Structural magnetic resonance imaging (MRI) studies have demonstrated aging-related anatomical alterations in the brain. As we age in general, the global gray and white matter volumes decrease and the volume containing cerebrospinal fluid increases (Good et al., 2001; Sowell et al., 2003; Lemaitre et al., 2012). Moreover, structural changes are heterogeneous across different brain regions in normal aging. The prominent aging effects were observed mainly in the frontal cortex and some parts of the temporal lobe (Fjell et al., 2009; Peelle et al., 2012; Ziegler et al., 2012; Fjell et al., 2013). Reports on these findings primarily interested in isolated regional/global brain structural changes utilizing univariate approaches, such as voxel-based morphometry (VBM) or regions of interest (ROIs). These studies laid important foundation to localize aging-associated brain areas or to investigate brain structural trajectories in the life span. Based on the knowledge gained from these previous works, it becomes possible to understand the aging-related changes by seeking out the covariance information of morphological features across various brain regions over its whole volume.

Multivariate analytical methods, which aim to determine the interrelationship among brain regions, have been extensively applied to neuroimaging studies to explore the brain systems in different population. Based on the existing literature, it is well-known that the human brain is intrinsically organized into complex functional and structural networks (Fox et al., 2005; Fox and Raichle, 2007; Bullmore and Sporns, 2009; Bassett and Gazzaniga, 2011). Recent multivariate MRI studies have revealed aging-related alterations in not only functional but also structural networks. For example, in addition to the resting-state functional magnetic resonance imaging (rs-fMRI) studies documenting age-related changes in functional connectivity in resting-state functional networks, such as the default-mode network (DMN), motor and visual networks (Ferreira and Busatto, 2013), several structural MRI studies have reported that age-associated regional covariance networks involve widespread areas of the cortex and subcortex (Alexander et al., 2006; Brickman et al., 2007; Bergfield et al., 2010). Reduced structural associations were also observed in the structural covariance networks subserving the language-related semantic network, the executive control network, and the DMN in older subjects compared with younger subjects (Montembeault et al., 2012). In addition, a graph theoretical analysis of brain structural networks demonstrated that the topological patterns exhibit a modular organization and that the aging brain exhibits reduced intra-/inter-modular connectivity in modules corresponding to the executive function and the DMN, resulting in the development of a more localized network organization in old age (Chen et al., 2011; Wu et al., 2012; Zhu et al., 2012). In contrast to the investigation of the effects of aging on the networks related to high-level cognitive functions, however, reports of the same nature but on the low-level sensory/motor systems, especially viewed as networks, are rare.

Bayesian networks (BN), as a powerful multivariate technique, enables investigation of specific conditional probability dependencies over a set of random variables, or the nodes in BN. In applying to the neuroimaging studies, these nodes are usually the predefined set of ROIs. The measurements obtained from each of the ROIs, either fMRI time-series or structural MRI data, are viewed as realizations of the corresponding random variable. At present, BN has been widely and successfully utilized to investigate brain connectivity relationships for both functional and structural MRI studies (Chen and Herskovits, 2006; Zheng and Rajapakse, 2006; Wu et al., 2011; Chen et al., 2012; Wang et al., 2014) and other studies (Bielza et al., 2011, 2013; Larrañaga et al., 2013). For fMRI data, BN modeling examines conditional dependencies of brain activity among the sets of the given brain regions for each individual subject. For structural MRI data, BN modeling investigates probabilistic associations of morphological feature based on morphometric estimations such as the average of gray matter volume of brain regions over multi-subjects at the group level either within or between groups. Our previous study examined the aging influence on the gray matter structural inter-regional linkage within the DMN utilizing BN modeling. Our findings showed that the older subjects exhibited significant structural association alterations among core brain regions within the DMN compared to their younger counterparts and indicated that the BN modeling approach can potentially serve as a useful tool for studying structural probabilistic dependence among multiple brain regions (Wang et al., 2014).

Overall, the DMN, visual, auditory, and motor networks are the four most common brain networks. However, to date, the majority of the aging-related studies focused on the DMN. The auditory, visual, and motor networks have received much less attention in normal aging research. The anatomical inter-regional relationships among distinct brain areas may potentially hold important implications on structural networks of the aging brain. In this study, we employed Gaussian BN to characterize the aging-related structural association alterations between core brain regions within each of these three structural networks from structural MRI data and further investigated the discriminability of the aging-associated BN models for the young and old groups.

