% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@INPROCEEDINGS{Domhof:893375,
author = {Domhof, Justin and Jung, Kyesam and Eickhoff, Simon and
Popovych, Oleksandr},
title = {{P}arcellation-induced {V}ariation of {E}mpirical and
{S}imulated {F}unctional {B}rain {C}onnectivity},
school = {Heinrich-Heine University Düsseldorf},
reportid = {FZJ-2021-02716},
year = {2021},
abstract = {Recent developments of large-scale whole-brain network
models have demonstrated their great potential when
investigating resting-state brain activity which is believed
to be highly personalised (Bansal 2018, Popovych 2019). So
far, however, it has not been systematically investigated
how alternating derivations of the empirical brain
connectivity from MRI data, which serves as the input to
these models, influence the modelling results. Here, we
study the influence from an element indispensable in such
calculations: the brain parcellation scheme that reduces the
dimensionality of investigated brain networks by grouping
thousands of voxels together into a few hundred brain
regions.MethodWe extracted empirical structural and
functional connectivities from the dwMRI and fMRI data of
200 subjects (96 males, age 28.5 ± 3.5 years) included in
the Human Connectome Project dataset (Van Essen, 2012, 2013)
using 19 different freely available brain parcellations.
Subsequently, for each combination of parcellation and
subject, we constructed two dynamical whole-brain models on
the basis of the empirical structural connectivity. The
first is a model of coupled phase oscillators, and the
second is a network of Wilson-Cowan neural mass models
(Wilson, 1972). Both models were used to simulate the
resting-state functional connectivity, which was correlated
to its empirical counterpart to obtain an index
characterising their similarity. By varying the free
parameters included in both models, we obtained the
maximised similarity index or goodness-of-fit for every pair
of parcellation and subject. To find an explanation for any
differences observed in the goodness-of-fit for varying
parcellations, we derived 9 graph-theoretical network
statistics characterising the degree and closeness
centrality distribution, the modularity and the global
efficiency from both the structural and functional
connectomes (Rubinov, 2010). Finally, through the use of
principal component analysis combined with linear
regressions, we related the goodness-of-fits to the
extracted networks statistics on both the level of the group
and individual subjects.ResultsOur study revealed large
deviations in the goodness-of-fit across parcellations (Fig.
1A). By regressing the group-averages corresponding to both
models, we showed that this heterogeneity does not depend on
the considered models. Next, we conducted a principal
component analysis of the parcellation-based group-averaged
network statistics and structure-function relationship of
the empirical connectomes. The scores of the first principal
component were regressed with the group-averaged
goodness-of-fits, which explained approximately $75\%$ and
$86\%$ of the parcellation-induced variation in the
goodness-of-fits of the phase oscillator and neural mass
model, respectively (Fig. 1B). Including more PCs in the
regression model only marginally increased this proportion
(Fig. 1C). However, applying the same approach to the
interindividual variation in the goodness-of-fit for every
parcellation in isolation resulted in low explanatory power
and its high variability across parcellations and models
(Fig. 2A). Finally, a multivariate linear model was used to
regress the network statistics directly with the
goodness-of-fits of the individuals. This change in strategy
led to an increase in explained variance, which was
nevertheless still smaller than $30\%$ and highly variable
across parcellations (Fig. 2B).ConclusionOur results
demonstrated that the goodness-of-fit of the model to
empirical data was influenced greatly by the parcellation
both at the level of the entire cohort and the individual
subjects. We furthermore provided evidence that
graph-theoretical network properties derived from the
empirical data can explain group-averaged but not
interindividual variations in this respect. These findings
may contribute to mechanisms explaining how the validation
of the whole-brain models works, and how the fitting results
depend on the choice of parcellation.},
month = {Jun},
date = {2021-06-21},
organization = {The 27th Annual Meeting of the
Organization for Human Brain Mapping,
Virtual (Virtual), 21 Jun 2021 - 25 Jun
2021},
subtyp = {After Call},
cin = {INM-7},
cid = {I:(DE-Juel1)INM-7-20090406},
pnm = {5231 - Neuroscientific Foundations (POF4-523) / HBP SGA2 -
Human Brain Project Specific Grant Agreement 2 (785907) /
HBP SGA3 - Human Brain Project Specific Grant Agreement 3
(945539)},
pid = {G:(DE-HGF)POF4-5231 / G:(EU-Grant)785907 /
G:(EU-Grant)945539},
typ = {PUB:(DE-HGF)24},
url = {https://juser.fz-juelich.de/record/893375},
}