Topological Alterations and Symptom-Relevant Modules in the Whole-Brain Structural Network in Semantic Dementia

Participants

We recruited 19 adults with SD (12 males; mean age: 61.26±8.63 y; mean education level: 11.53±3.34 y; disease duration: 3.03±1.46 y) and 20 healthy controls (8 males; mean age: 60.50±3.93 y; mean education level: 10.45±2.89 y) from Huashan Hospital in Shanghai. All participants were right-handed [24] native Chinese-speakers. The subject sample was the same as that in our recent study [25]. The participants underwent a set of comprehensive neuropsychological assessments. All patients satisfied the diagnostic criteria of SD [26], and presented selective dysfunction in semantic processing and atrophy in the ATL (see Table 1). This study was approved by the Institutional Review Board of the National Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University.

Collection and preprocessing of neuropsychological data

The methods used to collect and preprocess the neuropsychological data in the present study have been previously elaborated by Ding et al. [25]. For simplicity, these methods are not described in detail here.

Each subject was evaluated on the following three aspects of cognitive abilities. (1) Semantic processing ability. Semantic processing ability was assessed by the following six tasks: oral picture naming (n = 140; subjects were instructed to name the objects shown in pictures that appeared on a screen), oral sound naming (n = 36; subjects were instructed to name the objects that produced the sounds they heard), picture associative matching (n = 70; subjects were instructed to identify which of the two pictures of objects shown in the bottom was semantically closer to the top picture of the object that appeared on the screen), word associative matching (n = 70; this task is identical to the picture associative matching task except that the pictures were replaced with the corresponding written words), word-picture verification (n = 70; subjects were instructed to determine whether the object and the word on the screen were identical), and naming to definition (n = 70; subjects were instructed to name the objects whose definitions were visually and aurally presented). The tasks shared semantic components but varied in the modalities of the inputs and outputs. (2) Non-semantic processing ability. Non-semantic processing ability was evaluated by the following three tasks: sound perception (n = 44; subjects were instructed to determine whether the episode of sound has the same rhythm and pitch), visual perception (n = 30; subjects were instructed to determine whether two circles were the same), and proximity judgment (n = 3; subjects were instructed to determine which of two numbers was quantitatively closer to a third number that appeared on a screen). These tasks required minimal semantic involvement. (3) General cognitive ability was measured using the Chinese version of the Mini-Mental State Examination (MMSE) [27].

To obtain the cognitive measures, we first calculated the accuracy of each subject on each task. Each patient’s accuracy on each task was further corrected into a standardized t-score based on the distribution of scores of the healthy control group to rule out the influence of demographic variables. This method was developed by Crawford and Garthwaite [28] and has been widely adopted [25, 30]. Specifically, a regression model was first established for each task in the normal control group, in which the accuracy of the task was treated as the dependent variable and the demographic factors (i.e., age, gender, and education level) were treated as predictors. Then, the predicted accuracy of a patient was acquired by adding his or her demographic factors into the model. The discrepancy value was further extracted by the observed accuracy minus the predicted accuracy. Finally, we calculated the standardized t score of patients, (i.e., the quotient of the discrepancy value divided by the corrected standard error of the estimate).

To obtain a semantic index that could reliably reflect the degree of the semantic deficits in the SD subjects, we conducted a principle component analysis (PCA) using the above-mentioned 6 semantic tasks and 3 non-semantic tasks from the 19 patients. We used subcommands for varimax rotation, a plot of eigenvalues (>1), and a principal components extraction. The factor with the highest loading weight on all six semantic tasks was considered as the semantic factor, and the scores corresponding to this factor were considered to reflect the semantic performance of the SD subjects.

Collection and preprocessing of neuroimaging data

Construction of the networks in the SD and healthy subjects

We constructed whole-brain white-matter networks in the two subject groups using the following steps (see Fig. 1).

