Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
1
3
,
No.
2
,
Febr
uar
y
201
9
, pp.
837
~844
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
2
.pp
837
-
8
44
837
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Sh
ape analysis f
or classif
icat
i
on o
f breast
nodu
les on digit
al
ultrasou
nd
images
Ha
n
un
g Adi
Nugro
ho
1
,
He
sti Kh
uz
aima
h Nurul
Yu
su
fiy
ah
2
, Te
guh
Bhar
ata Adj
i
3
,
Widhi
a
K.Z
O
kt
oe
berz
a
4
1
,
2
,
3,
4
Depa
rtment
of
Elec
tr
ical Eng
ine
er
ing
and
Inf
orm
at
ion
T
ec
hn
olog
y
,
Fa
cul
t
y
o
f
Engi
n
ee
ring
,
Univer
sita
s Gad
j
ah
Mada
,
Yog
y
a
kar
ta,
Indon
esia
.
2
Depa
rtment of
Ph
y
sic, Fac
u
lty
of
Scie
n
ce a
nd
T
ec
hnolog
y
,
UIN
W
al
isongo Semara
ng,
Indone
sia
.
4
Depa
rtment of I
nf
orm
at
ic
s,
Facu
lty
of
Engi
n
ee
r
in
g,
Univer
si
ta
s B
engkul
u,
Indone
sia
.
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Sep
14
, 201
8
Re
vised
N
ov 15
, 2
018
Accepte
d
Nov 29
, 201
8
One
of
the
i
m
agi
ng
m
odal
ities
for
ea
r
l
y
det
e
ct
ion
of
br
ea
st
c
ancer
m
al
igna
nc
y
is
ul
tra
sonograph
y
(
US
G).
The
m
al
igna
nc
y
c
an
be
a
naly
s
ed
from
the
ch
aract
er
isti
c
of
nodule
sha
pe.
Th
is
stud
y
a
ims
to
deve
lop
a
m
et
hod
for
cl
assif
y
ing
the
s
hape
of
br
ea
st
n
odule
in
to
two
c
la
ss
es,
namel
y
r
egul
ar
and
irre
gul
ar
cl
asses.
The
input
image
is
pre
-
proc
esse
d
b
y
using
the
c
om
bina
ti
on
of
ada
p
ti
ve
m
edi
an
fi
lter
and
spec
kl
e
red
u
ction
bil
atera
l
fi
lt
e
rin
g
(SRB
F)
to
red
uce
spec
k
le
noises
and
to
el
iminat
e
th
e
image
la
b
el.
After
wards,
the
fil
tered
image
i
s
segm
ent
ed
base
d
on
ac
ti
ve
c
ontour
foll
owed
b
y
fea
tur
e
ext
ra
ct
ion
proc
e
ss
.
Nine
ex
tract
ed
fe
at
ur
es,
i
.
e
.
roundne
ss
,
sli
m
ness
and
seve
n
feature
s
o
f
inva
ri
ant
m
om
ent
s,
ar
e
used
to
cl
assif
y
nodul
e
shape
using
m
ult
i
-
lay
e
r
per
c
ept
ron
(MLP).
The
pe
rform
ance
of
th
e
propose
d
m
et
hod
is
eva
lu
at
ed
using
105
bre
ast
nodu
l
e
images
which
comprise
of
57
r
egul
ar
and
48
irre
gula
r
nod
ule
images.
The
result
s
of
cl
assific
a
ti
on
proc
ess
ac
hi
eve
th
e
le
ve
l
of
accura
c
y
,
sensi
ti
vi
t
y
an
d
spec
ifi
c
ity
a
t
9
6.
20%,
97.
90
%
and
94.
70%
,
respe
ctively
.
Th
ese
result
s
ind
icate
tha
t
the
pro
posed
m
et
h
od
succ
essfu
l
l
y
cl
assifi
es
the breast
nodul
e
imag
e
s ba
sed
on
shape
anal
y
s
is.
