Access to this full-text is provided by Wiley.
Content available from Human Brain Mapping
This content is subject to copyright. Terms and conditions apply.
RESEARCH ARTICLE
Preoperative function-specific connectome analysis predicts
surgery-related aphasia after glioma resection
Sebastian Ille
1,2
| Haosu Zhang
1
| Lisa Sogerer
1
| Maximilian Schwendner
1
|
Axel Schöder
1
| Bernhard Meyer
1
| Benedikt Wiestler
3
| Sandro M. Krieg
1,2
1
Department of Neurosurgery, Klinikum rechts
der Isar, School of Medicine, Technical
University of Munich, Munich, Germany
2
TUM-Neuroimaging Center, Technical
University of Munich, Munich, Germany
3
Department of Diagnostic and Interventional
Neuroradiology, Klinikum rechts der Isar,
School of Medicine, Technical University of
Munich, Munich, Germany
Correspondence
Sandro M. Krieg, Department of Neurosurgery,
Klinikum rechts der Isar Technische Universität
München, Ismaninger Street 22, 81675
Munich, Germany.
Email: sandro.krieg@tum.de
Abstract
Glioma resection within language-eloquent regions poses a high risk of surgery-
related aphasia (SRA). Preoperative functional mapping by navigated transcranial
magnetic stimulation (nTMS) combined with diffusion tensor imaging (DTI) is increas-
ingly used to localize cortical and subcortical language-eloquent areas. This study
enrolled 60 nonaphasic patients with left hemispheric perisylvian gliomas to investi-
gate the prediction of SRA based on function-specific connectome network proper-
ties under different fractional anisotropy (FA) thresholds. Moreover, we applied a
machine learning model for training and cross-validation to predict SRA based on pre-
operative connectome parameters. Preoperative connectome analysis helps predict
SRA development with an accuracy of 73.3% and sensitivity of 78.3%. The current
study provides a new perspective of combining nTMS and function-specific connec-
tome analysis applied in a machine learning model to investigate language in neu-
rooncological patients and promises to advance our understanding of the intricate
networks.
KEYWORDS
connectome, DTI, graphic analysis, nTMS, surgery-related aphasia
1|INTRODUCTION
Previous studies on brain structures involved in language networks
have demonstrated the inter-individual variations of cortical and sub-
cortical language regions (Dehaene et al., 1997; Shaywitz et al., 1995;
Tzourio-Mazoyer et al., 2004; Voets et al., 2019). Recent findings on
language refer more to the individual level analysis (Ardila et al., 2016),
Abbreviations: A, average degree; AAT, Aachener aphasia test; CPS, cortical parcellation
system; DCS, direct cortical stimulation; DTI, diffusion tensor imaging; EG, global efficiency;
EHI, Edinburgh Handedness Inventory; EL, local efficiency; FA, fractional anisotropy; FAT,
fractional anisotropy threshold; FLT, fiber length threshold; FT, fiber tracking; NA, no
aphasia; NEG, negative nTMS stimulation regions; nTMS, navigated transcranial magnetic
stimulation; POS, positive nTMS stimulation regions; SRA, surgery-related aphasia; VR,
visualized ratio.
Sebastian Ille and Haosu Zhang contributed equally to this study.
Received: 30 December 2021 Revised: 30 May 2022 Accepted: 22 June 2022
DOI: 10.1002/hbm.26014
This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any
medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
© 2022 The Authors. Human Brain Mapping published by Wiley Periodicals LLC.
5408 Hum Brain Mapp. 2022;43:5408–5420.
wileyonlinelibrary.com/journal/hbm
contributing to the modern concept of individualized treatment and
function-specific glioma resection, aiming to maximize glioma resection,
an important prognostic marker for these patients. However, for patients
with left-sided perisylvian glioma, surgery-related aphasia (SRA) is still a
common risk of surgery (Negwer et al., 2018; Sollmann et al., 2019).
When the glioma is localized in eloquent regions, different functional
outcomes were observed defined by critical areas preserved during sur-
gery (Sollmann et al., 2020). With this in mind, identification of bio-
markers that can predict the risk of SRA before surgery is pivotal to
improve treatment plans and to preserve language function, while still
aiming for a gross total resection.
The application of preoperative navigated transcranial magnetic
stimulation (nTMS)-based diffusion tensor imaging (DTI) tractography
combined with intraoperative direct cortical stimulation (DCS) lan-
guage mapping was increasingly adopted to investigate the correlation
of the network and language deficits with the aim of reducing risks of
SRA (Ille et al., 2016; Ille, Sollmann, Hauck, Maurer, Tanigawa, Ober-
mueller, Negwer, Droese, Boeckh-Behrens, et al., 2015a; Ille, Soll-
mann, Hauck, Maurer, Tanigawa, Obermueller, Negwer, Droese,
Zimmer, et al., 2015b; Picht et al., 2013; Tarapore et al., 2012). Previ-
ous studies on this topic focused on detecting the nodes and tracts
related to SRA. For instance, a cut-off distance between lesions and
the arcuate fascicle (AF) set at 16 mm was associated with SRA
(Sollmann et al., 2020). Likewise, the inferior fronto-occipital fascicle
(IFOF), the frontal aslant tract (FAT), and the superior longitudinal fas-
cicle (SLF)/AF were correlated with the corresponding status of lan-
guage function preoperatively, postoperatively, and at follow-up (Ille
et al., 2018).
The current view considers cerebral language function being pro-
duced and underpinned by dynamic interactions within and between
specialized brain networks (Shafto & Tyler, 2014). It demonstrated that it
is not enough to understand cerebral function from the composition of
language-related brain structures alone. It should also consider aspects
of the network performance as well. Mbwana's findings in 2009 demon-
strated that the individual language-related network was different at cer-
tain levels of network reorganization in localization-related epilepsy
(Mbwana et al., 2009). In 2019, language network's different individual
performances were detected by Voets and colleagues, showing the func-
tional connectivity of brain regions involved in language fluency. This
study identified “fingerprints”of brain plasticity in individual patients
(Voets et al., 2019). Their experimental method was based on functional
MRI (fMRI) to investigate the changes in the language network. In con-
trast, studies by Chang et al. were based on DTI tractography, which
developed the understanding of the anatomical and functional language
network (Chang et al., 2015). Notably, this study focused on the relation-
ship between the network and functional performances. However, it did
not investigate the characteristics of the network itself, including net-
work efficiency and average degrees.
nTMS language mapping defines positive and negative sites
according to whether the stimulation induces language errors. Previ-
ous studies mostly focused on positive sites and related tracts to
investigate their corresponding functional alteration. However, there
has been little agreement on understanding the networks’properties
corresponding to positive and negative mapping regions to date. We
investigated not only the intra-network properties of the mapped
sites, but also the network of brain regions connected to those sites.
