Functional, Anatomical, and Morphological Networks Highlight the Role of Basal Ganglia–Thalamus–Cortex Circuits in Schizophrenia

Abstract

Evidence from electrophysiological, functional, and structural research suggests that abnormal brain connectivity plays an important role in the pathophysiology of schizophrenia. However, most previous studies have focused on single modalities only, each of which is associated with its own limitations. Multimodal combinations can more effectively utilize various information, but previous multimodal research mostly focuses on extracting local features, rather than carrying out research based on network perspective. This study included 135 patients with schizophrenia and 148 sex- and age-matched healthy controls. Functional magnetic resonance imaging, diffusion tensor imaging, and structural magnetic resonance imaging data were used to construct the functional, anatomical, and morphological networks of each participant, respectively. These networks were used in combination with machine learning to identify more consistent biomarkers of brain connectivity and explore the relationships between different modalities. We found that although each modality had divergent connectivity biomarkers, the convergent pattern was that all were mostly located within the basal ganglia–thalamus–cortex circuit. Furthermore, using the biomarkers of these 3 modalities as a feature yielded the highest classification accuracy (91.75%, relative to a single modality), suggesting that the combination of multiple modalities could be effectively utilized to obtain complementary information regarding different mode networks; furthermore, this information could help distinguish patients. These findings provide direct evidence for the disconnection hypothesis of schizophrenia, suggesting that abnormalities in the basal ganglia–thalamus–cortex circuit can be used as a biomarker of schizophrenia.

Introduction

Schizophrenia is a severe mental disorder characterized by hallucinations, delusions, loss of initiative, and impairments in cognitive function. These abnormalities are considered to result from abnormal neuroanatomical connectivity and its functional consequences during cognition.¹ The emergence of in vivo neuroimaging techniques has provided new evidence supporting the disconnection hypothesis, which argues that schizophrenia is associated with abnormal or inefficient communication between functional brain regions, and disruptions in white matter structure.^2–6 Numerous functional magnetic resonance imaging (fMRI) studies have identified abnormalities in functional connectivity in patients with schizophrenia, mostly concentrated within the default mode network, salience network, central executive network, and the thalamocortical loop.^7,8 Diffusion tensor imaging (DTI) research has further verified that patients with schizophrenia exhibit abnormalities in multiple white matter tracts.^9,10 Furthermore, structural MRI (sMRI) studies have suggested that gray matter atrophy occurs in numerous brain regions in patients with schizophrenia,¹¹ and that these regions form an integrated structural network.¹² In addition, group-level morphological network analyses have demonstrated that patients exhibit a reduced ability to integrate information across different brain regions,^13,14 and delayed maturation of temporal–occipital connections.¹⁵

However, most previous studies have focused on single modalities only, each of which is associated with its own limitations. For instance, functional network analyses are highly sensitive to physiological noise and head motion, which can result in systemic false correlations.¹⁶ Anatomical network represents the white matter connection between the brain regions, namely, the axonal projection.¹⁷ Most DTI studies have used deterministic tractography algorithms to construct structural networks. Although such calculations can be performed quickly, they do not address the issue of fiber crossing. In contrast, the probability fiber tracking algorithm takes into account the uncertainty of fiber orientation estimates, which can improve the tracking accuracy.¹⁸ Morphological gray matter networks have provided valuable information regarding the coordination of gray matter (mainly including the soma of neuronal and dendrite structures) growth/atrophy,^19,20 but their construction is defined by the correlation of morphological data between brain regions at the group level. Thus, such analyses are unable to provide information regarding individual differences. Given the limitations of single modalities, multimodal analysis may provide a new avenue for investigating diseases that affect multiple aspects of the brain, such as schizophrenia.^21,22

Schizophrenia researchers have recently begun to incorporate multimodal imaging techniques. A common multimodal fusion method is to constrain one modality with another. For instance, Cocchi et al²³ first identified changes in whole-brain functional connectivity among patients, following which they evaluated the integrity of white matter pathways in the abnormally connected functional networks. Although this technology is powerful, it may be impractical for widespread/repeated use. However, multivariate methods such as joint independent component analysis²⁴ or multimodal canonical correlation analysis²¹ can be used to explore the associations between the feature sets of different modalities. This method only considers the features at the single-voxel level and does not directly utilize information regarding connectivity. In our recent study,²⁵ we also developed a novel multimodal method. Combining 3 modalities, we constructed the multi-index features of each region of interest (ROI) using Wilk’s lambda statistics and detected significant changes using multivariate analysis of variance. However, in this previous study, we only focused on local features such as gray matter volume, and did not consider the connectivity perspective.

