Patterns of white matter degeneration in remote brain areas from the basal ganglion lesion of ischemic stroke patients with motor impairment

Background: Diffusion tensor imaging (DTI) studies have revealed distinct white matter characteristics of the brain following diseases. Beyond the lesion-symptom mapping, recent studies have demonstrated extensive structural and functional alterations of remote areas to local lesions caused by stroke in the brain. Here, we further investigated the structural changes from a global level using DTI data through multivariate pattern analysis (MVPA) and network-based statistic (NBS). Methods: Ten ischemic stroke patients with basal ganglia lesions and motor dysfunctions and eleven demographically matched adults as controls underwent brain Magnetic Resonance Imaging scans. DTI data were processed to obtain fractional anisotropy (FA) maps and MVPA was used to explore brain regions that play an important role in classication based on FA maps. The white matter (WM) structural network was constructed by the deterministic ber tracking approach according to the Automated Anatomical Labeling (AAL) atlas. NBS was used to explore differences in structural networks between groups. Results: MVPA applied to FA images correctly identied stroke patients with a statistically signicant accuracy of 100% (P ≤ 0.001). Compared with the controls, the study patients showed FA reductions in the perilesional basal ganglia and brainstem, with a few showing reductions in bilateral frontal lobes. Using NBS, we found a signicant decrease in FA-weighted WM subnetwork in stroke patients. Conclusions: We identied some patterns of WM degeneration affecting brain areas remote to the ischemic lesion, revealing the abnormal organization of WM network in stroke patients, which may be helpful for the understanding of the neural mechanisms of stroke sequela.

Multivariate pattern analysis (MVPA) is a method of imaging data analysis based on machine learning and pattern recognition. It involves the use of pattern classi cation algorithms to extract spatial patterns from neuroimaging data for individual characteristics analysis [9]. MVPA has the advantage of considering interregional correlations and searching for abnormalities throughout the entire brain because it adopts an unbiased and whole-brain method without the region of interest (ROI) being arti cially set [10]. This method has been widely used in psychiatry related studies to analyze mild changes that can be observed in subclinical populations [11,12].
White matter structural connectivity can be modeled as a network. A network-based statistics (NBS) method is to control the family-wise error rate during mass univariate testing of every connection of the network. NBS is also performed to investigate the interregional correlations of the brain on a global level [13,14]. We utilized these methods to explore subtle relevant changes and abnormalities of structural networks of the brain for a further understanding of the intrinsic brain structural basis of ischemic stroke patients with residual motor dysfunctions.

Methods
Participants Ten right-handed stroke patients (mean age 56.7±10.5 years) from the Southeast University a liated Zhongda Hospital were recruited for this study from March 2019 to December 2019. The inclusion criteria for patients were as follows:(1) age ≥20 and age ≤80 years (2) rst-onset of ischemic stroke with the involvement of the basal ganglia; (3) pure motor de cits; (4) stable conditions after treatment of acute stroke without recurrence. The exclusion factors: (1) a history of neurological or psychiatric disorders, prior or subsequent symptomatic stroke; (2) brain abnormalities unrelated to the infarct lesions; (3) MRI contraindications. All of the affected extremities were evaluated for motor function. The motor outcome of the affected limbs was evaluated by the Fugl-Meyer assessment (FMA), including the upper and lower extremities [15]. Also, The Brunnstrom stage (BRS) was adopted. Recovery of the affected extremities was scored on a 6-point scale (1=severe; 6=normal) [16]. The clinical characteristics of the stroke patients are summarized in Table 2. Eleven demographically matched healthy controls (mean age 61.5 ± 7.8 years) were also recruited.

