., 2013; Raichle et al., 2001; Salami et al., 2014, 2016) shows greater activity during passive than active states (Buckner et al., 2008; Raichle et al., 2001). A core DMN structure in the medial parietal cortex encompasses the posterior cingulate cortex (PCC) and the retrosplenial cortex (RSC), often jointly referred to as posterior DMN (pDMN). Other cortical DMN regions are located in the lateral parietal and the medial prefrontal cortices. In addition, a medial temporal lobe (MTL) subsystem, including hippocampus (HC) and adjacent structures, is distinct from, yet related to, the major cortical DMN regions (Andrews‐Hanna et al., 2010; Buckner et al., 2008; Salami et al., 2014, 2016; Ward et al., 2015).

Interactions among the MTL subsystem and the pDMN nodes have been suggested to underlie a range of functions, including episodic memory (EM) in both younger (Touroutoglou et al., 2015; Voets et al., 2014; Vincent et al., 2006) and older adults (Wang et al., 2010; Mevel et al., 2013; McCormick et al., 2013; Salami et al., 2014). However, some studies, mainly based on older adults, have failed to observe such a relation (Andrews‐Hanna et al., 2007; Beason‐Held et al., 2009; Damoiseaux et al., 2016; Vidal‐Piñeiro et al., 2014). One possible reason for the discrepant findings is age‐related spatial reorganization of resting‐state networks (RSNs) and shifts in the location of functional regions (Chan et al., 2014; Littow et al., 2010; Goldstone et al., 2016; Sohn et al., 2015). Moreover, age‐related disruptions in MTL‐pDMN FC (Andrews‐Hanna et al. 2007; Salami et al., 2016) may contribute to variations in relation to cognition. The heterogeneity within the posterior DMN might also contribute to discrepant findings. Human and animal studies indicate that the medial parietal cortex, including PCC and RSC, is a heterogeneous region, and that it is important to distinguish RSC from PCC due to morphological and connectivity differences (Aggleton, 2012; Dastjerdi et al., 2011; Dillen et al., 2016; Kobayashi & Amaral, 2003, 2007; Leech et al., 2011; Mufson & Pandya, 1984; Vann et al., 2009). Critically, the density of HC‐RSC connections was found to be much higher than the density of HC connections to adjacent PCC areas (Kobayashi & Amaral, 2003, 2007; Mufson & Pandya, 1984). This suggests that, within the medial parietal cortex, RSC may also be involved in hippocampal‐based functions. Toward this end, animal research found that RSC in conjunction with HC is important to episodic memory (EM; for review, see Vann et al., 2009). However, human studies have mainly focused on the PCC‐MTL interaction and its link to EM (Mevel et al., 2013; McCormick et al., 2013; Smallwood et al., 2016; Wang et al., 2010). Thus, it remains unknown whether functional connectivity (FC; coherent spontaneous activity between distal brain regions) between MTL and RSC may be at least equally or even more strongly linked to EM processes than FC between MTL and PCC.

Axonal tracing studies suggest that RSC has a dense structural connection to both HC and other cortical DMNs and may serve as an intermediate layer that facilitates the connection between MTL and different cortical DMNs (Vann et al., 2009). Similarly, a diffusion‐tensor imaging study revealed a direct structural connection between MTL and RSC as well as between PCC and medial prefrontal cortex (mPFC) (Greicius et al., 2009). The latter structural‐connectivity pattern suggests hierarchical connectivity within the DMN, such that there is a flow of information from MTL through RSC to the adjacent PCC and mPFC. Although structural studies suggest that RSC is anatomically well placed to integrate information from the cortical and subcortical DMN, direct evidence regarding its functional role as an intermediate region is sparse. Given that RSC may be an intermediate layer for communication between MTL and other cortical DMNs, it is reasonable to hypothesize that RSC, compared to other DMN regions, mediates more neural traffic while controlling information transfer within the DMN (Tang et al., 2010). On this view, RSC should have a high betweenness centrality (BC) within the DMN. Whole brain graph‐theoretical analyses in young adults have demonstrated relatively high BC across the whole brain for the precuneus/dorsal PCC (Spreng et al., 2013), with dense structural and functional connectivity to many other regions across the whole brain (Hagmann et al., 2008). However, ventral PCC exhibited dense local FC primarily restricted to the DMN (Leech & Sharp, 2014; Leech et al., 2011, 2012; Spreng et al., 2013; Tomasi & Volkow, 2010). Although RSC may partly overlap with ventral PCC, as reported in some previous studies (Dastjerdi et al., 2011; Dillen et al., 2016; Leech & Sharp, 2014; Vann et al., 2009), its centrality has never been directly compared to the other cortical areas of DMN. Given previous evidence in support of age‐related functional dissociation between PCC and RSC (Mevel et al., 2013), it is indeed important to investigate the centrality roles of PCC and RSC in older age separately. RSC might have a relatively low degree of centrality in the context of the whole brain, but dense connectivity primarily restricted to DMN. On this view, RSC might be a provincial hub (Sporns et al., 2007; Van den Heuvel & Sporns, 2013) among older adults that facilitates interaction within, rather than outside, DMN.

The current research emanates from the Cognition, Brain, and Aging (COBRA) project, which includes assessment of several cognitive measures including EM as well as Magnetic Resonance Imaging (MRI) of 181 adults between 64 and 68 years of age (Nevalainen et al., 2015; Nyberg et al., 2016). COBRA is planned to be a longitudinal study over a 10‐year period with 3 equally spaced measurement occasions. The overall aim of the COBRA study is to investigate the extent of longitudinal changes in functional, structural, and molecular architecture of the brain and link those changes to cognitive decline. In this study, we used functional and cognitive data from baseline measurement. Key objectives of the current work were to examine whether the degree of functional coupling between MTL and RSC regions is stronger than that between MTL and other cortical DMN regions, and whether the MTL‐RSC link is especially important to EM performance. We used independent‐component analysis (ICA; Calhoun et al., 2001a, 2001b) on 6 min resting‐state functional MRI (fMRI) data to identify the DMN subsystems. A further aim was to test whether the RSC serves as a mediator of the cross‐talk between the MTL and other cortical DMN regions using a multilevel mediation approach (Kenny et al., 2003). We also examined whether the degree of mediation predicts individual differences in EM performance. Finally, we hypothesized that RSC as the potential intermediate layer between subcortical and cortical DMN areas would exhibit relatively higher centrality as compared to other DMN regions. Using graph‐theory analysis, a centrality measure (i.e., BC) (Freeman, 1977; Sporns et al., 2007) of RSC was computed and compared with BC for the other DMN regions. BC reflects how important a specific region is for information transfer within a network. We expected that RSC would exhibit relatively higher BC as compared to other DMN regions.