Early diagnosis and symptom prediction of mental diseases are an important focus of disease research. The imaging-based machine learning method has been increasingly applied in the research on brain disease, and several studies suggest that the brain-specific indexes extracted using various brain imaging techniques can well identify patients.^26,27 Compared with the recognition difference at a group level, modal classification at an individual level is more challenging.²⁸ Various biomarkers may carry complementary information. Therefore, the combination of multimodal features rather than a single modal feature is a promising option to obtain enhanced classification accuracy.²⁹

To overcome the aforementioned limitations and expand upon our previous study, we aimed to investigate abnormalities in connectivity among patients with schizophrenia, as well as the relationships between different modalities, by directly constructing 3 modal networks in conjunction with the machine learning method. To the best of our knowledge, our study is the first to directly construct 3 single-modality networks to investigate schizophrenia. The analytical method was used to determine (a) whether biomarkers discovered via a single modality can reliably distinguish patients from healthy controls; (b) whether fusion of multiple modalities improves the rate of accurate classification; and (c) whether there were any associations between biomarkers discovered via different modalities.

Materials and Methods

Participants

One hundred and thirty-five patients with chronic schizophrenia and 148 healthy control subjects were recruited for this study and the patients were identified according to the DSM-IV diagnostic criteria by qualified psychiatrists at the Taipei Veterans General Hospital, Taipei, Taiwan. The patient and control groups were well matched by gender (chi-square test, P = 0.77) and age (2 samples t test, P = 0.14). Patient and healthy control demographics are shown in table 1. In order to remove the possible influence of age and gender, they were used as the covariates for regression in the subsequent group comparison and discriminant analysis.

Table 1.

Subject Demographics

Characteristics	Patient	Control	P value
	(n = 135)	(n = 148)
Male/female	57/78	65/83	.77
Age (year)	43.9 ± 11.0	41.9 ± 11.1	.14
Education (year)	12.4 ± 3.5	16.1 ± 3.5	<.001
Duration (year)	16.4 ± 10.3	—	—
PANSS positive scale score	9.5 ± 3.1	—	—
Negative scale score	9.8 ± 4.7	—	—
General psychopathology score	20.5 ± 4.4	—	—
Total score	39.8 ± 9.8	—	—
Chlorpromazine equivalents (mg)	468.8 ± 405.2	—	—

Note: PANSS, Positive and Negative Symptom Scale.

Resting-state fMRI data, DTI data, and T1-weighted 3D high-resolution brain images were acquired for each subject on a 3T MR system (Siemens Magnetom Tim Trio, Erlangen, German) at National Yang-Ming University, equipped with a high-resolution 12-channel head array coil (see supplementary material for more sample information description and acquisition parameters).

Data Preprocessing

Statistical Parametric Mapping package (SPM12: http://www.fil.ion.ucl.ac.uk/spm) and Data Processing Assistant for resting-state fMRI³⁰ were used for preprocessing functional images. The PANDA toolbox³¹ was adopted to preprocess DTI data. The T1-weighted structural data were preprocessed using the Diffeomorphic Anatomical Registration using Exponentiated Lie algebra toolbox³² in SPM12 software. Detailed descriptions of the data preprocessing procedures were in the supplementary material.

Network Construction

Functional Networks

For the preprocessed data, the Human Brainnetome (BN) Atlas³³ was used to parcellate the brain into 246 ROIs (supplementary table S1 lists the names and abbreviations for these regions). The time series of each ROI was extracted by averaging the time series of all voxels within it, and then we computed the Pearson correlation coefficient across subjects between the time signals of every pair of ROIs as the functional connectivity, which resulted in 30 135 (246 × 245/2) links.