Ke
yw
or
d
s
:
Breast
n
odules
Invar
ia
nt m
o
m
ents
Sh
a
pe
a
naly
sis
Ultraso
und i
m
age
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
:
Hanu
ng Adi
N
ugr
oho
,
Dep
a
rtm
ent o
f El
ect
rical
En
gi
neer
i
ng and
Inf
or
m
at
ion
Tec
hnol
og
y,
Faculty
of E
ngineerin
g, U
nive
rsita
s G
a
djah
Ma
da,
Jl. Grafi
ka 2
K
a
m
pu
s
UG
M
, Yo
gyaka
rta
55281, I
ndonesi
a.
Em
a
il
:
adinu
gr
oho
@ugm
.ac.id
1.
INTROD
U
CTION
Nowa
days,
t
he
ultras
ound
e
xam
inati
on
re
sul
ts
hav
e
a
lo
w
accu
racy
rate
of
diag
no
sis
du
e
to
t
he
diff
e
re
nt
inter
pr
et
at
io
ns
of
s
onogram
read
i
ng
s
am
on
g
radi
olo
gists.
This
issue
is
cau
se
d
by
t
he
prese
nce
of
sp
ec
kle
noise
in
the
s
onogr
a
m
i
m
age
[1
,
2]
.
Ther
e
f
or
e,
a
decisi
on
s
upport
syst
em
that
can
m
ini
m
is
e
the
diff
e
re
nces
a
m
on
g
ra
dio
lo
gist
interp
reta
ti
on
s
is
deem
ed
nece
ssary
to
obj
ect
i
vely
disti
nguish
be
tween
ben
i
gn
an
d
m
al
ign
ant
nodule
s.
A
c
om
pu
te
r
ai
ded
dia
gnos
i
s
(CA
D
)
has
be
en
dev
el
oped
to
assist
ra
di
olo
gists
in
m
aking
a
diag
nosis
in
wh
ic
h
the
res
ults
of
CA
D
are
able
to
prov
i
de
an
ob
j
e
ct
ive
inform
ation
t
o
rad
i
ologist
s [
1]
. Th
e g
e
ner
al
principle o
f
CA
D
co
ns
ist
s o
f
the f
ollo
wing st
ages: p
re
-
pr
oce
ssing, seg
m
entat
ion
,
featur
e
ex
t
racti
on and
featu
re
sel
ect
ion
a
nd the
cl
assifi
cat
io
n.
Var
i
ou
s
resear
ch
w
orks
relat
ed
to
en
ha
nce
br
east
ultraso
und
im
age
hav
e
been
dev
el
op
e
d.
Wu
et
al
.
us
e
d
sp
ec
kle
reducti
on
bila
te
ral
filt
ering
(S
RB
F)
to
ov
e
rc
om
e
sp
eckle
noise
s
wh
il
e
pr
ese
rv
i
ng
t
he
inf
or
m
at
ion
of
i
m
age.
Howe
ver,
the
SRB
F
m
et
h
od
can
no
t
el
i
m
inate
the
i
m
age
la
bel
[3
]
.
Othe
rw
ise
,
i
m
ag
e
la
bel
can
be
re
m
ov
ed
by
usi
ng
ada
ptive
m
edian
filt
er
as
s
how
n
in
[4,
5]
.
For
se
gm
enta
ti
on
,
m
any
research
works
ha
ve
pr
ov
e
n
that
the
a
ct
ive
con
t
our
had
well
perform
ance
to
be
app
li
ed
i
n
ultra
so
un
d
im
ages
[6
-
10]
.
Fo
r
cl
assifi
cat
ion,
analy
sis
of
te
xtu
re
featu
r
es
su
c
h
as
histogram
sta
ti
s
ti
c
,
grey
le
vel
co
-
occ
urre
nce
m
at
rices
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
3
, N
o.
2
,
Fe
bru
ary
201
9
:
8
3
7
–
8
4
4
838
(G
LCM
)
a
nd
gr
ey
le
vel
run
le
ng
th
m
at
rice
s
(G
LRLM
)
w
ere
us
e
d
by
N
ugr
oho
et
al
.
fo
r
cl
assify
in
g
thyr
oid
nodule
s
into
s
olid
an
d
cy
sti
c
cl
asses
[11].