Investigation of the alterations in the global properties in the SD network

To determine whether SD disrupts the healthy structural network, we compared the global network attributes of whole-brain networks between the SD and healthy groups. Small-world attributes have been widely used to depict the topological properties of networks [37, 38]. There are two common small-world measures, i.e., γ and λ, which represent the local and global efficiencies in a given network, respectively. These measures are calculated using the following formula: γ= C^real / C^rand and λ= L^real / L^rand, where C^real and L^real are the observed clustering coefficient and characteristic path of brain networks, and C^rand and L^rand are the corresponding measurements of 10,000 random networks preserving the same number of nodes and edges and degree of distribution as a real network [39]. The cluster coefficient quantifies the connection density between the neighbors of a node ( C = 1 N ∑ i E i 0 . 5 * k i * ( k i - 1 ) . N is the number of nodes. E_i is the number of connections among the neighbors of node i. K_i is the number of neighbors of node i.). The characteristic path is the average of shortest path lengths between each node pair in the network ( L = 1 N * ( N - 1 ) ∑ i , j , i ≠ j d ij . N is the number of nodes. i and j are the node pair in the network. d_ij is the shortest path length between node i and j.) [35, 41]. In addition, other global network measures (i.e., mean network degree, assortativity and betweenness) used in previous studies [20, 42] were also adopted in our study. The network degree is the average connectivity strength among all nodes. Assortativity is calculated by correlating the nodal degree and its neighbors’ mean degree. High assortativity value implicates that the nodes are tightly connected. Network assortativity is the mean value of all nodes. Betweenness is the number of shortest paths through a node divided by the number of all possible shortest paths, indicating the influence of a node on a network. The network betweenness is the average value of all nodes [20, 43].

Using the above-mentioned methods, we calculated the two small-world measures (γ and λ) and three global measures (network degree, assortativity and betweenness) for each SD and healthy subject. These processes were completed using a graph theoretical network analysis toolbox for imaging connectomics (GRETNA) [44]. Then, two-sample t-tests were adopted to compare the measures between the subject groups.

Identification of modules in the structural network in SD

To determine whether the anatomical networks in the SD patients and healthy controls were organized into different modules and whether the modules had group difference, we conducted a modularity analysis on the mean network matrix of the two subject groups using the spectral optimization algorithm in GRETNA [44]. To avoid false-positive connections, the threshold was defined as FA > 0.15 for the mean networks of the SD patients and controls (see similar methods in Agosta et al. [20]) The number of modules in the network was determined by acquiring the maximal modularity Q, which represents a comparison of the within-module and inter-module connections [45]. To determine whether the Q value was significantly higher than that in the random networks, a Z score was calculated by dividing the Q value in the real network by the Q values generated from 10,000 random networks that preserve the same number of nodes and connections, and degree of distribution [39].

The distribution of the node number in different modules was assessed by a series of chi-square tests comparing the SD and control cohorts to determine whether there is significant difference between the SD and control groups. To investigate the damage of all modules, we first extracted the masks of GMV and diffusion metrics in each module. The GMV and diffusion masks were composed by the regions of the module in the AAL atlas and all white matter tracts in the module, respectively. Second, the modular GMV values were extracted by averaging GMV of all voxels in the masks. The 4 mean diffusion metrics of each module were computed by averaging the diffusion metrics of all voxels in the mask of the module and dividing by the number of possible connections of the module in all subjects. Third, we compared the GMV and diffusion metrics between the groups.

Identification of modules associated with semantic deficits in SD

To find which modules of SD observed in the above analyses were related with the semantic disruptions in SD, we calculated the correlation between the four diffusion metrics of each module and the semantic PCA scores across the 19 patients. To further determine whether the semantic-relevant effects of the modules observed in the above analyses were driven by other confounding variables, we carried out partial correlation analyses between the observed module’s FA values and the semantic PCA scores by additionally partialling out the influence of the following potential confounding variables: (1) total GMV (GMV of all voxels in the whole-brain gray-matter mask), (2) modular GMV (GMV of all voxels in the mask of the corresponding module), (3) general cognitive state (the corrected t-score of the MMSE test), and (4) the atrophy lateralization (a dichotomic variable: left- or right-hemispheric predominate atrophy was coded as 1 or 0, respectively, see a similar method in Ding et al. [25]).

Neuropsychological and atrophy profiles of participants

Table 1 shows the demographic information, neuropsychological performance, and cerebral atrophy of the SD patients relative to those of the healthy subjects. There were no differences between these two groups in terms of the demographic variables (p values > 0.15). According to the raw accuracies, the SD patients exhibited an impaired semantic ability (t-values < –8.02, p values < 0.001), general cognitive ability (t = –6.98, p < 0.001), and other cognitive abilities (word reading: t = –4.97, p < 0.001; sound perception: t = –3.45, p < 0.002; Rey-O recall: t = –3.08, p < 0.004), but the abilities of oral repetition (t = –0.65, p = 0.52), visual perception (t = 1.93, p = 0.06), sentence comprehension (t = –1.11, p = 0.28), executive control (t = 1.46, p = 0.16), and arithmetic (t = –1.49, p = 0.15) were spared. These results were also observed using the standardized t scores of the SD group. Moreover, the SD patients showed severe atrophy in the bilateral ATL and total GMV (t values > –4.55, p values < 0.001; Table 1), extending into other temporal and frontal areas (Fig. 2A; FDR corrected p < 0.05). In total, 13 patients exhibited the left-hemispheric predominant atrophy.

