Centraliens - N°632 - La revue des Arts et Manufactures - Mars/Avril 2014 - 40

Dossier

Machine Learning

for Neurological Disorders
The last two decades have seen tremendous advances in our understanding of human brain
structure and function, particularly at the level of systems neuroscience, where neuroimaging
methods have led to better delineation of brain networks and brain modules.
Brain understanding is one of the greatest challenges of our century with enormous
potential impact in a number of fields, including medicine.

ecent progress in the hardware
side has made possible the in-vivo
acquisition on top of structural/
anatomical data, functional information (through emerging image modalities like functional magnetic resnonance
imaging (fMRI), diffusion tensor imaging
(DTI), magneto encephalogram (MEG),
electro encephalogram (EEG), etc.) in a
non-invasive manner depicting task-specific states of the brain. Such information can
be of great interest towards understanding
of neurogenerative diseases, and providing
means of assessing the impact of different
therapeutic strategies.

Figure 1
Caption: A visualization of the areas of the brain identified as being discriminative between cocaine addicted
and control subjects. The rostral anterior cingulate cortex is prominent in the selected areas implicating a role
in cocaine addiction (Gkirtzou et al., ISBI 2013).

These data are by necessity high dimensional and complex, driving the widespread
application of machine learning techniques
for their analysis. Machine learning in this
context works by learning a regressor from
the space of brain recordings (e.g. voxels of a
fMRI scan) to an output variable describing
a condition to be explained. The form of the
regressor is important as it encodes the class
of interrelationships within the brain that
can be learned, e.g. linear relationships, or
non-linear interactions encoded in a graph
structure. Equally important is the choice of
output variable, e.g. the presence or absence
of a neural disorder, or the identity of a
stimulus or task undertaken by the subject
during recording. Through the form of the
regression task, a large number of neuroscientific questions may be elucidated, e.g.
the neural correlates of visual processing,
emotion, addiction, Alzheimer's, or autism.
In the Center for Visual Computing, our
research in this area has been guided by
the translation of neuroimaging problems
into rich mathematical representations
such as graphs for use in machine learning

algorithms. By employing novel machine
learning algorithms and representations,
we expand our ability to understand brain
function in a way that is reproducable, interpretable, and statistically sound.
A key area of neuroscience that we have
been working on is the understanding of the
neural mechanisms of cocaine addiction.
In collaboration with neuroscientists at the
Icahn School of Medicine in New York and
computer scientists at Stony Brook University, we have employed recordings of cocaine addicted and control subjects to find
regions of the brain that discriminate the
two groups based on functional activations
in a task carefully designed to differentially
invoke associations with word meanings.
The work has been successful not only in
showing the applicability of novel machine
learning techniques, but in its confirmation
of the deactivation of the rostral anterior
cingulate cortex (rostral ACC) in cocaine
addicted vs. control groups. This is an important result as it provides evidence that
certain categories of theraputic interventions previously applied in other addictions
may be effective in cocaine addiction as
well. As a result of this research, Katerina
Gkirtzou successfully defended her doctoral
dissertation entitled "Sparsity regularization
and graph-based representation in medical
imaging" in December, 2013.
A focus of research in our group has been
on the development of methodologies for
neuroscientific studies, rather than simply
providing an answer to a specific question.
We are therefore addressing applications
to several neurological disorders, including
Alzheimer's disease and autism. Our work
is greatly assisted through collaboration and
integration with multiple partners on the
Centraliens no632 [mars/avril 201
4]

Centraliens - N°632 - La revue des Arts et Manufactures - Mars/Avril 2014

Table des matières de la publication Centraliens - N°632 - La revue des Arts et Manufactures - Mars/Avril 2014