It provides a comprehensive perspective of the function of these
mapped sites in the network and facilitates our understanding on the
role of intra- and cross-hemispheric connectivity and their corre-
sponding networks in language performance.
The objectives of this present study are to determine preopera-
tive risk factors for early postoperative SRA in patients with language-
eloquent glioma based on (1) the changes of structural network char-
acteristics, (2) the differences between nTMS language-positive and
language-negative sites in patients with and without SRA, (3) differ-
ences between the network characteristics of nTMS language-positive
and language-negative sites, and (4) the comparisons of the network
of brain regions connecting mapping sites with the internal network
of mapping sites.
2|MATERIALS AND METHODS
2.1 |Ethical approval
The current study followed the Declaration of Helsinki and its later
amendments and was approved and supervised by the local ethic
committee of the Technical University Munich (registration number
222/14 and 192/18). All patients gave written informed consent to
participate in this study.
2.2 |Inclusion criteria and demographic data
collection
The present study is a post hoc analysis of 60 patients with unaf-
fected preoperative language function. Thirty patients developed SRA
lasting more than 1 month after the operation (Group SRA) and
30 patients did not show postoperative aphasia (Group NA).
The following inclusion criteria were defined for the current
study:
1. Age above 18 years.
2. German as first language.
3. Left-sided gliomas inside or adjacent to left-sided perisylvian
regions or the arcuate fasciculus
4. Preoperative MRI-imaging: T1-weighted image and a DTI scan
with 32 diffusion directions.
5. Preoperative nTMS language mapping.
6. Clinical language follow-up performed more than 1 month after
the operation.
The exclusion standards were as follows:
1. Neurological or psychiatric diseases (except for the diagnosis of
glioma).
2. Preoperative aphasia.
3. Biopsy only.
ILLE ET AL.5409
2.3 |MRI and clinical testing
Preoperative MRI scans (Achieva 3T, Philips Medical System) were
acquired during routine imaging, including DTI (TR/TE: 5000/78 ms,
voxel size of 2 22mm
3
, 32 diffusion gradient directions, b-value
1000 s/mm
2
) and T1-weighted with contrast agent (TR/TE: 9/4 ms,
1mm
3
iso-voxel; Dotagraf 0.5 mmol/ml, produced by Jenapharm
GmbH & Co. KG).
According to the Aachener aphasia test (AAT), the aphasia level
was rated before the operation and 1 month postoperatively (Biniek
et al., 1992). For the present analysis, data were reviewed by a speech
therapist, leading to the categorization of aphasia levels according to
the classification in the AAT (0 =no impairment of language function;
1=slight impairment of daily communication; 2 =moderate impair-
ment of language function, daily communication possible; 3 =severe
impairment of language function, daily communication not possible;
A=nonfluent; B =fluent). Afterwards we binarized the patients’lan-
guage status to SRA and NA depending on the comparison of evalua-
tions pre- and postoperatively. The handedness of patients was
analyzed using the Edinburgh Handedness Inventory (EHI;
Oldfield, 1971).
2.4 |nTMS language mapping
Preoperative nTMS language mapping was performed following pro-
cesses established in previous studies and according to our standard
protocol (Ille, Sollmann, Hauck, Maurer, Tanigawa, Obermueller,
Negwer, Droese, Boeckh-Behrens, et al., 2015a; Ille, Sollmann, Hauck,
Maurer, Tanigawa, Obermueller, Negwer, Droese, Zimmer,
et al., 2015b; Picht et al., 2013). T1 images were imported to the Nex-
stim eXimia NBS system (version 5.1.1; Nexstim Plc) for navigating
the targeted stimuli (Ille & Krieg, 2021; Picht et al., 2013). According
to the cortical parcellation system (CPS), stimulation with 100% rest-
ing motor threshold intensity was targeted at 46 regions predefined
on the left hemisphere. Comparing the patients’object naming task
performance in the session under stimulation (stimulation session) to
the session without nTMS stimulation (baseline session), the regions
corresponding to language errors were identified as language-positive
regions (POS) and language-negative regions (NEG). Both were then
exported as DICOM files. With respect to the intraoperative use of
nTMS data, the mapping of the tumor hemisphere had priority.
2.5 |Connectome construction
Image files were transformed to NIFTI format (dcm2niix. https://
github.com/rordenlab/dcm2niix). For individualized analysis, B0 was
extracted from the DTI file for registration. First, each gradient direc-
tion from the DTI was linearly registered onto the B0 image (Figure 1).
Second, the T1w image was skull-stripped (HD-bet, https://github.
com/MIC-DKFZ/HD-BET/), then linearly registered to the B0 image
(Isensee et al., 2019), during which transformation datasets of the
conversion process were generated. Because the POS and NEG were
constructed based on the space as T1w, POS and NEG were regis-
tered to the B0 image applying the same transformation datasets. In
the end, the atlas template AAL90 was deformably also registered to
the B0 image. This registration enabled the cerebral structures to be
individually identified, through which atlas regions corresponding to
mapping points can be identified in POS and NEG images,
respectively.
Next, whole brain tractography was constructed (Python library
DIPY, Version 1.2.0, https://dipy.org). The constrained spherical
deconvolution (CSD) model was applied (Figure 1l). Then a determinis-
tic algorithm was used for fiber tracking (FT) under a series of the
fractional anisotropy thresholds (FAT) that were started from 0.0 and
increased by 0.01 with a fiber length threshold (FLT) set at 30 mm
(Song et al., 2014) to exclude U-fibers. Individual FT stopped when a
minimum count of fibers was visualized, and the corresponding FAT
was at the maximum value (FAT
max
). During this process, the maxi-
mum FA for every single fiber was recorded as fa
max
, and its FA ratio
(VR) was calculated by the formula:
VR ¼
famax
FAmax
100%
Connecting fibers with VR ≥25% and ≥50% (Figure 1) were selected
according to the corresponding nodes used in the following seven
connectomes:
1. M
whole
: the matrix (M) with nodes from both hemispherical regions
according to the template AAL90 at the size of 90 nodes 90
nodes.
