Indonesi
an
Journa
l
of El
ect
ri
cal
Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
1
4
,
No.
1
,
A
pr
il
201
9
, p
p.
478
~
489
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
4
.i
1
.pp
478
-
489
478
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Hybrid
enh
ance
d
ICA &
KSVM b
ased br
ain
tumor im
age
segment
atio
n
Thri
vikra
m
B
at
hini
,
Ba
sw
ar
aj G
adgay
Depa
rtment
o
f
E
le
c
troni
cs
&
Co
m
m
unic
at
ion
En
gine
er
ing
,
Visve
svara
y
a
Technol
ogic
a
l
Univer
si
t
y
,
Indi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
un
1
7
, 2
018
Re
vised
O
ct
30
, 2
018
Accepte
d
D
ec
12
, 201
8
Medic
a
l
image
proc
essing
is
an
important
aspe
c
t
in
di
agnosis
an
d
tre
a
tmen
t
strat
eg
y
.
Th
e
tr
e
m
endous
volum
e
of
m
edi
c
al
d
ata
has
a
ccel
er
ate
d
the
n
ee
d
for
aut
om
at
ed
ana
l
y
sis
of
thi
s
image,
m
ore
so
in
the
ca
se
Magne
tic
Resonanc
e
Im
a
ging
(MRI).
An
improved
K
-
m
ea
ns
al
gorit
h
m
and
EM
al
gorit
hm
have
bee
n
combined
i
n
the
proposed
a
pproa
ch
to
prod
uce
a
h
y
brid
strat
eg
y
for
be
t
te
r
cl
uster
ing
a
nd
segm
ent
atio
n
using
Enh
anced
ICA.
A
cl
assifi
er
for
bas
ed
on
Support
Vec
tor
Ma
chi
ne
(
SV
M)
has
bee
n
form
ula
te
d
and
emplo
y
ed
for
the
cl
assifi
cat
ion
of
bra
in
tumors
in
Magne
ti
c
Resonanc
e
Im
age
s
(MRI).
The
proposed
SV
M
cl
assifie
r
u
sed
a
k
ern
e
l
in
the
form
of
Gauss
ia
n
ra
dia
l
basis
func
ti
on
k
ern
el
(GRB
ker
n
el
)
to
improve
t
he
cl
assifi
er
per
form
anc
e
.
Th
e
per
form
an
ce
o
f
the
class
ifi
er
has
bee
n
validated
throu
g
h
expe
rt
clinical
o
pini
on
and
c
al
cu
la
ti
on
of
p
erf
or
m
anc
e
m
ea
sures
.
Th
e
re
sul
ts
ampl
y
illus
tra
t
e t
he
suit
abi
l
ity
o
f
t
he
proposed
class
ifi
er.
Ke
yw
or
d
s
:
EICA
EM
GRB
Kernel
MRI
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Thr
i
vikram
Bath
ini
,
Dep
a
rtm
ent o
f
Ele
ct
ro
nics
&
Com
m
un
ic
at
io
n
E
ngineeri
ng
,
Visv
es
va
raya
Tech
no
l
og
ic
al
Un
i
ver
sit
y
,
Be
la
gav
i,
Karn
at
aka
,
India
590018.
Em
a
il
: t
hr
ivikram
bathini@
gm
ai
l.co
m
1.
INTROD
U
CTION
The
num
ber
of
m
edical
i
m
agi
ng
te
ch
niques
that
al
ign
with
com
pu
te
r
and
segm
entat
ion
al
gorithm
s
hav
e
inc
rease
d.
T
hese
te
ch
ni
qu
e
s
are
exam
ined
a
nd
valid
at
ed
by
resea
r
cher
s
th
rou
gh
the
ye
ars.
A
na
ly
t
i
c
m
et
ho
dolo
gies
[1
]
ha
ve
been
app
li
ed
f
or
cl
inica
l
analy
sis
l
ike
cance
r
tum
or
sta
ging.
Hum
an
interp
retat
ion
of
i
m
ages
us
e
d
to
gr
a
de
t
he
tum
or
s
m
ay
var
y
with
eac
h
rea
de
r
as
it
de
pe
nds
on
the
visu
al
featu
res
of
le
s
ion
s
.
Hen
ce
,
these
appr
oach
es
pl
ay
an
i
m
po
rta
nt
ro
le
i
n
ide
ntifyi
ng
a
nd
gr
a
ding
the
t
um
or
.
The
refo
re,
a
n
autom
at
ed
i
m
a
ge
a
naly
ti
c
process
is
us
e
d
t
o
cl
assify
t
he
br
ai
n
tum
or
s
wh
ic
h
a
re
ca
pa
ble
of
quantit
at
ively
assist
ing
in
a
bette
r
ob
j
ect
iv
e
diag
nosis.
S
ince
a
tum
or
do
e
s
not
ha
ve
a
pr
e
-
de
fine
d
char
act
erist
ic
,
it
is
i
m
po
rtant
that
these
le
sions
be
diff
e
re
ntia
te
d
into
a
norm
al
and
ca
nce
rous
tum
or
,
accurat
el
y.
Thu
s
,
the
br
ai
n
tum
or
i
m
age
analy
sis
is
a
chall
eng
i
ng
ta
sk
.
T
he
proce
ss
us
ed
for
im
age
analy
sis
to
diagnose
cancer
is
diff
e
re
nt fro
m
the r
ece
nt
wor
ks
on the
g
e
net
ic
an
al
ysi
s of tissue sam
ples [2]
.
The
m
or
ph
m
et
ric
m
e
tho
ds
use
m
e
tho
ds
tha
t
find
the
c
orre
la
ti
on
s
be
twee
n
br
ai
n
sha
pe
and
diseas
e
sever
it
y,
by
st
at
ist
ic
ally
identify
ing
an
d
c
ha
racteri
zi
ng
str
uctu
ral
dif
fer
e
nces
am
on
g
pe
op
le
.
O
wing
to
the
i
m
pr
ovem
ent
in
res
olu
ti
on
of
a
natom
ic
al
hum
an
br
ai
n
sca
ns
a
nd
i
m
age
processi
ng
te
ch
niques,
m
any
appr
oach
es
to
char
act
e
rize
the
diff
ere
nces
i
n
sh
a
pe
an
d
ne
uro
-
a
natom
ic
al
con
fig
urat
io
n
of
diff
e
ren
t
br
ai
ns
hav
e
em
erg
ed
in
the
rece
nt
tim
es.
The
m
or
ph
m
et
ric
analy
sis
of
brai
n’s
m
agn
et
ic
resonan
ce
im
ages
(MRI)
hav
e
gaine
d
popula
rity
to
exam
ine
the
neu
r
oa
natom
ic
al
correla
te
s
of
a
no
rm
al
br
a
in
de
velo
pm
e
nt
an
d
neur
ologica
l
diso
r
de
rs.
Si
nce,
MR
I
is
the
best
m
edical
i
m
ag
ing
te
ch
no
l
og
y
fo
r
diag
no
si
ng
patho
l
og
ic
al
brai
ns
by
pro
vid
in
g
detai
ls
MR
i
mages
f
ro
m
different
m
od
el
it
i
es.
These
MR
i
m
ages
can
be
us
ed
t
o
stu
dy
an
d
com
par
e
brai
n
tum
or
s
on
pre
op
e
rati
ve
a
nd
po
st
operati
ve
,
in
ad
diti
on
to
be
us
e
d
to
det
erm
ine
the
res
ect
ion
extent [
3].