Anatomical Networks

The anatomical network of each subject was constructed using the probability tracking algorithm. First, the T1-weighted structural images and the corresponding fractional anisotropy (FA) map of each subject were linearly registered through affine transformation, and then each single transformed structural image was standardized through nonlinear transformation, with ICBM-152 as the template. Subsequently, the BN template was registered into the individual DTI space of each subject through inverse transformation. Afterwards, these 246 nodes were used as the seed points, respectively, to carry out probability tracking using the FDT of FSL package¹⁸; typically, the connection probability of seed region i to another region j was defined as the number of fibers passing the region j divided by the total number of fibers sampled in region i (5000 × n, n is the number of voxels in region i). Noteworthily, the connection probability from i to j was not necessarily equal to that from j to i; therefore, these 2 probabilities were averaged to define the nondirectional connection probability P_ij between region i and region j.

Morphological Networks

According to the method used by Wang et al,³⁴ the Kullback–Leibler (KL) divergence-based similarity measurement (KLS) was utilized to quantify the morphological connection between 2 regions.³⁵ For each subject, the gray matter volume values of all voxels in each brain region were first extracted; subsequently, the probability density function of these values was estimated using the kernel density estimation.³⁶ Then, the probability distribution function (PDF) was calculated for the obtained probability density function. Afterwards, the KL divergence between any pair of ROI was calculated as:

$${\displaystyle \begin{array}{l}{\displaystyle D_{KL}(P,Q)={\displaystyle \sum _{i=1}^n}(P\left(i\right)\log \frac{P\left(i\right)}{Q\left(i\right)}+Q\left(i\right)\log \frac{Q\left(i\right)}{P\left(i\right)}),}\end{array}}$$

where P and Q are 2 PDFs, n is the number of sample points (here we choose n = 2⁷, as the same in Wang et al³⁴).

Finally, the KLS was computed as:

$${\displaystyle \begin{array}{l}{\displaystyle KLS(P,Q)=e^{-D_{KL}(P,Q)}}\end{array}}$$

The value range of KLS is [0,1]. Higher KLS value was obtained when the gray matter density distribution of 2 brain regions was closer (supplementary figure S2). The KLS-based morphological brain network could be generated for each subject after all these analyses.

Hub Regions for Each Modality

For each modality, the individual brain networks of the control group were averaged to obtain the group-level network. Subsequently, the weight of each brain region was calculated based on this group-level network, and the top 5% of brain regions in terms of weight were selected as the hub regions of the network.

SVM Classifier

In our study, the LIBSVM toolbox for MATLAB was applied to implement the support vector machine (SVM) classification.³⁷ Dimensionality reduction is necessary prior to discriminant analysis.^38–40 Consequently, rank-sum tests comparing every connection across the classes were performed separately for the training set, yielding a P value for each connection. Statistically significant features (P < 0.1 Bonferroni corrected) were selected, whereas the rest were eliminated. We used a nested 10-fold cross-validation strategy⁴¹ with inner cross-validation to determine the optimal parameters (including different kernel types: linear, t = 0; polynomial, t = 1; radial basis function, t = 2; and different trade-off parameter C: 0.001, 0.01, 0.1, 1, 10, 100, 1000, 10 000) and outer cross-validation was used to determine the classification performance as we did in our previous study.²⁵ Finally, the experiment was repeated 100 times to avoid any bias introduced by random partitioning in the cross-validation, thus determining the average accuracy on the test set. We had trained and summarized 7 models to compare the judgment effects of single modes and different modal combinations (including fMRI, DTI, sMRI, fMRI + DTI, fMRI + sMRI, DTI + sMRI, and fMRI + DTI + sMRI).⁴²

Consensus Features for Each Modality

When using the cross-validation method to evaluate classification performance, the feature set selected as the final output was slightly different, and different connections provided different contributions to classification. The consensus features refer to the connections that could be selected as the feature input in each cross-validation.^40,43 In this study, we focus on consensus features for each modality.

Robustness of the Network Analysis

The robustness of our findings was estimated by employing another parcellation scheme to construct the networks, known as the automated anatomical labeling atlas (AAL),⁴⁴ which contains 90 ROIs (supplementary table S2). The discriminant analyses were then repeated as described earlier.