2. METHODS

2.1. Participants

All participants were part of the COBRA project (Nevalainen et al., 2015; Nyberg et al., 2016). The COBRA dataset includes brain imaging in 183 healthy older adults (64–68 years old; mean = 66.2; SD = 1.2; 81 women), who were randomly selected from the population in Umeå, a city in Northern Sweden. From the initial sample, 180 participants underwent comprehensive assessment of different cognitive domains including episodic memory, working memory, and perceptual speed.

One individual was excluded because of extensive distortion in functional scans. Sample characteristics, including education (mean = 13.3 years; SD = 3.5), body‐mass index (mean = 26.1 kg/m2; SD = 3.5), systolic blood pressure (mean = 142 mmHg; SD = 17), and diastolic blood pressure (mean = 85 mmHg; SD = 10) indicated that the sample was representative of the target population in Umeå (Nevalainen et al., 2015). Statistical outliers (criterion: mean ± 3.29 × SD) were excluded in all correlational, t‐test, and ANOVA analysis.

2.2. Cognitive tests

EM was assessed using three different tasks: word recall, number‐word recall, and object‐position recall (Nevalainen et al., 2015). During word recall, 16 concrete Swedish nouns with different first three letters were presented on the screen. At encoding, each noun was presented for 6 s, with an interstimulus interval (ISI) of 1 s. After a list of 16 nouns had been presented, participants were asked to report all words they could recall. Following each practice trial, two test trials were conducted. Then, the sum of the scores on the two test trials was computed. The reliability of word recall was high (Spearman–Brown coefficient (Sc) = 0.79).

In the number‐word recall task the participants were asked to memorize pairs of 2‐digit numbers and a concrete plural noun (e.g., 46 dogs). At encoding, 8 number‐word pairs were presented for 6 s each, with an ISI of 1 s. Then, participants were asked to type the 2‐digit number associated with each word presented on the screen (e.g., how many dogs?). After each practice trial, two test trials were given. Similar to the word recall test, the sum of the scores for the two test trials was computed. The reliability was 0.63 for the number‐word recall task.

In the object‐position recall task, a 6 × 6 squares grid was shown on the screen on which any of 12 objects were presented for 8 s at different locations each time, with an ISI of 1 s. Then, subjects were asked to show the correct position of each object by moving objects with the computer mouse. Similar to the other two EM tests, there was a practice trial before performing two test trials. The sum of the scores on the two test trials was computed. The reliability for the object‐position task was 0.69.

The EM composite score was the average of T‐scored summary measures (across trials) of the three EM tasks. First, a summary score of each task was computed across trials of that particular task. Then, the summary scores were standardized and averaged to form one unit‐weighted measure for each task. Finally, by averaging the T‐scored measures of each task, a summary score was computed for EM domain. This composite score has been used in some previous studies from our group (e.g., Nevalainen et al., 2015; Nyberg et al., 2016; Lövdén et al., 2017). Analyses of the reliability of the EM composite measure revealed a good level. To handle missing data (<1.2% for all three variables), the participant mean on the other dependent measures of the construct was used to impute the missing values, so that the final composite score did not have any missing values. The resulting EM composite measure was standardized into z‐scores (mean = 0; SD = 1).

Working memory was assessed with a combination of verbal, numerical, and figural tasks. For the letter‐updating task, participants were asked to update and remember the three last shown letters from a presented letter sequence. Presentation of letters was lasting for 1 s, with an ISI of 0.5 s. Subjects were asked to type the last three letters of the sequence, at an unknown time. The 16 test trials consisted of a random combination of 7, 9, 11, or 13 letter sequences. Four practice trials were carried out before performing test trials. The total number of correct answers on the test trials was computed. The reliability measure was 0.76 (Cronbach's alpha).

During the numerical 3‐back task, 3‐digit numbers were shown in three boxes on screen. First, left digit was shown for 1.5 s, then each of the second and third digits were presented with an ISI of 0.5 s. This task included two practice trials followed by four test trials, while 30 digits were presented for each trial. Subjects were asked to indicate whether the appeared digit in each box was the same as the last digit which was displayed in that specific box (3‐back). The numbers of correct responses (while ignoring the responses to the first three digits of each trial) were summed across all trials. The estimated reliability measure was 0.90.

In the spatial‐updating task, three different 3 × 3 squares grids were shown on the screen on which three circular objects were presented at random positions in each grid for 4 s, all at the same time. Then, an arrow appeared below each grid (sequentially from left to right) for 2.5 s with an ISI of 0.5 s. The arrow was pointing at the direction of goal position for the circle (where each circle should be mentally moved). These spatial updating operations were performed twice for each of the three grids before subjects were asked to use the computer mouse and show the correct position of each circle. There were 5 practice trials followed by ten test trials. The computed score, which was simply the accuracy across all trials, had a reliability of 0.73. Similar to the EM composite score, a summary measure was also computed for working memory.

Perceptual speed was assessed with a combination of verbal, numerical, and figural comparison tasks. During the tasks, subjects were asked to put their index fingers on two keyboard buttons. After items were presented, the subjects were asked to quickly decide whether they were identical and press either right or left buttons representative of “yes” or “no” answers, respectively. In letter‐comparison task, subjects were supposed to compare two 4‐letter strings. For the cases in which letter strings were not identical, only one letter was different. Participants had 5 s to answer, otherwise the sequence disappeared. Forty item‐pairs were presented in each trial, with an ISI of 0.5 s. One practice trial was performed followed by two test trials. Number of correct answers in each trial was divided by the total response‐time (correct answers per minute). Total score which was the sum of computed scores across two trials, had a reliability estimate of 0.95. For the number and figure comparison tasks, presented items were either 4‐number strings or figures, with reliability of 0.94 and 0.88, respectively (see Nevalainen et al., 2015 for details). Similar to EM and working memory, a composite score was computed for perceptual speed.