The
b
rain
s
hap
e of
d
iffer
e
nt p
at
ie
nts su
f
fer
i
ng
from
sch
iz
op
hre
nia, au
ti
sm
,
alz
heimer
,
dysle
x
ia
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Hyb
ri
d
e
nhanc
ed ICA
&
KSVM base
d br
ain
tumor im
age s
egm
e
nta
ti
on
(
T
hr
iv
ik
ram B
athi
ni
)
479
Turne
r'
s
syndr
om
e
has
bee
n
stud
ie
d
by
res
earche
rs.
Segm
entat
ion
is
c
omm
on
ly
req
ui
red
as
a
prel
im
inary
sta
ge
w
hile
an
al
ysi
ng
the
m
e
dical
i
m
ages
f
or
c
om
pu
te
r
-
ai
ded
diag
nosis
and
t
her
a
py.
H
ow
e
ve
r,
the
i
ntrin
sic
natu
re
of
the
i
m
ages
m
akes
m
edical
i
m
age
segm
entat
ion
,
a
com
plex
and
c
halle
ngin
g
ta
sk
.
T
he
preci
se
segm
entat
ion
of
the
brai
n
i
s
im
po
rtant
due
to
it
s
com
plica
te
d
str
uct
ur
e
.
S
om
e
of
the
pr
om
inent
i
m
age
processi
ng
te
c
hn
i
qu
e
s
that
ha
ve
been
pro
pose
d
f
or
t
he
s
egm
entat
ion
of
brai
n
MR
I
a
re
th
reshold
,
r
e
gion
-
grow
i
ng
an
d
cl
us
te
rin
g.
T
he
di
stribu
ti
on
of
ti
ssu
e
inte
ns
it
ie
s
in
br
ai
n
im
ag
es
seem
s
ver
y
com
plex
that
create
diff
ic
ulti
es
in
t
hr
es
hold
deter
m
inati
on
.
T
his
m
akes
the
th
r
esh
old
m
et
ho
ds
restrict
ive
an
d
the
refo
re,
it
has
t
o
be
com
bin
ed w
it
h
oth
e
r
m
et
ho
ds
. T
he regi
on
grow
i
ng tech
niq
ue
, is an
e
xte
ns
io
n of
the t
hresh
old
m
et
ho
d, that
is,
it
is
a
c
om
bin
at
ion
of
t
hr
es
ho
l
d
m
et
ho
d
a
nd
c
onnecti
vity
co
nd
it
io
ns
or
re
gion
hom
og
eneit
y
crit
eria.
Here
,
the
anat
om
ic
al
inf
or
m
at
ion
s
hould
be
preci
se
so
a
s
to
su
c
cessf
ully
l
ocate
the
sin
gle/m
ulti
ple
seed
pi
xels
f
or
each
reg
i
on
a
nd
c
om
bin
e
it
with
th
ei
r
ass
oc
ia
te
d
hom
og
e
neity
.
A
nother
popula
r
m
et
ho
d
use
d
f
or
m
edical
i
m
age
segm
entat
ion
is
cl
ust
er
ing
.
S
om
e
ty
pi
cal
cl
us
te
rin
g
m
et
ho
ds
i
nclu
de
f
uzzy
c
-
m
e
ans
(F
CM
cl
ust
ering
and ex
pe
ct
at
io
n
m
axi
m
iz
ation
(EM)
al
gorithm
s.
In
t
he
pr
opose
d
al
gorithm
,
i
m
pr
oved
K
-
m
eans
al
gorithm
is
co
m
bin
ed
wit
h
EM
al
gorithm
,
to
pro
du
ce
a
hybri
d
strat
egy
f
or
an
e
nh
a
nce
d
cl
us
te
rin
g.
T
his
al
gorithm
aim
s
to
us
e
the
wel
l
distrib
uted
cl
us
te
r
that
is
der
ive
d
from
the
K
-
m
e
ans
al
ong
with
the
com
pactness
of
cl
ust
ers
that
is
pr
ovide
d
by
EM.
T
he
init
ia
l
cl
us
te
rs
are
pr
ov
i
ded
by
i
m
pro
ved
K
-
m
ea
ns
al
gorithm
.
These
cl
us
te
r
s
pr
od
uce
cent
ers
w
hich
are
widely
sp
rea
d
in the given
data. T
he
centers ar
e the
n
us
e
d
as init
ial v
ari
able for
E
M t
o
find
the local
m
axi
m
a
t
hro
ug
h
diff
e
re
nt
it
erati
on
s
.
T
his
is
f
ol
lowed
by
m
od
ific
at
ion
of
G
aussian
m
ixture
m
od
el
an
d
e
nh
a
ncem
ent
of
ICA
segm
entat
ion
appr
oach.
Ne
xt
,
the
tum
or
s
are
cl
assifi
ed
as
ben
i
gn
a
nd
m
al
ign
ant
us
i
ng
Kernel
SVM
.
The
discre
te
wa
vel
et
transfor
m
s
(DWT
)
is
us
e
d
to
e
xtract
th
e
featur
e
s
[
4],
[
5]
w
hich
a
r
e
furthe
r
re
duced
to
i
m
pr
ove
the
cl
assifi
cat
ion
ac
cur
acy
us
in
g
t
he
P
rinci
pal
Com
po
ne
nt
A
na
ly
sis
(P
CA)
.
Af
te
r
the
featu
res
a
r
e
extracte
d, S
V
M cl
assifi
er is
us
e
d
to
classi
f
y t
hem
.
In
this
pap
e
r,
the
Gaussi
an
rad
ia
l
basis
functi
on
ke
rn
e
l
(G
RB
kerne
l)
is
us
ed
t
o
analy
se
th
e
cl
assifi
cat
io
n.
Var
i
ou
s
perfor
m
ance
m
easur
es
[6
]
-
[
8]
an
d
exp
e
rt
cl
inica
l
op
i
nions
ha
ve
been
us
e
d
to
va
li
date
the cla
ssific
at
ion
m
et
ho
d
/
al
gorithm
.
2.
REVIEW
OF
LIT
ERATUR
E
Var
i
ou
s
segm
entat
ion
a
ppr
oa
ches
hav
e
bee
n
re
porte
d
in
t
he
li
te
ratur
e
an
d
ha
ve
bee
n
st
ud
ie
d
ov
e
r
a
per
i
od
of
ti
m
e.
E
m
ph
asi
s h
a
ve
b
een
giv
e
n
to
the seg
m
entat
i
on
of
br
ai
n
MR
i
m
ages.
So
m
e o
f
the r
ece
nt
works
m
entioned
i
n
the
li
te
ratur
e
hav
e
be
en
A
p
ixel
cl
assi
ficat
ion
ba
sed
br
ai
n
m
agn
e
ti
c
reso
na
nce
i
m
ages
segm
entat
ion
ha
s
been
prese
nt
ed
by
A.
In
[9]
,
wh
ere
a
uto
m
at
ic
seg
m
entat
i
on
of
brai
n
int
o
fou
r
cl
asses
li
kes
backg
rou
nd,
cereb
rospi
nal
fluid
,
gr
ey
a
nd
wh
it
e
m
atter
is
perform
ed.
A
n
uns
up
e
r
vised
and
k
no
wled
ge
based
sk
ull
st
rip
ping
al
gorithm
fo
r
br
ia
n
m
agn
et
ic
i
m
aging
te
r
m
ed
as
S
3
ba
ses
on
br
ai
n
anatom
y
and
i
m
age
intensit
y
char
a
ct
erist
ic
s
was
perform
ed
in
[
10
]
.
In
[11]
,
P.
Moesk
ops
et
.
al
.
us
ed
a
da
ptiv
e
intensit
y
thre
sh
ol
d,
after
wh
ic
h
m
orp
ho
l
og
ic
al
opera
ti
ons
is
ca
rr
ie
d
to
boos
t
rob
us
tnes
s.
