Three hundred sixteen right-handed, healthy subjects were selected from the Open Access Series of Imaging Studies (OASIS) cross-sectional database (http://www.oasis-brains.org) (Marcus et al., 2007). Data from seven subjects were excluded from further analysis because of poor image quality or image preprocessing. Because of relatively few subjects aged 30–60 years, we first selected 80 subjects from this age range as the middle-aged group (30 males and 50 females). Then, we selected 80 separate subjects around 75 years of age as the old group. Finally, we selected the youngest 80 subjects matched for gender as the young group. The names and characteristics of the groups are shown in Table 1. All subjects were evaluated using the Mini-Mental State Examination (MMSE) (Folstein et al., 1975) and Clinical Dementia Rating (CDR) scales (Morris, 1993; Morris et al., 2001). MMSE scores were higher than 29, and CDR scores were all zero. For demographic data on all subjects, see Marcus et al. (2007). This dataset has been used in several previous studies (Bakkour et al., 2009; Fjell et al., 2009; Salat et al., 2009; Li et al., 2011).

For each subject, three to four individual T1-weighted magnetization-prepared rapid gradient echo (MP-RAGE) images were acquired on a 1.5T Vision scanner (Siemens, Erlangen, Germany) within a single session. Head movement was minimized with cushioning and a thermoplastic facemask. Images were motion corrected and averaged to create a single image with a high contrast-to-noise ratio. MP-RAGE parameters were empirically optimized for gray/white contrast: TR = 9.7 ms; TE = 4 ms; flip angle = 10°; slice number = 128; resolution = 256 × 256 (1 × 1 mm); thickness = 1.25 mm.

Structural MR images were processed using a technical computing software program (MATLAB 2010; The MathWorks Inc., Natick, Mass) and Statistical Parametric Mapping software (SPM 8; The Wellcome Department of Imaging Neuroscience, London, UK). Following the inspection of image artifacts, image preprocessing was performed with the VBM8 toolbox (http://dbm.neuro.uni-jena.de/vbm/). Briefly, all native-space MRIs were segmented to extract GM based on an adaptive maximum a posteriori technique (Rajapakse et al., 1997) and partial volume estimation method (Tohka et al., 2004) without the need for a priori tissue probability information. In addition, a spatially adaptive non-local denoising filter (Manjon et al., 2010) and a hidden Markov random field model (Rajapakse et al., 1997) were applied to minimize the level of noise in the resulting GM segments. Subsequently, the high-dimensional DARTEL (diffeomorphic anatomical registration using exponentiated lie algebra) approach provided non-linear deformation to normalize the images to the DARTEL template in Montreal Neurological Institute (MNI) space, which was derived from 550 healthy control subjects (Ashburner, 2007). Non-brain tissue was also removed. Additionally, the Jacobian determinants derived from the spatial normalization were used to modulate the GM value for each voxel to preserve the total amount of GM from the original images (Good et al., 2001). We used non-linear components only, which allowed us to analyze relative (i.e., corrected for individual brain size) differences in regional GM volume. Finally, the resulting modulated and normalized images were smoothed with a 12 mm full width at half maximum isotropic Gaussian kernel.

To assess the structural covariance pattern of each large-scale network, we extracted individual GM volumes from eight seed regions of interest (ROIs). These ROIs included the primary visual, auditory, and motor cortex, as well as language-related speech and semantic areas, areas related to salience and executive control, and the default-mode network (DMN). We based our ROIs on previous studies (Zielinski et al., 2010; Montembeault et al., 2012). For each region, seeds were defined with the MarsBar ROI toolbox (http://marsbar.sourceforge.net/) as 4-mm radial spheres centered at the following MNI coordinates: right calcarine sulcus (9, −81, 7), right Heschl's gyrus (46, −18, 10), right precentral gyrus (28, −16, 66), left inferior frontal gyrus, pars opercularis (IFGo) (−50, 18, 7), left temporal pole (−38 10, −28), right frontoinsular cortex (38, 26, −10), right dorsolateral prefrontal cortex (DLPFC) (44, 36, 20), and right angular gyrus (46, −59, 23). In addition, we selected contralateral seeds by changing the sign on each seed's x coordinate to perform similar SCNs analyses (Table A1, and Figure A1).