2.3. Image acquisition

MR images were collected using a 3 T Discovery MR 750 scanner (GE), with a 32‐channel phased‐array head coil. Pillows, blankets, earplugs, and headphones were used to make the scanner comfortable. Cushions were put inside the head coil to minimize head movements. By applying 3D fast‐spoiled gradient‐echo sequence, high‐resolution T1‐weighted structural images were acquired (TR = 8.2 ms, TE = 3.2 ms, flip angle =1 2°, slice thickness = 1 mm, number of slices = 176, field of view = 25 × 25 cm). The functional data were acquired for 6 minutes during resting‐state conditions, where participants were instructed to keep their eyes open and fixate on a cross while remaining awake. For whole‐brain functional data, blood oxygen level‐dependent (BOLD) signals were recorded using T2*‐weighted single‐shot gradient echo‐planar sequence (37 transaxial slices, 3.4 mm slice thickness, 0.5 mm slice gap, TR = 2,000 ms, TE = 30 ms, flip angle = 80°, slice thickness = 3.4 mm, slice gap = 0.5 mm, number of volumes = 170; field of view = 25 × 25 cm). To avoid signals arising from progressive saturation and to allow for steady‐state imaging, 10 dummy scans were acquired in the beginning of the MRI session.

2.4. Data analysis

2.4.1. MRI data preprocessing

Prior to the FC analysis, resting‐state functional and structural MRI data were preprocessed using the Statistical Parametric Mapping software (SPM12; Welcome Department of Cognitive Neurology, University College London, London, United Kingdom). After applying slice‐time correction, data were motion corrected by rigidly aligning each volume to the first image volume. Then, fMRI data were co‐registered to T1‐weighted images using within‐subject rigid registration. Co‐registered T1 images were segmented into gray matter (GM) and white matter (WM). GM/WM images were imported into Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) space (Ashburner, 2007), followed by a 6‐step iterative procedure, which produced a group‐level template. Finally, the realigned fMRI images were nonlinearly normalized to the group‐level template, affine‐aligned into Montreal Neurological Institute (MNI) template space, and smoothed using a Gaussian kernel with a full‐width at half‐maximum (FWHM) of 8 mm.

2.4.2. ICA analysis

Following preprocessing, ICA, using the GIFT toolbox (Allen et al., 2011; Calhoun et al., 2001a, 2001b), was employed to identify RSNs in general and DMN components, in particular. ICA is a multivariate data‐driven approach, which estimates maximally independent components based on higher order statistics. Without the need for an a priori temporal model, ICA is perfectly suited for resting‐state fMRI data. In addition, compared with seed‐based approaches, ICA avoids the need to arbitrarily choose regions of interest and can successfully identify different DMN subregions (Leech et al., 2011, 2012). Spatial ICA decomposes data into spatially independent and temporally coherent component maps and their corresponding time courses (TCs), which reflect the dynamics of the BOLD signals within each component (Calhoun et al., 2001a, 2001b). Details of ICA have been provided in previous work (Allen et al., 2011; Salami et al., 2014, 2016). In short, time series of each voxel was divided by its average intensity (i.e., intensity normalized) to increase accuracy and reliability of the ICA decomposition. Then intensity‐normalized data for all participants were concatenated across time. After temporal concatenation, the ICA model order was set to 50. This number of components was used in our previous work (Salami et al., 2014, 2016) and provides a fine‐grain parcellation of DMNs. The dimension of data was reduced by applying two‐step data reduction using principal component analysis (PCA1: 100 components; PCA2: 50 components). Then, the Infomax algorithm (Bell & Sejnowski, 1995) was employed to identify temporally coherent networks. ICA was repeated 20 times with different initial points using ICASSO, which results in compact clusters of components in different runs (Himberg et al., 2004). All ICs exhibited acceptable reliability (0.76 < Iq < 0.99). After excluding visually inspected artefactual components, 39 RSNs were identified. Selected RSNs showed peak activation in gray matter and exhibited low spatial overlap with the topology of potential artifacts (e.g., vascular, ventricle, motion, and susceptibility artifacts). This inspection was carried out by two independent researchers. Moreover, the DMN components exhibited high spatial overlap with templates of DMN components reported in a previous study of older adults (Salami et al., 2014). These RSNs included the MTL network, the RSC network, and the anterior/posterior DMNs (mainly including mPFC, PCC, and precuneus). Finally, subject‐level spatial maps and TCs were estimated using the GICA3 back‐reconstruction method.

Subject‐specific TCs were detrended to remove drifts. Then, TCs were despiked, and low‐pass filtered (Butterworth, fifth‐order, cutoff frequency equal to 0.15 Hz). Finally, after removing the mean of each TC, the matrix of correlation coefficients (i.e., internetwork connectivity) was computed using all ICA components' TCs. Then, the Fisher's r‐to‐z transformation was applied to each subject's connectivity matrix.

Previous studies reported spatial reorganization of RSNs and shifts in the location of functional regions with advancing age (Chan et al., 2014; Goldstone et al., 2016; Littow et al., 2010; Sohn et al., 2015). Thus, we avoided using most commonly used brain parcellations (e.g., Yeo et al., 2011; Andrews‐Hanna et al., 2010), which are based on young samples. Instead, ICA was primarily carried out to define nodes of DMN networks. The use of ICA has recently been motivated by a simulation study which showed that the data‐driven approach for node definition yields more accurate results compared to using predefined ROIs (Yu et al., 2017).

2.4.3. Further preprocessing

After using ICA to identify DMNs, it is important to reduce unwanted variability from DMN nodes before computing internetwork correlations (on the nodal‐level) using an average TC from each network/subnetwork. To do so, additional preprocessing steps were carried out to reduce the effects of motion and physiological noise on TCs of ICA‐derived nodes. First, to suppress local spikes in TCs, co‐registered fMRI data were despiked using the 3dDespike function in AFNI (http://afni.nimh.nih.gov/pub/dist/doc/program_help/3dDespike.html). Similar to what was done before ICA, the usual preprocessing steps were performed (segmentation, template generation and normalization using DARTEL, smoothing, and reslicing images at voxel‐size of 3 mm). Next, the effect of motion and different kinds of physiological noise were removed from the data using the regression model including 24 motion parameters (Yan et al. 2013), as well as global signal, white‐matter signal, and cerebrospinal fluid signal (extracted from unsmoothed images), and also linear and quadratic trends. Then, the data was band‐pass filtered (0.01–0.1 Hz). Finally, spherical ROIs with a radius of 4 mm were centered on the nonartefactual local maxima (peaks, t > 29; this threshold ensures the distinctiveness of all local maxima) of five DMN‐related spatial ICA maps (Table 1). This size of the sphere has been used in studies targeting RSC (Park & Chun, 2009). Overall, 21 nonoverlapping nodes were generated and TCs from these nodes were extracted from nuisance‐corrected images and used for further analyses. Also, nodal‐based FC between DMN regions was computed by averaging significant FCs (p < .05, FDR corrected) among all node‐pairs belonging to different RSNs. We then examined the internetwork FC of MTL to RSC and other posterior DMN areas (using both ICA‐based and nodal‐based FC). Partial correlation was used to explore the relation between FC and offline measures. Frame‐wise displacement (FD) (Power et al., 2012), FD squared, age, and sex were used as covariates of no interest in all statistical analyses. Finally, to compare correlation results, we used Bayesian analysis with an uninformed prior and calculated the posterior probability distribution for the difference in the strength of correlations.