A
n
a
uto
m
at
ic
s
egm
entat
ion
m
et
hod
wh
ic
h
is
based
on
the
Co
nvol
ution
Ne
ural
Netw
orks
(C
N
N)
an
d
e
xp
l
ores
s
m
al
l
3×3
ke
rn
el
s
is
desc
ribed
i
n
[12]
.
He
re
int
ensity
norm
ali
zat
ion
is
us
e
d
with
data
a
ugm
entat
ion
,
as
the
processi
ng
ste
p
that
prov
i
des
eff
ect
ive
resu
lt
s
for
brai
n
t
umor
se
gm
entat
ion
in
m
agn
et
ic
resona
nce
im
a
ges.
T
he
detect
ion
of
hu
m
an
br
ai
n
tum
or
us
in
g
m
agn
et
ic
resonan
ce
im
age
segm
entat
ion
and
m
or
ph
ologi
cal
op
e
ra
tor
s
has
bee
n
pr
es
ented
i
n
[13]
.
Her
e
the
tum
or
cel
ls
w
ere
se
par
at
e
d
f
ro
m
norm
al
ce
ll
us
in
g
m
or
ph
ologica
l
operat
or
s
com
bin
e
d
with
conve
ntion
al
i
m
age
processi
ng
te
c
hn
i
ques.
Mi
cro
wa
ve
i
m
aging
was
use
d
in
[14]
to
detect
the
brai
n
tum
or
and
local
iz
at
io
n
of
a
dee
p
br
ai
n
RF
so
ur
ce.
Her
e,
Le
ve
nb
e
r
g
-
Ma
r
qu
adi
it
erati
ve
schem
e
was
us
ed
as
m
ic
ro
wa
ve
im
agin
g
te
ch
niqu
es
to
so
l
ve
th
e
inv
er
se
scat
te
rin
g
pro
blem
for
the
he
ad
of
the
phant
om
in
403.5MHz
m
e
dical
rad
io
ba
nd.
It
was
seen
thr
ough
the
si
m
ulati
on
res
ults
that
at
le
a
st
45
dB
SN
R
wa
s
require
d
f
or
s
m
al
l
tu
m
or
detect
ion
.
Her
e
,
a local
iz
at
ion
m
et
hod
ba
sed
on
m
ic
ro
wa
ve
im
agin
g
is
use
d
f
or
d
ee
p
br
ai
n
RF sou
rc
e. A
sem
i
-
su
pe
rv
ise
d
cl
us
te
ri
ng tech
nique tha
t uses th
e c
on
c
ept o
f
m
ulti
o
bject
ive opti
m
izati
on
for
se
gm
entat
i
on
of
m
agn
et
ic
resona
nce
brai
n
im
age
in
intensit
y
sp
ace
is
p
r
opos
e
d
by
A.K.
In
[
15]
the
intensit
y
value
s
we
re
use
d
as
featur
e
s
a
nd
a
m
od
ern
obj
ect
i
ve
op
ti
m
iz
a
ti
on
te
ch
nique
w
hich
us
es
t
he
c
on
ce
pt
of
sim
ulate
d
ann
eal
in
g
is
im
plem
ented
to
op
ti
m
iz
e
the
t
hr
ee
cl
us
te
r
va
li
dity
ind
ic
es.
Its
pe
rfor
m
ance
wa
s
com
par
ed
with
ap
proac
hes
li
ke
FCM
,
E
xpe
ct
at
ion
m
axi
m
iz
at
ion
,
fu
z
zy
-
VGAP
S
cl
ust
erin
g
te
ch
nique
s.
An
e
m
pirical
wavel
et
transfor
m
(E
WT)
m
et
ho
d
to
e
xtract
the
featur
e
s
of
brai
n
S
PECT
im
age
an
d
assist
in
br
ai
n
tum
or
detect
ion
wa
s
pr
op
os
e
d
in
[16]
.
T
he
i
m
age
is
decom
po
sed
int
o
a
nu
m
ber
of
s
ub
-
ba
nd
im
ages
us
i
ng
E
W
T,
w
hile
s
egm
entat
ion
is
done
usi
ng
t
he
fu
zzy
C
-
m
eans
cl
us
te
ri
ng,
to
achie
ve
bette
r
acc
ura
cy
.
The
Suppor
t
vec
t
or
m
achine
cl
assifi
er
wa
s
us
ed
.
A
sp
at
ia
l
f
uzz
y
C
-
m
eans
(SPFC
M)
al
gorithm
was
prese
nt
ed
in
[17]
for
segm
e
nt
at
ion
of
m
agn
et
ic
resonan
c
e
i
m
ages.
Her
e
,
the
sp
at
ia
l
info
rm
at
ion
from
the
neig
hborh
ood
of
each
pi
xel
is
e
m
plo
ye
d
an
d
reali
zed
by
de
fining
a
probabil
it
y
fu
nctio
n.
T
he
SP
FC
M
al
go
rithm
helped
t
o
so
lve
the
prob
l
e
m
relat
ed
to
s
ensiti
vity
to
noise
and
intensit
y
in
ho
m
og
ene
it
y
in
m
agn
et
ic
resonan
ce
im
a
ging
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
4
, N
o.
1
,
A
pr
il
201
9
:
478
–
489
480
data
an
d
hen
c
e
i
m
pr
ove
the
segm
entat
ion
resu
lt
s.
T
he
auth
or
s
pro
ve
d
that
the
SP
FCM
had
a
bette
r
perform
ance
than
so
m
e
of
t
he
FCM
based
al
gorithm
s.
A
cl
assifi
cat
ion
m
et
ho
d
t
o
cl
as
sify
br
ai
n
m
agn
et
ic
resona
nce
im
a
ges
as
norm
al
and
ab
norm
al
us
in
g
wa
vel
et
s
te
xtu
re
fea
tures
a
nd
k
-
m
eans
cl
assifi
er
was
pro
po
se
d
in
[
18
]
.
Her
e,
t
he
Eucli
dean
dis
ta
nces
we
re
m
easur
e
d
betwe
en
featu
re
vec
tor
of
te
st
m
a
gn
et
ic
resona
nce
i
m
age
an
d
k
-
m
ea
ns
cl
assifi
er
w
as
fed
with
refe
ren
ce
m
agn
e
ti
c
reso
na
nce
i
m
ages.
A
f
our
ph
ase
hybri
d
ap
proa
ch
was
pr
ese
nt
ed
in
[19]
w
hi
ch
can
be
us
e
d
f
or
br
ai
n
t
um
or
detect
ion
and
cl
assifi
cat
i
on
in
m
agn
et
ic
reson
ance
i
m
ages.
The
first
phase
includes
,
i
m
a
ge
pre
-
processi
ng
th
rou
gh
no
i
se
filt
ering
an
d
sk
ull
detect
ion
.
I
n
th
e
seco
nd
phase
,
the
featur
e
s
a
re
ext
racted
usi
ng
gr
ay
le
vel
co
-
occ
urren
ce
m
at
rix.
N
or
m
al
an
d
abno
rm
al
c
la
ssific
at
ion
of
i
nputs
us
in
g
l
east
sq
ua
re
s
upport
vecto
r
m
achine
cl
as
sifie
r
with
m
ulti
la
yer
per
ce
ptio
n
ke
r
nel
is
dealt
in
the
thir
d
phase
.
The
f
o
ur
th
ph
ase
was
se
gm
e
ntati
on
of
tum
or,
f
or
wh
ic
h
a
uthors
us
e
d
fa
st
bo
unding
box.