2. M
left
: M with nodes from the left hemispherical regions according
to the template AAL90 at the size of 45 nodes 45 nodes.
3. M
right
: M with nodes from the right hemispherical regions accord-
ing to the template AAL90 at the size of 45 nodes 45 nodes.
4. M
pos-rela
: M with nodes from the positive language mapping
regions and the regions connected to them according to the tem-
plate AAL90.
5. M
neg-rela
: M with nodes from the negative language mapping
regions and the regions connected to them according to the tem-
plate AAL90.
6. M
pos
: M with nodes from the positive language mapping regions
according to the template AAL90.
7. M
neg
: M with nodes from the negative language mapping regions
according to the template AAL90.
Because the individual brain and tumor sizes were variable, binari-
zation was applied to minimize their impacts. Likewise, connections
between two respective regions in the connectome were transformed
depending on whether the fiber counts were ≥3 or not.
The metrics representing the graphic properties of the connec-
tomes were analyzed through the NetworkX 2.5 library (https://
networkx.org/) in Python 3.7 (https://www.python.org/), consisting
of (1) the average degree (AD): the number of edges compared to the
5410 ILLE ET AL.
number of nodes representing the connecting density (Latora &
Marchiori, 2001; Wang et al., 2010); (2) the average shortest path
lengths (AL): the average of the smallest number of edges in the path
between two nodes in the whole network, indicating its intensity and
density of connections (Rutter et al., 2013; Zhang et al., 2021);
(3) global efficiency (EG): to measure the efficiency of parallel informa-
tion transfer and integrated processing (Latora & Marchiori, 2001;
Wang et al., 2010); and (4) local efficiency (EL): to estimate the
FIGURE 1 This figure shows the workflow of MRI processing in network construction. Figure A represents DTI scans with 32 directions. Figure C is
the skull-stripped MRI (T1 with contrast) derived from Figure B through HD-bet. Figure D is the AAL90 template, and figures E and F show nTMS data.
Figure A, C, D, and F are registered to the B0 (Figure G). Figure H contains the registered DTI, Figure I contains registered T1 with contrast, Figure J
shows the registered AAL90 template, and Figure K shows the registered nTMS data (positive and negative regions). Then, tractography based on the
CSD model applying deterministic fiber tracking thresholding at a visualization ratio (VR) of 25% and 50% was performed to generate respective
connectomes. In the next step, connectomes for further analysis were constructed based on different nodes and connections as the following: (1) M
whole
(a): the matrix (M) with nodes from both hemispherical regions accordingtothetemplateAAL90atthesizeof90nodes*90nodes.(2)M
left
(b): M with
nodes from the left hemispherical regions according to the template AAL90 at the size of 45 nodes * 45 nodes. (3) M
right
(c): M with nodes from the right
hemispherical regions according to the template AAL90 at the size of 45 nodes * 45 nodes. (4) M
pos-rela
(d): M with nodes from the positive language
mapping regions and the regions connected to them according to the template AAL90. (5) M
neg-rela
(e): M with nodes from the negative language
mapping regions and the regions connected to them according to the template AAL90. (6) Mpos (f): M with nodes from the positive language mapping
regions according to the template AAL90. M
neg
(g): M with nodes from the negative language mapping regions according to the template AAL90
ILLE ET AL.5411
efficiency of communication between neighbors of a node when that
node is eliminated (Latora & Marchiori, 2001; Wang et al., 2010). Con-
nectomics figures were created in Matlab (R2016b, Academic License
to TUM) using BrainNetViewer (Xia et al., 2013).
2.6 |Statistical analysis
The following statistical analyses were performed with GraphPad
Prism (Version 8.4.3): (1) Baseline characteristics: the comparisons of
the NA and SRA groups using the Chi-square test for handedness,
gender, and histopathological diagnosis. Independent t-testing was
applied for the comparison of age and glioma size. (2) Intra- and inter-
group mapping region analysis was performed by Chi-Square or Fisher
Exact testing. (3) Calculation of the intra-group ratio of language map-
ping regions. (4) Inter-group comparisons of the network properties
from different connectomes. (5) Connectome parameters were used
to train a six-fold cross-validated multi-layer perceptron (with
12>6>1units and sigmoid output) to predict SRA (Figure 2).
3|RESULTS
3.1 |Demographic analysis
There were no differences of age between the NA group (59.7
± 15.1 years) and SRA group (58.8 ± 15.2 years; p=.327). The com-
parison of tumor size in NA group (2.438 ± 2.630 cm
3
) and SRA group
(2.829 ± 2.853 cm
3
) showed no difference (p=.583; Table 1). The
Chi-square test showed no difference for handedness, gender, and
WHO grade between the NA and SRA group (Table 1). The FAT
max
values of both groups were comparable with 0.531 ± 0.065 for the
NA and 0.545 ± 0.053 for the SRA group (p=.378; Table 1).
3.2 |Mapping region analysis
The number of POS regions was significantly lower than NEG regions
in both groups (p< .001, Figure 3), without differences between the
NA and SRA groups regarding the number of POS and NEG regions.
There were no differences of objects numbers detected after regard-
ing baseline tests (p=.921).
The intragroup analysis of overlapping regions detected in more
than 20 patients per group (>66.7%) is presented in Table 2. The fron-
tal superior gyrus was positively mapped in 20 patients from the NA
group but only in 15 patients of the SRA group. The triangular parts
of the frontal inferior gyrus and the parietal inferior gyrus were nega-
tively mapped in 20 and 22 patients of the SRA group but only in
16 and 18 NA patients, respectively (Table 2). There were no signifi-
cant differences for the number of mapping regions between NA and
SRA groups (Table S2).
3.3 |Average degree and shortest path length
analysis
Generally, ADs were higher in the NA group than in the SRA group
thresholding at 25% and 50% VR, in which the comparisons of ADs
from the M
right
,M
pos-rela
, and M
pos
between the NA group and the
SRA group showed statistically significant differences (25% VR:
p=.044, p=.040, p=.045; 50% VR: p=.031, p=.018, p=.046;
Table 3; Figure 4). Only at 50% VR, the M
whole
and M
left
showed
FIGURE 2 This figure presents the
workflow of the sixfold cross-validation
machine learning for predicting the
postoperative language functions. First,
common features are defined as to be
significantly correlated to the
postoperative language levels in more
than 80% of the six training sets. Second,
those parameters were used for the
cross-validation machine learning using
neural network, in which the model was
generated, and its performance was
assessed, including accuracy, precision.