So
m
e
featur
es
includi
ng
r
oundne
ss,
c
onve
xity
,
so
li
dity
and
aspec
t
rati
o,
play
im
p
or
ta
nt
ro
le
f
or
recogn
isi
ng
nodule
sh
a
pe.
I
n
seve
ral
stud
i
es,
Zer
nik
e
an
d
inv
a
riant
m
ome
nts
wer
e
al
so
us
ed
as
featur
es
f
or
nodule
sh
a
pe
analy
sis
[5
,
12
-
15]
.
How
e
ve
r,
not
al
l
featur
es
ha
ve
sign
i
f
ic
ant
correla
ti
on
to
yi
el
d
accurate
cl
assifi
cat
ion
.
So
m
e
cl
assifi
e
rs,
su
c
h
as
sup
port
vecto
r
m
a
chine
(
SV
M)
[
16,
17]
arti
fici
al
neu
r
a
l
networ
k
(
ANN)
[1,
12]
,
k
-
near
est
neig
hb
our
(
K
NN),
ra
ndom
fo
rest
a
nd
Naïve
Ba
y
es
[12]
hav
e
b
ee
n wid
el
y used f
or cla
ssifyi
ng u
lt
ras
ound im
ages.
The
obj
ect
ive
of
t
his
stu
dy
is
to
de
velo
p
a
m
et
ho
d
for
cl
assify
ing
the
nodule
sh
a
pe
s
of
breast
ultraso
und
im
a
ges
into
tw
o
cl
asses,
i.e.
re
gu
l
ar
an
d
irre
gu
la
r
cl
asses
as
il
lu
strat
ed
in
Fig
ure
1.
Cha
racte
r
ist
ic
s
of
breast
nodu
le
determ
ines
the
m
al
ign
ancy
le
vel
of
breas
t
cancer
.
T
he
r
egu
la
r
nodule
ind
ic
at
es
the
be
nign
cancer
w
hile
irre
gu
la
r
no
du
l
e
ind
ic
at
es
the
m
al
ign
ant
on
e.
The
c
om
bin
at
ion
of
a
dap
ti
ve
m
edian
filt
er
an
d
SRB
F
is
pr
opose
d
to
prese
rv
e
the
in
f
orm
at
io
n
qu
al
it
y
of
im
ages.
M
or
e
ove
r,
nin
e
sh
a
pe
f
eat
ur
es
c
onsist
ing
of
rou
ndness,
sli
m
ness
and
sev
en
feat
ur
es
of
inv
a
riant
m
o
m
ents,
are
e
xtra
ct
ed
to
cl
assif
y
the
nodule
sh
ape
of
br
east
ultras
ound im
age.
(a)
(b)
Figure
1. The
c
har
act
erist
ic
of
breast
no
du
le
(a)
Re
gula
r (b)
I
r
re
gu
la
r
2.
RESEA
R
CH MET
HO
D
The
dig
it
al
sca
nn
i
ng
of
br
e
as
t
ultraso
und
i
m
ages
ta
ke
n
from
the
Ra
dio
l
og
y
De
par
tm
e
nt
of
Sar
dj
it
o
and
Ha
rdj
oluk
it
o
Hospita
ls,
Yogyaka
rta,
Indonesia,
were
us
e
d
in
this
st
ud
y.
The
data
wer
e
ac
quire
d
us
in
g
USG
L
ogic
c5
Pr
em
iu
m
,
Voluson.
The
dataset
co
ns
ist
ed
of
10
5
RGB
im
ages
in
bitm
ap
form
at
,
with
57
regular
a
nd
48
irre
gu
la
r
nodule
im
ages.
Th
e
57
regular
nodule
im
ages
com
pr
ise
d
of
11
oval
a
nd
46
rou
nd
sh
a
pes.
A n
umber
of
67 im
ag
es w
e
re
us
e
d
a
s d
at
a trai
ning
and 38 im
ages
as d
at
a test
in
g.
