Elucidating the structural–functional connectome of language in glioma‐induced aphasia using nTMS and DTI

Abstract Glioma‐induced aphasia (GIA) is frequently observed in patients with newly diagnosed gliomas. Previous studies showed an impact of gliomas not only on local brain regions but also on the functionality and structure of brain networks. The current study used navigated transcranial magnetic stimulation (nTMS) to localize language‐related regions and to explore language function at the network level in combination with connectome analysis. Thirty glioma patients without aphasia (NA) and 30 patients with GIA were prospectively enrolled. Tumors were located in the vicinity of arcuate fasciculus‐related cortical and subcortical regions. The visualized ratio (VR) of each tract was calculated based on their respective fractional anisotropy (FA) and maximal FA. Using a thresholding method of each tract at 25% VR and 50% VR, DTI‐based tractography was performed to construct structural brain networks for graph‐based connectome analysis, containing functional data acquired by nTMS. The average degree of left hemispheric networks (M left) was higher in the NA group than in the GIA group for both VR thresholds. Differences of global and local efficiency between 25% and 50% VR thresholds were significantly lower in the NA group than in the GIA group. Aphasia levels correlated with connectome properties in M left and networks based on positive nTMS mapping regions (M pos). A more substantial relation to language performance was found in M pos and M left compared to the network of negative mapping regions (M neg). Gliomas causing deterioration of language are related to various cerebral networks. In NA patients, mainly M neg was impacted, while M pos was impacted in GIA patients.


| INTRODUCTION
Gliomas are the most common primary intracranial tumor, representing 81% of malignant primary brain tumors (Ostrom et al., 2013(Ostrom et al., , 2014. Brain tumors within language eloquent regions can cause aphasia depending on tumor size, edema, and localization per se. However, the localization of language eloquent fiber tracts shows significant interindividual variability, especially in patients suffering from lesions in eloquent brain areas. Understanding the mechanisms of aphasia in glioma patients will improve the understanding of the neural function and structural plasticity, which is essential for preserving brain functions in individualized tumor treatment concepts.
Preoperative identification of language-related regions is essential to preserve functionality during tumor resection in eloquent locations.
Navigated transcranial magnetic stimulation (nTMS) and functional magnetic resonance imaging (fMRI) have both been widely applied preoperatively to identify the individual language-related regions.
However, fMRI lacks precision, especially in the vicinity of brain lesions based on inaccuracies due to pathological tumor vascularization (Lin et al., 2017;Silva, See, Essayed, Golby, & Tie, 2018). nTMS has shown to be highly predictive regarding negative stimulation sites, and comparisons to the gold standard of intraoperative direct cortical stimulation (DCS) showed a high accuracy of nTMS (Ille et al., 2015a(Ille et al., , 2015bPicht et al., 2013). Therefore, nTMS mapping of motor-and language-eloquent brain areas is routinely used preoperatively to identify language-related fiber tracts and for preoperative risk assessment combined with fiber tracking (FT).
Connectomes are used to represent the functional composition of nodes-relevant brain areas as mapped by nTMS-and their connections as visualized by DTI within the network. The combination of graph theory, nTMS, and connectome analysis of DTI data is a novel multidisciplinary paradigm considering the brain as a complex network of individual components interacting through continuous communication. Thereby it offers further insight into both local and global effects of gliomas on these complex networks (Hart, Romero-Garcia, Price, & Suckling, 2019). Previous studies using nTMS-based FT only focused on single neural tracts in patients with glioma-induced aphasia (GIA; . However, recent studies have shown that gliomas have a global impact on the whole brain (Derks et al., 2017;Hart et al., 2019). Therefore, there is a necessity of analyzing global nTMS mapping-based networks to investigate neuroplasticity and remodulation mechanisms related to aphasia.
As shown by graph theory in previous studies, the organization of cerebral structures is compatible with the hypothesis that the brain evolved to maintain the dynamic balance between maximization of the efficiency in transferring information and minimization of connection cost (Betzel et al., 2014;Bullmore & Sporns, 2009;Gargouri et al., 2016). Graph-based network analysis enables to derive properties of the brain's "connectome" and delivers information on the topological architecture of human brain networks, such as average degree (AD), global efficiency (EG), and local efficiency (EL), which has already been introduced for functional MRI (Betzel et al., 2014) and EEG analysis (Gu et al., 2020). AD is the basic character to present the intensity of connections across all nodes in the network (Cohen & D'Esposito, 2016). EG measures the capacity in parallelly transferring and comprehensively processing information (Wang, Zuo, & He, 2010). EL indicates the fault-tolerant capacity of the network and the efficiency of the communication between immediate neighbors of the local node (Wang et al., 2010). Previous studies on tractography networks in glioblastoma patients demonstrated significant differences in tract volumes based on nTMS positive and negative mapping sites . However, graph theory metrics and their differences have not yet been measured to analyze the impact of tumors on cerebral structural network properties. Further investigations are still lacking to distinguish the network based on eloquent language regions identified by nTMS from the left and right hemispherical networks.
This study aimed to evaluate the global and local properties of function-specific connectomes derived from nTMS language mapping.
Graph properties of aphasic and nonaphasic glioma patients of the resulting connectomes based on positive and negative nTMS mapping regions were analyzed and correlated with different states of aphasia.
Connectomes were constructed and analyzed based on tractography thresholding at different levels to assess the robustness and efficacy of networks.

