IDENTIFICATION OF A LARGE-SCALE FUNCTIONAL NETWORK IN FUNCTIONAL
MAGNETIC RESONANCE IMAGING
Pierre Bellec
1,2
, Guillaume Marrelec
1,2
, Vincent Perlbarg
1,2
, Saˆ ad Jbabdi
1,2
,
Odile Jolivet
1,2
, M´ elanie P´ el´ egrini-Issac
1,2
, Julien Doyon
1,3
, Habib Benali
1,2
1
INSERM U494, Paris, France
2
IFR 49, Orsay, France
3
Universit´ e de Montr´ eal, Montr´ eal, Canada
ABSTRACT
In functional magnetic resonance imaging (fMRI), cerebral
activity has been increasingly considered as the consequence
of a network activation. Selecting the brain regions relevant
for the network has thus become a key issue. We propose to
define the so-called large-scale functional network involved
in a particular task as a set of regions exhibiting strong in-
trinsic homogeneity, as well as at least one strong long-
distance inter-regional interaction. We develop a method to
identify such a network, and we validate it on a real dataset,
in a context where the existence of a distributed network
has already been demonstrated. Our results are compatible
with previous studies. This new tool is thus promising for
selecting regions when analyzing functional connectivity in
fMRI.
1. INTRODUCTION
Functional magnetic resonance imaging (fMRI) is a recent
technique for studying non-invasively the hemodynamic changes
in the working human brain. In the past decade, cognitive
processes have been increasingly considered as the conse-
quence of network activation, and estimating inter-regional
relationships has therefore become a key issue [1]. most
techniques addressing this issue make use of the functional
connectivity between brain regions [2], which is the corre-
lation between the mean time courses of these regions.
The regions of interest are usually selected on the ba-
sis of either anatomic information or functional activation
maps [3]. However, such approach is not fully satisfactory
for at least three reasons: time courses of the voxels belong-
ing to a same region can be highly heterogeneous [4]; there
is sometimes neither strong anatomic a priori at hand, nor an
explicit paradigm, for example for resting-state datasets [5];
even worse, some regions can be used as fundamental infor-
mation relays or modulate activity of other regions, with-
out following the activation pattern, and wrongly excluded
from the network. Thus, the question of what the functional
network of brain regions involved in a task is, and how it
can be identified remains open in many cases. Recent pa-
pers, e.g. [6], highlight the importance of homogeneity, i.e.
voxels belonging to the same region must exhibit similar
activity patterns. This could be sufficient to define a net-
work, but only those regions which are correlated with each
other are likely to be selected for a connectivity study, even
though correlation of two brain regions could be merely due
to their spatial proximity. Our attention is therefore focused
on regions that exhibit a strong functional connectivity with
a distant region of the brain. As it is not clear how ”distant”
is ”distant” enough, we will set this notion a priori and re-
fer to it as large-scale. We can then define the large-scale
functional network associated with a dataset as the set of re-
gions exhibiting strong intrinsic homogeneity together with
at least one strong large-scale correlation.
In this paper, we first introduce the background of appli-
cation of our method to a real dataset, in order to exemplify
the concept of large-scale functional network. Assuming
knowledge of what is large-scale, we propose a two-step
method to identify the large-scale functional network in a
fMRI dataset (section 3). First, we develop a region growing
algorithm, using the concept of mutual nearest neighbour
[7] to find some regions of the brain whose voxels have ho-
mogeneous time courses (section 3.1). Then, we seek which
of these regions exhibit strong large-scale interactions (sec-
tion 3.2). Finally, in section 4, we apply this new tool to the
real dataset.
2. THE ISSUE: A MOTOR LEARNING STUDY
Psychophysical studies have demonstrated that motor learn-
ing follows two distinct stages: first, an early, fast stage in
which considerable improvement can be seen within a sin-
gle training session; and second, a later, slow stage in which
further gains can be observed across several sessions (and
848 0-7803-8388-5/04/$20.00 ©2004 IEEE