OPEN IN VIEWER

Fig.2

Atrophy pattern in the SD patients and white-matter networks in the normal control and SD groups. Panel A shows the cerebral atrophy in SD (FDR corrected p < 0.05). Panels B and C denote the modular and connectivity pattern in SD and NC. Panel D shows the labels of the regions, and their corresponding names are provided in Table 3. NC, normal controls; SD, semantic dementia; Temporal_L, left temporal/occipital/parietal module; Frontal_L, left frontal module; Frontal_R, right frontal/parietal module; Temporal_R, right temporal/occipital module.

The PCA analysis extracted three distinct components that accounted for 78% of the overall variance (i.e., component 1 = 47%, component 2 = 16%, and component 3 = 15%). The loading weights of all tasks on each component are listed in Table 2. Component 1 was treated as a semantic component since it had high loading weights on all semantic tasks (>0.61). Meanwhile, components 2 and 3 were treated as perceptual and arithmetic components because their heaviest loading weights were on the two perceptual tasks (>0.65) and proximity task (0.96), respectively.

Alterations in the global properties in the SD network

We found that the small-world attribute, i.e., the λ value, in the SD patients was significantly higher than that in the healthy subjects (SD: 1.08±0.01; controls: 1.08±0.01; t = 2.11, p < 0.03). However, the values of γ were comparable between the two groups (SD: 2.84±0.28; controls: 2.70±0.19; t = 1.84, p = 0.07). These findings demonstrate that the SD patients had a reduced global processing efficiency, but their local processing efficiency in the network was unaffected. We did not find any difference between the two cohorts in the other 3 global network measures (degree: t = –1.3, p = 0.20; betweenness: t = 1.52, p = 0.14; assortativity: t = 0.43, p = 0.67).

Modules associated with semantic deficits in SD

When we used an FA value > 0.15 as the threshold, the mean network densities in the SD patients and controls were 0.14 and 0.15, respectively. For each subject group, our modularity analysis divided the 90 nodes into the following four modules (SD: Q = 0.47, Z = 27.40; controls: Q = 0.44, Z = 24.64; see Figs. 2 & 3): the left temporal/occipital/parietal module (SD: 17 nodes; controls: 22 nodes; χ² = 0.64, p = 0.42), the left frontal module (SD: 28 nodes; controls: 27 nodes; χ² = 0.02, p = 0.89), the right temporal/occipital module (SD: 22 nodes; controls: 23 nodes; χ² = 0.02, p = 0.88), and the right frontal/parietal module (SD: 23 nodes; controls: 18 nodes; χ² = 0.61, p = 0.44). The label of each node is presented in Table 3.

OPEN IN VIEWER

Fig.3

Four modules in the SD white-matter network and the correlations between the diffusion metrics of each module and the semantic performance in the SD patients. The labels and names of the nodes in each module are provided in Fig. 2D and Table 3, respectively. *p < 0.05. FA, fractional anisotropy; MD, mean diffusivity; AD, axial diffusivity; RD, radial diffusivity; PCA, principle component analysis.

Seventy-four of the 90 nodes were divided into identical modules between the two groups. Relative to random level, the number of overlap nodes between the groups (74/90) was significantly higher than that at the random level (25%) (χ² = 157, p < 0.001), indicating that the SD patients had a module pattern in the node distribution that was highly similar to that in the controls. However, the gray-matter volumes of all modules were lower in the SD patients than those in controls (t-values < –  2.74, p-values <0.009). In contrast, SD disease led to an increase in the three white-matter integrity measures (i.e., MD, AD, and RD) in the left temporal/occipital/parietal module (t values > 2.17, p values < 0.04) and the RD values in the right temporal/occipital module (t = 2.12, p = 0.04). No other group differences reached a significant level (t-values <1.86 or > –0.76, p values > 0.07).

Figure 3 displays the correlations between the diffusion metrics in each module and the semantic PCA scores in the SD individuals. Significantly positive correlations were observed in the left temporal/occipital/parietal module in all metrics (r values > 0.54, p values < 0.02) and in the right frontal/parietal module in the RD values (r = 0.46, p = 0.05). However, a negative correlation was observed in the left frontal module in the FA value (r = –0.50, p < 0.03). Other correlations were not significant (r values < 0.44 or > –0.34, p values < 0.06).