| Ethics
The current study was performed in accordance with the Declaration of Helsinki and its later amendments, and its protocol was approved and supervised by the local ethics board (registration number: 222/14, 338/16, 2793/10, 5811/13, 223/14, and 336/17). Written informed consent was obtained from all patients before enrolling in the current study.

| Study eligibility
The following inclusion criteria were considered: (a) age above 18 years, (b) mother tongue German, (c) primary diagnosis being glioma with following pathological confirmation, (d) tumor within a left perisylvian region adjacent to the arcuate fasciculus-related cortical and subcortical regions, (e) no previous cranial surgery, and (f) written informed consent. Patients with contraindications for MRI or nTMS examinations such as pregnancy, intracranial metallic implants, cochlear implants, and pacemakers were excluded. Overall, 30 patients with no aphasia (NA) and 30 patients with GIA were considered eligible from our database of patients undergoing nTMS language mapping in our department from 2016 to 2019.
Contrast-enhanced T1 weighted images were used for neuronavigation.
Stimulator output was set at 100% of the individual resting motor threshold. Stimulation was performed using 5 pulses at 5 Hz on individually predefined 46 targets according to the cortical parcellation system (Krieg et al., 2017;Sollmann et al., 2018). nTMS language mapping consisted of a baseline session without stimulation and a stimulating session with stimulation during the patient performing an object naming task (ONT), during which audios and videos were recorded for post-hoc analysis. Each naming performance with stimulation was compared with the individual baseline to identify naming errors, which were categorized into five types: no-response, performance, phonological paraphasias, semantic paraphasias, and neologism (Krieg et al., 2017;Sollmann et al., 2018).

| Network construction
Contrast-enhanced T1 images were skull-striped using HD-Bet (Isensee et al., 2019). In the next step, they were linearly co-registered to b0 images derived from the DTI data set, and the POS and NEG stimulation sites derived from nTMS language mapping were transferred into the DTI space. The anatomic atlas template AAL90 (Tzourio-Mazoyer et al., 2002) was co-registered to the T1 image using the SyN algorithm from ANTs (https://github.com/ ANTsX/ANTs; Avants, Epstein, Grossman, & Gee, 2008) and diffusion-weighted b1000 images were linearly registered to the b0 image. The gradient vector table was rotated accordingly and corrected for eddy currents (Figure 1). Finally, nTMS POS and NEG sites, as well as AAL90 atlas locations, were co-registered to the DTI space, and the total number of regions in the AAL90 anatomic template corresponding to POS or NEG points was counted for each subject.
Next, whole-brain tractography was conducted through the Python library DIPY (Version 1.2.0, https://dipy.org; Garyfallidis et al., 2014), during which each voxel in the brain was used as region of interest (ROI; Figure 1). Then, constrained spherical deconvolution (CSD) was performed on the dataset (Tournier, Calamante, & Connelly, 2007). In the next step, a deterministic algorithm was applied to tractography with fractional anisotropy thresholds (FAT) starting from 0.0 continuously increased by 0.01 with a fiber length threshold (FLT) at 30 mm and stopping at a maximal FAT (FAT max ).
During this process, the maximal FA (fa max ) for every single fiber was identified. The following formula calculated the visualized ratio (VR) for each fiber: The tractography based on the fibers with VR above 25 and 50% were selected (Figure 1), from which the following five matrices Since brain size is variable among individuals, connections between two regions in matrices containing more than three fibers were considered as connected and binarized to 1, otherwise as they were considered as disconnected and binarized to 0 ( Figure S1).
The connectome properties, including AD, EG, and EL, were assessed in each of the five binarized matrices under VR thresholds of 25 and 50%, respectively, using algorithms from the NetworkX 2.5 library (https://networkx.org/) in Python 3.7 (https://www.python. org/; Latora & Marchiori, 2001). The differences of each property calculated for 25% VR and 50% VR thresholds were recorded for each group as AD-diff, EG-diff, and EL-diff.