5412 ILLE ET AL.
higher AD in the NA group (p=.023, p=.037; Table 3; Figure 4).
Regarding the path length, AL from the M
right
is significantly smaller in
NA group (p=.021; Table 3; Figure 4).
3.4 |Efficiency analysis
The connectome efficiencies EG and EL averages were higher in the
NA group than those of the SRA group. Efficiencies from M
neg-rela
showed no significant difference between the two groups. The NA
group had statistically higher EG and EL based on M
whole
under both
25% and 50% VR as compared to the SRA group (Figure 4; Table 3).
The EL of M
left
showed a difference between the two groups, while
the difference in EG did only appear with 25% VR (Figure 4; Table 3).
The M
pos-rela
and M
pos
showed more EG under both VR thresholds in
the NA group. However, their ELs were without significant differ-
ences (Figure 4; Table 3). EL from M
neg
with 50% VR was significantly
lower in the SRA group (Figure 4; Table 3).
3.5 |Correlation and machine learning analysis
Based on the ROC analysis, EG from M
whole
and M
left
, EL from M
left
,
and AL from M
right
under 25% VR were statistically significant
(p=.020, p=.027, p=.013, and p=.025; Table 4). Moreover, at
50% VR, additional graphic properties were found to be statistically
significant (M
pos-rela
-EG, M
pos
-EG, M
whole
-EG, M
right
-EG, M
left
-EL,
M
whole
-EL, M
whole
-AD, M
left
-AD, M
right
-AD, and M
pos-rela
-AD;
Table 4). M
whole
-EG and M
pos-rela
-EG showed the highest sensitivity
(0.733) and specificity (0.600) compared to other graphic properties.
M
whole
-AD, M
left
-AD, M
right
-AD, and M
pos-rela
-AD showed high sensi-
tivity (>0.800) at a lower specificity between 0.433 and 0.567, leading
to the high risk of false-negative differentiation.
In the cross-validation machine learning model, the average of
Accuracy, Precision, Sensitivity, and area under the curve (AUC)
across six cross-validation folds are all above 70% (76.7%, 75.2%,
80.3%, and 79.9%; Table S2).
4|DISCUSSION
The present study provides a new perspective to advance the investiga-
tion of language-related network properties through function-specific
connectome analysis. It offers a new analysis to identify patients in the
preoperative phase at risk of potentially developing postoperative SRA,
using graphic analysis and a machine learning model.
4.1 |Prediction of SRA
Predicting SRA before surgery is a difficult point in current neurosur-
gery practice. Comprehensive and effective methods are needed to
TABLE 1 Patient characteristics Items Chi-square p-value
Handedness Left NA 4 Right NA 26 0.131 .718
SRA 5 SRA 25
Gender Male NA 6 Female NA 24 2.052 .152
SRA 11 SRA 19
WHO I–II NA 8 III–IV NA 22 0.089 .766
SRA 7 SRA 23
Age NA: 59.7 ± 15.1 years SRA: 58.8 ± 15.2 years P: .327
T: .988
Note: This table presents the Chi-square testing of patient characteristics, including handedness, gender,
and WHO (World Health Organization) grade of glioma. No difference was detected between the two
groups without aphasia (NA) and with surgery-related aphasia (SRA).
FIGURE 3 The figure shows that the number of positive mapping
regions (in red) was significantly lower than the negative mapping
regions (in blue) in both groups compared by paired t-testing. No
differences were found for the number of mapping regions between
the NA (no aphasia, show as “o”) and the SRA (surgery-related
aphasia, show as “●”) group.
ILLE ET AL.5413
TABLE 2 Analysis of mapping regions
No aphasia (NA) Surgery-related aphasia (SRA)
Ratio of overlapping positive regions Ratio of overlapping negative regions Ratio of overlapping positive region Ratio of overlapping negative regions
Frontal middle gyrus: 93.3%
Precentral gyrus: 76.7%
Temporal middle gyrus: 76.7%
Postcentral gyrus: 70.0%
Frontal superior gyrus: 67.7%
Paracentral Lobule: 100%
Fusiform gyrus: 100%
Precuneus gyrus: 100%
Lingual gyrus: 100%
Cuneus gyrus: 100%
Frontal inferior gyrus (Orb): 100%
Insula: 100%
Frontal superior medial gyrus: 96.7%
Occipital superior gyrus: 96.7%
Supplementary motor area: 86.7%
Heschl gyrus: 86.7%
Rolandic gyrus (Oper): 70.0%
Frontal middle gyrus: 83.3%
Temporal middle gyrus: 80.0%
Precentral: 73.3%
Postcentral: 66.7%
Occipital superior gyrus: 100%
Fusiform gyrus: 100%
Precuneus gyrus: 100%
Frontal superior medial gyrus: 100%
Insula: 100%
Fronta l inferior gyrus (Orb): 100%
Paracentral Lobule: 100%
Lingual gyrus: 100%
Cuneus gyrus: 100%
Heschl gyrus: 96.7%
Supplementary motor area: 83.3%
Rolandic gyrus (Oper): 73.3%
Parietal inferior gyrus: 73.3%
Frontal inferior gyrus (Tri): 66.7%
Note: This table shows the intragroup analysis of overlapping positive and negative mapping regions identified in more than 20 patients (>66.7%) after being registered to the AAL90 template (Mid: middle; Sup:
superior; Inf: inferior; Supp: supplementary; Oper: opercular part; Tri: triangular part; Orb: orbital part). All mapping regions were located in the left hemisphere.