Lesion mapping
Lesion-side normalization was performed to place lesions on the left side of the brain. For the patients with lesions in the right hemisphere, the images were ipped from the right to the left along the midsagittal line to simplify the comparison with the other patients with stroke. According to previous investigators, lesion masks of each patient were manually segmented on individual structural MRI images (T1-weighted MP-RAGE and FLAIR images) using MRIcron software (http://www.mricro.com). After the spatial normalization of all individual lesion masks, a lesion overlap image for all patients was constructed ( Fig. 1) [17,18,19].
DTI Data Pre-Processing and Network De nition DTI data analysis was performed by a pipeline toolbox for analyzing brain diffusion images (PANDA, http://www.nitrc.org/projects/panda) [20].The main procedure includes: (1) correcting for the head motion and eddy current effects using the FMRIB's Diffusion Toolbox (FDT); (2) calculating the diffusion tensor (DT) metric-----FA for each voxel using the DTIFIT tool; (3) normalizing: registrations of all the individual FA images to the FMRIB58_FA template by calling the FNIRT tool [21]. The FA maps were obtained and the FA values of bilateral CST were computed. Whole-brain tractography was performed for each subject in the native diffusion space based on ber assignment by continuous tracking (FACT) algorithm [22]. All voxels with fractional anisotropy (FA) ≥0.2 were used as seed points; the FA and curvature thresholds of path tracing were set to 0.2 and 45°, respectively.
FA weighted networks were constructed from 116 nodes de ned according to the Automated Anatomical Labeling (AAL) atlas [23]. The entire cerebral cortex was automatically partitioned into 116 anatomical regions of interest (ROIs) (45 ROIs for each cerebral hemisphere, 26 ROIs for the cerebellum) using the AAL algorithm. The weight of the edges in the network was de ned as the mean FA value of the connected bers between each pair nodes. In order to reduce false-positive connections, two nodes were considered structurally connected only when at least 3 bers were reconstructed between them [24].

Multivariate pattern analysis
Pattern classi cation analysis could be used to examine the differences in the FA values between groups. A speci c MVPA approach known as support vector machine (SVM) was implemented using PRoNTo (Pattern Recognition for Neuroimaging toolbox) software (http://www.mlnl.cs. ucl.ac.uk/pronto/) [25]. Individual FA maps were treated as points located in a high dimensional space. A linear decision boundary in this high dimensional space was de ned by a hyperplane that separated the individual brain images according to a class label (i.e., patients vs. controls)( Fig. 2A). A more detailed description of the SVM can be found in the previous reports [11]. The receiver operating characteristic curve (ROC), sensitivity and speci city of the FA classi cations and the weight of each brain region in the classi cation analysis were obtained. The brain regions with voxels that have values ≥30% of the maximum weight vector value of the discrimination map were taken as the key areas.

Statistical Analysis
The two-sample t tests were employed to test the group differences in demographic data and imaging measures. Testing differences in FA values of the CST between the affected and unaffected sides were accomplished with a paired t test. Correlations among the FA in key areas derived from MVPA and the CST were computed using Pearson's correlation coe cients. These statistical tests were performed using the Statistical Package for Social Science (SPSS) 22.0 software. The two-sample t test was adopted for the inter-nodal connection comparisons followed by the NBS method using GRETNA (v2.0.0) to analyze the FA networks between groups. Visualization of the results was performed by the BrainNet Viewer [26,27]. The comparisons of the two sets of FA values in the key areas derived from MVPA were performed in SPM12 with false discovery rate (FDR)-corrected (P FDR < 0.05, cluster size > 5) [28].

Demographic data and Patient characteristics
The demographic data of the patients and the controls are presented in Table 1. No signi cant differences in age or gender were observed between the two groups. The mean interval from stroke onset to DTI scans was 10.3 ± 9.0 weeks ( Table 2). The lesions were mainly located in the basal ganglia region. For the 3 patients with lesions in the right hemisphere, the imaging data were laterally ipped. Stroke lesions for each patient were overlaid onto a T1 template in MNI standard space (Fig. 1).
The FA values of the ipsilesional CST in the patients were statistically signi cant reduced than that of the contralateral side (t=5.257, P=0.001). Compared with the controls, the FA values of the left (injured side) CST in the stroke group were signi cantly decreased (t=6.511, P<0.001). And no signi cant differences in the FA values of the right CST (t=0.153, P=0.88) were observed between the two groups.
Overall Classi er Performance Figure 2 A shows the result of the MVPA classi cation between the 10 stroke patients and the 11 controls based on FA values. The overall accuracy was 100% , and was signi cant at P≤0.001 (P= 0.001). Both the sensitivity and the speci city were 100% (with ROC shown in Fig. 2B). This overall classi cation accuracy of the algorithm measures its ability to correctly sort the two groups.