| Connectome analysis and statistical analysis
The statistical analysis was performed using SPSS Statistic (IBM SPSS Statistics for Mac, Version 23.0. IBM Corporation, Armonk, NY) and GraphPad Prism (Version 8.4.3, San Diego, CA).
The chi-square test was applied to compare the demographic data between both groups, including handedness, pathological diagnosis, tumor locations, and gender. Furthermore, independent t-testing was applied to compare age and glioma size between the two groups.
For the analysis of mapping regions, the number of patients in each group with the same positively or negatively mapped regions was summarized, respectively. The intra-group proportion of being positively or negatively mapped was calculated for each region in the mapping template. The nonparametric test was used to analyze the difference of the mapping points between the NA and GIA groups.
ANCOVA testing with covariates was used to compare network properties between the NA and GIA groups, consisting of AD, EG, EL, 3 | RESULTS

| Demographic analysis
Sixty subjects were enrolled from the database of patients receiving treatments in our department between 2016 and 2019, consisting of F I G U R E 1 Workflow of the current study. This figure illustrates the process of network construction. DTI scans with 32 directions (A), T1 images with contrast (B), T1 images with contrast without skull and skin (C), the anatomic atlas template AAL90 (D), and nTMS language mapping images (E) were registered to a B0 image (F). A deterministic algorithm was used for fiber tracking after applying constrained spherical deconvolution (CSD). The minimal FA for each fiber to be visualized was identified, from which its visualization ratio (VR) was calculated. The fibers with VR values above the thresholds of 25% (G) and 50% VR (H) were respectively used to construct five matrices: M whole , matrix (M) derived from nodes from both hemispheres and edges (fibers) connecting them. M left and M right , respective matrices with nodes from the left (M left ) or the right hemisphere (M right ) and intra-hemispheric edges (fibers). M pos and M neg , matrix with nodes from the positive language mapping regions and edges from their corresponding edges (fibers), and matrix with nodes from the negative language mapping regions and edges from their corresponding edges (fibers) 30 patients in the NA group and 30 patients in the GIA group. Age was 57.7 ± 15.1 years for the NA group and 63.9 ± 12.4 years for the GIA group (p = .093). There were no significant differences between the two groups regarding handedness, gender, and World Health Organization (WHO) grading (Table 1). FAT max was calculated for the NA (0.531 ± 0.065) and GIA (0.541 ± 0.063) groups. No significant differences were found between the two groups (p = .774).
Notably, glioma size in the NA group (2.4 ± 2.6 cm 3 ) was significantly smaller than in the GIA group (4.7 ± 4.3 cm 3 ; Table 1; p = .015). However, glioma size did not correlate with aphasia levels in the GIA group (p = .060, R = .249). As the tumor size was different between the two groups, glioma size was regarded as a covariate for the analysis of covariance (ANCOVA) in the following comparisons between the two groups for investigating the performance of structural networks. In the GIA group, the tumor affected the precentral gyrus, frontal inferior gyrus, and Rolandic operculum more often (Table S1).

| Analysis of nTMS mapping regions
There were significantly less positive than negative stimulation sites in both the NA (Average counts of mapping regions: 16.600 (NEG) vs.
More POS regions were found in the GIA group compared to the NA group (9.567 vs. 8.400, p = .083; Figure 2). In the GIA group, the total count of the POS regions (R = .226, p = .230) and NEG regions (R = À.226, p = .230) did not correlate with aphasia levels.  (Table 3).
There was a higher difference of EG (EG-diff) and EL (EL-diff) between both VR settings for the GIA group compared to the NA group (Table 3) p = .008), which were higher in the GIA group (Table 3). However, comparisons of other matrices were without significance.

| Correlation analysis
The correlation analysis between aphasia levels and network properties included all cases from NA and GIA groups.   M pos (R = À.337, p = .008) were negatively correlated with aphasia levels, while EL-diffs from M left (R = .341, p = .008) and M pos (R = .365, p = .004) were positively correlated with aphasia levels (Table 4).

| DISCUSSION
The present study focused on a graph-based, joint structural-   (Hertrich, Dietrich, & Ackermann, 2016). The difference of overlapped ratios between two groups can be reasoned by the disconnection of SMA in the language network, which has been suggested in a study on functional connectivity properties of the language network in children with an autistic spectrum disorder, detecting a profound loss of functional connections between the SMA and modulatory control of the dorsolateral prefrontal region (Verly et al., 2014).
The angular gyrus is a critical area for language perception, as confirmed in previous studies (Hartung et al., 2020). In NA patients, there is a higher probability of the angular gyrus to be negatively mapped (30.0% in the GIA group and 56.7% in the NA group), supposing that there is a network of collaboration related to the angular gyrus (Papathanassiou et al., 2000). Hence, when nTMS targets the angular gyrus, the language-related network in NA patients can better compensate for its inhibition.