Results

Networks

Figure 1 depicts the multimodal networks obtained for healthy controls at the group level. The functional network exhibited potent homotopic connections, whereas the anatomical network was sparse, and its connections were mainly short-range connections to adjacent regions. The morphological network also exhibited strong homotopic connections, indicating that the gray matter density distribution in the contralateral homotopic region was the most similar, consistent with group-level networks constructed based on cortex thickness.¹⁶ Moreover, there was a great difference in the gray matter density distribution between the subcortical and cortical regions. Consequently, the KLS values calculated for most (but not all) subcortical and cortical regions were relatively low.

Fig. 1.

Group networks in healthy controls.

Figure 2 illustrates the hub regions in each network. The hub regions of the 3 major networks differed greatly: That of the functional network was located mainly within the temporal and frontal lobes, that of the anatomical network was located mainly within the subcortical region, and that of the morphological network extensively distributed throughout the cortical region but mainly within the temporal lobe.

Fig. 2.

Hub regions for the 3 networks in healthy controls.

Classification

In this study, SVM was utilized to classify whether a sample belonged to the patient group or the control group (table 2). For single-modality analyses, the anatomical network constructed using DTI exhibited the highest accuracy rate (84.06%), followed by the morphological network constructed using gray matter volume, which achieved an accuracy rate of 80.41%. The accuracy rate of the functional network was 71.96%. Typically, classification accuracy improved after combining the network features of the 3 modalities, achieving an accuracy of up to 91.75% with a sensitivity and specificity of 90.90% and 92.53%, respectively. The classification accuracy obtained using 3 modalities was significantly higher than that obtained using 2 modalities, whereas that obtained using 2 modalities was markedly higher than that of a single modality (P < .0001 for all comparisons, rank-sum test), suggesting that each modality is essential for achieving a favorable classification.

Table 2.

Discrimination Accuracy of Different Models

	Accuracy (mean/std) (%)	Sensitivity (%)	Specificity (%)
fMRI	71.96/1.75	69.64	74.09
DTI	84.06/1.54	83.47	84.59
sMRI	80.41/1.99	79.8	80.97
fMRI + DTI	86.31/1.29	85.33	87.21
fMRI + sMRI	85.23/1.75	82.79	87.47
DTI + sMRI	90.61/1.21	89.70	90.14
fMRI + DTI + sMRI	91.75/1.05	90.90	92.53

Note: DTI, diffusion tensor imaging; fMRI, functional magnetic resonance imaging; sMRI, structural magnetic resonance imaging.

Consensus Features

We examined the consensus connections for the classification model that combine the 3 modal networks (figure 3, supplementary table S4). The functional network yielded 96 such connections: 4 (4.2%) of these connections, which were connections between the thalamus and cortex, were enhanced in the patient group; in patients with schizophrenia, most of the connections were weakened, especially those involving the frontal lobe (SFG, OrG), insular, subcortical (BG), temporal (STG, PhG), and parietal (IPL) regions, particularly the subcortical–cortical connections. Frontal abnormal connections were revealed primarily with the insular, temporal, and basal ganglia region. The abnormal subcortical–cortical connections with the basal ganglia and cingulate connection were the most significant.

Fig. 3.

The “consensus” features identified via the different modalities. (A) Functional network, (B) anatomical network, and (C) morphological network.

The anatomical network yielded 27 connections, which were extensively distributed throughout the brain. Four (15%) of these connections were enhanced in the patient group. However, most connections were weakened in patients with schizophrenia, including 9 within the subcortex (mainly in the thalamus) and 1 from the right globus pallidus to the inferior frontal gyrus. Cortical connections are concentrated within the ipsilateral cortex, especially the connections within the right temporal lobe (4 connections) and the right parietal lobe (4 connections).

The morphological network yielded 8 connections, 4 of which were enhanced in the patient group, all of which were the connection between subcortex and cortex. Four connections were weakened in patients with schizophrenia, and all weakened connections were between the frontal lobe with other regions.

