Coupling analysis between functional and structural brain networks in Alzheimer's disease

1.   Introduction

Alzheimer's disease (AD) is one kind of neurodegenerative disease, which is clinically accompanied by typical progressive cognitive decline and behavioral impairments. It not only affects the daily life of patients but also brings a heavy burden on families, public healthcare systems and society. The AD affected more than 15% of people aged 65 years and older and its annual healthcare cost had reached more than $ 300 billion worldwide ^[1].

Alzheimer's changes typically begin in the human brain with the slow accumulation of abnormal structures, namely plaques and tangles, which was suspected to be the prime contributor to the damage and death of nerve cells in the human brain. The plaques in extracellular β-amyloid protein and tangles in intracellular twisted fibers of tau protein somehow played an important role in blocking the communication among nerve cells and disrupting processes needed for cell survival ^[2]. Previous studies confirmed that these pathological changes started decades ago before the first signs of memory loss ^[3].

Besides the structural MRI, the diffusion tensor imaging (DTI) and resting-state functional MRI (rsfMRI) had been proposed to evaluate the brain morphological and functional changes. Recently, the DTI and rsfMRI brain networks based on graph theory had been applied in the AD studies. According to the structural connectivity (SC) and functional connectivity (FC), the human brain was assumed as one assembly of highly interconnected networks or connectomes. The brain networks of SC were reconstructed by abundant cortical-cortical and subcortical WM axonal pathways, while the brain FC networks in vivo mapped the statistical connections between the cortical nodes in the complex neural systems, i.e., functional connectome ^[4,5]. The graph-based network analyses promised to reliably quantify brain SC and FC with a cluster of neurobiologically meaningful and easily computable brain network measures. Therefore, the relevant studies revealed that the capability of information transmission across brain networks could be assessed by common network characteristics of path length (L_p), clustering coefficient (C_p), global efficiency (E_g), local efficiency (E_l) and small-worldness (σ) ^[6,7,8].

DTI-SC networks demonstrated unique values in the evaluation of AD ^[9]. The global DTI-SC demonstrated that there were increased L_p and reduced E_g for AD ^[10]. Moreover, the DTI-SC within the intra-occipital lobe gradually became weakened during the progression of AD ^[11]. Another study found that the hemispheric asymmetry of brain networks could be taken as a potential biomarker for AD ^[12]. Again, WM seeded networks had approximately more than 400 connections and presented stronger connections than GM seeded networks for both AD and control groups ^[13]. Accordingly, the network of rsfMRI-FC showed specific characteristics of random network topology with powerless connections at anterior, posterior, and prefrontal brain regions. Nevertheless, the FC at the occipital and frontal lobes did not change with significant large-scale SC alterations during the transition from mild cognitive impairment (MCI) to AD ^{[14,15,16,17,18]}. Related studies verified that the cognitive and memory impairments were strongly correlated with the decreased σ, increased L_p and C_p for AD ^[19].

Till now, the analysis of FC-SC coupling based on the graph theory has become a hot spot of brain studies. Since the default mode network (DMN) had been highlighted as a large-scale network of brain regions that were highly correlated with each other in terms of neural activity and the changes within this network had been described in healthy aging as well as in AD ^[20]. It was proved that the core regions of the DMN were associated with brain regions having memory function ^[21]. At the mode of DMN, the coupling strength between DTI-SC and rsfMRI-FC was significantly correlated, and the comparison of brain FC and SC networks for healthy participants indicated that pairs of nodes had stronger SCs and FCs ^[22]. The decreased coupling of FC-SC might indicate more stringent and less dynamic brain function for AD ^[23]. The coupling between rsfMRI-FC and DTI-SC was obviously increased for the connections of DMN in AD ^[24]. It was evident that the correlations between FC and SC could provide more comprehensive and sensitive biomarkers, and helped to reveal the variations of network connectivity and functional interactivity of AD ^[25,26].

In our study, an optimized scheme for brain network coupling was proposed with the aim to investigate the coupling correlation between DTI-SC and rsfMRI-FC. Via the measured brain network characteristics, the coupling of rsfMRI-FC/DTI-SC was evaluated by Pearson's correlation analyses and multiple comparison with Bonferroni correction. Furthermore, the correlations between the rsfMRI-FC/DTI-SC coupling and four neuropsychological scores, including mini-mental state examination (MMSE), clinical dementia rating-sum of boxes (CDR-SB), functional activities questionnaire (FAQ) and montreal cognitive assessment (MoCA), were inferred by partial correlation analyses, respectively. This study would provide new insights to explore the changes in FC-SC coupling during AD progression.