| Different network performance between NA and GIA group
The current study applied the deterministic algorithm to track fibers.
Previous studies have shown that deterministic and probabilistic FT have respective advantages and disadvantages. Their sensitivity and specificity were affected by many factors, such as edema and/or hemorrhage around the tumor, resulting in uncertainty in the number of fibers obtained (Schlaier et al., 2017;Zhan et al., 2015). Therefore, this study was not only based on the exact number and volumes of connections but also on a binarized matrix to present the existence of connectivity among brain regions, which can meanwhile minimize the confounding effects of individual brain volume and tumor size in the analysis of network properties.

| Higher ADs in the NA group
Thresholding at both VRs, ADs from the matrices in the NA group were higher than in the GIA group, except for M pos under 25% VR, which suggested that the NA group showed more connections of higher VR values (Table S2)  usually regarded as regions that can "safely" be resected without causing damages to language function and demonstrated a great consistency with intraoperative DCS (Ille et al., 2015a(Ille et al., , 2015bPicht et al., 2013). Previous studies using nTMS preferred to focus on positively mapped brain regions and their related subcortical structures due to language errors directly being induced by nTMS targeting these regions, which lead to an insufficient understanding of network structures and functions of NEG regions. Further studies are necessary to analyze those negatively mapped areas. However, the malignant growth of glioma also leads to difficulties in Note: This table shows the average degree (AD), average global efficiency (EG), local efficiency (EL), and differences between a visualization rate of 25% (25% VR) and 50% (50% VR) for AD (AD-diff), EG (EG-diff), and EL (EL-diff) of each matrix for patients with no aphasia (NA) and glioma-induced aphasia (GIA). Most matrices in the no aphasia group (NA) show a higher AD than the glioma-induced aphasia group (GIA), except for the matrices of nTMS positive regions (M pos ) thresholding at a VR of 25%. The average EG and EL in the NA group were higher than that in the GIA group under 25% and 50% VR, except for EG and EL of M pos under 25%VR, which were higher in GIA. The EG-diff and EL-diff were higher in GIA patients than in the NA patients for all matrices. Furthermore, results of the ANCOVA analysis on the average degree and efficiency of each matrix in both groups using tumor size as a covariate are shown. Under 25%VR, there were significant differences of EG in whole-brain matrices (M whole ), left hemispheric matrices (M left ), and matrices of negative nTMS regions (M neg ) between NA and GIA group, while EL in the M left was significantly different between both groups. Under 50% VR, EG, and EL of M pos , and EL of M left and EG of right hemispheric matrices (M right) showed a significant difference between both groups. Regarding the changing levels (DIFF), EG-diff, and EL-diff of M pos were significantly higher in the GIA groups. *p < .05; ** p < .01. observing the long-term effect after surgery in those negatively mapped regions.

T A B L E 4 Correlation analysis of connectome properties and aphasia levels in all cases
The intergroup analysis showed no difference in the ADs from the POS-based networks between NA and GIA groups. Notably, the count of POS regions was significantly smaller than the count of NEG regions, leading to a bigger size of M neg compared to M pos . The ADs from M pos were higher than the ADs of M neg in the GIA group, which was not found in the NA group (Table 3). It demonstrates that the higher density in intra-network connections within M pos compared to M neg in aphasia patients (Figure 4). It supports previous studies and indicates the enhanced connectivity profile within the POS-related network (Sollmann, Zhang, Kelm, et al., 2020). Thus, POS regions might be the hubs of coordinating and processing information from different brain regions.
Previous studies have shown that gliomas led to a reduction in FA (Tropine et al., 2004). However, the impact of intracranial glioma on brain structural networks based on tractography constructed under different FA thresholds still has not been investigated. The current study found more fibers with a VR > 50% (Table S2)  cortices (Jiang et al., 2015). Although experience in this field is scarce, our findings still demonstrate that glioma has a structural and functional influence on language function. Moreover, it also indicates the distinguished importance of the POS-based network properties for language performance.