T
his
i
m
ple
m
entat
ion
was
fou
nd
to
be
96.
3%
acc
ur
at
e.
T
he
perf
or
m
ance
analy
sis
of
diff
e
re
nt
m
et
ho
ds
of
tum
or
de
te
ct
ion
was
de
scribe
d
in
[
20]
.
A
com
par
at
i
ve
stud
y
bet
we
en
dif
fer
e
n
t
m
et
hods
for
tum
or
dete
ct
ion
was
desc
ribe
d
by
the
a
uthors
w
hile
em
ph
asi
zi
ng
on
the
ro
le
of
se
gm
entat
ion
in
m
edical
i
m
aging
.
T
hey
sh
owe
d
that
segm
entat
ion
works
ef
fici
en
tl
y
in
detect
ing
an
d
e
xtracti
ng
t
he
tum
or
from
m
agn
et
ic
r
es
on
ance im
aging
.
3.
PROP
OSE
D
APP
ROAC
H
Ther
e
a
re
m
an
y
i
m
po
rtant
ap
plica
ti
on
s
of
bl
ind
source
se
par
at
io
n
thr
ou
gh
in
de
pende
nt
co
m
po
ne
nt
analy
sis
in
m
any
areas
li
ke
signa
l
proc
essing
an
d
m
edical
sig
nal
processi
ng.
M
any
com
pu
te
r
ai
ded
al
gorithm
s
ha
ve
bee
n
pr
opose
d
ove
r
the
ye
ars,
that
can
be
im
ple
m
e
nted
to
a
naly
ze
m
agn
et
ic
reso
na
nc
e
i
m
ages,
li
ke
Ei
gen im
age an
al
ysi
s,
pr
i
ncipal
com
po
ne
nt an
a
ly
sis
(P
CA)
and fuzzy
C m
et
ho
d et
c. Eig
en
i
m
age
analy
sis
is
better
wh
e
n
it
com
es
to
of
fe
rin
g
a
n
eff
ect
ive
se
gm
entat
ion
and
featur
e
e
xtracti
on.
H
ow
e
v
er
,
wh
e
n
it
com
es
to
sat
isfact
or
y
se
gme
ntati
on
of
br
a
in
ti
ssu
es
,
the
neural
netw
ork
base
d
m
et
ho
ds
pro
ve
t
o
perf
or
m
bette
r.
The
se
m
et
ho
ds
pr
ov
i
de
bette
r
pe
rfor
m
ance
wh
e
n
com
par
ed
to
the
cl
assic
al
m
axi
m
u
m
li
kelihood
m
et
ho
ds.
T
here
has
bee
n
a
ri
se
in
t
he
m
ult
i
-
sp
ect
ral
im
ag
es,
du
e
to
w
hich
m
any
segm
entat
ion
an
d
a
naly
si
s
proce
dures
ha
ving
their
ba
se
on
ort
ho
gonal
pro
j
ect
io
n,
Kalm
an
filt
er,
et
c
ha
ve
been
sub
j
ect
ed
to
enh
a
ncem
ent
ov
e
r
the
ye
ar
s.
These
proc
edures
howe
ve
r,
posse
ss
a
dr
a
wb
ac
k,
t
ha
t
they
req
ui
r
e
pr
io
r
knowle
dge.
T
her
e
fore,
the
i
nd
e
pe
nd
e
nt
c
om
po
nen
t
a
nal
ysi
s
(I
C
A)
se
gm
entat
ion
m
et
hod
pro
vid
e
s
bette
r
perform
ance in
the se
gm
entat
i
on of
br
ai
n
ti
ss
ues.
ICA
ca
n
be
use
d
to
i
de
ntify
li
near
non
-
ort
hogo
nal
co
ordin
at
e
syst
e
m
s.
The
data’s
sec
ond
an
d
highe
r
orde
r
sta
ti
sti
cs
determ
ine
the
directi
on
of
a
xis
c
orres
pond
ing
t
o
t
he
c
oo
rd
i
nate
syst
em
s.
T
he
tr
ans
f
orm
ed
var
ia
bles
that
are
f
ound
by
the
li
near
tran
sform
at
ion
of
data
by
ICA
are
s
uch
that
they
are
sta
ti
s
ti
cal
l
y
ind
e
pende
nt
from
each
oth
er
as
far
as
poss
ible.
This
m
e
ans
that
the
I
CA,
just
li
ke
pr
i
ncipal
com
pone
nts
analy
sis
(P
CA
)
te
chn
i
qu
e
,
is
us
ef
ul
in
fin
di
ng
the
data
struct
ur
e.
T
he
I
CA,
ho
wev
e
r,
has
a
dr
a
w
ba
ck
of
assum
ing
the
s
ources
to
be
in
dep
e
ndent.
D
ue
to
this
the
c
oncept
of
m
ixtur
e
m
od
el
s
have
been
i
ntr
o
du
ced
s
o
that
the
data
obser
ve
d
ca
n
be
char
a
ct
erized
into
diff
e
re
nt
m
utu
al
l
y
exclu
sive
cl
asses.
O
ne
of
t
he
im
per
at
ive
ste
p
in
IC
A
is
choosi
ng
a
pro
per
sea
rch
sp
ace.
Hen
ce
,
the
ge
ner
al
i
zat
ion
co
ns
i
de
rati
on
s
a
re
ge
ner
al
ly
pr
ece
de
d
by
di
m
ension
al
it
y
red
uctio
n
pr
oce
dures
in
IC
A.
This
i
m
pr
oves
the
gen
er
al
iz
at
ion
pe
rfor
m
ance
of
ICA
al
on
g
wit
h
a
re
du
ct
io
n
i
n
the
c
om
pu
ta
ti
on
al
com
plexity
.
Exp
ect
at
io
n
m
axi
m
iz
at
io
n
al
gorithm
(EM)
is
us
e
d
to
est
im
a
t
e
the
pro
bab
il
i
ty
den
sit
ie
s
th
r
ough
e
xpect
at
ion
m
axi
m
iz
a
tio
n
cl
us
te
rin
g.
The
EM
al
gori
thm
is
base
d
on
the
s
earch
of
m
axi
m
u
m
chan
ces
of
pa
ram
et
er
e
stim
at
es
wh
ic
h
can
be
m
ade
wh
e
n
the
data
m
od
el
is
dep
e
ndent
on
certai
n
la
te
nt
var
ia
bles.
I
n
t
he
al
gorithm
pr
ese
nted
in
t
his
pa
pe
r,
the
init
ia
l
cl
us
te
rs
are
identifie
d
us
in
g
K
-
m
ean
s
al
gorithm
after
wh
ic
h
ex
pectat
ion
(E
)
an
d
m
axim
iz
at
ion
(M)
ste
ps
are
pe
rfor
m
ed
al
te
rn
at
ively
to
converge
int
o
a
resu
lt
thr
ough
it
erati
on.
Th
e
la
te
nt
var
ia
bl
es
are
us
e
d
in
t
he
sam
e
way
a
s
they
wer
e
obser
ve
d
to
com
pu
te
th
e
exp
ect
at
io
n
of
c
ha
nces
(E
).
The
r
es
ults
of
the
la
st
E
ste
p
are
us
e
d
to
c
om
pu
t
e
the m
axi
m
u
m
pro
bab
il
it
y i
n
the m
axi
m
iz
ati
on (
M)
step
.
Ma
them
a
ti
call
y for a
giv
e
n
t
r
ai
nin
g datase
t
(
1
)
,
(
2
)
.
.
.
.
.
.
.
