Altered connectivity among the triple brain networks in patients with mild cognitive impairment: a source-based morphometry study with a large elderly population

Abstract

Purpose

Previous research indicates brain network alterations in the default mode network (DMN), salience network (SN), and central-executive network (CEN) in individuals with mild cognitive impairment (MCI). However, replication has been inconsistent due to small, varied samples. We aimed to explore intra- and inter-networks alternations among DMN, SN, and CEN in individuals with MCI using multivariate source-based morphometry (SBM) with a larger, population-based sample.

Materials and methods

This cross-sectional study included 1,997 participants (median age: 69 years; 61.9% female) who underwent three-dimensional (3D) T1-weighed imaging with 3 T magnetic resonance imaging (MRI). They were classified into 1,236 healthy controls (HC) and 761 individuals with MCI. SBM was used to extract triple brain networks as structural networks, and Z-scores were calculated. Intra-network comparisons of DMN, SN, and CEN between HC and MCI groups were conducted using logistic regression analysis. Inter-network comparisons among the triple brain networks were performed using structural equation modeling (SEM).

Results

Connectivity (median Z score) of each network was lower in the MCI group than in the HC group: DMN (0.08 vs. − 0.12), SN (0.08 vs. − 0.23), and CEN (0.07 vs. − 0.06). Logistic regression showed significant association of SN connectivity with MCI (odds ratio 0.862, p < 0.05). SEM analysis revealed a significant group difference in the model where SN mediated input from CEN to DMN.

Conclusion

We found altered network patterns in individuals with mild cognitive impairment, suggesting a transformation in network connectivity among DMN, SN, and CEN, particularly compensating for degraded SN connectivity.

Introduction

Cognitive impairment in older adults is a growing social concern in many countries [1]. Mild cognitive impairment (MCI) is characterized by a more serious cognitive decline than normal aging and is associated with a high incidence of Alzheimer’s disease (AD) [2]. Early intervention for MCI is appropriate to prevent the progression of cognitive decline [3], which is crucial to delay the progression of MCI to AD. Therefore, the identification of individuals with MCI has been a persistent focus of research [4].

Resting-state functional magnetic resonance imaging (rs-fMRI) allows for the non-invasive detection of spontaneous neural activity at baseline. It is a useful tool to investigate the properties of functional brain networks, both locally and globally [5]. Many previous rs-fMRI studies have focused on the differences between individuals with MCI and healthy controls (HC) and have indicated altered patterns of brain networks related to cognitive decline. As cognition-related networks, default mode networks (DMNs), salience networks (SNs), and central-executive networks (CENs) are widely recognized [6–8].

A schematic review of rs-fMRI studies has indicated decreased connectivity in the DMN in individuals with MCI and AD [9]. Furthermore, studies have reported abnormal connections in the brains of patients with AD, not only in the DMN but also in the CEN and SN. Sorg et al. [10] reported reduced connectivity in the DMN and CEN hubs in individuals with MCI (n = 24). Meanwhile, Agosta et al. [11] reported decreased connectivity in the DMN among individuals with MCI (n = 12) and AD (n = 13) and increased mean connectivity in the CEN among patients with AD. Furthermore, Zhou et al. [12] reported increased functional connectivity in the SN of patients with AD (n = 12). Although these reports suggest the importance of investigating the roles of these three cognition-related networks in individuals with cognitive impairment, there have been no consistent findings. Furthermore, the DMN and CEN have consistently been reported to exhibit antagonistic activity in the resting state and during task performance in healthy individuals [13–15].

Interactions between these networks have been argued to be important for the effective maintenance of normal cognitive status [16–18]. In recent studies, MCI has been associated with disrupted interactions among the DMN, SN, and CEN [8]. Thus, connections among these networks have recently been proposed as a triple network model [17–20] to understand altered network patterns in cognitive impairments. The triple network model suggests that the SN dynamically controls network interactions between the DMN and CEN and that the core role of the SN is to initiate network-mediated dynamic interactions between the CEN and DMN [8, 19, 21]. To the best of our knowledge, population-based studies with large sample sizes have not been conducted to evaluate connectivity among the triple networks in MCI.

