DiagnosisofBipolarDisorderbasedonPrincipal - - PowerPoint PPT Presentation

Diagnosis
of
Bipolar
Disorder
based
on
Principal
 Component
Analysis
and
Support
Vector
Machines


• ver
the
MRI
deforma@on
Jacobian


M.
Termenon,
M.
Graña,
A.
Besga,

J.
Echeveste,
J.
M.
Pérez,
 A.
Gonzalez‐Pinto


Grupo
de
Inteligencia
Computacional,
UPV/EHU,
 
Unidad
de
Inves@gación
en
Psiquiatría,
Hospital
San@ago
Apostol,
Vitoria‐ Gasteiz.

 Departamento
de
Resonancia
Magné@ca,
Osatek‐Vitoria.

 Servicio
de
Neurología,
Hospital
San@ago
Apostol,
Vitoria‐Gasteiz.


1
 CORES
2013,
Milkow,
Poland


Contents


• Introduc@on
and
mo@va@on

• Materials

• Methods

• Results

• Conclusions


2
 CORES
2013,
Milkow,
Poland


Introduc@on


• Bipolar
disorder
(BD)



– psychiatric
disorder



• at
least
one
episode
of
mania
or
hypomania


• or
a
mixed
episode

<‐>
a
depressive
episode,


• changes
in
mood
states
and
psycho@c
symptoms.

• It
is
associated
with
cogni@ve,
affec@ve
and
func@onal


impairment.


• A
diagnosis
BD


– symptoms,

 – course
of
illness
and,

 – family
history,

 – neuroimaging



• iden@fied
several
regions
that
are
affected
by
the
disease




3
 CORES
2013,
Milkow,
Poland


Introduc@on


• we
compare
brain
structural
MRI
of
healthy
controls
with


pa@ents
with
bipolar
disorder,



– to
discriminate
between
both
groups

 – selec@ng
relevant
informa@on
embedded
in
the
images.



• Machine
Learning
(ML)



– of
feature
vectors
extracted
from
the
deforma@on
of
the
 structural
MRI
images

 – computer
aided
diagnosis
(CAD)
tools.



• Preprocessing,



– registra@on
of
the
volumes,

 – affine
and
non‐linear
registra@ons
to
a
standard
MNI
template.


• The
Jacobian
of
the
deforma@on
at
each
voxel
will
be
used
to
extract


the
relevant
features



4
 CORES
2013,
Milkow,
Poland


Materials


• Pa@ents
recruited
at
the
psychiatric
unit
at
Alava
University
Hospital,


Vitoria
(Spain)



• All
pa@ents
were
living
independently
in
the
community.



– clinical
evalua@on,

 – a
cogni@ve
and
a
neuropsychological
evalua@on,

 – and
brain
imaging
(MRI).



• Forty
men
and
women
elderly
subjects
were
included
in
the
present
study.



– The
healthy
control
group
included
20
subjects
without
memory
complaints



• (mean
age
74.10
(SD:8.03
years))



– and
BD
group
included
20
subjects
fulfilling
DSM
IV’s
criteria



• (mean
age
70.37
(SD:
9.07
years)).


• Subjects
with
psychiatric
disorders
(i.e.
major
depression)
or
other


condi@ons
(i.e.
brain
tumors)
were
not
considered
for
this
study.


• Structural
MRI
3D
data
(T1‐weighted).



CORES
2013,
Milkow,
Poland
 5


Methods


• Image
preprocessing


– Affine
and
elas@c
registra@on
 – Tensor
deforma@on
map
 – Jacobian
at
each
voxel


• Feature
selec@on


– Voxel
sites
with
dis@nc@ve
values
of
the
deforma@on
 Jacobian
 – Dimension
reduc@on:
PCA


• Classifica@on:


– Linear
SVM
 – Valida@on:
Leave
One
Out



CORES
2013,
Milkow,
Poland
 6


Methods


CORES
2013,
Milkow,
Poland
 7


Methods


• Feature
selec@on


– Compute
voxel‐wise
means
of
each
class



• healthy
controls

• BD
pa@ents


– Compute
histogram
of
class
means
differences
 – Select
the
tail
percen@le
according
to
a
threshold



CORES
2013,
Milkow,
Poland
 8


Methods


• Classifica@on
performance
measures


CORES
2013,
Milkow,
Poland
 9


Results


CORES
2013,
Milkow,
Poland
 10


Voxel
loca@ons
for
feature
selec@on
with
threshold
0.4



Results


CORES
2013,
Milkow,
Poland
 11


3
first
PCA
before
(lel)
and
aler
(right)
feature
selec@on


Results


CORES
2013,
Milkow,
Poland
 12


Results


CORES
2013,
Milkow,
Poland
 13


Conclusions


• Free
of
circularity


– Feature
selec@on
is
performed
in
each
LOO
step
 – PCA
is
unsupervised


• Selected
voxels
are
located
in



– thalamus
and
angular
gyrus,



• precuneous
cortex,
precentral
and
postcentral
gyrus,
supramarginal
gyrus,


rightlateral
ventricle,
superior
parietal
lobe,
inferior
temporal
gyrus
and
 cerebellum.


• Thalamus
is
one
of
the
most
relevant
biomarkers
in
bipolar
disorder


• Also
superior
parietal
lobe
,
precuneous
cortex,
precentral
gyrus
and


Cerebellum.



• Valida@on
of
the
approach



– with
no
disease
a
priori
informa@on
we
find
brain
discriminant
regions
 consistent
with
known
biomarkers
of
BD



CORES
2013,
Milkow,
Poland
 14