2.   Materials and methods

2.1.   Participant population

A cohort of 112 participants enrolled from the the Alzheimer's disease neuroimaging initiative (ADNI) database was divided into three groups: NC (62 cases), MCI (31 cases) and AD (19 cases). The demographic and neuropsychological information of the participants is shown in Table 1. The inclusion criteria were listed as follows: (1) age between 55 and 90 years, (2) visual and auditory acuity sufficient for neuropsychological testing, (3) good general health and no diseases that could interfere with testing, (4) at least sixth-grade education or good work history (excluding mental retardation), (5) no contraindications to MRI, and (6) all participants who were willing and able to take part in relevant follow-up studies, including blood collection for genomic analysis and biomarker testing, cerebrospinal fluid collection and neuropsychological testing.

2.2.   Imaging acquisition

Three MR protocols of rsfMRI, DTI and T1-weighted imaging (T1WI) were conducted by Siemens 3.0T MR scanners. The rsfMRI scans were acquired using echo planar imaging sequence with the following parameters: flip angle (FA) = 90°; time of repetition (TR) = 3000 ms; time of echo (TE) = 30 ms; slice thickness = 3.4 mm; pixel spacing = 3.4 × 3.4 mm²; matrix size = 448 × 448; and time points = 197. The imaging parameters for DTI were listed as follows: gradient directions = 54 (94 cases) and 30 (18 cases); FA = 90°; TR = 7200/12, 400 ms; TE = 56/95 ms; slice thickness = 2.0 mm; pixel size = 2.0 × 2.0 mm²; and matrix size = 1044 × 1044. The main imaging parameters for T1WI were: FA = 90°; TR = 2300 ms; TE = 3 ms; slice thickness = 1.0 mm; voxel size = 1.0 × 1.0 mm²; and matrix size = 240 × 256.

2.3.   Overview of brain network coupling

The pipeline of brain network coupling of rsfMRI-FC/DTI-SC mainly involved four steps: data preprocessing, construction of brain networks, coupling analysis of brain networks and measurement of network characteristics, as shown in Figure 1.

2.3.1.   Data preprocessing

For the construction of rsfMRI and DTI-SC networks, the data preprocessing of rsfMRI and T1WI was performed by the package of data processing assistant for resting-state fMRI (DPARSF) ^[27]; and the preprocessing of DTI and T1WI was conducted via the software of pipeline for analyzing brain diffusion images (PANDA) ^[28], respectively.

The preprocessing of rsfMRI data included: removal of the first 10 time points, time correction, correction of head movement, regression of interference signal, normalization of voxel size to 3 × 3 × 3 mm³, spatial smoothing with a kernel of 4 × 4 × 4 mm³ and high-pass filtering (0.01-0.1 Hz). Participants were excluded in cases of head translation greater than 3 mm and head rotation greater than 3°. At the preprocessing of DTI data, it consisted of format conversion [(digital imaging and communications in medicine) DICOM to (neuroimaging informatics technology initiative) NIFTI], resampling, removal of brain tissues, segmentation, local diffusion homogeneity, eddy current calibration, standardization, smoothing and deterministic fiber tracking (DFT). During the DFT, the tracking termination criteria were set as: FA < 0.2 or tracking angle > 45° ^[29].

2.3.2.   Construction of brain networks

The construction of brain networks of rsfMRI-FC and DTI-SC mainly composed of three steps: brain parcellation, definition of brain nodes and edges, and connection of brain nodes via edges. Firstly, the brain MR images were parcellated into 90 cerebral regions by the atlas of anatomical automatic labeling (AAL) from montreal neurological institute (MNI) ^[30]. Secondly, the nodes and edges of the rsfMRI-FC network were defined by extracting the time series of the 90 cerebral regions and the normalized correlation matrix with Fisher Z-transformation was constructed by Pearson's correlation coefficients ^[31]. Similarly, the nodes and edges of the DTI-SC network were defined as mentioned above. The T1WI images registered to the original B0 image were mapped to the MNI space by nonlinear transformation. Then, the AAL template was inverted from the MNI space to the original DTI space with the implementation of nearest neighbor interpolation. Here, the product of fractional anisotropy (FA) matrix and fiber number (FN) matrix was used to define the nodes and edges. To reduce the effect of pseudo-junctions during whole-brain fiber tracking, only the nodes connected with fiber bundles (edges) having more than three fibers were regarded as effective connections. Thirdly, the non-directional and weighted brain networks were delineated by defined nodes and edges.