Importantly, the correlation between damage to the left frontal module and the semantic impairments remained significant even after controlling for the total GMV, modular GMV, general cognitive state, and atrophy lateralization (partial r-values < –0.52, p-values < 0.03). Similarly, the effects in the left temporal/occipital/parietal module remained significant or marginally significant after controlling for all of the above-mentioned confounding factors (r values > 0.41, p values < 0.09) except for the modular GMV (r values < 0.22, p values > 0.38). Note that the GMV values in this module were significantly correlated with the semantic impairments in the SD patients (r = 0.73, p = 0.004). This finding suggests that the gray matter in this module might be involved in semantic processing. The effects of the right frontal/parietal module could not be accounted for by all confounding factors (r values > 0.44,p values < 0.07) except for the total volume (r = 0.37, p = 0.13). Moreover, the GMV values of the module were associated with the raw scores of the shape trail test (r = –0.58, p = 0.02), indicating that this module might be engaged in general executive control.

The global topological changes in the white-matter network of SD

We observed that the anatomical network in the SD subjects relative to that in the healthy controls had an increased λ value but no significant differences in γ or the other global network measures (degree, betweenness, and assortativity). This finding indicates that SD patients show a declined global efficiency with spared local efficiency. The two efficiencies reflect the different properties of a network. The global efficiency refers to the ability to transform information between any two nodes in the network and can be measured by the characteristic path length (λ value). The global efficiency mainly relies on the edges connecting far-distance nodes. In contrast, the local efficiency represents the ability to transform information between neighbor nodes, which can be measured by the cluster coefficient. The local efficiency mainly depends on the strength of local connections [38, 41]. Thus, our results indicate that distant white-matter pathways in the SD network are disrupted but the local white-matter tracts remain intact. This brain network pattern satisfies the deterioration rules in degenerative diseases [42]. Notably, Agosta et al. [15] also found that SD altered other global network measures (e.g., degree and assortativity) in the functional network. However, a similar pattern was not observed in our study. This difference might be due to the following differences between the two studies. One difference is the type of neural network type (functional versus structural network). The changes in SD might occur earlier in a functional network than in a structural network [46] or the brain atrophy might be more severe in a gray-matter network than in a white-mater network [47]. Therefore, the functional gray-matter network reported by Agosta et al. [20] presented more alterations than the structural white-matter network presented in our study. Other possible reasons include differences in the number of patients (13 versus 19) or brain lesion types (patients with predominant atrophy only in the left hemisphere versus those with predominant atrophy in the left and right hemispheres).

The semantic deficit-relevant modules in the SD network

We found that the anatomical network in SD could be divided into four independent modules. The modular patterns were quite similar in the SD and control groups, although there was reduced gray-matter atrophy in all the modules and significant white-matter changes in the left and right temporal/occipital modules of the SD patients. Moreover, the abnormalities in three modules (i.e., the left temporal/occipital/parietal module, the left frontal module, the right frontal/parietal module) were correlated with the semantic impairments in the SD patients. We speculate that the left temporal/occipital/parietal and frontal modules might be dedicated to the following critical components of semantic cognition: semantic representation and control, respectively [48–50]. However, the right frontal/parietal module might be crucial for general executive control.

Limitations

This study has some caveats as follows: (1) We only inspected the general semantic processing ability in SD subjects, but the finer semantic processing abilities (e.g., different categories or modalities [71, 72]) were not investigated. (2) We recruited an unequal number of SD patients with left and right dominant atrophy. This uneven sample may have led us to miss the effect of right temporal modules in semantic processing, because the semantic system is thought to depend on both hemispheres [50]. (3) The subject number was relatively low for the PCA analysis. (4) Some of our cognitive assessments were developed in house. (5) The influence of the fiber crossing on the white-matter integrity values (e.g., FA value) was not considered. (6) Our brain gray-matter template is not a good representation of functional regions. (7) Because the atrophy of patients did not affect all regions in the modules, the effect of the modules might be associative with only part of the modules.

Conclusion

After constructing the whole-brain neuroanatomical network of SD, we observed that the network had a decreased global efficiency. The network was separated into the following distinct modules: left temporal/occipital/parietal, frontal, right frontal/parietal and temporal/occipital modules. The first two modules were associated with the semantic deficits in SD. The third module might be crucial for general executive control. These findings can deepen our understanding of the structure of the SD neural network, which may have important clinical implications for the treatment of this disease.