Discrimination Map
In the whole-brain voxel weight maps, the weight vector value indicates the relative importance of the voxel in the decision function ------discrimination between patients and controls (Fig. 3). Note that all voxels in the white matter mask contribute to the decision function since the analysis is multivariate. The spatial distribution of the weight vector provided information about the contribution of different areas to classi cation. The brain regions that contributed the most to the discrimination between stroke patients and controls were identi ed by setting the threshold to ≥30% of the maximum weight vector scores, consistent with previous studies using MVPA for disease classi cation [11,29]. The most identi ed classi ed features of the FA maps included the perilesional basal ganglia and brainstem, with a few in the bilateral frontal lobes ( Fig. 3 and 4 A), which were taken as the key areas.
By comparing the FA values of the two groups in the discrimination map by SPM12 (P FDR < 0.05, cluster size > 5), we found that the stroke group showed reduced FA in the key areas, shown in Figure 4A and reported in Table 3. According to the white matter atlas, the clusters were mainly located in the CST pathway. Moreover, the FA values of these areas have a positive correlation with that of the ipsilesional CST (r=0.888, p<0.001) (Fig. 4 B).
Decreased Connection of the Component Network in stroke patients FA weighted networks were constructed from the nodes (brain areas) de ned according to the AAL atlas.
The weight of the edges in the network was de ned as the mean FA value of the connected bers between each pair nodes. The NBS approach was used on the structural networks constructed by deterministic tractography. We identi ed several signi cantly decreased connections of a component network (subnetwork) in stroke patients (P<0.01, P=0.008) including 26 regions and 32 connections (Table 4 and Fig. 5 A). Figure 5 showed the subnetwork that reveals deterioration in the brain structure of the stroke patients. Table 3 reported the regions as nodes in the subnetwork. The backbone matrix of the group structural matrixes for each group of the subnetwork was calculated, including the 26 regions in Table 4 ( Fig. 5 B and C). The subnetwork was composed of the cerebral cortical regions (the frontal lobe, the parietal lobe and the occipital lobe), the subcortical areas (the basal ganglion), and parts of the cerebellum. Compared with the results of MVPA, there were changes involving the frontal parietal lobe and the basal ganglia, and those in the brainstem region may overlap with some of the bers connected to the cerebellum.