5414 ILLE ET AL.
assist clinical treatment decisions, in order to reduce the risk of post-
operative aphasia. In the current study, network properties shown in
Table 4are significant predictors of SRA. However, using one of them
alone as a predictor cannot result in satisfactory predictive perfor-
mance. When comprehensively combining them into a machine learn-
ing model, a sensitivity in predicting SRA of 78.3% (specificity of
TABLE 3 Analysis on average degree
and efficiency analysis
Items
Efficiency
M
whole
M
left
M
right
M
pos-rela
M
neg-rela
M
pos
M
neg
25% VR
AD
NA 7.635 5.430 5.833 1.896 2.991 2.073 2.392
SRA 7.073 4.958 5.326 1.624 2.880 1.619 2.205
p-value .054 .087 .044 .040 .457 .045 .344
AL
NA 2.079 1.927 1.861 2.235 2.28 1.251 1.901
SRA 2.121 1.958 1.921 2.254 2.329 1.203 1.953
p-value .091 .427 .021 .674 .264 .737 .587
EG
NA 0.537 0.566 0.593 0.300 0.296 0.704 0.595
SRA 0.520 0.535 0.571 0.271 0.292 0.596 0.548
p-value .014 .025 0.078 .040 .743 .050 .084
EL
NA 2.863 1.883 2.344 0.733 1.105 0.682 0.633
SRA 2.358 1.466 1.914 0.552 0.960 0.556 0.586
p-value .037 .005 .089 .122 .225 .122 .247
50% VR
AD
NA 2.863 1.883 2.344 0.733 1.105 0.933 0.766
SRA 2.358 1.466 1.914 0.552 0.960 0.661 0.583
p-value .023 .037 .031 .018 .182 .046 .154
AL
NA 2.804 2.424 2.547 2.437 2.807 0.942 1.474
SRA 2.934 2.437 2.411 2.288 2.884 0.617 1.48
p-value .092 .929 .214 .319 .614 .090 .976
EG
NA 0.330 0.297 0.371 0.105 0.121 0.437 0.227
SRA 0.289 0.246 0.314 0.079 0.105 0.323 0.160
p-value .036 .064 .017 .026 .215 .041 .078
EL
NA 0.367 0.332 0.373 0.133 0.127 0.338 0.174
SRA 0.306 0.261 0.307 0.098 0.087 0.187 0.086
p-value .048 .037 .070 .095 .053 .095 .026
Note: This table shows differences of average degree (AD), global efficiency (EG) and local efficiency (EL)
between the no aphasia (NA) group and the surgery-related aphasia (SRA) group under visual ratio (VR)
thresholds of 25% and 50%. AD of matrices of regions located in the right hemisphere (M
right
), matrices
of regions connected to nTMS positive regions (M
pos-rela
), and of matrices of nTMS positive regions
(M
pos
) was higher in the NA group than in the SRA group. Under both 25% and 50% VR thresholds, EGs
from matrices of whole brain regions (M
whole
), matrices of left hemispheric regions (M
left
), matrices of
regions connected to nTMS positive regions (M
pos-rela
), matrices of nTMS negative regions (M
neg
), and
matrices of nTMS positive regions (M
pos
) were higher in the NA group than in the SRA group, except
from 50% VR, EGs of matrices from right hemispheric regions (M
right
) were higher in the NA group than
in the SRA group instead of EGs from M
left
at a VR of 50%. ELs of M
whole
and M
left
under both 25% and
50% VR thresholds as well as ELs from M
neg
at a VR of 50% were higher in the NA group as well. Values
written in bold highlight p-values less than .05.
ILLE ET AL.5415
FIGURE 4 This figure presents the tractography of connectomes thresholding at 25% and 50% visualization ratio (VR). (a) Shows connections
tracked in more than 10 patients (>33.3%) for the networks: M
whole
,M
left
,M
right
,M
pos-rela
, and M
neg-rela
. (b) Shows connections tracked in more
than three patients (>10.0%) for M
pos
and M
neg
. Results are divided into the NA (no aphasia) and the SRA (surgery-related aphasia) group each.
5416 ILLE ET AL.
66.7%) can be accomplished (Table S2). While encouraging, these
results need to be validated in further studies.
Most of the mapped regions are similar between the two groups.
Furthermore, the number and location of positive and negative map-
ping regions were similar between both groups. It is reasonable to see
those similar cortical mapping results considering that the NA and
SRA group were without clinically detectable aphasia in the preopera-
tive testing. Hence, the impact of subcortical connections on postop-
erative outcome as analyzed can definitely be related to the SRA
since both groups underwent tumor resection.
4.2 |Higher connectome efficiencies in the NA
group
The lower network efficiencies were calculated from connectomes
from the whole brain and left hemisphere in the SRA group at both
global and local levels. The efficiencies of connectomes from
M
whole
and M
left
of higher VR (>50%) and lower VR (>25%) are both
affected. The impact of the left-sided glioma was not limited to one
hemisphere but could also be found at the global level. In the study
by Ries and colleagues, it was demonstrated that patients with
brain lesions relying more on interhemispheric cooperation to
select words to complete the object naming task (Ries et al., 2016),
which supported that SRA patients’undermined language perfor-
mance was related to reduced connections among regions. The
dual-stream theory on language function has pointed out that lan-
guage processing is not a simple function of local brain regions but
involves information processing and signal coordination among
multiple cortical and subcortical structures (Saur et al., 2008), also
emphasizing that multi-regional network collaboration is needed
for executing language functions. Furthermore, a study by Schup-
pert and colleagues indicated that the global, fragmented neural
substrates underlying local and global musical information were
processed in the melodic and temporal dimensions (Schuppert
et al., 2000).
Only EGs were significantly higher in M
pos-rela
and M
pos
in the NA
group, again demonstrating the importance of the language-positive
mapping regions and connections among them in maintaining global
functional integrity. It is consistent with the results of AD analysis,
which indicates that language-positive regions in terms of nTMS map-
ping are a part of the connectome to transfer information during the
language production process. This might also explain the difference
between DCS and TMS positive brain regions found in previous stud-
ies, suggesting that nTMS mostly discovers nodes and connections of
language-related networks, while DCS focuses more on detecting the
importance of certain regions and related fibers for language perfor-
mance (Sollmann et al., 2014). The interaction of language-related cor-
tical areas and subcortical fibers might be investigated more entirely
through nTMS-based function-specific connectome analyses, which is
of great significance for further research on language's structural
composition.