Table 1.

ICA‐derived ROIs, which are local maxima of different DMN components (t > 29). All coordinates are in standard MNI space

Network label	MNI coordinates			t value	Area specification
MTL	24	−10	−16	32,96	Anterior right hippocampus
	−20	−14	−18	32,77	Anterior left hippocampus
	18	−22	−10	30,44	Posterior right hippocampus
	−14	−24	−10	29,89	Posterior left hippocampus
RSC	6	−54	14	33,64	Retrosplenial (BA30)
	−4	−54	14	33,58	Retrosplenial (BA30)
	−6	−46	2	31,57	‐
	8	−44	4	30,77	‐
pDMN	−2	−28	30	41,52	Left posterior cingulum (BA23)
	−2	−44	22	42,35	Left posterior cingulum
	−4	−70	32	41,02	Left cuneus
	10	−58	30	38,04	Right precuneus
paDMN	0	−56	26	46,85	Precuneus (BA 23)
	4	62	22	38,48	Right medial frontal
	−8	64	16	33,3	Left medial frontal
	−8	54	38	33,19	Left medial frontal
	2	62	−4	32,66	Right middle orbitofrontal
aDMN	−4	44	8	43,32	Left anterior cingulum
	6	42	10	42,86	Right anterior cingulum
	−4	52	−2	41,92	Left middle orbitofrontal
	−4	58	10	38,51	Left medial frontal

Note. Abbreviations: aDMN = anterior default mode network; BA = Brodmann area; MTL = medial temporal lobe; pDMN = posterior default mode network; paDMN = posterior–anterior default mode network; RSC = retrosplenial cortex.

2.5. Multilevel mediational analysis

In mediational models, the relationship of a predictor variable X to the outcome variable Y may be influenced by a third variable M (Figure 4a). The association of X and Y, which is the superposition of the direct effect (c′) and the mediational effect, is called total effect (c in Figure 4a). The mediational effect can be split into two parts. The first part concerns the association of the predictor variable with the mediator variable (path a in Figure 4a) and the second part addresses the relationship of the mediator variable to the outcome variable (path b in Figure 4a), controlling for the effect of the predictor variable. The mediational effect is computed by multiplying the two paths (paths a and b in Figure 4a).

Figure 4.

(a) The leftmost figure illustrates the mediational model. In the middle and rightmost panels, the average TC for RSC served as the mediator (M), whereas the average TCs from MTL regions (4 nodes) and cortical DMNs (13 nodes) were either predictors (X) or outcome variables (Y), respectively (ps < 1e‐5 are shown by ***, while larger p values are italicized and shown in parentheses). (b) The leftmost figure illustrates direct effects (i.e., c′ path) between MTL and cortical DMN while controlling for the RSC signal; the middle figure reflects total effects for the associations between MTL and cortical DMN adding the mediational effect to the model; the rightmost figure illustrates mediational effects (i.e., ab path) of RSC. All paths shown are significant at p < 0.05, FDR‐corrected across 104 models. Both number and strength of connections from MTL to cortical DMN areas increase after adding the RSC mediational effect, such that the average of the MTL‐cortical DMN FC increased greatly after adding RSC as mediator. List of abbreviated network names is given in caption for Figure 1

Mediational analysis was used to measure the average mediation effect of RSC on the MTL‐cortical DMN link. Such mediation, if exists, indicates that RSC accounts for some of the association between MTL and cortical DMNs. Multilevel mediational analysis on the FC patterns was performed using the MATLAB functions available in (http://wagerlab.colorado.edu/tools; Wager et al., 2008). In this type of two‐level analyses, level 2 refers to individuals, whereas level 1 refers to different observations across time. Thus, level 1 units are nested within level 2 units (Kenny et al., 2003). To determine mediational path coefficients (a, b, and c) at the first level, ordinary least squares (OLS) regression models were conducted at the subject‐level and path coefficients were estimated (Baron & Kenny, 1986). Then, the mixed‐effect model was used considering subject‐level path coefficients as random variables with random intercepts and slopes, for the group‐level inference.

Mean‐centered TCs extracted from MTL (4 nodes) and cortical DMN (13 nodes) areas were averaged to have a representative TC for the MTL and cortical DMN, respectively. Then, multilevel mediation analysis was performed for MTL‐cortical DMNs, while using the average BOLD signals from RSC (4 nodes) as the mediator (Figure 4a). The group‐level path coefficients are illustrated in (Figure 4a). In another attempt, to see mediational effect of RSC at nodal‐level, nodal‐level TCs extracted from the MTL and cortical DMN were set as either predictor or outcome variables, resulting in 104 different models.

Multilevel mediational analysis was performed for each pair of nodes with the average BOLD signals from RSC (4 nodes) serving as the mediator. Group‐level mediation measures (mediation effect, direct effect, and total effect) were FDR‐corrected (false discovery rate set at 0.05) across all 104 different models and illustrated by color‐coded links between the predictor and outcome nodes (Figure 4b).

Also, the FDR‐corrected subject‐level total effect (c) and subject‐level direct effect (c′) values, were averaged across 52 models, to be representative of either the MTL to cortical DMN or cortical DMN to MTL FC with and without (i.e., controlling for mediator's) mediation effect, for each subject.

As a control analysis, a similar model was run with the pDMN (primarily PCC and precuneus) and parahippocampus (pHC) (Ward et al., 2014) as the mediator variables. Furthermore, the association between the mediational effects of RSC (i.e., standardized mediation coefficients) and EM was investigated.