Discussion
Beyond the well-known concept of lesion-symptom mapping, some lesions in a single location in the brain could disrupt brain functions routed to widespread neural networks [1,30]. Our analysis con rmed that local destruction of certain anatomic regions could affect remote areas in the brain.
White matter degeneration in the CST pathway of stroke patients The degree of anisotropy depends on the level of organization, the integrity of the white matter tract, and the degree of freedom of water diffusion movement caused by the oriented axonal membranes and myelin sheaths [31]. Reduced anisotropy along the CST far from the original lesions has been interpreted as Wallerian degeneration (WD) [32]. DTI can quantify the FA values to evaluate the pathology change of the white matter, such as WD. Using the good classi cation ability of MVPA, we reported that apart from the basal ganglia region which has direct infarcts, brain areas with signi cantly decreased FA values were also located in the brainstem of the lesioned hemisphere, and a few in the bilateral frontal lobes, which might be indicative of degenerative lesions caused by WD.
The acute and chronic phases of stroke probably differ in the white matter changes since neural changes can contain anterograde and retrograde degeneration, or refactoring. However, no matter what kind of alteration is present, it should generate speci c structural changes and affect corresponding functions. Therefore we performed the pattern recognition classi cation using the whole brain FA map and explored the key brain regions important for distinguishing the stroke patients from the controls.
Our study veri ed degenerative changes in the white matter of stroke patients. The infarcted lesions were mainly located in the basal ganglia region, but the FA reduction of some remote areas has reached the point where they can be differentiated from the controls. In conclusion, the damaged structual anatomy of the subcortical areas may induce the deterioration of the key white matter areas in the brain.
Decreased WM connections were widely distributed across brain regions of stroke patients Many patients were left with motor dysfunctions after the occurrence of cerebral infarction. Usually, it is related to the injury of the corticospinal tract. The CST originates from multiple motor and somatosensory cortices including premotor cortex, supplementary motor cortex (SMA), primary motor cortex, as well as primary and secondary somatosensory cortices. The CST is crucial for proper execution of a volitional movement [33,34]. Apart from the CST and the motor areas of the cortex, the proper execution of movements involving balance and coordination also requires the extrapyramidal tract and other brain regions such as the nucleus in basal ganglia and the cerebellum [35]. The NBS analysis showed that the structural subnetwork connection of the stroke group was weaker than that of the control group, which indicate that the nerve bers involved in the subnetwork were affected. When the integrity and order of the brain structure were destroyed, it can be re ected in the white matter. Not only the brain regions directly related to the motor commands (e.g. precentral and SMA) but also other regions that might participate in the regulation of motor control were involved in the subnetwork.
In addition, since the MVPA analysis is multivariate, all voxels in the white matter mask contribute to the decision function in the processing stage. We observed that the area where the left CST passes is reddish, and there is a light green distribution in the corpus callosum and some other areas (Fig. 3). Combined with the NBS results, the affected brain tissues were distributed across a wide range of brain regions (Fig.   5). In short, the local abnormalities spread among the whole network as time elapses after the acute stage of the stroke onset. There is implicational information contained in the patterns of white matter degeneration.
Next, we brie y discuss the brain regions involved in the subnetwork ( Table 4).

The bilateral frontal lobe
Similar to previous studies results showing infarct-related focal thinning of the motor area in remote cortex via degeneration of inter-hemispheric connection ber of the corpus callosum [36,37], we found changes in the connection between the frontal hemispheres, as well as reduced FA values in a small area located in the contralesional frontal cortex. It has been reported that secondary degeneration occurred in the ipsilesional precentral gyrus after subcortical stroke involving the CST at the 6-month follow-up in stroke patients by calculating the mean kurtosis (MK) value of manually drawing ROI from diffusion kurtosis imaging (DKI) imaging study [38]. We speculated that the degeneration in remote parts of the CST occurred in subacute phases of stroke.
In addition to motor-related brain areas, the affected brain areas in the frontal lobe may generate mild or long-term cognitive changes in patients. Recently, a research reported that the Reading the Mind in the Eyes Test (RMET) was associated with damages to the white-matter tracts connecting the frontal and the temporo-parietal components of the RMET functional network [39]. Cognitive impairment still requires sensitive detection of subtle changes by complex experimental design or long-term observations [40].

The basal ganglia region
The patients were infarcted mainly in the basal ganglia region. The CST section passing through the posterior limb of the internal capsule may receive regulatory information from the nucleus of basal ganglia or constitute loops within them. Due to the cortico-basal ganglia-thalamocortical 'motor' loop, any impact on the circuit constituent can lead to a shift in the balance between neural interactions in the direct and indirect pathways and subsequently lead to variations in the brain functions [41,42]. In the early stage of rehabilitation, stroke patients with hemiplegia often have synergistic movements, which are thought to be related to this loop.

The occipital lobe
In our study, changes in white matter connections involving the occipital (cuneus and precuneus) region were most likely related to visual effects. Voluntary actions modulate perception that follows the anatomical-functional bias of the motor system. Therefore stroke patients with dysfunction of normal voluntary movements may develop corresponding abnormal sensory modulations that gradually affect brain structure [43]. A study indicated a reduction in functional connectivity (FC) between the motor and executive control and the visuospatial networks in patients with motor de cits vs. healthy controls [44]. This demonstrated the existence of FC between the visual cortex, the ipsilateral/ contralateral motor cortex and the cerebellum. We conformed that this decline in FC is probably accompanied by a decrease in white matter structural connectivity.