TABLE 4 ROC analysis of graphic
properties Items AUC Cut-off value Sensitivity Specificity Youden index p-value
25% VR
M
whole
-EG 0.674 0.533 0.733 0.600 0.333 .020
M
left
-EG 0.667 0.572 0.70 0.600 0.300 .027
M
left
-EL 0.687 0.722 0.867 0.500 0.367 .013
M
right
-AL 0.669 1.866 0.733 0.600 0.333 .025
50% VR
M
pos-rela
-EG 0.650 0.095 0.733 0.600 0.333 .046
M
pos
-EG 0.652 0.419 0.667 0.600 0.267 .044
M
whole
-EG 0.663 0.322 0.633 0.700 0.333 .030
M
right
-EG 0.677 0.370 0.733 0.567 0.300 .019
M
left
-EL 0.650 0.315 0.700 0.633 0.333 .046
M
whole
-EL 0.654 0.352 0.700 0.633 0.333 .040
M
whole
-AD 0.681 2.894 0.900 0.467 0.367 .016
M
left
-AD 0.647 1.822 0.800 0.567 0.367 .049
M
right
-AD 0.654 2.522 0.900 0.433 0.333 .040
M
pos-rela
-AD 0.647 0.784 0.867 0.467 0.333 .048
Note: This table presents the corresponding graphic properties’area under the curve (AUC), as well as
the sensitivity and specificity with p< .05 in receiver operating characteristic (ROC) curve analysis. The
corresponding sensitivity and specificity of average degree (AD), global efficiency (EG) and local
efficiency (EL) in predicting surgery-related aphasia (SRA) were summarized under different visual ratio
(VR) thresholds for different matrices (M) including matrices of whole brain regions (M
whole
), matrices of
left hemispheric regions (M
left
), matrices of right hemispheric regions (M
right
),matrices of regions
connected to nTMS positive regions (M
pos-rela
), matrices of nTMS positive regions (M
pos
), matrices of
regions connected to nTMS negative regions (M
neg-rela
), and matrices of nTMS negative regions (M
neg
).
ILLE ET AL.5417
4.3 |More high-VR fibers in the NA group
ADs were generally higher in the NA group than in the SRA group.
M
pos-rela
and M
pos
under both VRs thresholds were different between
the two groups but not for the M
neg
and M
neg-rela
.M
pos-rela
shows the
matrix consisting of positive mapping regions and their connected
regions. First, from a functional point of view, it demonstrated that
positive regions are essential for connecting other regions and serving
as interaction centers to integrate language performance signals. Sec-
ond, there are two potential reasons for higher AD in the NA group
from the structural point of view. As the tumor size did not show dif-
ferences between the two groups, one attribution is that patients in
the SRA group rely on a weaker organization of connections initially,
which ultimately leads to functional decompensation after resection.
Similarly, patients in the NA group have more capability for
compensation—potentially induced by glioma—to generate more con-
nections. With this in mind, previous studies have concluded similar
results indicating a higher risk of SRA in case of fewer interhemi-
spheric connections (Negwer et al., 2018; Sollmann et al., 2017).
The tumor was localized within the left hemisphere in both
groups leading to less AD as compared to the right hemisphere. The
AD from fibers with VR between 25% and 50% were similar in both
groups (Table S3). Significant differences in AD were shown for the
connections with VR >50% (Table 3). Moreover, the investigation on
the FAT
max
shows no difference between the NA and SRA groups
(p=.378). It indicated that the fibers with higher VR (>50%) were
more in the NA group than in the SRA group, which was related to
the higher FA of the language-related connections. The higher FA was
related to the healthy subjects’structural remodeling receiving mem-
ory learning training (Darquie et al., 2001; Sagi et al., 2012). Further-
more, it supports a study on Baduk players, who developed increased
FA values in the white matter of frontal, cingulum, and striato-
thalamic areas after long-term training that are part of the functional
networks for attentional control, working memory, executive regula-
tion, and problem-solving (Lee et al., 2010). In the present study, the
tumor's remodulation was at different levels, being lower in the SRA
group than in the NA group, leading to different language prognoses.
4.4 |Limitation
First, no healthy subjects were enrolled as controls for the present
analysis, and a bigger sample size for machine learning for predicting
the postoperative language levels is still in need. Second, we have not
used FMRI to identify the dominant language hemisphere, the results
from left side should not be directly interpreted as dominant side
effects and is only to represent the intrahemispheric effects in tumor
side. Third although the Chi-square test presented no difference
between the two groups in the present study, the intragroup WHO
grades of glioma varied and therefore showed difference in edema
size, tumor growth rate, and different extent of infiltration. Its impact
on the function-specific connectome needs further studies. Fourth,
further tests analyzing the neurocognitive functionality of patients
pre- and postoperatively were not conducted in the current study.
We will add tests on cognitive function in the subsequent data collec-
tion for a more comprehensive analysis in order to improve the under-
standing of correlations to language function. Additionally, as
described in the methods part of the manuscript, we finally rated
aphasia levels of patients according to the final rating of aphasia in
the AAT. We did not perform the entire AAT in the whole period of
patient enrolment. With respect to data consistency, we used data
available for all patients (according to AAT subtests “spontaneous
speech”and “naming”; Table S4). Further analyses should also contain
data on a more comprehensive pre- and postoperative evaluation of
language function as well as a differentiation of preoperative risk fac-
tors for different levels of SRA.
5|CONCLUSION
Preoperative connectome analysis can perform risk assessments pre-
dicting the development of SRA even prior to surgery. The current
study provides a new perspective of function-specific connectome
analysis to investigate language function in neurooncological patients.
Connectome properties are a potential indicator for predicting SRA
and their comprehensive combination using machine learning models
predicts SRA with a sensitivity of 78.3%.
ACKNOWLEDGMENT
Open Access funding enabled and organized by Projekt DEAL.
FUNDING INFORMATION
This research did not receive any specific grant from funding agencies
in the public, commercial, or not-for-profit sectors. This study was
funded entirely by institutional grants from the department.
CONFLICT OF INTEREST
The authors declare the existence of a financial/nonfinancial competing
interest in the cover letter. All authors declare no other relationships or
activities that could appear to have influenced the submitted work.
DATA AVAILABILITY STATEMENT
Data and processing scripts involved in this study are available upon
reasonable request.
ORCID
Sebastian Ille https://orcid.org/0000-0003-2065-6464
Sandro M. Krieg https://orcid.org/0000-0003-4050-1531
REFERENCES
Ardila, A., Bernal, B., & Rosselli, M. (2016). How localized are language
brain areas? A review of Brodmann areas involvement in oral language.
Archives of Clinical Neuropsychology,31(1), 112–122. https://doi.org/
10.1093/arclin/acv081
Biniek, R., Huber, W., Glindemann, R., Willmes, K., & Klumm, H. (1992).
The Aachen aphasia bedside test—Criteria for validity of psychologic
tests. Nervenarzt,63(8), 473–479.