In this study, we constructed the BN models of the auditory, visual, and motor networks based on regional gray matter volumetric data from the old and young groups. We then examined the between-group differences in connection with weighted coefficients by a nonparametric permutation test. The results of the BN modeling showed that, for both young and old groups, there were 7 connections in the auditory network and 5 connections in the motor network, but 5 connections for the old group and 7 connections for the young group in the visual network. We further examined the discriminability of the aging-associated BN models for each network.

The results of the BN modeling demonstrated that the structural associations exist between two homotopic brain regions from the left and right hemispheres in each of the three networks, such as the bilateral heschl gyri (HES), STG and SMG in the auditory network, bilateral LING, MOG, and CAL in the visual network, and bilateral PreCG and PoCG in the motor network (Figures 1–3 and Table 2). Some previous structural MRI studies indicated that the morphology of the human brain covaried in gray matter density/volume or cortical thickness; moreover, such structural covariance would change owing to different clinical conditions (Mechelli et al., 2005; Andrews-Hanna et al., 2007; Evans et al., 2008; Seeley et al., 2009; Zielinski et al., 2010). Mechelli et al. reported structural correlations in the human cortex and observed homotopic effects in the inferior frontal, primary auditory and somatosensory cortex that are consistent with our findings, although with the exception of primary visual cortex (Mechelli et al., 2005). The existing diffusion tensor imaging (DTI) studies reconstructed and visualized prominent white matter tracts in the human brain (Wakana et al., 2004; van den Heuvel et al., 2009). Van den Heuvel et al. reconstructed the fiber tract pathways between the functionally linked brain areas of several resting-state networks and demonstrated the existence of direct white matter connections in these resting-state networks, such as the well-known DMN, primary motor and visual networks (van den Heuvel et al., 2009). The above MRI studies support our findings of structural associations between two homotopic regions from two hemispheres in each of the three networks.

We found that there exist 7 connections in the auditory network for both the old and young groups. Compared with the young subjects, the old subjects had significant connection reductions from the left STG to the right STG and from the left SMG to the right SMG. Using a cortical thickness and graph theory analysis, Chen et al. investigated age-related alterations in the modular structure networks and found a reducing trend for the intra-modular connectivity in the auditory module in the aging group (Chen et al., 2011), which is in agreement with our findings. Hearing loss is one of the major impairments in the older population and causes difficulties in speech comprehension and communication (Peelle et al., 2011; Eckert et al., 2012). Lower gray matter volumes were observed in the primary auditory cortex (Husain et al., 2011; Peelle et al., 2011), moreover, Peelle et al. demonstrated a significant linear relationship between the hearing ability and the gray matter volume in primary auditory regions. Hence, such findings support the view that age-related hearing loss is associated with neuroanatomical alterations in the auditory cortex.

For the visual network, there were 5 connections for the old group but 7 connections for the young group. The old subjects had significant connection reductions from the right CAL to the right LING, from the left CAL to the left MOG and from the left CAL to the left LING and connection increases from the right LING to the right MOG. Based on the regional gray matter volume, Wu et al. applied graph theory to investigate age-related changes in the topological organization of structural brain networks and depicted that the old group showed a prominent decrease in the connector ratio and the intermodular connections in comparison to the young and middle groups; thus, the modular organizations associated with memory, auditory and visual systems in the old group are quite different from those in the younger groups (Wu et al., 2012), in agreement with our findings. In addition, Wu et al. also characterized the age-related longitudinal changes in the structural brain networks and found significant negative correlations between the baseline age and the annual rate of change in nodal strength in the brain regions primarily involved in the visual and motor systems (Wu et al., 2013). Van den Heuvel et al. found that tracts of the splenium of the corpus callosum connected the bilateral brain regions of the primary visual network (van den Heuvel et al., 2009), and Ciccarelli et al. demarcated the optic radiations as the major fiber tract in the visual network (Ciccarelli et al., 2003). The existence of these white matter tracts provides the subserving foundation for structural associations in gray matter volume.