2.3.3.   Coupling analysis of brain networks

At the DMN, 32 common brain regions of rsfMRI-FC and DTI-SC were selected as interested nodes to reconstruct the functional and structural connection matrices. The two FC and SC matrices were quantitatively evaluated by the metric of coupling strength.

2.3.4.   Measurement of network characteristics

Five brain network characteristics, including path length (L_p), clustering coefficient (C_p), global efficiency (E_g), local efficiency (E_l) and small-worldness (σ), were measured for evaluating the two brain networks across there groups. The definition of network characteristics was summarized as follows: (1) σ reflected the transmission efficiency of human brain network, (2) E_g quantified the efficiency of parallel information transfer for global brain network, (3) E_l denoted the mean value of the sums of neighboring sub-networks from E_g, (4) L_p represented the shortest path length between brain neighboring nodes, and (5) C_p indicated the degree of node clustering. The network characteristics of rsfMRI-FC and DTI-SC were measured by GRETNA (https://www.nitrc.org/projec-ts/gretna/) ^[32] at the sparsity threshold of 0.05-0.4 with stable step size of 0.01.

2.4.   Statistical analyses

The statistical analyses were performed in four parts: demographic evaluation, brain network characteristic analyses, coupling analysis of brain networks and evaluation of network coupling with neuropsychological scores. The statistical analyses were performed using the software of SPSS ^[33]. The independent-sample t tests were performed to assess the statistical difference for any two groups among NC, MCI and AD. The p value < 0.05 indicated statistically significant difference for all tests.

Firstly, the differences of demographic and clinical variables were evaluated. The chi-square tests were used for the gender variable, and the variance analyses were used for other variables. Secondly, the brain network characteristics of rsfMRI-FC and DTI-SC were also compared. The network characteristics of rsfMRI-FC and DTI-SC were evaluated by variance analyses. If the results were significantly different between two groups, the Bonferroni tests would be further used to evaluate the differences. Thirdly, the coupling strength between rsfMRI-FC and DTI-SC were evaluated by Pearson's correlation analyses and multiple comparison with Bonferroni correction. Especially, the area under the curves (AUC) derived from receiving operating characteristic curves were used for the intergroup comparison of rsfMRI-FCs because of the sensitivity for detecting topological differences in brain FC networks ^[34]. Finally, the correlations between the coupling of rsfMRI-FC/DTI-SC and neuropsychological scores (MMSE, CDR-SB, FAQ and MoCA) were assessed by the partial correlation analyses, respectively.

3.   Results

3.1.   Evaluation of demographics and clinical variables

It was clear that there were no significant differences for age and gender ratio, and significant differences for the education level among the three groups. Furthermore, the scores of MMSE, CDR-SB, FAQ and MoCA presented significant differences across the three groups (Table 1).

3.2.   Comparison of global brain network characteristics

In Table 2, the brain network characteristics of rsfMRI-FC demonstrated that the σ and E_g of NC group were greater than those of MCI and AD groups, but the L_p of NC group was significantly lower than those of MCI and AD groups. Moreover, there existed significant differences for L_p and E_g between NC and AD groups and between NC and MCI groups, and no statistically significant differences were observed for all network characteristics between MCI and AD groups.

Accordingly, the brain network characteristics of DTI-SC were shown in Table 3. There were no significant differences for σ, E_l and C_p; the L_p of AD group was higher than those of MCI and NC groups, and the E_g of AD group was lower than those of other two groups. No significant differences were found for E_l between any two of the three groups. However, there were significant differences for L_p and E_g between MCI and AD groups, C_p and σ between MCI and NC groups, and L_p and E_g between MCI and AD groups.

