The effect of the total small vessel disease burden on the structural brain network

Different cerebral small vessel disease (SVD) lesion types have been shown to disrupt structural brain network individually. Considering that they often coexist, we investigated the relation between their collective effect using the recently proposed total SVD score and structural brain network on MRI in 95 patients with first transient ischemic attack (TIA) or ischemic stroke. Fifty-nine patients with and 36 without any SVD lesions were included. The total SVD score was recorded. Diffusion tensor imaging was performed to estimate structural brain connections for subsequent brain connectivity analysis. The global efficiency and characteristic path length of the structural brain network are respectively lower and higher due to SVD. Lower nodal efficiency is also found in the insular, precuneus, supplementary motor area, paracentral lobule, putamen and hippocampus. The total SVD score is correlated with global network measures, the local clustering coefficient and nodal efficiency of hippocampus, and the nodal efficiency of paracentral lobule. We have successfully demonstrated that the disruption of global and local structural brain networks are associated with the increase in the overall SVD severity or burden of patients with TIA or first-time stroke.

Global and local structural brain networks. The structural brain network of the two patient cohorts are shown in Fig. 1. They resemble the properties of small-world network with larger clustering coefficient (patients without SVD: 2.90 ± 0.43; with SVD: 3.03 ± 0.63) and equivalent characteristic shortest path length (without SVD: 1.25 ± 0.03; with SVD: 1.26 ± 0.04) as compared to random network. Figure 2 shows the measurement of network measures that are significantly different between the two cohorts. The characteristic path length (p = 0.003) and global efficiency (p = 0.005) of the brain network of patients with SVD are respectively higher and lower than patients without SVD (Fig. 2a).
Association between total SVD score versus brain networks. The   brain network of patients with SVD are associated with the total SVD score (Fig. 3a). The local clustering coefficient (r = −0.429, p = 0.001) and nodal efficiency of hippocampus (r = −0.316, p = 0.015), as well as the nodal efficiency of paracentral lobule (r = −0.325, p = 0.012) are also associated with the total SVD score (Fig. 3b).

Discussion
With brain connectivity analysis, brain network can be systemically characterized at the global scale 20 . Normal brain has a high capacity for global information flow, also known as functional integration, for supporting cognition and behavior thanks to the underlying large-scale structural brain network that is characterized by high global efficiency and short characteristic path length 26 . We have demonstrated in this study that such optimal structural network architecture is disrupted in patients with SVD ( Fig. 2). That the key consequences of confluent white matter lesions, such as periventricular and deep WMHI, are demyelination and axonal loss of the long projection fibers 27 suggests that SVD likely has significant bearing on the long-distance structural connections between clusters of brain regions 28 . Global network changes similar to our study were also demonstrated by Lawrence et al. 29 whom showed that the global structural brain network of patients with SVD was impaired, and the extent of which was related to cognitive impairment. In another study, patients with SVD who developed incident all-cause dementia in 5-year follow-up were found to exhibit lower global efficiency than SVD patients without dementia, and that the risk of incident dementia was independently and negatively associated with global efficiency 30 . The presence of multiple CMBs was also reported to alter the characteristic path length and global efficiency of patients with Alzheimer's disease 25 . Other neurological diseases or disorders that reduce the global efficiency of the structural brain network include Alzheimer's disease 31 , multiple sclerosis 32 and schizophrenia 33 .
Reduction of global efficiency was found to be associated with poorer behavior 32 and cognitive 31 performance. Brain connectivity analysis also allows the characterization of the components of brain network at the local scale. The relation between a single brain region and its neighborhood can thus be investigated 34 . We have observed that various nodal characteristics, which reveal the communication capacity of a brain region with its neighborhood 35 , of hippocampus, precuneus, insular, supplementary motor area, paracentral lobule and putamen were reduced due to SVD (Figs 1 and 2). These brain regions are largely responsible for spatial and long-term memory 36 , higher-order cognitive functions 37 , and motor functions 38 . Structural disconnections in the local network of these brain regions may underscore the cognitive and motor impairments that are typically observed in patients with SVD 2 . The role of some of these brain regions has been implicated in previous studies 39,40 . For example, the volume of hippocampus is found to be related to the development of incident dementia 41 , and the cortical thickness of supplementary motor cortex to cadence 39 . In a different study of age-related cognitive decline, the nodal efficiency of insular is related to visuospatial function, and that of precuneus to executive function 42 . The distribution of nodal degree, the number of connections that a node has, is 'heavy-tailed' for human brains 43 . In other words, there are only a small number of brain regions, known as hubs, that have large number of structural connections to other regions, and that they are essential to efficient global brain communication 43 . Disruption of the characteristics of which is previously shown to be associated with brain dysfunction 44 . That we observed no change in brain hubs due to SVD likely suggests that the overall impact of SVD on hub organization is limited.
Since different neuroimaging biomarkers of SVD often coexist, the total SVD score should provide a better assessment of the overall severity of SVD than an individual MRI biomarker 11 . Our findings ( Fig. 3) showed that the overall SVD severity has insidious effect on most of the measures of the global network, and the local clustering coefficient and nodal efficiency of hippocampus, as well as the nodal efficiency of paracentral lobule. The total SVD score can explain up to 15% of the variance of the global network measure, and up to 18% and 11% of the variance of the local network measure of hippocampus and paracentral lobule, respectively. Our findings on the global network are similar to a previous study of cerebral amyloid angiopathy which showed an association between the overall SVD severity, measured using a different score, and global efficiency, with the former explaining only 7% of the variance of the latter 45 . In two other studies, relations between global network and individual SVD lesions, such as WMHI, lacunar infarcts and CMB, were demonstrated 29,46 . The global efficiency was also found to mediate the association between SVD lesions and cognition 46 . In the main, our results (Figs 2 and 3) suggested that the capacity of a structural brain network for functional integration may be most vulnerable to change in SVD burden. So may the case for the communication of the entire brain network with hippocampus and paracentral lobule as well as the connections between the neighboring regions of hippocampus be. The role of SVD lesions on the neurodegeneration within hippocampus, a structure that is vulnerable to Alzheimer's Disease, has been implicated 47 . Freeze et al. have demonstrated an interaction effect between cerebral amyloid-beta protein deposition, an Alzheimer's Disease pathology, and SVD pathology on hippocampal neurodegeneration only when there is abnormal amyloid-beta deposition, suggesting that patients with the two pathologies are at higher risk of faster disease progression. Together with our results, it may be plausible that the disruption of the network connected to and the subnetworks around hippocampus due to SVD lesions are exacerbated by that due to amyloid-beta deposition.
A major limitation of our study is the lack of cognitive and behavioral assessments that could allow us to study the relation amongst the overall SVD severity, structural brain network and brain functions. Such relation was previously demonstrated 29,46 but was limited to individual MRI biomarker of SVD, not total SVD score, for patients with large overall SVD burden. In addition, although deterministic tracking algorithm was the conventional method for estimating the edges of structural brain network 29,32,48 , it has several limitations, chief of which is its failure to resolve fiber crossing. The recently developed probabilistic tracking methods are more robust to fiber crossing, but may also yield spurious connections 49 . More advanced methods such as high angular resolution diffusion imaging and diffusion spectral imaging could be used to potentially resolve these issues in future studies. We also plan to investigate the relation between total SVD score and the change in the coupling between structural and functional brain networks in future studies in a hope to elucidate the underpinnings of the brain functional changes due to SVD.