Five connections were present in the motor network for both groups. In the old subjects, significant connection reductions were found from the right PoCG to the right PreCG compared to the young subjects, and connection increases from the left PreCG to the right PreCG and from the left PoCG to the left PreCG were also found. Older adults exhibit motor performance deficits with advanced aging, and these deficits are accompanied by age-related atrophy of the motor cortical regions and corpus callosum and biochemical changes; moreover, larger gray matter volume and better white matter integrity are correlated with better motor performance for review see Seidler et al. (2010). Some studies have indeed depicted accelerated loss of gray matter volume in the local sensorimotor cortex, including the SMA, and the bilateral precentral and postcentral gyri in older adults (Good et al., 2001; Resnick et al., 2003; Kalpouzos et al., 2009). Van den Heuvel et al. found that tracts of the body of the corpus callosum connected the primary sensorimotor network (van den Heuvel et al., 2009), and some other DTI studies demonstrated the degenerative changes in the corpus callosum in normal aging (Kochunov et al., 2009; Sala et al., 2012). Using the cortical thickness as a morphometric measurement in structural MRI, Chen et al. found significant correlation strength increases among the bilateral frontal regions involved with the bilateral precentral gyri in older adults in comparison to their younger counterparts (Chen et al., 2011), which is consistent with our findings.

In this study, the BN model depicts the conditional probability dependencies based on the average of gray matter volume of core brain regions within each of the three networks, and the differences in such conditional dependencies exhibit the aging effects at a global level. Once the BN model is constructed, it can be used as an effective tool for classification. Hence, we used the derived BN models to identify the membership between the older and younger groups. The results showed that the auditory network had the highest statistical discriminability with 89.02% sensitivity, 90.83% specificity and 90.05% accuracy. The sensitivity, specificity, and accuracy were relatively low for the visual network, but all three were above 70%. Overall, our findings suggest that the aging-related BN models could predict the group distinction with reliable statistical power.

It should be noted that the changes in gray matter volume of the brain regions within these three networks have not always been consistent from the existing structural MRI studies. Some publications in the literature reported these brain regions to be preserved with aging, whereas others have indeed reported gray matter atrophy in most of these brain regions. This discrepancy may result from the different methods applied or the samples selected in these studies. In the present study, we applied a multivariate BN approach and focused on the probabilistic associations over a set of gray matter volume estimations in a specific network but not the single brain region alterations (e.g., volume increases or decreases). A number of publications have suggested that structural covariances or associations between functionally connected brain regions may result from mutually trophic influences by white matter fiber connections or common experience-related plasticity (Ferrer et al., 1995; Mechelli et al., 2005; Soriano-Mas et al., 2013). The alterations of dependence between brain regions may arise from a lack of mutually trophic influences in different clinical conditions (He et al., 2008; Seeley et al., 2009).

Most of the previous MRI studies have documented structural association decreases in the structural networks that promote high-order cognitive functions, such as the language-related semantic network, the executive control network and the DMN, in older adults compared with younger adults (Chen et al., 2011; Montembeault et al., 2012; Wu et al., 2012). Our previous study also examined the aging influence on gray matter structural associations within the DMN utilizing BN (Wang et al., 2014). As we all know, the auditory, visual, and motor function is related to basic abilities in life, and these impairments will lead to a diminished quality of life in the older. Although our present study does not provide the direct biological evidence of such structural inter-regional associations, we expect these preliminary findings of age-related alterations to help elucidate the anatomical changes in these three networks.

About BN methodology, BN can be learned in an unsupervised manner from data without prior knowledge. By compactly representing the joint probability distribution over a set of random variables, BN model can be used to examine probabilistic associations among these variables in an exploratory manner. Till now, BN approach has been successfully utilized in brain neuroimaging studies. BN modeling, as a valuable method of mining association relationships between continuous variables, can be used to investigate the association dependency based on regional gray matter volumes. In our study, the directed edge representing a conditional probabilistic dependence cannot be viewed, without any direct biological and medical evidence, as the biological causal relationship (Chen and Herskovits, 2006). In addition, the BN model constructed can be used as an effective tool for discriminating the group membership. We hope that our study provided additional, but preliminary or exploratory findings in the existing investigations.

In conclusion, we characterized the influence of aging on interregional gray matter structural associations within each of the auditory, visual, and motor networks. Our findings suggest that the BN model can detect between-group differences with reliable statistical power. These differences in structural associations may further provide new insights into the neuronal mechanisms of morphometric changes in normal aging.

Designed the work: Xiaojuan Guo, Kewei Chen, and Li Yao. Analyzed the data: Yan Wang, Xia Wu, and Jiacai Zhang. Interpreted the data: Xiaojuan Guo, Ke Li, Zhen Jin, and Li Yao. Wrote the paper: Xiaojuan Guo and Kewei Chen. Revised the paper: Xiaojuan Guo, Kewei Chen, and Li Yao.