(
)
x
x
x
m
(1)
And
m
od
el
(
,
)
p
x
z
(2)
Wh
e
re z
is the
la
te
nt v
aria
ble,
w
e
ha
ve:
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Hyb
ri
d
e
nhanc
ed ICA
&
KSVM base
d br
ain
tumor im
age s
egm
e
nta
ti
on
(
T
hr
iv
ik
ram B
athi
ni
)
481
1
l
o
g
(
;
)
m
l
p
x
(3)
1
l
o
g
(
;
;
)
m
z
p
x
z
(4)
Wh
e
re,
x,
y,z
and
re
pr
ese
nt
the
log
li
ke
li
ho
od.
Since
z
is
an
unknow
n
la
te
nt
var
ia
ble,
appr
ox
im
at
ion
s
ha
ve
bee
n
us
e
d,
wh
ic
h
are
f
or
m
ed
in
E
a
nd
M
ste
ps
de
scrib
e
d
ab
ove.
It
can
be
m
at
he
m
at
ic
a
lly wr
it
te
n
as:
E
Step, f
or
eac
h
i
:
(
)
(
)
(
)
/;
i
i
i
i
Q
z
p
z
x
(5)
M Ste
p, f
or all
z:
(
)
(
)
()
()
l
o
g
.;
:
a
r
g
m
a
x
(
)
ii
i
i
i
iz
i
p
x
z
t
Q
z
Qz
(6)
Wh
e
re
i
Q
the
pos
te
rior
distri
bu
ti
on of is
()
i
z
‘s
g
i
ve
n
the
()
i
x
s
.
Con
ce
ptu
al
ly
to
inc
orp
or
at
e
t
he
s
patia
l
inf
orm
ation
int
o
G
MM
,
as
a
ty
pi
cal
var
ia
ti
on
of
GMM
is
pro
po
se
d
by
usi
ng
t
he
MR
F
m
od
el
as
a
pr
ior.
Dif
fer
e
nt
f
ro
m
GMM,
ea
ch
pix
el
i
i
n
MOD
IF
I
ED
G
MM
is
char
act
e
rized
by
it
s p
r
ob
a
bili
ty
v
ect
or
12
,
,
.
.
.
.
.
.
.
,
c
i
i
i
i
w
her
e
k
i
denotes the
pro
bab
il
it
y o
f
the
i
th
pix
el
belo
ng
i
ng to
t
he
k
th
cluster
.
In m
od
ifie
d G
MM
, th
e c
orre
sp
on
ding m
ixtur
e
m
od
el
o
f
xi
is assum
ed
as
1
/
,
/
c
k
i
i
i
k
k
p
x
p
x
(7)
wh
e
re
/
p
x
i
k
is a
Gau
s
sia
n dist
ri
bu
ti
on
with
pa
r
a
m
et
ers
,
k
u
k
k
.
To
ta
ke
the
spa
ti
al
dep
en
de
nc
e
into
acc
ount
,
the
pr
io
r
dis
tribu
ti
on
of
π
is
giv
e
n
by
the
MR
F
m
od
el
thr
ough a
Gibb
s
de
ns
it
y f
un
ct
i
on
1
e
x
p
/
i
N
N
i
p
V
Z
(8)
wh
e
re Z
is a
norm
al
iz
ing
con
sta
nt and β
is re
gu
la
rizat
ion
pa
ram
et
er.
The
cl
iq
ue
pot
entia
l
functi
on
of
t
he
pix
el
l
abel
vecto
rs
π
m
is
giv
en
by
V
Ni
(π
)
w
hich
li
es
within
the
neig
hborh
ood
Ni of t
he
it
h
pi
xel
i
i
Z
N
i
m
m
c
N
V
(9)
No
ti
ce
that
t
he
1
,
2
,
.
.
.
.
.
k
in
GMM
is
s
har
e
d
by
al
l
pi
xels,
wh
e
reas
in
MO
DIFIE
D
GMM
i
is
dif
fer
e
nt
f
or
each
pix
el
i
a
nd
dep
e
nds
on
it
s
nei
ghborin
g
pix
el
s.
I
n
M
O
DI
F
IE
DG
MM
,
th
e
m
od
ifie
d
EM
al
gorithm
is u
ti
li
zed to
ob
ta
i
n t
he
m
axi
m
u
m
a posterio
ri (
M
AP
)
esti
m
a
ti
on
of the
p
a
ram
eter
s.
Subseque
ntly
the
ICA
is
do
ne
at the
fo
ll
owin
g
le
vels
1.
En
han
cem
ent
of
e
nergy
crit
e
ria:
Eigen
valu
es
are
use
d
t
o
determ
ine
the
eff
ect
ive
ness
of
the
f
eat
ur
e
s.
These
ei
ge
n
va
lues
a
re
in
dic
at
ive
of
t
he
ori
gin
al
data;
the
ei
ge
n
value
s
pectr
um
on
t
he
ot
her
ha
nd
ind
ic
at
es
the
e
nergy
of
t
he
or
iginal
data.
T
he
inform
at
ion
t
hat
represe
nts
the
or
i
gin
al
da
ta
need
s
to
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
4
, N
o.
1
,
A
pr
il
201
9
:
478
–
489
482
be
prese
ver
e
d
t
o
the
m
axi
m
um
extent
w
hile
trans
fo
rm
ing
f
ro
m
a
hig
h
di
m
ension
al
s
pa
ce
to
a
lo
wer
dim
ension
al
spac
e.
2.
Im
pr
ov
em
ent
of
m
agn
it
ude
c
rite
ria:
In
orde
r
to
im
ple
m
ent
the
ICA
i
n
this
reg
a
rd,
it
sho
uld
be
do
ne
in
a
re
duced
P
CA
s
pace
s
o
t
hat
the
sm
al
l
valued
t
rail
ing
ei
gen
value
s
a
re
not
incl
ud
e
d.
T
h
e
refor
e
,
these
crit
eria
favo
ur
s
lo
w
dim
ension
s
pac
es
wh
e
reas
th
e
s
m
al
l
valued
trai
li
ng
ei
ge
n
val
ues
are
exclu
ded.
An
im
pr
ove
d
perform
ance
is
achieve
d
in
the
pro
posed
work
by
usi
ng
dim
ension
al
it
y
red
uc
ti
on
proce
dures
t
ha
t
aim
at
balancing
t
he
e
nerg
y
and
m
agn
it
ude
crit
eri
on.
I
CA,
a
n
e
xtensi
on
to
t
he
co
-
va
riance
base
d
PC
A
is
us
e
d
to
so
l
ve
BSS
pr
ob
le
m
s
li
ke
coc
ktail
pa
rty
pro
blem
.
The
obser
ve
d
sign
al
s
a
re
m
ade
up
of
li
near
c
om
bin
at
ion
of origin
al
sig
nals and a
m
at
rix
m
ixtur
e
.
It can
b
e
r
e
pr
es
ented
a
s
X
A
S
(10)
W=A
-
1
in
vers
e is com
pu
te
d t
o ob
ta
in
m
ixing
m
at
rix
A. Th
e IC is
obta
ine
d
as:
,
ˆ
WX
S
(11)
ˆ
S
S
(12)
The
s
uspic
iou
s
reg
i
on
s
nee
d
t
o
be
segm
ented
f
ro
m
an
IC.
Assum
ing
ti
j
to
be
t
he
(i,
j)
‘t
h
el
em
ent
of
a
co
-
occ
urre
nc
e
m
at
rix
W
t
ha
t
co
ns
ide
rs
t
he
gr
ay
le
vel
tra
nsi
ti
on
s
betwee
n
t
wo
ad
j
ace
nt
pi
xels.