Conclusion
We have successfully demonstrated that SVD lesions disrupt both global and local structural brain networks for patient with TIA or first-time stroke. We also showed that the overall SVD severity has insidious effect on most of

Material and Methods
Participants. We retrospectively reviewed the clinical and MRI data of consecutive patients admitted due to acute onset of neurological symptoms, and informed consent was obtained from all subjects. Patients with intracerebral hemorrhage, prior history of ischemic stroke, cerebral tumor, and other non-cerebrovascular neuropsychological conditions were excluded. A total of 95 patients with confirmed diagnosis of first TIA or acute ischemic stroke were included. Of these, 59 were found to have SVD lesions (see Total Small Vessel Disease Score below for how each type of SVD lesions was defined), and a corresponding control group of 36 patients without any SVD related lesions were included by matching age, gender and confirmed diagnosis of either TIA or firsttime acute stroke. The classification of acute ischemic stroke was recorded based on the Trial of Org 10172 in Acute Stroke Treatment (TOAST) 50

Total small vessel disease score and infarct quantification. Both CMBs and lacunar infarcts were
assessed based on the international consensus 4 . Focal lesions that were rounded, well-defined and hypointense with diameter less than 10 mm on SWI were considered as CMBs. CMBs mimics, such as blood vessel, mineralization in the globi pallidi or dentate nuclei, hemorrhage within infarct or small hemorrhage close to large infarct, air-bone interface and partial volume artifact, were excluded. Lesions that were round or ovoid with diameter between 3 and 20 mm, and signal intensity similar to CSF on FLAIR images were considered as lacunar infarcts 51 . WMHI were assessed based on the Fazekas scale 52 . Periventricular WMHI with Fazekas score of 3 (irregular periventricular signal extending into the deep white matter) and/or WMHI in deep brain with Fazekas score of 2 (confluent areas) were considered as high-grade WMHI 4 . PVS were defined as punctate or linear hyperintense lesions smaller than 3 mm in diameter on T2-weighted images, and were rated on a semi-quantitative scale from 0 to 4 10 . Moderate to severe (scale 2-4) PVS in the basal ganglia was defined as high-grade PVS 4 . The total SVD score was subsequently obtained by giving one point each to the presence of lacunar infarct, CMBs, high-grade WMHI, or high-grade PVS 12 . Patient with total SVD score higher than 0 was assigned to the SVD cohort.
Multi-slice binary mask of the ischemic infarct for each patient was manually defined on the mean of all DWIs with b-values = 1000 s/mm 2 . Only pixels with intensity distinctly higher than those in the contralesional hemisphere were included in the infarct mask.
Image processing and brain network construction. The image pre-processing steps for brain connectivity analysis were performed using the FMRIB's Software Library (FSL) version 5.0 53 .
Brain parcellation. T1-weighted images were used for parcellating the brain into 90 cortical and subcortical regions based on the Automated anatomical labeling (AAL) atlas with the following steps: (1) T1-weighted images were first coregistered to DWIs in the native space using FLIRT of FSL 54 . (2) The coregistered T1-weighted images were subsequently mapped to the ICBM 152 template using FNIRT of FSL 54 . (3) Estimated transformation parameters were inverted and applied to the AAL atlas to warp all brain regions-of-interests from the MNI space to the native diffusion space.
White matter tractography. Corrections for eddy current and head motions were performed by coregistering all DWIs to b 0 using FMRIB's Diffusion Toolbox 53 . Diffusion Toolkit 55 was also used to estimate the diffusion tensor using the linear least-squares fitting method. To construct the structural connections between the 90 brain regions, deterministic whole-brain fiber tracking was performed using the Fiber Assignment by Continuous Tracking (FACT) algorithm of TrackVis (http://trackvis.org) with fractional anisotropy threshold of 0.2 and turning angle threshold of 45°. Brain network. In a brain network, each brain region is considered as node whilst the connections between regions as edge. Structural connections between two regions are considered as edges only when WM fiber tracts originate from one region and terminate in the other. The number of fibers between regions was then used to weight each edge using the UCLA Multimodal Connectivity Package in the native diffusion space 56 . An inter-regional undirected network with weighted connections was subsequently obtained for each subject.