A voxel-wise statistical analysis was performed on the GM images using the GLM as applied in SPM8. Eight multiple regression models were used to test the strength of the structural covariance between each seeds and all other regions across whole-brain GM for each age group. In each regression model, the extracted mean GM volume from each ROI was entered as a covariate of interest, and gender as a confounding covariate. Because of the unequal sample size across the genders, we removed the effects of gender on the SCN patterns in the correlation analysis. These statistical analyses identified voxels that had a positive covariance with each a priori selected ROI in each group. The resulting correlation maps were using height and extent thresholds at P < 0.05 with family-wise error (FWE) correction. They were displayed on the MNI template to allow qualitative comparisons between the age groups using BrainNet Viewer software (http://www.nitrc.org/projects/bnv/). To quantify differences in normal aging trajectories across networks, we calculated the total number of significant positive ipsilateral, contralateral, and whole-brain voxels and plotted these across the age groups.

To further examine the effects of age on specific regional covariance, we performed between-group difference analyses of SCN patterns for any two groups according to the scientific literature (Lerch et al., 2006). For any pair of voxels in two groups, their structural correlation may have different slopes, and the difference in slope may represent the difference in their structural association. The difference in slope was tested using a classic interaction linear model: Vi=β0+β1Vj+β2Group+β3(Vj×Group)+ε where V_i and V_j represented the GM volumes of a pair of voxels in two groups. The Group component was modeled using treatment contrasts, and significance tested using the Student's t statistic. Specific t-value contrasts were established to map the voxels that expressed a significantly different structural association between any two groups. The threshold for the resulting statistical parametric maps was given at P < 0.05 for the height and extent thresholds, with FWE multiple comparisons correction.

Seeds within primary visual cortex (right calcarine sulcus) produced SCNs with a relatively preserved pattern (Figure 1). However, there were small changes in the distributions of the covariance maps. In the young and middle-aged groups, the covariance regions mainly included the bilateral calcarine sulcus, lingual gyrus, cuneus and right lateral occipital gyrus, whereas only included bilateral calcarine sulcus, lingual gyrus, and cuneus in the old group. Primary auditory cortex (i.e., the right Heschl's gyrus) covaried with the bilateral insula and precuneus, the right posterior cingulate cortex (PCC), parahippocampus, and inferior frontal gyrus, and the left orbital-frontal cortex in the young participants, and underwent significant contraction in the middle-aged and old groups to include only the bilateral insula regions. There was a flat transition in the covariance maps during aging. Primary motor cortex (right precentral gyrus) correlated with the ipsilateral precentral regions in the young participants, which progressed at middle age to include the contralateral precentral and supplementary motor areas. In old participants, these structural associations were similar to the young participants.

In addition, within the primary sensory and motor networks, there were no significant differences observed in the structural association between any two groups when we compared the slopes of the structural associations on a voxel-by-voxel basis.

The two language-related speech and semantic networks followed similar variation patterns (Figure 2). From the young participants to the middle group, the covariance regions showed abrupt decreases. They then showed smaller changes at the following ages. For the language-related speech network, the covariance map involved the bilateral anterior insula, medial frontal, cingulate cortex, and temporal regions in the young participants, and then contracted to include seed autocorrelation in the older participants. The semantic regions correlated with the bilateral temporal cortices, cingulate gyrus, insula and frontal regions in the young participants, and shrank sharply to include only the anterior temporal cortices during normal aging.

Comparing the other SCNs, language-related speech and semantic networks showed more age-related changes. Specifically, the left supplementary motor and superior temporal areas showed significant positive associations with the speech seed region in young adults, whereas this covariance disappeared in the middle-aged group (Table 2 and Figure 3). Decreased positive associations between the left superior temporal region and left temporal pole were found in the middle-aged group compared with the young group (Table 2 and Figure 4). There were significant positive associations between the left anterior and the PCC and left IFG in the young participants, and this covariance became a negative correlation in the old group (Table 2 and Figure 4). When compared with the old group, there was a significant positive association between the right PCC and the semantic seed region in the young participants (Table 2 and Figure 5).

The three networks associated with social–emotional and cognitive function showed similar patterns to the language-related speech and semantic networks, with more distributed structural covariance in the young participants and limited covariance patterns throughout the later stages (Figure 5). Specifically, the right frontoinsular cortex anchored covariance maps included extensive areas of the bilateral lateral and medial frontal cortex, temporal cortices, and cingulate regions. This SCN underwent significant shrinkage in the middle-aged group to include only bilateral insular regions. In the older subjects, this network somewhat extended to include the bilateral medial prefrontal regions. The right DLPFC seed covaried with the bilateral frontal, temporal, and cingulate cortices in the young participants, but shrank into a more focal distribution in the middle-aged and older subjects. A seed in right angular gyrus produced an aging SCN, including the bilateral angular gyrus, middle temporal, cingulum, and prefrontal regions in the young participants, and contracting to include the bilateral PCC and angular gyrus in the middle-aged group, and only the bilateral angular gyrus in the old group.