5418 ILLE ET AL.
Chang, E. F., Raygor, K. P., & Berger, M. S. (2015). Contemporary model of
language organization: An overview for neurosurgeons. Journal of Neu-
rosurgery,122(2), 250–261. https://doi.org/10.3171/2014.10.
JNS132647
Darquie, A., Poline, J. B., Poupon, C., Saint-Jalmes, H., & Le Bihan, D.
(2001). Transient decrease in water diffusion observed in human
occipital cortex during visual stimulation. Proceedings of the National
Academy of Sciences of the United States of America,98(16), 9391–
9395. https://doi.org/10.1073/pnas.151125698
Dehaene, S., Dupoux, E., Mehler, J., Cohen, L., Paulesu, E., Perani, D., van de
Moortele,P.,Lehéricy,S.,&leBihan,D.(1997).Anatomicalvariabilityin
the cortical representation of first and second language. Neuroreport,
8(17), 3809–3815. https://doi.org/10.1097/00001756-199712010-
00030
Ille, S., Engel, L., Kelm, A., Meyer, B., & Krieg, S. M. (2018). Language-
eloquent white matter pathway tractography and the course of lan-
guage function in glioma patients. Frontiers in Oncology,8, 572.
https://doi.org/10.3389/fonc.2018.00572
Ille, S., & Krieg, S. M. (2021). Functional mapping for glioma surgery, part
1: Preoperative mapping tools. Neurosurgery Clinics of North America,
32(1), 65–74. https://doi.org/10.1016/j.nec.2020.08.004
Ille, S., Kulchytska, N., Sollmann, N., Wittig, R., Beurskens, E.,
Butenschoen, V. M., Ringel, F., Vajkoczy, P., Meyer, B., Picht, T., &
Krieg, S. M. (2016). Hemispheric language dominance measured by
repetitive navigated transcranial magnetic stimulation and postopera-
tive course of language function in brain tumor patients. Neuropsycho-
logia,91,50–60. https://doi.org/10.1016/j.neuropsychologia.2016.
07.025
Ille, S., Sollmann, N., Hauck, T., Maurer, S., Tanigawa, N., Obermueller, T.,
Negwer, C., Droese, D., Boeckh-Behrens, T., Meyer, B., Ringel, F., &
Krieg, S. M. (2015a). Impairment of preoperative language mapping by
lesion location: A functional magnetic resonance imaging, navigated
transcranial magnetic stimulation, and direct cortical stimulation study.
Journal of Neurosurgery,123(2), 314–324. https://doi.org/10.3171/
2014.10.JNS141582
Ille, S., Sollmann, N., Hauck, T., Maurer, S., Tanigawa, N.,
Obermueller, T., Negwer, C., Droese, D., Zimmer, C., Meyer, B.,
Ringel, F., & Krieg, S. M. (2015b). Combined noninvasive language
mapping by navigated transcranial magnetic stimulation and func-
tional MRI and its comparison with direct cortical stimulation. Jour-
nal of Neurosurgery,123(1), 212–225. https://doi.org/10.3171/
2014.9.JNS14929
Isensee, F., Schell, M., Pflueger, I., Brugnara, G., Bonekamp, D.,
Neuberger, U., Wick, A., Schlemmer, H. P., Heiland, S., Wick, W.,
Bendszus, M., Maier-Hein, K. H., & Kickingereder, P. (2019). Auto-
mated brain extraction of multisequence MRI using artificial neural
networks. Human Brain Mapping,40(17), 4952–4964. https://doi.org/
10.1002/hbm.24750
Latora, V., & Marchiori, M. (2001). Efficient behavior of small-world net-
works. Physical Review Letters,87(19), 198701. https://doi.org/10.
1103/PhysRevLett.87.198701
Lee, B., Park, J. Y., Jung, W. H., Kim, H. S., Oh, J. S., Choi, C. H., Jang, J. H.,
Kang, D. H., & Kwon, J. S. (2010). White matter neuroplastic changes
in long-term trained players of the game of “Baduk”(GO): A voxel-
based diffusion-tensor imaging study. NeuroImage,52(1), 9–19.
https://doi.org/10.1016/j.neuroimage.2010.04.014
Mbwana, J., Berl, M. M., Ritzl, E. K., Rosenberger, L., Mayo, J.,
Weinstein, S., Conry, J. A., Pearl, P. L., Shamim, S., Moore, E. N.,
Sato, S., Vezina, L. G., Theodore, W. H., & Gaillard, W. D. (2009). Limi-
tations to plasticity of language network reorganization in localization
related epilepsy. Brain,132(Pt 2), 347–356. https://doi.org/10.1093/
brain/awn329
Negwer, C., Beurskens, E., Sollmann, N., Maurer, S., Ille, S., Giglhuber, K.,
Kirschke, J. S., Ringel, F., Meyer, B., & Krieg, S. M. (2018). Loss of sub-
cortical language pathways correlates with surgery-related aphasia in
patients with brain tumor: An investigation via repetitive navigated
transcranial magnetic stimulation-based diffusion tensor imaging fiber
tracking. World Neurosurgery,111, e806–e818. https://doi.org/10.
1016/j.wneu.2017.12.163
Oldfield, R. C. (1971). The assessment and analysis of handedness: The
Edinburgh inventory. Neuropsychologia,9(1), 97–113.
Picht, T., Krieg, S. M., Sollmann, N., Rösler, J., Niraula, B., Neuvonen, T.,
Savolainen, P., Lioumis, P., Mäkelä, J. P., Deletis, V., Meyer, B.,
Vajkoczy, P., & Ringel, F. (2013). A comparison of language mapping
by preoperative navigated transcranial magnetic stimulation and direct
cortical stimulation during awake surgery. Neurosurgery,72(5), 808–
819. https://doi.org/10.1227/NEU.0b013e3182889e01
Ries, S. K., Dronkers, N. F., & Knight, R. T. (2016). Choosing words: Left
hemisphere, right hemisphere, or both? Perspective on the lateraliza-
tion of word retrieval. Annals of the New York Academy of Sciences,
1369(1), 111–131. https://doi.org/10.1111/nyas.12993
Rutter, L., Nadar, S. R., Holroyd, T., Carver, F. W., Apud, J.,
Weinberger, D. R., & Coppola, R. (2013). Graph theoretical analysis of
resting magnetoencephalographic functional connectivity networks.
Frontiers in Computational Neuroscience,7, 93. https://doi.org/10.