Recently, a source-based morphometry (SBM) technique utilizing 3D T1-weighted imaging (T1WI) data was introduced. This multivariate approach pools information across different voxels and identifies unpredictable patterns thereof [22, 23]. SBM applies independent component analysis (ICA) to segmented images, arranges voxels into sets containing similar information [22], and acquires common morphological features among individuals at the brain network level. Importantly, the underlying hypothesis of morphological network analysis is that brain regions exhibiting similar structural characteristics—such as gray matter volume or cortical thickness—may be developmentally, genetically, or functionally related. This morphological covariance is thought to reflect shared neurobiological processes, suggesting that these regions could be interconnected at a systems level [24]. Three-dimensional (3D) T1WI has also been included in the standard brain magnetic resonance imaging (MRI) protocol. Conversely, functional MRI has been widely used for research, not in clinical situations. Therefore, the SBM method using 3D T1WI data is convenient to identify novel networks and may be a reasonable tool for brain MRI studies in larger populations. Moreover, structural MRI imaging, compared to rs-fMRI, has the advantage of high spatial resolution.

We reviewed data from a population-based prospective study on dementia in a large population of older Japanese individuals (The Iki-Iki Health Promotion Project: Iki-Iki study) to acquire and analyze the brain networks using SBM. We then aimed to assess the network connectivity of individuals with MCI compared with cognitively normal older adults as HC among the three cognition-related networks by applying SBM. Furthermore, we evaluated the differences in network interactions among the DMN, SN, and CEN between HC and individuals with MCI.

Materials and methods

Study population and design

This study was conducted in accordance with the ethical guidelines of the Declaration of Helsinki. The use of data from The Iki-Iki study was approved by the Ethics Committee. Written informed consent was obtained from all participants. The Iki-Iki study was established in 2016 as a population-based 12-year prospective study of cerebrovascular and cardiovascular disease and dementia in an older Japanese population from the Iwaki area of Hirosaki City, Aomori Prefecture, Japan. This is one study field of the Japan Prospective Studies Collaboration for Aging and Dementia (JPSC-AD) that is a multisite, population-based prospective cohort study of dementia [25, 26]. A total of 2,390 participants aged > 64 years participated in the screening surveys in 2016 and 2017, and 2,226 (93.1%) participants underwent brain MRI. We excluded 50 participants with image distortions (seven with metal artifacts, 13 with excessive motion artifacts, and 30 for whom brain volume or structural covariance intra-networks could not be measured accurately for various reasons) and 43 participants without available MRI data (T1-weighted images). Of the 2,133 participants, 761 (35.8%) with Mini-Mental State Examination (MMSE) scores of 24–27 were defined as MCI [27–29]. A total of 1,236 (57.9%) participants with MMSE scores of 28–30 were defined as HC, and 136 (6.3%) participants with ≤ 23 MMSE were excluded due to significant cognitive decline [28]. Thus, 1,997 participants (761 with MCI and 1,236 HC) were enrolled in this study (Table 1).

Table 1.

Clinical characteristics of participants

	HC	MCI	P Value
	(N = 1236)	(N = 761)
Age, median (IQR)	68 (64–80)	70 (64–80)	< 0.01
Sex: male/female	449/787	312/449	0.04
Education: High school/Junior high school	1091/145	572/189	< 0.01
MMSE, mean (SD)	29.26 (0.77)	25.79 (1.05)	< 0.01
DMN, median (95% CI)	0.08 (− 0.73, 0.88)	− 0.12 (− 0.89, 0.63)	0.18
SN, median (95% CI)	0.08 (− 0.77, 1.03)	− 0.23 (− 1.10, 0.69)	< 0.01
CEN, median (95% CI)	0.07 (− 0.91, 1.08)	− 0.06 (− 1.08, 0.89)	0.10
HP/ICV ratio, mean (SD)	0.0055 (0.00049)	0.0054 (0.00052)	< 0.01
GMV/ICV ratio, mean (SD)	0.42 (0.023)	0.42 (0.024)	< 0.01
WMV/ICV ratio, mean (SD)	0.32 (0.019)	0.32 (0.020)	< 0.01