| Difference of efficiency between the two groups
Regarding the global performance, EG from M whole , M left , M right , M pos , and M neg involved statistical differences under different VR thresholds. The correlation analysis on EGs also showed a significant correlation to aphasia levels ( Table 4). The impairment of the cerebral structural organization by gliomas leads to reduced signal transmission and interaction efficiency at a global level. Na et al. (2018) analyzed white matter networks of adults suffering from brain tumors and found that their EG was significantly lower than in the healthy adult control group proving the long-term effects of tumors on brain structures after resection. The structural network of patients suffering from mild cognitive impairment without anatomical structure damages also showed significantly lower EGs than normal subjects, and it was related to the functional score, which provided the correlation between changes in structural network properties and changes at a functional level (Berlot, Metzler-Baddeley, Ikram, Jones, & O'Sullivan, 2016). Results of the current study regarding the EGs from between the NA and GIA group and negatively correlated to their aphasia levels (Table 3). Koenis et al. (2018) reported that higher local efficiencies of the tractography-based network were related to higher intelligence and indicated efficiency of a local area to be related to the overall performance. Regarding neurosurgical treatment, this functional correlation indicates the necessity to avoid damaging positively mapped regions and their tracts, which has been demonstrated beneficial for resection planning and clinical outcomes (Raffa et al., 2018).
In our study, ELs and EL-diffs from left hemispheric and POS-based networks correlate with aphasia levels (Table 4), further emphasizing the importance of preserving POS regions.
In the intragroup analysis in NA patients, EG-diff (p < .001, corrected) and EL-diff (p = .014, corrected) of M neg were the largest among matrices ( Figure S2). In the GIA intragroup analysis, however, EG-diff (p < .001, corrected) and EL-diff (p < .001, corrected) of M pos were largest ( Figure S2). Intergroup analysis showed that EG-diff and EL-diff of M pos in the NA group were smaller than in the GIA group. Although the current study cannot explain the exact mechanism between global and local efficiency in language function, it proves that gliomas interfere with the information interaction of networks at local and global levels. Moreover, the alteration of connectome properties between different VR thresholds enhances functional performance investigations.
Alterations in the efficiency of various networks between NA and GIA patients reflect different capability information interactions and transmission in these brains. To some extent, this supports the theory of language function being based on the collaboration between dorsal and ventral streams as proposed in previous studies (Chang, Raygor, & Berger, 2015;Lou et al., 2019;Saur et al., 2008). It is essential to notice the importance of robustness of the POS-related network performance in maintaining language function. The relationship between the dorsal and ventral streams and the degree of aphasia also reflects the impaired efficiency of information processing in cerebral networks impacted by tumors (Milner, 2017;van Polanen & Davare, 2015).

| Limitations
The current study is based on the nTMS-based language connectome applying a deterministic FT approach. However, some points should be taken into account when interpreting the results.
First, tumor sizes varied between the two groups. Yet, this is unavoidable concerning the induction of preoperative aphasia due to larger tumor volume. Moreover, glioma size did not significantly correlate to the GIA group's aphasia levels and was adopted in the ANCOVA analysis.
Second, data on cognitive impairment were not acquired in this study. Understanding the functional plasticity might further improve when combining the complexity of the language network with cognitive analysis. Yet, the application of combining cognition testing and nTMS mapping still needs to be further investigated.
Third, the ONT combined with nTMS language mapping cannot be performed by patients with complete aphasia. Thus, data of patients with extensive aphasia were not collected in the current study due to data consistency requirements. Furthermore, the ONT applied in nTMS mapping only involves the most basic language functions and is incapable of identifying logic and grammar-related errors.
POS and NEG regions found in ONT do not fully represent the complex organization of language function, which should be noticed when interpreting the results of the current study. Intracranial connectome changes in those patients could, however, further deepen the understanding of the language networks.

| CONCLUSION
To the best of our knowledge, this is the first study to combine nTMS language mapping and connectome analysis to investigate the brain's glioma-related functional changes. SMA is more likely to be involved in language production in GIA patients than in NA patients, indicating nTMS language mapping enables to investigate regions related to language superordinate controlling levels. Intracranial glioma interference causes differences in the distribution and performance of the brain's various networks mainly impacting M neg in NA patients and M pos in GIA patients. GIA patients were without sufficient M neg compensation, this is related to their language dysfunction. nTMS mapping can identify unstable connectomes related to functional deficits such as aphasia.
Individualized FT threshold settings like VR in this study and changes in local and global connectome properties under different thresholds can reflect the lesions' impact on the language network.
This new approach is suggested to identify language-involved network components via a function-specific connectome approach not only in glioma patients. Moreover, it could be applied to various diseases including stroke, traumatic brain injury, and neurodegeneration.

ACKNOWLEDGMENT
This study was funded entirely by institutional grants from the Department of Neurosurgery, Technical University of Munich, Germany, School of Medicine, Klinikum Rechts der Isar. Open access funding enabled and organized by Projekt DEAL.