Brain connectivity analysis. Each weighted connectivity matrix was normalized to its maximum fiber count to minimize the overall differences in the connectivity strength between patients 20 . The characteristics of global network, such as the properties of small-world network and network efficiencies, as well as those of the local network for each brain region, also known as nodal characteristics, for each subject were estimated using the Brain Connectivity Toolbox 20 . For a detailed review on the measures of brain network, please refer to ref. 20 .
Small-world properties. Measures of small-world network were first introduced by Watts and Strogatz 57 . Clustering coefficient of a node, also known as local clustering coefficient, measures the likelihood that its neighborhoods are connected to each other. The clustering coefficient of a network reflects the extent of local cluster in that network. Thus, network with high clustering coefficient has high level of functional segregation. Characteristic shortest path length measures the shortest geodesic length between a node and any other node. The characteristic shortest path length of a network is the average of the shortest path length of all pairs of nodes in that network. Network with low characteristic shortest path length promotes parallel information propagation for functional integration. To determine whether a network is a small-world network, the clustering coefficient and characteristic shortest path length of the network are compared to 100 random networks that are generated based on the same number of nodes, edges and degree distribution as that of the network-of-interest 20 . A small-world network has normalized clustering coefficient larger than one and the normalized characteristic shortest path length close to one 57 .
Network efficiencies. Global efficiency is computed as the average of the inverse of the shortest path length of all node pairs in a network 58 . It estimates the ability of a network for integrating and coordinating information. Local efficiency of a network can be estimated by averaging the global efficiency of all subnetworks. It measures how well each subnetwork exchanges information when the most-connected node in the subnetwork is removed 59 .
Nodal characteristics. The degree of a node is the number of edges connecting it to rest of the network 20 . Nodal efficiency of a node is defined as the inverse of the harmonic mean of the characteristic shortest path length between other nodes in a network and itself 34 . It reflects the importance of the node for communication within a network. The betweenness centrality of a node quantifies the number of shortest paths between other nodes in an entire network and itself. It is a measure of the importance of the node for the integration of all the connections in a network 35 .
Identification of hubs. Brain regions that play key roles in the coordination of information flow in a network are considered as hubs. Brain hubs are characterized by their large number of connections or high degree of connectivity to other regions 43 . They occupy a central placement in the network and contribute significantly to functional integration. To identify hubs, brain regions are ranked by hemisphere-averaged betweenness centrality and degree. The two ranks are then summed to obtain a hub score. Brain regions with the top 20% score are considered as network hubs 60 .
Statistical analysis. Sex, history of diabetes mellitus, hypertension, ischemic heart disease and smoking, presence of infarct in 7 brain areas, namely basal ganglia, brain stem, cerebellum, frontal lobe, parietal lobe, temporal lobe and occipital lobe, and treatment of antiplatelets and lipid-lowering drugs were recorded as binary variables. Stroke classifications were recorded as categorical variable. Age, infarct size and serum total cholesterol were recorded as continuous variables. Demographics were compared between patients with and without SVD using independent samples t-test for continuous variables and Fisher's exact test for proportions. The nodal characteristics for each brain region from the two hemispheres were averaged before statistical analyses. Independent-samples t-test were performed to evaluate the group difference in all global and local network measures between patients with and without SVD. The association between all network measures and total SVD scores of patients with SVD were performed using Spearman rank correlation. All analyses were adjusted for age, sex, stroke classification, infarct size and location, and vascular risk factors (hypertension, diabetes, serum total cholesterol). Bonferroni correction for the problem of multiple comparisons was performed for all statistical analyses described above. SPSS 22.0 (SPSS, Chicago, IL) was used for all the statistical analyses. A significance level of p < 0.05 was set for all statistical tests.