MCI mild cognitive impairment, HC healthy controls, DMN default mode network, SN salience network, CEN central-executive network, HIP hippocampus, ICV intracranial volume, GMV gray matter volume, WMV white matter volume, CI confidence coefficient, SD standard deviation, IQR interquartile range, MMSE Mini-Mental State Examination

MRI

All brain MRI data were obtained using the same protocol on a single 3 T MRI scanner (Signa EXCITE 3 T; GE Healthcare, Waukesha, WI) with an eight-channel brain phased-array coil. Original T1-weighted images were obtained in the steady state using a 3D fast spoiled gradient recalled sequence with the following parameters: repetition time, 10 ms; echo time, 4.1 ms; inversion time, 700 ms; flip angle, 10; field of view, 26 cm; section thickness, 1.2 mm; resolution, 1.0 × 1.0 × 1.2 mm; Number of Excitations, 1; and imaging time, 3 min 20 s.

Image processing for voxel-based morphometry

Image preprocessing was identical to that used for classical voxel-based morphometry analyses. A fully automatic technique for the computational analysis of differences in regional brain volume throughout the brain was conducted using SPM12 software (Institute of Neurology, London, UK) [30, 31]. The 3D fast spoiled gradient-echo images in the native space were spatially normalized; segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid images; and modulated using the Diffeomorphic Anatomical Registration Through Exponential Lie Algebra (DARTEL) toolbox in SPM12 [32]. DARTEL was devised as an alternative method for normalization in the SPM package [30]. To preserve the GM and WM volumes within each voxel, we modulated the images using Jacobean determinants derived from spatial normalization using DARTEL. The resulting modulated GM images were smoothed using an 8-mm full-width-at-half-maximum Gaussian kernel. The analysis was conducted using the CAT12 toolbox (C. Gaser, Structural Brain Mapping Group, Jena University Hospital, Jena, Germany; http://dbm.neuro.uni-jena.de/cat/), which is implemented in SPM12 (Wellcome Trust Center for Neuroimaging, London, UK; http://www.fil.ion.ucl.ac.uk/spm/software/spm12/) [30, 31]. Intracranial volume (ICV), total gray matter volume (GMV), total white matter volume (WMV), and bilateral hippocampal volume (HV) were calculated using the 40-brain LPBA40 atlas [33]. GMV/ICV, WMV/ICV, and HV/ICV ratios were calculated. The HV/ICV ratio is used as an indicator of hippocampal atrophy.

Image processing for SBM

SBM is a multivariate technique that utilizes ICA [22] and considers information across different voxels and identifies unpredicted, naturally occurring patterns of covariance across brain regions. The GIFT toolbox (http://icatb.sourceforge.net) was used for image processing in the SBM analysis [22]. The minimum description length principle was used to estimate the number of independent components, identifying 147 reliable independent components (GM structural networks). We performed ICAs using a neural network algorithm (Infomax). Through ICAs, we attempted to minimize the mutual information of network outputs to identify natural groupings and maximally independent sources [34]. ICAs were repeated 20 times in ICASSO (http://research.ics.aalto.fi/ica/icasso/), and the obtained components were clustered to confirm consistency and reliability. Consequently, we selected a matrix in which 1,997 rows represented 1,997 subjects (1,236 HC and 761 MCI), and each column represented a voxel. This matrix was decomposed into two matrixes using ICA, the “mixing matrix” and the “source matrix.” The first matrix (“mixing matrix”) comprised one subject per row and IC per column and included loading coefficients that indicated the contribution of each structural component to the participants. Thus, information on the relationship between each participant and each component was included in the mixing matrix. Thus, the loading coefficients revealed the contribution of each participant to the specific brain network components. The second matrix was called the “source matrix” and prescribed the relationship between the independent components and the voxels. The loading coefficients were transformed into Z-scores using Fisher’s z-transformation. We visualized the information gained from the independent components. For visualization, the source matrix was reshaped back to a 3D image and scaled to unit standard deviations (Z maps); the threshold was set at Z > 2.0.