T
he
eq
uation
is wr
it
te
n
a
s:
11
(
,
)
MN
ij
ik
t
l
k
(13)
Wh
e
re
1
,
1
(
1
1
,
)
,
(
,
)
,
(
,
1
)
/
(
1
,
)
,
(
1
1
,
)
0
,
{
d
i
f
k
i
I
I
l
k
i
I
I
K
a
n
d
o
r
I
k
I
k
j
Othe
rw
ise
(14)
The pr
obabili
ty
o
f
a tra
ns
it
io
n of t
his
gr
ay
l
evel f
ro
m
i to
j can
be de
fine
d as
11
00
ij
ij
LL
ij
ij
t
P
t
(15)
Her
e
,
t‘re
pr
ese
nts
the
thre
sho
ld
that
par
ti
ti
ons
the
c
o
-
occurre
n
ce
m
at
rix
t
hat
is
def
ine
d
by
(13).
Th
e
co
-
occ
urren
ce
m
at
rixes
are
f
irst
gro
up
e
d
i
nto
4
qua
dr
a
nt
s
nam
ely
A,
B,
C
and
D,
after
w
hich
t
he
y
are
gro
up
e
d
i
nto
f
or
e
gro
und
a
nd
bac
kgr
ound
obj
ect
s
.
T
he
on
es
ha
ving
pix
e
ls
with
i
ntensit
y
le
vel
great
er
than
thres
ho
l
d
f
al
l
into
t
he
f
oreg
round
obj
e
ct
cat
e
gory
w
hile
th
ose
with
pix
el
i
ntensity
le
ss
th
an
th
res
ho
l
d
fa
ll
into
the
backgro
und
cat
e
gory.
T
he
tran
sit
ion
s
within
bac
kgr
ound
a
nd
f
ore
gro
und
a
re
re
pr
ese
nted
by
A
a
nd
C
qu
a
drants
w
hile
the
transiti
ons
acro
ss
bo
unda
ries
betwee
n
f
or
e
gro
und
an
d
back
gr
ounds
a
re
represe
nted
by
B
and D q
ua
dr
a
nt
s.
The
foll
owing eq
uatio
ns
re
pr
ese
nt the
pr
obabili
ti
es asso
c
ia
te
d
with
each
of these
qua
drants
.
00
tt
t
A
i
j
ij
PP
1
01
tL
t
B
i
j
i
j
t
PP
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Hyb
ri
d
e
nhanc
ed ICA
&
KSVM base
d br
ain
tumor im
age s
egm
e
nta
ti
on
(
T
hr
iv
ik
ram B
athi
ni
)
483
1
10
Lt
t
C
i
j
i
t
j
PP
11
11
LL
t
D
i
j
i
t
j
t
PP
(16)
The
cel
l
pro
ba
bili
ti
es can
be
us
e
d
to
obtai
n
t
he pr
obabili
ti
es in eac
h q
uadr
ant:
/
ij
t
i
j
A
t
A
P
P
p
/
ij
t
i
j
B
t
B
P
P
P
//
i
j
i
j
tt
i
j
C
i
j
D
tt
CD
pp
PP
pp
(17)
In
the
pro
pose
d
wor
k,
a
loca
l
entr
op
y
base
d
th
res
ho
l
d
ha
s
bee
n
em
plo
ye
d
to
se
gm
ent
the
tum
or
.
Since
the
local
transiti
on
s
f
rom
back
gro
und
to
backgro
und
(BB)
an
d
obj
e
ct
s
to
ob
j
ect
s
(
FF)
are
re
pres
ented
by qua
dr
a
nts
A
and C
res
pecti
vely
, th
e m
at
hem
at
ic
al
d
ef
init
ion o
f
l
ocal ent
r
op
ie
s
can
b
e
writ
te
n
as:
//
00
(
)
.
l
o
g
tt
tt
B
B
i
j
A
i
j
A
ij
H
t
P
P
(18)
11
//
11
(
)
.
l
o
g
LL
tt
F
F
i
j
C
i
j
C
i
t
j
t
H
t
P
P
(19)
The
sec
ond
-
or
der
l
ocal
entr
opy
can
be
ob
t
ai
ned
by
su
m
m
ing
up
the
l
ocal
withi
n
-
cl
a
ss
transiti
on
entr
op
ie
s
of th
e f
or
e
gro
und
a
nd the
bac
kgr
ound a
nd ca
n be
w
ritt
en
as
(
)
(
)
(
)
L
E
B
B
F
F
H
t
H
t
H
t
(20)
Ma
xim
iz
ing
th
e ab
ov
e
equati
on r
es
ults in
a t
hr
es
hold
base
d o
n
local
e
ntr
opy,
us
in
g w
hich
the
tum
or
can
be
se
gm
ented
m
a
x
a
r
g
(
)
t
L
E
L
E
t
H
t
(21)
3.1.
Me
thod
The
pro
pose
d
appr
oach
is
to
be
co
ded
us
in
g
Ma
tl
ab.
The
ste
ps
invol
ved
i
n
the
process
c
an
be
li
ste
d
as b
el
ow.
a.
Popu
la
ti
ng
t
he
re
qu
i
red
im
ag
es
f
or
a
naly
sis
from
database
s
li
ke
B
rain
T
um
or
Disease
database
an
d
i
m
ages s
ource
d from
o
ther
s
uper
sp
eci
al
ty
hosp
it
al
s.
b.
Perfo
rm
ing
pr
e
processi
ng
ope
rati
on
s
li
ke
no
i
se
rem
ov
al
,
e
dge
detect
ion
a
nd
th
res
ho
l
ding
to
rem
ov
e
the b
ac
k g
rou
nd a
nd o
t
her cl
inica
ll
y i
rr
el
eva
nt thin
gs i
n
t
he
context
of the
pro
po
se
d
a
naly
sis
c.
Desig
ning a
n En
han
ce
d ICA
Mi
xtu
re
Mo
de
l (EI
C
AMM)
f
or aut
om
atic segm
entat
ion
of
br
ai
n.
d.
Desig
ning
a
C
la
ssifie
r
base
d
on
Sup
port
Vecto
r
Ma
chi
ne
(
SV
M)
f
or
cl
assify
ing
a
nd
co
rr
el
at
in
g
diff
e
re
nt n
e
uro
log
ic
al
d
is
orde
rs.
e.
Desig
ning
a
Gr
a
phic
al
Use
r
I
nterf
ace
(
G
UI)
us
i
ng
Ma
tl
ab
for
loa
din
g
t
he
re
quire
d
im
age
fo
r
analy
sis, pr
oce
ssing t
he res
ults an
d pr
e
sentin
g
the
inter
pret
at
ion
.
f.
Vali
dating t
he
resu
lt
s
by c
ompari
ng w
it
h st
and
a
r
d data set
s
and e
xp
e
rt cli
nical
opini
on
g.
Perfo
rm
ing
Re
cei
ver
O
per
at
i
ng
Cha
racteri
s
ti
c
(ROC)
a
na
ly
sis
to
valida
te
the
pe
r
form
ance
of
the
cl
assifi
er.
4.
RESU
LT
S
&
DISCU
SSI
ONS
The
f
ollow
i
ng
sect
ion
su
m
m
a
rizes
the
resu
lt
s
of
the
pro
pos
ed
segm
entat
ion
ap
proac
h.
Th
e
pr
op
os
e
d
was
co
de
d
in
Ma
tl
ab
R
20
12
a
a
nd
the
validit
y
of
the
segm
entat
ion
is
de
m
on
strat
ed
with
the
he
lp
of
evaluati
on
pa
r
a
m
et
ers.