Note that in the typical mediation model, a mediating system translates into how an input leads to a response (Baron & Kenny, 1986). However, in our model, neither the mediator (RSC signal) nor the output (e.g., MTL signal) is caused by the input (e.g., DMN signal). Rather, these variables are only statistically related, which implies no causality.

2.6. Graph‐theory analysis

To obtain a centrality measure of RSC, graph‐theory based analysis was carried out (Sporns, 2002). Using this approach, the DMN was defined as a graph in which nodes represent different ICA‐derived regions and edges represent interregional FC. Within a graph, some nodes, known as hubs, are more important than others in terms of information transfer (Sporns et al., 2007; Van den Heuvel & Sporns, 2013). Various centrality measures have been introduced to measure the importance (hubness) of nodes within a network (Sporns et al., 2007). Among them, BC (Freeman, 1977; Sporns et al., 2007) measures the proportion of short paths among all pairs of nodes in a network that pass through a particular node. Thus, a node with high BC mediates a higher proportion of traffic while controlling information passing between others. Then, BC was the centrality measure of interest in this study.

Graph analysis was performed using 21 defined nodes within DMN‐related RSNs. FC between nodes (i.e., edges) was measured by computing the Pearson correlation coefficient between the TCs of each pair of nodes, which resulted in an FC matrix for each subject. Nonsignificant connections (those that did not survive p < .05, FDR‐corrected) were removed from the subject‐level matrices. For group‐level thresholding, a binary matrix was created such that each element of the matrix (representative of each connection) is equal to 1 if the connection exists consistently in more than 70% of participants. The binary matrix was used as a mask and applied to individual connectivity matrices to retain only robust connections across the sample. We assume that connections that are consistently detected in a large percentage of individuals are more likely to exist than connections detected in only a few subjects. Group‐level thresholding method (de Reus & van den Heuvel, 2013a) minimizes false‐positive and false‐negative edges in the network architecture, at the expense of reducing interindividual differences in FC. Thus, this approach is mainly useful for healthy age‐homogeneous samples, as it may conceal interindividual variability because of age or disease. To check the robustness of the method, a range of different group‐level thresholds from 60% to 80% for creating the agreement matrix led to similar findings (data not shown). Finally, BC measures were computed for the average connectivity matrix and for each subject separately, using the “betweenness_wei” MATLAB function available in the Brain Connectivity Toolbox (https://sites.google.com/site/bctnet/). BC was then normalized by dividing it into (N − 1) × (N − 2), where N is the number of nodes in the network (21 nodes for ICA‐derived nodes). For graphical illustration, BrainNet Viewer (Xia et al., 2013, http://www.nitrc.org/projects/bnv/) was used.

BC is an ideal measure to find the most important node for communication control and information transfer within the DMN. However, other centrality measures might rank the regions in different orders (e.g., Zuo et al., 2012). Thus, as a complementary analysis, we also computed a composite hub score using average of four different unity‐based normalized measures reported in the literature (degree, strength, eigenvector centrality, and BC). Compared to nodal degree which is simply the number of direct connections between a given node and the other nodes in the network, nodal strength measures the sum of connection‐weights of that node. Eigenvector centrality is computed by considering the overall connectivity pattern of the network and based on the concept that being connected to the nodes with higher centrality is more beneficial than nodes with lower centrality. Therefore, a node with high eigenvector centrality represents a highly connected node which is connected to the other highly connected nodes (Van den Heuvel & Sporns, 2013; Zuo et al., 2012).

2.7. Effect of parcellation on DMN connectivity

Although in the main analyses, we avoided using predefined ROIs reported in the most commonly used brain parcellations (e.g., Andrews‐Hanna et al. 2010; Yeo et al., 2011), which are based on young samples, control analyses were carried out using 11 DMN nodes as reported in (Andrews‐Hanna et al., 2010). Specifically, all previous analyses were rerun using DMN nodes (instead of ICA‐derived nodes) reported in the aforementioned study.

3. RESULTS

3.1. Mapping resting‐state DMNs

Using ICA, we identified 5 DMN components. These included pDMN (mainly PCC [Brodmann Area (BA) 23] and precuneus [BA 31]), aDMN (medial prefrontal cortex [BA 32 and BA 10]), RSC (BA 30), and the MTL network (Figure 1). The coordinates of the medial parietal peaks in the RSC component (Table 1) were in good agreement with coordinates of the RSC reported in previous studies (Dillen et al., 2016; Holzschneider et al., 2012; Park & Chun, 2009). Note that we only focused on the most commonly used DMN subsystems (the key players in understanding MTL‐cortical DMN interactions). Therefore, we did not include cerebellar regions, though they have been suggested, to some extent, to be part of the DMN (Buckner et al., 2011; Habas et al., 2009). ICA‐based FCs of interest (among MTL and any of the other DMN‐related components) were significantly different from zero (ts (181) > 6, ps < 8.54e‐09). Among the DMN components, degree of internetwork connectivity between MTL and RSC (mean ± SD = 0.32 ± 0.27) was reliably stronger (t (181)s > 5.44, ps < .001) than for MTL‐pDMN (pDMN: mean ± SD = 0.22 ± 0.23; paDMN: mean ± SD = 0.19 ± 0.18) and MTL‐aDMN (mean ± SD = 0.10 ± 0.23). Similar to the results from ICA‐based FCs, nodal‐level analysis revealed that the degree of MTL‐RSC connectivity (mean ± SD = 0.25 ± 0.25) was greater (pDMN: t (181) = 11.48, p = 0; paDMN: t (181) = 7.21, p = 1.49e‐11; aDMN: t (181) = 5.36, p = 2.50e‐07) compared to both MTL‐pDMNs (pDMN: mean ± SD = −0.04 ± 0.29; paDMN: mean ± SD = 0.10 ± 0.25) and MTL‐aDMN (mean ± SD = 0.12 ± 0.26). Nodal‐based MTL‐RSC FC (t (181) = 13.14, p = 3.87e‐28), MTL‐paDMN FC (t (181) = 5.55, p = 9.80e‐08) and MTL‐aDMN FC (t (181) = 6.14, p = 4.98e‐09) were significantly >zero, except MTL‐pDMN FC, which was only marginally significant (t (181) = −1.71, p = .09).

Figure 1.