As a result, ICA generated 147 independent components (GM structural networks). From these components, two experienced neuroradiologists (S.K. with 24 years of experience, and K.W. with 12 years of experience) identified the three key networks: the DMN, SN, and CEN (Fig. 1). This identification was performed by visually inspecting each component against established patterns documented in prior research, specifically the atlases of brain networks tailored for an elderly cohort [35]. These atlases provide detailed anatomical regions typically associated with these networks, aiding in accurate identification; DMN: We selected components that prominently included the posterior cingulate cortex and the angular gyrus, which are critical regions of the DMN, SN: Components that contained the anterior insula were chosen, as this area is central to the SN, CEN: We identified components that involved the dorsolateral prefrontal cortex, a key region in the CEN. These selections were made by consensus between the two neuroradiologists.

Fig. 1.

The main images of each network. (A) default mode network (DMN), (B) salience network (SN), and (C) central-executive network (CEN)

The Z-scores in SBM allowed us to identify the sources that exhibited differences between the HC and MCI groups. Multivariable linear regression was used to investigate the association between cognitive decline and the loading coefficients (Z-scores) for each component. Regression coefficients within groups were determined using dummy variables for HC and MCI as reference groups. The analyses were adjusted for age, sex, and education level (below junior high school, higher than high school graduate, or equivalent).

Statistical analyses

All statistical analyses were performed using EZR (Saitama Medical Center, Jichi Medical University, Saitama, Japan) [36], and statistical significance was set at p < 0.05. The clinical characteristics of the participants within the HC and MCI groups were compared using logistic regression analysis, Welch’s t-test, and Pearson’s chi-square test, as appropriate. Nominal variables are expressed as percentages and continuous variables as mean ± standard deviation, median (interquartile range), or range based on distribution. In the analysis of the loading coefficients calculated from the SBM for each component, we compared the two groups with the loading coefficients, as each participant had a loading coefficient for each component.

Using SPSS AMOS (version 27.0; SPSS Inc., Chicago, IL) for structural equation modeling (SEM) analysis, we examined the role of the SN in mediating communication between the DMN and CEN in the HC and MCI groups. The constructed model posited that the SN coordinates communicate between the DMN and CEN [21]. To compare and examine the interactions among these three networks, we created two path diagrams based on a previous study [8] (Fig. 2). First, we created Model A in which the SN mediated network input from the DMN to the CEN. The other model (Model B) assumes that the SN mediates network input from the CEN to the DMN. Path diagrams were created for both models in the HC and MCI groups, and path coefficients were obtained. In the path diagram of Model A, the networks from the DMN to the CEN were compared by setting direct and indirect routes. The direct path reflected the input from the DMN to the CEN. Indirect pathways are models in which the SN intervenes in the network input from the DMN to the CEN. The error variable setting represented the variation left unexplained by the explanatory variables of the path model. In contrast, in the path diagram of Model B, we compared the network inputs from the CEN to the DMN by setting direct and indirect paths. The direct path reflected the input from the CEN to the DMN. An indirect pathway is a model in which the SN intervenes in the network input from the CEN to the DMN. In Model B, the error variable setting represented the variation left unexplained by the explanatory variables of the path model.

Fig. 2.

Path diagrams of the network model. Two path diagrams were constructed using a model in which the SN coordinated communication between the DMN and CEN. Model A: This diagram illustrates a network model assuming a direct connection from the Default Mode Network (DMN) to the Central-Executive Network (CEN) and an indirect connection mediated by the Salience Network (SN). The error variables indicate unexplained variance in each network in Model A. Model B: This diagram depicts a network model where the CEN has direct connections to the DMN, with the SN mediating additional indirect connections. Similar to Model A, error variables capture the unexplained variance in each network in Model B

Models A and B were subjected to an SEM analysis. Several model fit indices should be considered to appropriately interpret the model. The model’s goodness of fit was tested using the Goodness of Fit Index (GFI) and the comparative fit index (CFI). When the GFI and CFI were greater than 0.95, they were considered good indices [37]. Briefly, the GFI is an absolute fit measure that reflects the proportion of variance explained by the model. In contrast, the CFI compares the fit of the proposed model to an independent baseline model, adjusting for model complexity. Both indices help assess model suitability but from different perspectives. In each model, path coefficients were gained, and a chi-square difference test was performed to evaluate differences in path coefficients between the HC group and the MCI group.

