Synopsis
A great challenge in neuroscience is the understanding of
the dynamic manner in which brain regions interact with one another in both
task-based and resting-state brain imaging studies. In this work, we introduce
a novel statistical method, called Network Change Point Detection (NCPD), which
dynamically clusters brain regions by their functional connectivity. NCPD promises
to offer deeper insight into the large-scale characterizations and mechanisms
of the brain as it can be used for the dynamic modelling of a very large number
of voxels or brain regions. We apply this new method to a resting-state fMRI
study.PURPOSE
Recently, in
functional magnetic resonance imaging (fMRI), there has been an increased
interest in quantifying changes in connectivity between brain regions over an
experimental time course to provide a deeper insight into the fundamental
properties of brain networks. The application of graphical modeling has been
instrumental in these analyses and has enabled the examination of the brain as
an integrated system. In this work, we
propose a new statistical method, called Network Change Point Detection (NCPD),
which is a data-driven technique for detecting changes in the functional connectivity (FC) structure over
time where the location and number of changes are unknown
a priori. The method provides important insights
into the time varying nature of the FC of brain regions while
subjects are at rest. NCPD uses spectral
clustering to study the network structure between brain regions and uses a
non-parametric metric to detect the change in the structures across the time
course. NCPD promises to offer important insights into the inner workings of
the whole brain. We apply NCPD to
simulated data and to a resting-state fMRI data set. The temporal features of
this novel connectivity method will provide a more accurate understanding of
the large-scale characterizations of disorders such as Alzheimer's
disease and may lead to better diagnostic and prognostic indicators.
METHODS
In
this work, the objective is to understand the FC dynamics of
the entire brain, the FCdynamics in the situations where
the number of brain regions exceeds the number of time points in the
experiment. The idea originates from community detection in an undirected
network G = (V,E), where V and E are the collections of vertices and edges,
respectively. In this framework, the community detection problem can be
formulated as finding the true disjoint partition of the community. One
of the most widely used community detection techniques is spectral clustering
1. In
a brain imaging context, detecting FC between brain
regions can be seen as a community detection problem. Here, vertices and
edges represent brain regions and FC, respectively. Moreover, our target
is not only to find the underlying clustering structure, but also to detect the
structural change along the experimental time course. We use a
non-parametric metric, the cosine of the principal angles between the two
partitions, to detect the change in the structures across the time course.
Finally, at each change point, we carry out inference on the cosine
of the principal angles using the stationary bootstrap
2.
We apply NCPD to a resting state fMRI data set, as described in
3.
Participants are instructed to rest in the scanner for 9.5 min, with the
instruction to keep their eyes open for the duration of the scan. The
individual time series data are bandpass-filtered and motion corrected. The
voxel time courses at white-matter and CSF locations are submitted to a
Principal Components Analysis and, together with the motion parameters, we use
all components with an eigenvalue > 1 as independent variables in a
subsequent nuisance regression. Each voxel’s time series is residualized with
respect to those independent variables. The residual time series images are
then smoothed with an isotropic Gaussian kernel (FWHM=6mm). We apply the
Automated Anatomic Atlas Labeling
4 to the adjusted voxel-wise time series
and produce time series for 116 ROIs for each subject.
RESULTS
NCPD detects
evidence of state changes in the resting-state networks for all subjects bar
one (Figures 1 and 2). In addition, NCPD allows us to identify common
functional states across subjects (Figure 3). Hence, NCPD not only allows for estimation of whole
brain characterizations of the functional organization of the brain but it also can be used to characterize differences in
the functional architecture for subjects with/out neurological disorders.
Discussion
NCPD is very
flexible as there is no a priori
assumption on where the changes in the FCoccur. The new
method has a major advantage over moving window-type methods as we do not have
to choose the window length.
NCPD will lead to the robust identification
of cognitive states at rest for both controls and subjects with brain disorders.
It is hoped that the large-scale temporal features resulting from the accurate
description of FC from our novel method will lead to
better diagnostic and prognostic indicators of these disorders. More
specifically, by comparing the change points and the community network
structures of FC of healthy controls to patients with
these disorders, we can understand the key differences in functional brain
processes. In particular, NCPD allows us to find common cognitive states that
recur in time, across subjects, and across groups in a study.
Acknowledgements
Funding for Ivor Cribben was provided
by the Pearson Faculty Fellowship, Alberta School of Business, and a Alberta
Health Services (AHS) Grant.References
1. Von Luxburg, U. A tutorial on spectral clustering.
Statistics and computing. 2007;17:395-416.
2. Politis,
D. N. and Romano, J. P. The stationary bootstrap. Journal of American
Statistical Association. 1994;89:1303-1313.
3.
Habeck, C. G., Steffener, J., Rakitin, B., Stern, Y. Can the default-mode
network be described with one spatial-covariance network? Brain Research.
2012;1468:38-51.
4. Tzourio-Mazoyer, N., Landeau, B.,
Papathanassiou, D., Crivello, Etard, O., Delcroix, N., Mazoyer, Joliot, B.M.
Automated Anatomical Labeling of activations in SPM using a Macroscopic
Anatomical Parcellation of the MNI MRI single-subject brain. NeuroImage.
2002;15:273-289.