Networks derived from independent‐component analysis (ICA) and their regions of interest (ROI) peaks (t > 29). These networks include posterior default mode network (pDMN), posterior–anterior DMN (paDMN), anterior DMN (aDMN), retrosplenial cortex (RSC), and the medial temporal lobe (MTL)

3.2. MTL‐RSC and cognitive performance

In relating FC between MTL and the different DMN components to interindividual differences in the EM composite score, we found a significant EM association (r = .19, p = .01) (Supporting Information, Figure 1) for MTL‐RSC FC. Measures of association with EM for the connectivities between MTL and pDMNs were not significant (pDMN: r = .13, p = .09; paDMN: r = .04, p = .59).

Results from the Bayesian inference suggested a high posterior probability (>97%) in favor of stronger degree of correlation between EM and MTL‐RSC FC than EM association with FC of MTL and other cortical DMNs (Supporting Information, Figure 2a). Note that the mean of the posterior probability distribution of the correlation difference is larger than zero (Supporting Information, Figure 2a). Similar to ICA analyses, a positive association with EM was found for nodal‐level MTL‐RSC FC (r = .19, p = .01) (Figure 2a, left column). No significant associations with EM were found for MTL‐pDMNs (pDMN: r = .07, p = .37; paDMN: r = .08, p = .27) or MTL‐aDMN (r = .03, p = .67). Similar to the ICA‐based analyses, we found a high posterior probability (>98%) that the degree of correlation with EM for MTL‐RSC FC is greater than its correlation with FC between MTL and other cortical DMNs (Figure 3a). Further control analysis was performed by applying median split on EM measure to define low and high EM performers. Then bootstrapping (10,000 resampling) was carried out to explore whether low and high EM performers differ in terms of MTL‐RSC FC, using 95% bootstrap confidence interval (CI) based on Efron and Tibshirani (1993). We found significant difference for MTL‐RSC FC (CI: 0.23–0.73). The same procedure was performed for FC between MTL and three other cortical DMNs, but no significant difference in these connections between two groups was found (Figure 2a, right column). Data for different measures across the whole sample and the two groups (i.e., low and high EM performers) are reported in Table 2.

Figure 2.

(a) The scatter plot shows a positive association with EM for nodal‐based MTL‐RSC FC. FC values are z‐transformed. The rightmost plot illustrates bootstrap resampling of the average difference of MTL connectivity to different DMNs between low and high performers of EM. Only the confidence interval for MTL‐RSC FC does not include zero, suggesting that MTL‐RSC FC significantly differs between these two groups. However, there is no evidence that MTL FC to other DMNs is different between the two groups (i.e., the confidence intervals cross zero). (b) The scatter plot shows positive associations between the standardized RSC mediational effects and EM in both directions (i.e., MTL–>DMN and DMN—>MTL). The rightmost plot illustrates bootstrap resampling of the average difference in mediation effect of RSC between low and high performers. The confidence intervals do not include zero and show that the mediation effects of RSC differ significantly between the two groups. For abbreviations, see Figure 1

Figure 3.

The posterior probability distribution for the difference in correlations using nodal‐based FCs. (a) The degree of association between episodic memory and MTL‐RSC FC was greater than that between MTL and other cortical DMNs, with a high posterior probability (>98%). (b) There was high posterior probability (> 78%) that the strength of the MTL–RSC FC association with episodic memory was different from the MTL–RSC FC link with working memory and speed. See Figure 1 for decoding of the abbreviated network names. EM, episodic memory; WM, working memory

Table 2.

Results for different measures across the whole sample and for low and high performers in episodic memory. t tests were conducted to compare low and high performers across measures

	Age	FD	EM	WM	Speed
Whole sample	66.13 ± 1.21	0.19 ± 0.11	50.05 ± 7.82	50.09 ± 7.46	49.90 ± 8.58
Low performers	66.27 ± 1.18	0.20 ± 0.13	44.08 ± 3.99	47.35 ± 7.77	48.96 ± 8.85
High performers	65.99 ± 1.22	0.19 ± 0.10	55.96 ± 5.96	52.81 ± 6.04	50.82 ± 8.24
	t = 1.56, p = 0.12	t = 0.73, p = 0.47	t = −15.65, p = 3.12e‐35	t = −5.25, p = 4.30e‐07	t = −1.45, p = 0.15

Note. Abbreviations: EM, episodic memory; FD, frame‐wise displacement; speed, perceptual speed; WM, working memory.

All values are given in mean ± SD.

No significant association was found between MTL‐RSC FC with either working memory (r = .10, p = .17) or perceptual speed (r =.06, p = .41). Using posterior probability distribution, we found that there is high (>78%) probability that the correlation between MTL‐RSC FC and EM is larger than the corresponding correlation with working memory and speed (Supporting Information, Figures 2b and 3b for ICA‐based and nodal‐based FCs, respectively).

We found no significant difference (t (178) = −0.45, p = .65) in the degree of FC between RSC with the anterior (r = .23) or posterior (r = .22) MTL clusters. Similarly, we found significant associations with EM for both anterior MTL‐RSC (r = .15, p = .05) and posterior MTL‐RSC (r = .18, p = .02) FCs. Based on the posterior probability distribution (Supporting Information, Figure 3), there is high (> 88%) probability that the correlation between posterior MTL‐RSC FC and EM is greater than the correlation between anterior MTL‐RSC FC and EM.

3.3. RSC mediates FC between MTL and cortical DMN

Next, we tested the hypothesis that RSC serves as an intermediate region, interrelating MTL with cortical DMNs. First, in the multilevel mediation analysis on the average mean‐centered TCs for the MTL, cortical DMN, and RSC, we found significant mediation effects (ab > 8; ps < 1e‐5, second and third columns) regardless of the direction of MTL‐cortical DMN connectivity (Figure 4a). Second, we generated 104 different models for all possible nodal interactions between MTL regions (4 nodes) and cortical DMN regions (13 nodes), using a multilevel mediation approach. An average signal from 4 nodes within RSC served as mediator in all 104 models (Figure 4a, first column). Significant positive group‐level path coefficients (p < .05, FDR‐corrected) are portrayed in Figure 4b. A few significant connections between MTL and the cortical DMNs were observed when the RSC signal was regressed out (Figure 4b, left). By contrast, several regions in the cortical DMN exhibited connectivity with the MTL, when RSC was added as a mediator to the direct effect between MTL and cortical DMNs (Figure 4b, middle). Finally, the direct comparison between the MTL‐cortical DMNs with and without RSC as a mediator confirmed a very strong mediational effect of RSC (Figure 4b, right). Consistent with the latter observation, a one‐sample t test on the average value of RSC mediation (i.e., ab (p < .05, FDR‐corrected) in Figure 4a) revealed significant effects (MTL to cortical DMN: t (181) = 12.51, p = 2.73e‐26; cortical DMN to MTL: t (181) = 12.58, p = 1.77e‐26). Taken together, these results support the notion that RSC is a gateway that facilitates MTL–cortical DMN interactions.