Results

Table 1 presents the demographic characteristics of the participants. There were significant differences in the age, sex, and years of education between the HC and MCI groups. Connectivity across the various networks was generally diminished in the MCI group compared to the HC group. Median Z-scores for the DMN were 0.08 in the HC group versus − 0.12 in the MCI group; for the SN, 0.08 versus − 0.23; and for the CEN, 0.07 versus − 0.06 (Table 1).

Logistic regression analysis showed that SN connectivity (Model 2 in Table 2) was a significant independent predictor of MCI (p < 0.01) after adjusting for potential confounders (age, sex, and education level), but not DMN or CEN connectivity (Models 1 and 3 in Table 2). The GMV/ICV, WMV/ICV, and HV/ICV ratios were not significant independent predictors of MCI after adjustment (Models 4, 5, and 6 in Table 2), although the MCI group showed significantly reduced GMV/ICV, WMV/ICV, and HV/ICV ratios compared to the HC group (Table 1).

Table 2.

Multiple regression analysis

Predictor	Model 1		Model 2		Model 3		Model 4		Model 5		Model 6
	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds ratio (95% CI)	p	Odds Ratio (95% CI)	p
Age	1.04 (1.02, 1.07)	< 0.01	1.04 (1.01, 1.06)	< 0.01	1.05 (1.02, 1.07)	< 0.01	0.998181 (0.997978, 0.998384)	< 0.01	0.997913 (0.997728, 0.998097)	< 0.01	0.999952 (0.999948, 0.999957)	< 0.01
Sex	0.796 (0.658, 0.963)	0.02	0.725 (0.595, 0.884)	< 0.01	0.791 (0.654, 0.958)	0.02	1.023310 (1.021518, 1.025048)	< 0.01	1.003676 (1.002128, 1.005226)	< 0.01	1.000312 (1.000273, 1.000351)	< 0.01
Education [High School Graduate]	1.28 (1.03, 1.59)	0.02	1.27 (1.02, 1.57)	0.03	1.29 (1.04, 1.60)	0.02	1.002779 (1.000902, 1.004659)	< 0.01	1.000586 (0.998883, 1.002293)	0.50	1.000034 (0.999990, 1.000077)	0.13
Education [below junior high school]	2.69 (2.03, 3.57)	< 0.01	2.64 (1.99, 3.51)	< 0.01	2.70 (2.03, 3.58)	< 0.01	1.005155 (1.002561, 1.007757)	< 0.01	1.000778 (0.998429, 1.003133)	0.52	1.000066 (1.000006, 1.000126)	0.03
DMN	0.93 (0.865, 1.03)	0.18
SN			0.862 (0.80, 0.929)	< 0.01
CEN					0.948 (0.89, 1.01)	0.10
GMV / ICV							0.999229 (0.997504, 1.000957)	0.38
WMV / ICV									0.999864 (0.998293, 1.001437)	0.87
HV / ICV											0.999963 (0.999923, 1.000003)	0.07

MCI mild cognitive impairment, DMN default mode network, SN salience network, CEN central-executive network, ICV intracranial volume, GMV gray matter volume, WMV white matter volume, HV hippocampus volume, CI confidence coefficient

In the SEM analysis, the results of testing the goodness of fit of Models A and B both showed GFI = 1.00, and CFI = 1.00. In both models, the goodness of fit indices were acceptable. While Model A (network model from DMN to CEN via the SN) did not show any significant differences between the HC and MCI group (Fig. 3), Model B (network model from CEN to DMN via the SN) showed notable disparities between the HC group and MCI group (Fig. 4) (p < 0.05).

Fig. 3.