The cerebellum
The cerebellum is often related to balance adjustment and patients with hemiplegia after stroke usually have problems of stability and coordination. The connections between the cerebrum and the cerebellum pass through the cerebral peduncle in the brainstem. So we assume that the weakening of those connections is partially consistent with the decrease in FA in the brainstem. Previous DTI studies also found decreased FA in midbrain of the stroke patients by manually plotting ROI [38]. A recent study indicated that the cerebellum plays a role on residual motor output by facilitating cortical excitability in chronic stroke [45]. Changes in cortico-cerebellar structural connectivity are probably caused by brain injuries related to the neural bers connecting the cerebellum. The inability to perform normal movements might gradually lead to abnormal balance that is re ected in the decrease of cortico-cerebellar connectivity.

Limitations and Expectations
Despite our signi cant ndings, this research is not free of limitations. First, the sample size is small. Second, the interval of DTI acquirement from stroke onset ranging from 2-24 weeks. Recovery of FA in penumbra regions occurs most rapidly during the rst two weeks following stroke, with continued slow increases in FA for many weeks thereafter [46][47]. The phases of stroke rehabilitation may in uence white matter organization considering that we only discussed stroke patients with motor impairments in a cross-sectional insight. Future work can be further explored by expanding the sample size and dynamically observing imaging changes from the acute phase to the recovery phase. Additionally, emphasis should be placed on the classi cation and re nement of clinical behaviors of stroke patients, with speci c brain feature changes that correspond to the functional outcome.
Brain hemisphere dominance varies with different activities. Differences in the left and right hemisphere function often exist in emotion [48], language [49] and attention [50]. More studies are needed to fully understand all of the factors that affect this. Brain hemisphere dominance was little known in the motor movement. A study elucidated the symmetrical characteristics of the CST in bilateral hemispheres using probabilistic ber tracking method. They clearly showed that the hand CST had symmetric structures in bilateral hemispheres [51]. When the researches emphasized different functions of the brain, especially the language, the two hemispheres are asymmetrical. In the future study, we should classify various functions of the brain and analyze the left or the right hemisphere separately.
A study on 132 stroke patients using rest-fMRI has revealed that although structural damage from stroke is focal, remote dysfunction can occur in regions of the brain distant from the lesions. These results link key organizational features of the brain networks to behavior in stroke [52]. To date, there is no largescale post-stroke structural network analysis similar in quantity and method to the above functional analysis. In this study, only stroke patients with motor dysfunction were recruited for analysis using DTI data. By MVPA and NBS, we detected the reduced FA values and abnormal white matter connections in the brain of those patients from a global level.
We need to further investigate the structure changes of the brain or the functional network in the natural process of different types of brain injuries. Then the diagnosis and prognosis can be more accurately assessed paving the way for better treatment selection and research methods.

Conclusions
In summary, we found that there are multiple WM structural abnormalities in the affected brain areas of stroke patients with motor impairment. Our study may provide a basis for further exploration of the neural mechanism involved in residual motor de cits in stroke patients.
Looking forward to the future, we will further compare the neural changes of well-recovered patients to provide a basis for the development of adaptive rehabilitation training strategies.

Declarations
Availability of data and materials The data in the current study are available on reasonable request to the corresponding author.

Ethics approval and consent to participate
The study was approved by the local Ethics Committee of the Southeast University a liated Zhongda Hospital. All participants gave written informed consent to participate according to the Declaration of Helsinki.

Consent for publication
Not applicable.

Competing interests
The authors declare that they have no competing interests Funding This work was funded by the National Natural Science Foundation of China (grant number:6590000127, 81801680), Nanjing health science and technology development special fund project (grant number:     The result of the MVPA classi cation. Classi cation plot (A) and Receiver Operating Characteristic (ROC) curve (B) for the comparison between stroke patients and controls using FA maps derived from DTI data, which yielded an accuracy of 100% (100% sensitivity, 100% speci city), statistically signi cant at P≤0.001.

Figure 3
Whole-brain voxel weight map. It shows the white matter regions contributing to discrimination between groups based on FA values. The colour bar indicates the weight vector value of the voxel, which is also indicated in the intensity eld of the Anatomical image (white ber atlas 'JHU-ICBM-FA-2mm') panel.