Network Analysis Robustness

We repeated our network analysis using the same process via the AAL atlas. Supplementary figure S5 depicts the multimodal networks obtained for healthy controls at the group level. The judgment analysis results also indicated that the combination of multiple modalities could markedly enhance the accurate judgment rate (supplementary table S5). Typically, the accurate rate of 3-modality testing reached 89.94%: the accurate judgment rate of the morphological network was 82.78%, the functional network was 70.84%, and the anatomical network was 69.83%. And we also detected the consensus features for each modality (supplementary figure S6). The analysis results of the networks defined by 2 templates were basically similar, and the greatest difference lay in the thalamus region, because the AAL template did not segment the thalamus into different subregions, leading to the lack of numerous thalamus results.

Discussion

In this study, we utilized brain imaging data from 3 different modalities to construct functional, anatomical, and morphological networks for each participant, to identify biomarkers of brain connectivity in patients with schizophrenia. The biomarkers identified using these 3 networks were mostly located in the basal ganglia–thalamus–cortex circuit, achieving a discriminant accuracy of up to 91.75%. Relative to single-modality features, integration of the 3 modalities significantly improved classification. These results highlight the critical role of the basal ganglia–thalamus–cortex network in the pathophysiology of schizophrenia.

In our study, the functional network was constructed using resting-state fMRI, and the observed functional connections were used as features for classification, yielding a diagnostic accuracy of approximately 72%. The judgment effects of the functional networks constructed using the 2 atlases were both mediocre; however, the connections detected were far more significant than those detected by the other 2 modalities. This can be easily explained because structure is generally considered as the basis of function,⁴⁵ and multiple functional connections and even the entire functional network may be affected when one connection in the structure is affected. However, the precise relationship between structure and function may not be simple, and a general consensus is that, resting-state functional connection strength is positively correlated with the structural connection strength.⁴⁶ However, functional connection can also be observed between regions with a few structural connections or no connections, which suggests the indirect structural connection-mediated functional association (namely, through a third region).⁴⁷ Consequently, a functional network may be more complex than a structural network, and the analysis method used in our research may not well explain the rich information encompassed in this network. Abnormal functional connections include abnormal enhancement of the thalamocortical connections and abnormal reduction of the basal ganglia–cortical connections, which are consistent with previous studies, including studies on the construction of functional networks from whole-brain voxel-level⁴⁸ and studies specifically targeting thalamocortical connections⁸ and striatum–cortical connections.⁴⁹ Furthermore, the BN templates used to build the network showed multiple connections within the insular region, which is a rare finding. The AAL atlas, used to build the functional network, detected only a small number of connections involving the insular region, indicating that this region may have been neglected because the template used did not subdivide this region, causing important information to be dismissed.

The diagnostic accuracy of the anatomical network reached as high as 84%. In addition, the consensus connections identified using DTI were extensively distributed throughout the brain, consistent with our previous voxel-based analysis results obtained using the same data, which revealed that FA values are reduced in most brain regions in patients with schizophrenia.¹⁰ In a recent study,⁵⁰ the investigators collected 4322 samples from 29 datasets, also reporting that FA values were generally lower among patients than among controls. In addition to extensive alterations to internal cortical connections, patients with schizophrenia also exhibited numerous internal subcortical connections. Interestingly, the AAL template did not segment the subregions of the thalamus; as a result, the AAL template can only identify connection differences within the cortical lobe, whereas the connection information within the thalamus was lacking, which greatly reduces the accuracy of classification (<70%), thus further highlighting the importance of treating the difference in networks within the thalamus as a significant marker to distinguish a patient.

The individual morphological network was constructed based on KL divergence to measure similarities in the distribution of gray matter volume. Such methods were first utilized by Kong et al^35,51 in 2014 and further developed by Wang et al.³⁴ These authors discovered that morphological network analysis at the individual level is a significant and reliable method for characterizing human brain structure. Our results indicated that this method attained relatively high accuracy using both atlases. The results obtained using the BN template mainly focused on the thalamus–cortex circuit, whereas those obtained using the AAL mainly focused on the basal–ganglion–cortical circuit. Because both templates had relatively high judgment rates, we believe that there should be a significant difference between patients and normal subjects for both of them. The abnormality in subcortical volume has been verified in 2 large sample studies.^52,53 Moreover, a previous group-level network based on gray matter volume also detected subcortical–cortical gray matter coupling abnormalities.¹² Our results suggest that an abnormal increase in subcortical–cortical connectivity may reflect a synchronous decrease in gray matter integrity.