Furthermore, we explored whether interindividual differences in the RSC mediation effect (ab in Figure 4a) was associated with EM performance. Results revealed a positive association (MTL to cortical DMN: r = .20, p = .007; cortical DMN to MTL: r = .17, p = .02) (Figure 2b, left), suggesting that individuals with greater RSC mediation have an EM advantage. In addition, control analysis was performed by applying median split on EM measure to define two groups of high‐ and low‐performers. Then bootstrapping (10,000 resampling) was carried out to explore whether high and low performers differ in terms of the mediation effect of RSC, using 95% bootstrap CI. We found significant difference for mediation effect of RSC (CI of MTL to cortical DMN: 0.22–0.75; CI of cortical DMN to MTL: 0.13–0.16; Figure 2b, right), suggesting that high performers exhibit greater RSC mediation effect.

To investigate the specificity of the RSC mediation effect on MTL‐cortical DMN connectivity, a series of control analyses were carried out to investigate whether (a) the pDMN network (mainly including PCC and precuneus) mediates the interaction between MTL and cortical DMNs; and (b) the pHC network mediates the interaction between MTL and cortical DMNs (Ward et al., 2014). We found no mediation effect for PCC (Supporting Information, Figure 4), and relatively weak mediation only in one direction for pHC (Supporting Information, Figure 5). Based on the mediation effects illustrated in Supporting Information, Figure 5B, third column, this weak mediation effect seems to be driven by the inclusion of RSC among cortical DMN ROIs.

Figure 5.

(a) Normalized betweenness centrality (BC) in color format for each node listed in Table 1. The top and middle rows show lateral and medial views of the left and right hemispheres, respectively. The bottom row illustrates the ventral view of the left and right hemispheres. Values are unity‐based normalized for illustration. (b) Violin plots showing the distribution of average BC values in each DMN subsystem. BC was higher in RSC compared to the other DMN subsystems (t test: RSC and MTL: t (172) = 46.77, p = 8.23e‐99; RSC and pDMN: t (172) = 18.92, p = 2.79e‐43; RSC and paDMN: t (172) = 15.73, p = 8.23e‐35; RSC and aDMN: t (172) = 18.29, p = 9.19e‐42). Reported p values are FDR‐corrected across four comparisons. Asterisks and connecting lines denote significant t‐test comparisons: *p < .0001. See caption for Figure 1 for the list of abbreviations

3.4. Centrality of RSC within DMN

Given the important role of RSC in MTL‐DMN coupling, it is reasonable to assume that RSC has a central role in controlling information transfer within the DMN. To address this issue, we computed BC for all regions listed in Table 1. Results show that RSC nodes have the highest BC compared to other DMN nodes (Figure 5a; RSC is marked with an arrow). Figure 5b shows the distribution of average BC values of all nodes within each network (MTL: mean ± SD = 0.01 ± 0.01, pDMN: 0.03 ± 0.01, paDMN: 0.03 ± 0.02, aDMN: 0.03 ± 0.01, RSC: 0.06 ± 0.02). ANOVA with repeated measures revealed a main effect of DMN subnetwork (F = 420.37, p = 1.27e‐200). Follow‐up t tests confirmed that BC is significantly larger for the RSC network compared to the other DMN subsystems (ts > 15.73, ps < 8.23e‐35, FDR corrected across four comparisons). Thus, RSC is located on the most traveled paths within the DMN, by having the highest BC.

Similarly, results from the composite centrality measure show that RSC had the highest hub score among all DMN nodes in the average connectivity matrix (Supporting Information, Figure 6a; RSC is marked with an arrow). We also computed hub score for each node at the subject‐level followed by averaging across subjects. Similar to the initial analysis, the first two nodes of the RSC network, located in BA 30, showed the highest composite hub score (Supporting Information, Figure 6b). ANOVA with repeated measures revealed a main effect of region (F = 1467.48, p = 0). Follow‐up t tests confirmed that there is significantly larger hub score for the first two nodes of RSC network compared to the other DMN nodes (ts > 10.79, ps < 2.78e‐20, FDR corrected across 2 × 19 comparisons).

3.5. Control analyses using an alternative DMN parcellation

A series of control analyses were conducted using 11 DMN nodes as reported by Andrews‐Hanna et al. (2010). Similar to the results reported above, we found that MTL‐RSC FC was still greater than MTL‐PCC FC (t (179) = 5.94, p = 1.44e‐08) (Supporting Information, Figure 7). In addition, mediation analyses revealed a significant mediation effect of RSC (MTL to cortical DMN: t (181) = 9.84, p = 1.47e‐18; cortical DMN to MTL: t (181) = 8.34, p = 1.84e‐14) (Supporting Information, Figure 8). Finally, the RSC node still showed the highest BC compared to other DMN nodes (ts > 14.5, ps < 4.79e‐31, FDR corrected across 10 comparisons) (Supporting Information, Figure 9). However, we did not find a significant relationship of MTL‐RSC FC and EM (r = .03, p = .68). Taken together, we replicated our main findings using the DMN fractionation reported in Andrews‐Hanna et al. (2010).

4. DISCUSSION

We sought to provide a characterization of FC within the DMN with a key focus on the potentially critical role of RSC in older adults. Specifically, we examined heterogeneity in FC between the MTL and different parietal regions within the DMN, and their relationships with EM in a sample of older adults. Using ICA, we identified the MTL component and two distinct posterior cortical DMN components, which primarily included PCC and RSC. Several resting‐state studies have found cortical and MTL DMN subsystems in younger (Andrews‐Hanna et al., 2010) and older (Salami et al., 2014, 2016) adults. Furthermore, past work reported structural and functional variations within the medial parietal DMN, particularly between RSC and neighboring PCC (Dillen et al., 2016; Kobayashi & Amaral, 2007; Leech et al., 2011; Summerfield et al., 2009; for a review, see Vann et al., 2009). Here, we demonstrate heterogeneity in degree of connectivity between MTL and different medial parietal DMNs, with greater FC of the MTL with the RSC as compared with the PCC. A previous study of functional fractionation of the DMN in younger adults revealed different subsystems, with RSC and PCC belonging to the medial temporal lobe and midline core subsystems, respectively (Andrews‐Hanna et al., 2010). However, evidence for heterogeneity in connectivity between MTL and different parietal DMNs in old age is scarce. Our observation of greater MTL‐RSC FC as compared with MTL‐PCC FC suggests that the overall functional hierarchies (i.e., heterogeneity in degree of connectivity) within DMN remain preserved in healthy older adults, although the degree of connectivity might change (Andrews‐Hanna et al., 2007; Avelar‐Pereira et al., 2017; Salami et al., 2016).