There were no significant differences observed in the structural associations between any two groups within the salience network. The left PCC showed a significant positive association with the executive control seed region and right angular region (DMN seed) in the young adults, whereas this covariance disappeared in the old group (Table 2 and Figure 5). Furthermore, a less robust positive association between the right prefrontal region and DMN seed region was found in the middle-aged group compared with the young participants (Table 2 and Figure 4).

The cerebral cortex is organized into networks of functionally complementary areas. New advances in modern neuroimaging techniques and quantitative analysis of complex networks have made the investigation of brain network topological organization possible. DTI is a useful tool for non-invasively mapping cortico-cortical anatomical connections by examining axonal integrity. Insight into the structural covariance of GM morphology (e.g., volume, thickness, and surface area) between brain regions is provided by sMRI. Resting-state fMRI has allowed for assessments of the strength of functional connections within a network by quantifying correlated activity [i.e., spontaneous, low-frequency fluctuations in the blood oxygen level-dependent (BOLD) signal] between brain regions at rest.

The current studies have demonstrated the underlying relationships among structural correlations, anatomical connectivity, and functional correlations. For example, structural networks based on cortical thickness measurements were consistent with known functional networks (He et al., 2007). A recent study reported agreement in the correlations in GM thickness and underlying fiber connections across brain areas, but more information was included for the thickness correlation network than the fiber connection network (Gong et al., 2012). In addition, SCNs determine using VBM recapitulated the canonical intrinsic connectivity networks topologies (Seeley et al., 2009). The consistency among the SCNs and other networks provides substantial support for the use of SCNs as a measure of network integrity for cross-sectional studies.

In this study, we employed three age groups to map the trajectory of SCN changes over age. Except the primary motor network, these SCNs appeared the similar non-linear pattern that had a distributed topology in young participants, a sharply localized topology in middle aged participants, and were relatively stable in older participants. This trajectory was similar to age-related changes of integrated local efficiency of brain network (Wu et al., 2011b). They employed the graph theory analysis method and reported that the local efficiency in the young group was significantly larger than those of the middle and old groups, whereas no significant difference was found between the middle and old groups. The shrinkage of SCNs as well as the reduction of local efficiency might be explained by the regionally distributed pattern of GM atrophy (Bergfield et al., 2010). As for primary motor network, we found inverted V-curve tendency among three age groups in the SCN of primary motor cortex. The previous study reported increased inter-regional correlations between bilateral primary motor cortex in the aging brain (Chen et al., 2011). In our study, we found the structural covariance of primary motor cortex between two hemispheres in the middle group. Alternatively, we found that the volume of right precentral gyrus reduced with age increasing (Figure A2), which might reflect a compensation mechanism that a network may need to work harder by becoming overactive due to regional volumetric atrophy with the network (Reuter-Lorenz and Cappell, 2008).

We found similar SCNs across both hemispheres (Figure A1), consistent with previous SCN analyses of the developing brain (Zielinski et al., 2010). However, the auditory and speech networks showed a notable exception. The SCN derived from the left seed of Heschl's gyrus was distinctly smaller than the one derived from the right seed. In contrast, the SCN derived from the left seed of the IFGo was distinctly larger than the one derived from the right seed, which may explain why speech networks are left-dominated (Powell et al., 2006; Xiang et al., 2010).

To build upon this study, several issues need to be addressed. First, we assigned subjects in their 30s to the middle-aged group because there were fewer subjects in their 40s and 50s in our sample. This resulted in wide age range in the middle-aged group. In future, this age range should be refined by including more subjects. Second, a previous study has reported the brain functional connectivity pattern could be affected by the sex factor (Biswal et al., 2010). In future, it is interesting to explore the sexual differences when the SCNs change with age. Third, it should be highlighted that this study used ROIs based on previous studies (Zielinski et al., 2010; Montembeault et al., 2012). Future studies using ROIs extracted from ICNs of the same subjects may obtain more accurate definitions of seed regions.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.