After merging the network connections of the 3 modalities as features, diagnostic accuracy was higher than that obtained using any single modality, suggesting that the combination of multimodal features provided complementary information carried by different biomarkers. Although the connections identified by each modality differed, the convergent pattern is that all were mostly located within the basal ganglia–thalamus–cortex circuit. Several previous studies specifically investigated subcortical gray matter injury,^53,54 thalamocortical dysconnectivity,⁸ and corticostriatal dysconnectivity^49,55–57 in patients with schizophrenia. However, these studies utilized hypothesis-based methods and were unable to determine the positions of abnormalities in the thalamus and basal ganglia in the patient group. In this study, we started with whole-brain networks derived using 3 different modalities, following which we utilized a purely data-driven method to identify abnormalities in the basal ganglia–thalamus–cortex circuit. Our findings suggest that such abnormalities can be used to distinguish patients from controls, further highlighting their core role in the pathophysiology of schizophrenia. However, as the effects of medication cannot be completely ruled out, our results may include samples of psychosis risk, not just schizophrenia.

Brain connectivity can be a confusing concept, as the term can refer to various types of “connections.”⁵ Thus, brain networks constructed via different modalities may reflect different aspects of brain architecture. Gong et al⁵⁸ suggested that the morphological network constructed using cortex thickness correlations differs significantly from white matter networks constructed using DTI. One recent study¹⁶ compared 4 types of trans-species and trans-modal networks, revealing that the congruity of these 4 networks was relatively poor. In this study, the consensus connections discovered using the 3 modalities did not overlap, and no significant correlations were observed among the consensus connections identified using different modalities (see supplementary materials). Such differences may be related to the influence of genetic risk factors on different brain tissues in different manners or at different stages.⁵⁹ Moreover, some previous studies have reported the coupling of functional structure,⁴⁶ but these studies were performed using a few restricted seed points, which may make it easier for the structural functional coupling correlation to pass the correction.

This study possesses some limitations of note. First, we utilized simple splicing methods only for the classification of the multimodal brain images, thus ignoring the relationships between different modalities. Future studies should take these relationships into account. Second, the methods used to construct the 3 networks remain to be improved: Our functional network was constructed based on the hypothesis of static network interaction; however, the interaction between large-scale functional brain networks was unstable. Some recent studies focused on deriving more valuable information by constructing dynamic functional networks.^60,61 However, there are many indices for the measurement of structural network edges. The FA value for the fiber pathway may be more suitable for measuring white matter integrity. Individual morphological networks have only been proposed within the last few years, and the physiological significance of morphological connectivity remains unclear. Third, we used the BN atlas to define the node for 3 modalities because it is a fine-grained, cross-validated atlas with a framework for integrating multimodal information.³³ We had used the AAL template to repeat our analysis, but there was a certain difference in the results between the 2 templates, particularly in the thalamus region and it is still unclear how to assess and control the influence of different node definition.^62,63 Fourth, patients with schizophrenia exhibited relatively low levels of education, which may be related to cognitive deficits before or after disease onset.¹³ Finally, most patients with schizophrenia enrolled in this study had taken antipsychotics. We carried out an analysis to explore the Spearman correlation between the chlorpromazine equivalents and the consensus connection strength; however, we did not find any significant correlation (supplementary table S4). Nevertheless, we cannot rule out the effect of drugs on our results. To exclude the influence of psychotropic drugs on network connectivity, further studies should investigate the patterns of connectivity in drug-naive patients.

Funding

National Natural Science Foundation of China (NSFC; 11671129 to Professor Guo); Ministry of Science and Technology (104-2218-E-010-007-MY3, MOST 107-2420-H-010-001, and MOST 107-2634-F-010-001 to Professor Lin); National Health Research Institutes, Taiwan (NHRI-EX106-10611EI to Professor Lin); Science and Technology Commission of Shanghai Municipality (17JC1404105 and 17JC1404101 to Professor Lin); Ministry of Science and Technology (MOST 107-2634-F-075-002 to Professor Tsai); Hunan Provincial Innovation Foundation for Postgraduate (CX2018B237 to Dr. Zhao).

Conflicts of Interest

All authors declare no competing interests.