Path analysis in Model A showing a comparison of path coefficients between HC and MCI regarding networks from DMN to CEN. This path model shows that the SN inter-events with the connection from the DMN to the CEN. Each number near the arrows represents a path coefficient indicating the strength and direction of each connection. Here, e1 through e6 mean error variable. A small number represents the square of multiple correlation coefficients (R²) for each error variable. No significant differences were observed between the HC and MCI groups. HC healthy controls, MCI mild cognitive impairment, DMN default mode network, SN salience network, CEN central-executive network

Fig. 4.

Path analysis in Model B showing a comparison of path coefficients between HC and individuals with MCI regarding networks from CEN to DMN. This path model shows that the SN interacts with the connections from the CEN to the DMN. Each number near the arrows represents a path coefficient. Here, e7 through e12 mean error variable. A small number represents the square of multiple correlation coefficients (R²) for each error variable. Significant differences in the network from the CEN to the SN were observed between the HC group and MCI group (*p = 0.01). HC healthy controls, MCI mild cognitive impairment, DMN default mode network, SN salience network, CEN central-executive network

Discussion

Our multiple regression analysis of the brain networks showed that only SN connectivity was a significant independent predictor of MCI, after adjusting for potential confounders. The SEM analysis revealed a significant difference between the HC and MCI groups in a model in which the SN mediated network input from the CEN to the DMN. To our knowledge, this is the first population-based study with large sample sizes to demonstrate brain network alterations in individuals with MCI.

As for brain networks in cognitive impairment, although most studies have focused on DMN, recent studies showed that cognitive abilities affect not only DMN but also other networks [6, 7]. Furthermore, Onoda et al. demonstrated that disruption of the SN is associated with cognitive decline in elderly individuals [38]. Brier et al. found reduced connectivity in DMN and SN of individuals with MCI as well as patients with AD [39], and He et al. found that both the structural and functional organization of the SN was impaired in individuals with MCI and AD [40]. These results are consistent with those of the present study.

We found a significant difference between the HC and MCI groups in a model where the SN intermediates the inter-network from the CEN to the DMN, suggesting altered SN function between the DMN and CEN inter-network connectivity. The DMN has been shown to be more internally activated during self-monitoring, directed cognitive activities, and social functions. The CEN is often activated for externally directed higher-order cognitive functions, such as attention, decision-making, and working memory [16, 18]. The DMN and CEN are known to interact competitively during cognitive information processing [14, 41], and the SN drives switching between the DMN and CEN [17, 21, 42]. Chand et al. [8] showed that the disruption of SN control over the DMN and CEN provides a neuronal basis for cognitive decline and may be a candidate neuroimaging biomarker for cognitive impairment. Moreover, Li et al. [43] showed abnormal SN function at intra-network connectivity in patients with AD and late MCI, particularly SN-DMN and SN-CEN connectivity. These findings support our results that the SN plays a key role in mediating networks between the CEN and DMN and that there is a significant difference between HC and individuals with MCI.

Brain atrophy, as measured using MRI, is a remarkable sign of neurodegeneration, and different morphological analysis methods have been used to investigate subtle structural alterations in individuals with MCI [44]. However, improvements are still needed for the accurate diagnosis of MCI stages. In this study, we found no significant differences in brain volume measurements (WMV, GMV, and HV) between the HC and MCI groups using logistic regression analysis. The SN connectivity of the MCI group with SBM was significantly lower than that of the HC group. This suggests that the SBM has great potential to provide prognostic information for cognitive decline. Therefore, further studies are needed to clarify whether altered connectivity might precede volume changes in the brain parenchyma and whether altered connectivity and reduced brain volumes may reflect different and complementary information for MCI pathology.

There are many reports on brain networks in cognitive dysfunction using rs-fMRI. However, we used structural MRI to investigate brain networks and their interactions. Regarding the DMN, our logistic regression analysis did not show a significant difference in the DMN between the HC and MCI groups. Previous studies established that DMN activity is impaired in individuals with MCI and AD [5, 39]. However, a more recent systematic review of 57 MCI vs. HC comparisons suggested substantial inconsistency in patterns of altered resting DMN connectivity. Nine studies showed no group difference, eight studies showed MCI > HC in DMN connectivity, 22 studies showed HC > MCI across the brain, and 18 studies showed regionally mixed directions of effect [45]. Our study has several strengths compared to previous rs-fMRI studies. Our community-based design and large sample size reduced sampling bias and the effects of individual variability and permitted a detailed evaluation of confounding factors. Our SBM analysis was performed using 3D T1WI, which has been widely used for MR examinations in clinical situations. Thus, this method may be useful for brain MRI studies in larger populations. Additionally, this study was conducted using images captured on the same 3 T MRI scanner with a consistent protocol. Previous analyses have suggested that volume measurements across platforms (vendors, MRI sequences, and scanner upgrades) introduce a difference bias [46]. By using the same MRI scanner and protocol, this bias could be avoided.