3389/fncom.2013.00093
Sagi, Y., Tavor, I., Hofstetter, S., Tzur-Moryosef, S., Blumenfeld-
Katzir, T., & Assaf, Y. (2012). Learning in the fast lane: New insights
into neuroplasticity. Neuron,73(6), 1195–1203. https://doi.org/10.
1016/j.neuron.2012.01.025
Saur, D., Kreher, B. W., Schnell, S., Kümmerer, D., Kellmeyer, P., Vry, M. S.,
Umarova, R., Musso, M., Glauche, V., Abel, S., Huber, W., Rijntjes, M.,
Hennig, J., & Weiller, C. (2008). Ventral and dorsal pathways for lan-
guage. Proceedings of the National Academy of Sciences of the
United States of America,105(46), 18035–18040. https://doi.org/10.
1073/pnas.0805234105
Schuppert, M., Munte, T. F., Wieringa, B. M., & Altenmuller, E. (2000).
Receptive amusia: Evidence for cross-hemispheric neural networks
underlying music processing strategies. Brain,123(Pt 3), 546–559.
https://doi.org/10.1093/brain/123.3.546
Shafto, M. A., & Tyler, L. K. (2014). Language in the aging brain: The net-
work dynamics of cognitive decline and preservation. Science,
346(6209), 583–587. https://doi.org/10.1126/science.1254404
Shaywitz, B. A., Shaywitz, S. E., Pugh, K. R., Constable, R. T., Skudlarski, P.,
Fulbright, R. K., Bronen, R. A., Fletcher, J. M., Shankweiler, D. P.,
Katz, L., & Gore, J. C. (1995). Sex differences in the functional organi-
zation of the brain for language. Nature,373(6515), 607–609. https://
doi.org/10.1038/373607a0
Sollmann, N., Fratini, A., Zhang, H., Zimmer, C., Meyer, B., & Krieg, S. M.
(2019). Associations between clinical outcome and tractography based
on navigated transcranial magnetic stimulation in patients with
language-eloquent brain lesions. Journal of Neurosurgery,1–10, 1033–
1042. https://doi.org/10.3171/2018.12.JNS182988
Sollmann, N., Negwer, C., Tussis, L., Hauck, T., Ille, S., Maurer, S.,
Giglhuber, K., Bauer, J. S., Ringel, F., Meyer, B., & Krieg, S. M. (2017).
Interhemispheric connectivity revealed by diffusion tensor imaging
fiber tracking derived from navigated transcranial magnetic stimulation
maps as a sign of language function at risk in patients with brain
tumors. Journal of Neurosurgery,126(1), 222–233. https://doi.org/10.
3171/2016.1.JNS152053
Sollmann, N., Tanigawa, N., Ringel, F., Zimmer, C., Meyer, B., & Krieg, S. M.
(2014). Language and its right-hemispheric distribution in healthy
brains: An investigation by repetitive transcranial magnetic stimulation.
NeuroImage,102(Pt 2), 776–788. https://doi.org/10.1016/j.
neuroimage.2014.09.002
Sollmann, N., Zhang, H., Fratini, A., Wildschuetz, N., Ille, S., Schröder, A.,
Zimmer, C., Meyer, B., & Krieg, S. M. (2020). Risk assessment by pre-
surgical tractography using navigated TMS maps in patients with
highly motor- or language-eloquent brain tumors. Cancers (Basel),
12(5), 1264. https://doi.org/10.3390/cancers12051264
ILLE ET AL.5419
Song, A. W., Chang, H. C., Petty, C., Guidon, A., & Chen, N. K. (2014).
Improved delineation of short cortical association fibers and
gray/white matter boundary using whole-brain thr ee-dimensional
diffusion tensor imaging at submillimeter spatial resolution. Brain
Connectivity,4(9), 636–640. https://doi.org/10.1089/brain.2014.
0270
Tarapore, P. E., Tate, M. C., Findlay, A. M., Honma, S. M., Mizuiri, D.,
Berger, M. S., & Nagarajan, S. S. (2012). Preoperative multimodal
motor mapping: A comparison of magnetoencephalography imaging,
navigated transcranial magnetic stimulation, and direct cortical stimu-
lation. Journal of Neurosurgery,117(2), 354–362. https://doi.org/10.
3171/2012.5.JNS112124
Tzourio-Mazoyer, N., Josse, G., Crivello, F., & Mazoyer, B. (2004). Interin-
dividual variability in the hemispheric organization for speech. Neuro-
Image,21(1), 422–435. https://doi.org/10.1016/j.neuroimage.2003.
08.032
Voets, N. L., Parker Jones, O., Mars, R. B., Adcock, J. E., Stacey, R.,
Apostolopoulos, V., & Plaha, P. (2019). Characterising neural plastic-
ity at the single patient level using connectivity fingerprints. Neuro-
Image: Clinical,24, 101952. https://doi.org/10.1016/j.nicl.2019.
101952
Wang, J., Zuo, X., & He, Y. (2010). Graph-based network analysis of
resting-state functional MRI. Frontiers in Systems Neuroscience,4, 16.
https://doi.org/10.3389/fnsys.2010.00016
Xia, M., Wang, J., & He, Y. (2013). BrainNet viewer: A network visualiza-
tion tool for human brain connectomics. PLoS One,8(7), e68910.
https://doi.org/10.1371/journal.pone.0068910
Zhang, N., Li, K., Li, G., Nataraj, R., & Wei, N. (2021). Multiplex recurrence
network analysis of inter-muscular coordination during sustained grip
and pinch contractions at different force levels. IEEE Transactions on
Neural Systems and Rehabilitation Engineering,29, 2055–2066. https://
doi.org/10.1109/TNSRE.2021.3117286
SUPPORTING INFORMATION
Additional supporting information can be found online in the Support-
ing Information section at the end of this article.
How to cite this article: Ille, S., Zhang, H., Sogerer, L.,
Schwendner, M., Schöder, A., Meyer, B., Wiestler, B., & Krieg,
S. M. (2022). Preoperative function-specific connectome
analysis predicts surgery-related aphasia after glioma
resection. Human Brain Mapping,43(18), 5408–5420. https://
doi.org/10.1002/hbm.26014
5420 ILLE ET AL.
Content uploaded by Sandro M. Krieg
Author content
All content in this area was uploaded by Sandro M. Krieg on Aug 04, 2022
Content may be subject to copyright.