The
gro
und
tr
uth
im
ages
f
or
va
li
da
ti
on
we
re
obta
ined
th
r
ough
m
anu
al
se
gm
ent
at
ion
.
The
f
ollow
i
ng
i
m
ages
l
ist
e
d
in
Figu
re
1
ha
ve
bee
n
co
nsi
der
e
d
for
te
sti
ng
an
d
valida
ti
on
.
T
o
hav
e
a
true
represe
ntati
on
the im
ages ar
e
of d
if
fer
e
nt siz
es an
d
i
ntensit
y values
.
The
histo
gr
am
prof
il
e o
f
the
i
m
ages
ser
ves
t
o
giv
e
a
tre
nd
in d
ist
rib
utio
n
of
inte
ns
it
y
valu
es
an
d
hel
p
in
the
init
ia
l
sta
ges
of
the
c
hoos
i
ng
the
th
r
esh
old
.
T
he
histogram
of
th
e
i
m
ages
is
il
lustrate
d
in
t
he
F
igure
2.
T
he
bel
ow
s
how
n
are
the
histo
gr
am
pr
ofi
le
s,
cl
early
il
l
us
trat
es
that
th
e
te
st
i
m
age
hav
e
dif
fer
e
nt
intensit
y
prof
il
e
an
d
va
r
ia
nt
pix
el
distri
bu
ti
on.
T
his
pi
xel
distrib
utio
n
is
al
so
influ
en
ced
the
ty
pe
an
d
the
locat
io
n
of
the
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
4
, N
o.
1
,
A
pr
il
201
9
:
478
–
489
484
tum
or
too
.
Sim
il
arly
the
size
of
the
tum
or
al
so
play
s
a
cru
ci
al
ro
le
in
def
inin
g
the
intensit
y
prof
il
e.
The
intensit
y
pr
ofi
le
of
a
par
ti
cual
r
reg
i
on
ca
n
al
so
giv
e
a
n
incli
nation
to
ward
s
per
c
e
ntage
of
scat
te
red
el
em
ents.
Ed
ge dete
ct
ion refe
rs
t
o
the
pr
ogressi
on of id
entify
and l
oca
te
sh
ar
p disc
onti
nu
it
ie
s in
a
n
i
m
age.
Ed
ge
is
a
basic
and
im
po
rtant f
eat
ur
e o
f
a
n
im
age.
Im
age
is
a
com
bin
at
ion
of
ed
ges
.
Dete
ct
ing
ed
ge
s
is on
e of the
m
ai
nly si
gn
i
ficant featur
e
s in
im
age seg
m
ent
at
ion
. E
dg
e de
te
ct
ion
is a v
it
al
ste
p
as it
is a
p
ro
ce
s
s
of
ide
ntifyi
ng
and
locat
es
s
ha
rp
dis
-
c
on
ti
nui
ti
es
in
a
rep
res
entat
ion
.
T
he
e
dg
e
s
of
the
te
st
i
m
ages
as
ide
ntifie
d
us
in
g
Pr
e
witt
edg
e
d
et
ect
or
i
s
il
lustrate
d
throu
gh
F
i
gure
3.
The
com
plex
it
y
of
m
edical
i
m
age
segm
ent
at
ion
can
be
cl
early
unde
rstood
fro
m
the
abo
ve
i
m
ages.
Eve
n
thou
gh
we
are
us
in
g
a
sim
il
ar
edg
e
de
te
ct
or
we
can
see
an
a
ppreci
able
di
ff
e
ren
ce
in
perform
ance
betwee
n
dif
f
eren
t
im
ages.
I
t
can
be
cl
earl
y
ob
se
r
ved
tha
t
the
edg
e
s
are
neatl
y
dem
arcate
d
in
im
age
(b)
w
her
e
as
in
im
a
ge
(c
)
t
he
e
dg
e
s
ap
pear
e
d
t
o
m
erg
e
a
nd
i
n
t
he
case
of
im
age
(a)
it
app
ear
s
to
be
cl
uttered
an
d
di
storted.
It
can
be
ob
se
rv
e
d
f
ro
m
the
intensit
y
pr
of
il
e
an
d
edg
e
s
that
te
st
im
age
s
prese
nt
a
ve
r
y
com
plica
te
d
ta
sk
for
se
gm
e
ntati
on
.
T
he
re
su
lt
s
of
t
he
se
gm
entat
ion
of
these
te
st im
ages u
si
ng the
pro
pose
d
a
ppr
oach are
dep
ic
te
d u
sin
g t
he
F
ig
ure
4.
Figure
1. Im
age (
a)
, im
age (
b) a
nd i
m
age (
c)
conside
red f
or evaluati
on
(a)
0
50
100
150
200
250
300
0
5
10
15
x
1
0
4
S
l
i
c
e
h
i
s
t
o
g
r
a
m
Fr
e
q
u
e
n
c
y
I
n
t
e
n
s
i
t
y
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Hyb
ri
d
e
nhanc
ed ICA
&
KSVM base
d br
ain
tumor im
age s
egm
e
nta
ti
on
(
T
hr
iv
ik
ram B
athi
ni
)
485
(b)
(c)
Fi
gure
2.
Histo
gr
am
s
of i
m
ag
e (a) Im
age (b)
i
m
age (
c)
Figure
3
.
The
e
dg
e
s ide
ntifie
d f
or
test
im
ages u
si
ng prewitt
op
e
rato
r
of
im
age
(a)
im
age (b) im
age (
c)
0
50
100
150
200
250
300
0
0
.
5
1
1
.
5
2
2
.
5
3
3
.
5
x
1
0
4
S
l
i
c
e
h
i
s
t
o
g
r
a
m
Fr
e
q
u
e
n
c
y
I
n
t
e
n
s
i
t
y
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
4
, N
o.
1
,
A
pr
il
201
9
:
478
–
489
486
Figure
4
.
Tum
or in im
age (
a)
, im
age (
b), a
nd im
age (
c)
are
seg
m
ented using t
he p
rop
os
e
d
a
ppr
oach
Thro
ugh
F
i
gur
e
4,
it
can
be
cl
early
ob
se
rved
th
rou
gh
vis
ual
ins
pecti
on
that
the
pro
pos
ed
ap
proac
h
ha
s
delivere
d
a
nea
t
and
cl
ean
se
gm
entat
ion
.
The
per
f
orm
ance
i
s
cl
e
arly
visible
in
i
m
age
(a)
and
im
age
(c)
,
wh
il
e
in
i
m
age
(b)
we
can
obser
ve
so
m
e
of
the
backgro
und
e
lem
ents
hav
e
al
so
bee
n
incl
ud
e
d.
T
o
il
lust
rate
the
eff
ect
ive
ness
of
the
se
gm
entat
ion
a
sam
ple
il
lustrati
on
of
i
ntensity
pro
fili
ng
of
the
se
gm
ented
t
um
or
i
m
ag
e
(b)
is
giv
e
n
in
the
Fig
ur
e
5
.
It
ca
n
be
cl
ea
rly
obser
ve
d
f
ro
m
the
fi
gure
there
is
a
nea
t
distrib
utio
n
of
the
segm
entat
ion
ind
ic
at
in
g
cl
ear
pr
ofi
li
ng.
Th
e
validit
y
of
t
he
segm
entat
i
on
is
evaluate
d
thr
ough
eval
uation
par
am
et
ers
dis
cusse
d
in
sect
i
on
4,
t
hes
e
a
re
com
pu
te
d
by
com
par
ing
the
segm
ented
im
age
with
t
he
gro
und
truth o
btaine
d usin
g
m
anu
al
s
egm
entat
ion
. T
he results
of e
va
luati
on
are li
s
te
d
usi
ng the
T
able 1
.