3.3.   Coupling analysis of FC and SC

It was demonstrated that the rsfMRI-FC was positively correlated with DTI-SC and there were significant differences for FC-SC coupling across the three groups and the coupling strength in AD group was greater than those of NC and MCI. With the coupling analysis, the coupling strengths (mean ± variance) were obtained with the NC of 0.272 ± 0.062, MCI of 0.297 ± 0.050 and AD of 0.328 ± 0.075, respectively. There were significant differences among the three groups for FC-SC coupling at the p-value of 0.003.

3.4.   Correlations between FC-SC coupling and neuropsychological scores

The correlations between FC-SC coupling and MMSE, CDR-SB, FAQ and MoCA scores were shown in Figure 2. The coupling of rsfMRI-FC/DTI-SC was negatively correlated with MMSE score, positively correlated with CDR-SB and FAQ scores, and no correlation with MoCA score, respectively.

4.   Discussion

In this study, the network coupling of FC-SC based on graph theory was evaluated via the three aspects of network characteristics, coupling analysis and correlations with the neuropsychological scores. The results showed that the coupling strength varied across three groups and this network coupling was proved to be correlated with neuropsychological scores in different trends.

According to the graph theory of regular and random networks, any small-world networks could facilitate the efficiency of information transmission and processing at lower processing cost and capacity consumption ^[35]. The inconsistency change of C_p among groups may be related to the differences in sample size, network node and threshold selection. During the AD progression, increased L_p and decreased E_g suggested that the information transmission rate and integration capacity were obviously reduced, and this would in turn lead to the observed impairments in cognitive and executive functions ^[36]. Indeed, the structural core of the brain played a great role in integrating information across functionally segregated brain regions ^[37,38]. The changes of the damaged FC and SC networks provided not only the critical imaging markers but also the potential clues to understand the underlying mechanism of cognitive deficits of AD ^[39].

For the rsfMRI-FC, the NC had larger σ and smaller L_p, while the AD had larger L_p. It was consistent with the previous neuroscientific findings ^[19]. In addition, the efficiency of the rsfMRI-FC decreased along with the smaller E_g in AD. It was indicated that the topology of damaged brain FC would certainly lead to the decrease of global network processing efficiency in AD ^[10]. Similarly for the DTI-SC network, there were no significant difference for the σ among the three groups. It was hinted that the human brain might maintain the optimal network topology to achieve the most efficient information transmission for the most efficient information transmission during progressive neurodegeneration. While the increased L_p and decreased E_g reflected the impairment of global connectivity of DTI-SC in AD ^[36].

With the coupling analysis, there were significant correlations for rsfMRI-FC/DTI-SC. The correlations between FC-SC coupling and neuropsychological scores (MMSE, CDR-SB, FAQ and MoCA) provided further support for the assumption that these changes in network coupling contributed to the cognitive decline in AD. It was reported that the coupling strength of rsfMRI-FC/DTI-SC was increased with the known greater severity of cognitive deficits in AD ^[40,41]. It might be explained by the reasons that the loss of plasticity within pathways connecting different brain regions (nodes) for constrained brain network dynamics ^[42]. The integrity of SC might reflect the ability of the cerebral cortex to maintain functional diversity and neural activity interactions ^[43]. Therefore, the increased coupling indicated that AD would produce more direct functional interactions related to anatomical connections, which implied that brain function became more rigid, less dynamic and adaptable. Moreover, it reduced the restructuring capacity of functional network and mitigated the effects of brain regional degeneration ^[44,45].

This study had several limitations. First, the gender effect of participants was not considered due to the limited sample size. Second, DTI-SC was certainly underestimated for the missed information of crossover fibers during DFT. More accurate SC would be achieved by the betterment of fiber tracking algorithms on higher resolution data sets. Finally, multiple modalities of brain networks should be involved for convincing results, such as the metabolism connectivity of positron emission tomography (PET) and FC of electroencephalography. It would be of further interest to investigate disease-related changes in these dynamic properties and their relationship to cognitive decline on more larger data samples.

5.   Conclusions

It was concluded that there existed FC/SC coupling and varied network characteristics for rsfMRI and DTI, and this would provide the clues to understand the underlying mechanisms of cognitive deficits of AD.

Acknowledgments

This study was funded by the grants of the National Key Research and Development Program of China (2020YFC2008700) and the National Natural Science Foundation of China (Nos 61971275, 81830052, and 82072228).

Conflict of interest

The authors declare no conflict of interest.