In this study, we conducted network research using structural MRI rather than resting-state functional MRI (rs-fMRI). Traditionally, most network studies utilize rs-fMRI, a technique that captures a series of 3D brain images over time to analyze temporal correlations between BOLD signal fluctuations in different brain regions during rest. This method is effective in reflecting neural synchrony and transient brain states. However, our approach using SBM examines structural correlations by analyzing the covariance of differences from mean images across multiple scans instead of relying on temporal information. This approach allows us to identify covarying structural patterns across individuals, which may infer underlying network connections based on brain morphology. Unlike the dynamic interactions observed in rs-fMRI, the use of structural networks through SBM provides insights into long-term formations within the brain architecture. While structural network analysis with SBM is not identical to the functional connectivity measured by rs-fMRI, previous studies have shown a direct association between functional and structural covariance networks across the entire brain, suggesting a complementary relationship between these methods [47]. Furthermore, studies have reported the extraction of well-known networks such as the DMN, CEN, and SN using SBM [48], as demonstrated in the current study. These findings indicate that SBM can effectively reveal key networks and may serve as a valuable tool, particularly when rs-fMRI data are not available.

Our study has some limitations. First, functional MRI was not performed to confirm or compare the networks extracted from structural MRI. Second, data from the Montreal Cognitive Assessment, which is considered to be highly sensitive to MCI [49], were not available in the Iki-Iki Health Promotion Project. The MMSE is a simple and useful bedside cognitive testing tool. However, it is subjective and often insensitive to the early stages of dementia [50]. It is difficult to separate HC from individuals with MCI using cognitive function evaluation tests, which are objective indicators. In this study, MCI was diagnosed using only MMSE scores. Therefore, individuals with MCI and normal MMSE scores might have been included in the HC group [4, 51]. In contrast, participants with partial symptoms, such as prodromal AD, according to the new AD diagnostic criteria [52, 53] may have been included in the MCI group. Third, we performed SEM analysis based on the hypotheses that there are direct or indirect causal influences from the SN to the DMN and CEN [42] and that the SN plays a role in network switching between the DMN and CEN [17, 21, 42]. Brain networks influence each other in a complex manner, and the results of the SEM analysis are evaluations based on hypotheses; therefore, care must be taken when evaluating causal relationships. Fourth, our sample was restricted to relatively well-educated ethnic Japanese individuals, which may limit the generalizability of our findings to other racial and ethnic groups, particularly to those in less-developed countries. Hence, further studies are required to assess altered networks in individuals with MCI using 3 T MRI datasets from other sources, such as the Alzheimer’s Disease Neuroimaging Initiative [54]. Finally, some relevant confounders may have been excluded from the multivariate linear regression analysis.

In conclusion, our study uncovered significant disturbances within the SN among individuals with MCI. Furthermore, we discovered impairments in the switching mechanism between the CEN and the DMN, mediated by the SN. These findings suggest a critical alteration in network dynamics, highlighting the role of SN in orchestrating the complex interplay and transition of neural activities among the key brain networks in individuals with MCI. Our findings provide a deeper understanding of the underlying neural mechanisms contributing to cognitive decline.

Abbreviations

AD

Alzheimer’s disease

CEN

Central-executive network

DMN

Default mode network

HC

Healthy controls

ICA

Independent component analysis

MCI

Mild cognitive impairment

SBM

Source-based morphometry

SEM

Structural equation modeling

SN

Salience network

Declarations

Competing interests

There are no conflicts of interest to declare.