The analysis of the relationship of MTL‐cortical DMN FC to EM revealed a selective pattern, such that the degree of MTL‐RSC FC was positively associated with EM performance (with no significant association between MTL‐RSC FC and working memory or perceptual speed). The observed associations with EM were modest. However, they were found in a highly controlled experimental context using an age‐homogenous sample. Our finding is in line with results showing that intrinsic connectivity between MTL and posteromedial DMN is predictive of EM performance (Kucyi & Davis, 2014; McCormick et al., 2013; Salami et al., 2014; Vincent et al., 2006; Wang et al., 2010), although the EM association for MTL‐RSC in older adults has not been reported before. Relatedly, reduced FC between MTL and pDMN has been reported in patients with mild cognitive impairment and Alzheimer's disease (Bai et al., 2009; Celone et al., 2006; Greicius et al., 2004; Petrella et al., 2011). HC and surrounding MTL structures have a well‐established role in EM (Eichenbaum, 2004; Salami et al., 2010; Squire et al., 2004). RSC has also been found to be engaged in studies of EM (Maguire, 2001a; Svoboda et al., 2006), although its differentiation from PCC regions has been less investigated (Maguire, 2001b; Vann et al., 2009). Similarly, neuropsychological studies of patients with RSC lesions indicate marked memory impairment (Maguire, 2001b; McDonald et al., 2001; Valenstein et al., 1987). Unsurprisingly, it has been difficult to rule out the possibility that lesions in neighboring PCC also contributes to memory impairment (Maguire, 2001b). Toward this end, the association between MTL‐PCC FC and EM did not reach statistical significance, which suggests that connectivity between MTL and RSC is critical to EM. Taken together, the current findings provide novel evidence of heterogeneity across medial parietal DMN in FC to the MTL, and in relation to EM.

It is however important to note that previous studies showed variations in FC along the anterior–posterior axis of MTL (Salami et al., 2016; Adnan et al., 2016; Poppenk et al., 2013). Our MTL component encompasses both the anterior and the posterior regions based on the apex landmark (Y = −21). We found no significant difference in the degree of FC between RSC with the anterior or posterior MTL clusters. The absence of a stronger association between posterior MTL and RSC as compared to anterior MTL and RSC may reflect age‐related reduced selectivity in the functional architecture of the brain (i.e., dedifferentiation; Grady, 2012) in aging.

Path modeling demonstrated that both strength and number of connections between MTL and the cortical DMN increased with RSC as a mediator. This finding suggests that RSC mediates the input and output streams of the MTL to cortical DMN regions. A previous study showed that FC between HC and the cortical DMN is mediated via the pHC, a region with strong connections to the RSC (Ward et al. 2014). However, we found little evidence for a mediational role of the pHC in this context (Supporting Information, Figure 5). Our finding that pHC only mediated the interaction between MTL and RSC may suggest that pHC and RSC are sequential interfaces between HC and cortical DMN. Conceivably, this discrepancy stems from differences in methodology (RSC was included as part of PCC area in Ward et al., 2014) and sample characteristics, such as participant age. Our analysis of the link between degree of mediation and EM showed a positive association. Thus, larger RSC mediation was associated with higher EM performance. Overall, our results provide novel evidence that RSC constitutes a gateway by linking MTL to other DMN subsystems. This finding concords with animal work on structural connectivity in which RSC is regarded as an intermediate cortex (Kobayashi & Amaral, 2000; Morris et al., 1999).

Graph analysis revealed higher BC and higher hub score for RSC nodes compared to other DMN nodes, substantiating the important role of RSC for global coordination of information flow across the DMN. This finding along with mediational role of RSC suggest that this region is a key hub within the DMN (He et al., 2009; Zuo et al., 2012). Previous research has failed to demonstrate a central role for RSC in the context of the whole brain. Precuneus and/or PCC have been identified as critical hubs (Achard et al., 2006; Leech et al., 2011; Utevsky et al., 2014; Zuo et al., 2012) with dense structural and functional connectivity to many other regions across the whole brain (Hagmann et al., 2008). Using a graph‐theoretical approach, a previous study in younger individuals reported relatively high BC across the whole brain for the dorsal PCC/precuneus (Spreng et al., 2013). In contrast, low BC across the whole brain was found for the ventral PCC (Spreng et al., 2013), with dense FC essentially restricted to DMN (Leech & Sharp, 2014; Leech et al., 2011, 2012; Spreng et al., 2013; Tomasi & Volkow, 2010). This finding along with our observation of higher BC for RSC within the DMN compared to the other DMN regions suggests that RSC is a provincial hub that facilitates interaction within, but not outside (i.e., across the whole brain), the DMN. Our finding is in contrast to previous findings where PCC exhibits the dense FC (Fransson & Marrelec, 2008; Tomasi & Volkow, 2010) and the highest BC within the DMN (Andrews‐Hanna et al. 2010). Such differences could have been partly accounted for by differences in sample demographics (i.e., age). Having access to a canonical model of connectivity in normal older adults may facilitate identifications of severe cognitive decline among older adults. Moreover, future studies may use alternative methods such as seed‐based approach, which has been shown to yield a stronger correlation between EM and MTL‐medial parietal FC (McCormick et al., 2013).

In sum, our results suggest that RSC is an important gateway to EM by linking subcortical and cortical subsystems of the DMN in older adults.

Supporting information

Additional Supporting Information may be found online in the supporting information tab for this article.

Supporting Information

Kaboodvand N, Bäckman L, Nyberg L, Salami A. The retrosplenial cortex: A memory gateway between the cortical default mode network and the medial temporal lobe. Hum Brain Mapp. 2018;39:2020–2034. 10.1002/hbm.23983