Fig
ure
5
.
I
nten
sit
y p
rofil
e d
ist
rib
ution o
f
se
gm
ented
im
age (
b)
Table
1.
E
val
ua
ti
on
par
am
et
e
rs
f
or the
pr
opos
e
d
se
gm
entation
a
ppr
oac
h
I
m
a
g
e
PRI
VO
I
G
CE
J
ID
JD
PSNR
I
m
a
g
e (
a
)
0
.99
0
.04
1
0
.00
4
0
.83
0
.16
6
2
.15
I
m
a
g
e (
b
)
0
.96
0
.23
1
0
.01
2
0
.87
0
.12
4
5
.58
I
m
a
g
e (
c
)
0
.99
0
.04
1
0
.00
4
0
.91
0
.05
6
0
.15
Fr
om
the
T
abl
e
1,
it
can
be
i
nf
e
rr
e
d
that
the
pro
po
se
d
m
eth
od
has
deliver
ed
in
te
rm
s
of
al
l
the
evaluati
on
par
am
et
ers
.I
t
is
al
so
interest
ing
to
obse
rv
e
that
the
i
m
age
(a)
w
hich
pr
oduce
d
cl
utter
ed
has
in
fact
been
segm
ented
bett
er th
a
n
t
he oth
er tw
o
im
ages as
evi
den
t
fro
m
the ev
al
uation pa
ram
et
ers
.
0
50
100
150
200
250
300
350
400
0
50
100
150
200
250
300
P
i
x
e
l
p
o
s
i
t
i
o
n
I
n
t
e
n
s
i
t
y
p
r
o
f
i
l
e
I
n
t
e
n
s
i
t
y
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Hyb
ri
d
e
nhanc
ed ICA
&
KSVM base
d br
ain
tumor im
age s
egm
e
nta
ti
on
(
T
hr
iv
ik
ram B
athi
ni
)
487
Figure
6
.
O
ver
l
ap
im
ages o
f
s
egm
ented
im
ag
es w
it
h g
rou
nd tru
t
h
im
ages
The
overla
p
i
m
ages
of
gro
und
truth
im
ages
a
nd
the
se
gm
ented
i
m
ages
il
lustrate
d
us
in
g
Fi
gure
6
al
so
cl
early
p
oi
nts t
o near
p
e
rf
ect
s
egm
entat
ion
ac
hieve
d wit
h
t
he
h
el
p o
f
t
he
propose
d
a
ppr
oa
ch.
In
the
pr
es
ente
d
w
ork,
a
cl
ass
ifie
r
base
d
on
Kernal
base
d
s
upport
vect
or
m
achine
(KSV
M)
has
bee
n
i
m
ple
m
ented.
Her
e
,
after
a
tu
m
or
is
seg
m
ented,
analy
sis
is
done
us
in
g
thi
rteen
di
ff
e
ren
t
par
am
et
ers
extracte
d
us
in
g
D
WTs.
Fo
r
this,
a
DB5
wa
velet
is
use
d.
A
fter
th
e
extracti
on
of
t
hese
par
am
et
ers,
a
PC
A
a
nal
ysi
s
is
e
m
plo
ye
d
that
reduce
d
the
dim
ension
al
it
y
of
t
he
par
am
et
ers
f
or
t
he
extracte
d
data
set
.
A
fter
th
e
PC
A
analy
sis,
a
Gau
ssian
rad
ia
l
basis
(G
RB
)
ker
ne
l
is
e
m
plo
ye
d
for
ke
rn
el
base
d
supp
ort
vector
m
achine
cl
assifi
cat
ion
.
Ther
e
are
m
ore
than
12
0
ty
pe
s
of
brai
n
tu
m
or
s
wh
ic
h
diff
e
r
in
ori
gin,
l
ocati
on,
siz
e,
c
har
act
erist
ic
s
of
the
tum
or
ti
ssu
es,
as
def
i
ne
d
by
the
Wo
r
ld
Healt
h
O
rganizat
ion
(
WH
O)
cl
assifi
cat
ion
syst
em
[
21
]
-
[
2
3
].
Ou
t
of
these,
we
hav
e
c
onsi
der
e
d
th
ree
ty
pes
of
m
al
ign
a
nt
tum
or
s.
T
he
first
on
e
is
G
li
ob
la
stom
a:
pr
im
a
r
y
m
al
ign
ant
br
ai
n
tum
or
s
cl
assifi
ed
as
Gr
a
de
IV
an
d
de
vel
op
e
d
f
ro
m
sta
r
-
sh
a
pe
d
cel
ls,
cal
le
d
ast
ro
cy
te
s
that
su
pp
or
t
ne
rv
e
cel
ls,
w
hich
usual
ly
sta
rts
f
rom
cerebr
um
.
S
arco
m
a
tum
or
has
a
gra
de
tha
t
var
ie
s
f
ro
m
1
to
IV.
This
tum
or
ari
ses
in
the
co
nnect
ive
ti
ssu
e
s
li
ke
bloo
d
ve
ssels.
T
he
ne
xt
on
e
is
Me
ta
sta
ti
c
br
onc
hogen
ic
carcin
om
a,
a
s
econda
ry
m
a
lig
na
nt
brai
n
t
um
or
that
sp
rea
ds
to
t
he
brai
n
from
br
onch
ogenic
ca
rcin
om
a
lu
ng
tum
or
.
66
hum
an
brai
n
MR
Is
al
on
g
with
22
norm
al
and
44
a
bnor
m
al
i
m
ages
m
ake
up
the
data
set
.
The
se
include
glio
blastom
a,
sarco
m
a
and
m
e
ta
s
ta
ti
c
br
on
c
hoge
nic
carcin
oma
tu
m
or
s
coll
ect
ed
from
Har
va
r
d
Me
dical
School
we
bs
it
e
(
ht
tp:
//m
ed
.
harv
ard.ed
u/AA
NL
IB/).
T
he
brai
n
MR
Is
we
re
in
axial
plan
e,
T2
-
weig
hted
a
nd
256 _
25
6 pixels
.
Fig
ure
7
s
hows
a
sam
ple o
f
the
data set
.
(a)
Met
ast
ic
br
on
c
hoge
nic
carcin
om
a
(b)
Sa
rc
om
a
(c)
Glio
blasto
m
a
Figure
7
.
Sam
ple o
f
d
at
a set
Tw
o
data set
s
wer
e
f
or
m
ed
to
v
al
idate
the
propose
d
cl
assifi
er.
T
hese t
wo
data set
s inclu
ded a total
of
40
im
ages
(
20
i
m
ages
each)
a
nd
we
re
hete
rogen
e
ous
with
both
norm
al
and
abno
rm
al
i
m
a
ges.
T
he
te
rm
’
Tr
ue
po
sit
ive
’
is
us
e
d
to
cat
e
gorize
correct
cl
assif
ic
a
ti
on
of
a
bnorm
ality
wh
il
e
‘tru
e
neg
at
i
ve’
i
s
us
e
d
to
cat
eg
or
iz
e
correct
cl
assifi
cat
ion
of
nor
m
al
i
m
age.
Sim
il
arly
an
incorrect
cl
assifi
cat
ion
of
ab
nor
m
al
i
ty
is
cl
ass
ifie
d
as
‘F
al
se
ne
gativ
e’
an
d
inc
orre
ct
cl
assifi
cat
ion
of
norm
al
ity
is
cat
ego
rize
d
as
‘F
al
s
e
posit
ive’.
The
T
able
2
sh
ows
the
pe
rfor
m
ance of the
prop
os
ed
class
ifie
r.
Evaluation Warning : The document was created with Spire.PDF for Python.