The
ra
dio
lo
gis
ts
wer
e
involv
ed
in
this
study
as
exp
ert
for
validat
ing
the
dataset
.
The
nodule
sh
a
pe
app
ea
rs
da
r
k
colle
ct
ively
than
oth
e
rs.
T
hi
s
stud
y
co
nsi
sts
of
fou
r
m
ai
n
sta
ges,
i.e.
pr
e
-
proces
sing,
segm
entat
ion
, feat
ur
e
s e
xtract
ion
a
nd classi
fi
cat
ion
as
sho
w
n
in
Fig
ure
2.
R
o
I
n
o
d
u
l
e
i
m
a
g
e
P
r
e
-
p
r
o
c
e
s
s
i
n
g
S
e
g
m
e
n
t
a
t
i
o
n
F
e
a
t
u
r
e
s
e
x
t
r
a
c
t
i
o
n
A
c
c
u
r
a
c
y
,
s
e
n
s
i
t
i
v
i
t
y
a
n
d
s
p
e
c
i
f
i
c
i
t
y
I
n
p
u
t
P
r
o
c
e
s
s
O
u
t
p
u
t
Figure
2. Bl
oc
k diag
ram
o
f
th
e pro
posed
sc
hem
e
2.1.
Pre
-
pr
oc
essing
In
t
he
br
east
ul
trasoun
d
im
ages,
no
du
le
sh
a
pe
a
pp
ea
rs
as
a
co
ncen
t
rated
dark
hole
area
.
Firstl
y,
the
rad
i
ologist
m
a
rk
e
d
the
s
pecific
nodule
are
a
of
the
or
i
gin
al
i
m
age
to
ob
ta
in
reg
i
on
of
inte
rest
(R
oI)
a
nd
fo
c
us
e
d
on
th
e
area
a
naly
sis
as
dep
ic
te
d
in
Fi
gure
3.
T
hen,
R
oI
R
GB
i
m
age
was
c
onve
rted
t
o
grey
scal
e
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
Shape
analysis
for
cl
as
sif
ic
ation of
br
e
as
t
nodu
le
s
on digit
al
u
lt
ra
s
ound im
ag
e
s
(
H
anung
Adi Nu
gro
ho
)
839
fo
ll
owe
d
by
filt
ering
proces
s
us
in
g
the
com
bin
at
io
n
of
a
da
ptive
m
edian
filt
er
an
d
SRB
F
to
ov
e
rc
om
e
la
bel
and s
peck
le
no
ise
s.
Ad
a
ptive
m
edian
filt
er
works
base
d
on
a
djacent
pix
el
s
and
is
ca
pa
ble
to
preser
ve
t
he
detai
le
d
inf
or
m
at
ion
of
the
f
ocu
se
d
obje
ct
wh
il
e
co
nd
ucting
noise
re
du
ct
io
n
[14,
15
]
.
Si
m
il
ar
to
Gau
ssian f
il
te
r,
SRBF
is
sta
rted
by
m
easur
in
g
we
igh
t
of
Gau
s
s
ia
n
in
sp
at
ia
l
do
m
ai
n
and
range
of
intensit
y.
The
SRB
F
is
m
at
he
m
at
ic
a
lly
expresse
d
by
(
1
)
.
He
re,
f
is
the
in
pu
t
im
age
an
d
h
is
t
he
ou
tp
ut
im
age.
The
s
patia
l
ad
j
a
cent
of
coor
din
at
e
p
pi
xel
is
descr
ib
ed
by
Ω
(p),
wh
il
e
ξ
is
the
var
ia
ble
com
bi
nation
represe
nting
of
Ω
c
oor
din
at
e
pix
el
s.
T
he
s
pa
ti
al
weigh
t
of
Eucli
dea
n
dist
ance
bet
ween
c
and
ξ
is
f
unct
ion
e
d
by
c,
w
hi
le
s
is
the
weig
ht
that
o
pe
rates inte
nsi
ty
d
om
ai
n
(w
e
igh
t i
nte
ns
it
y).
Figure
3. The
RoI
m
ark
e
d by ra
dio
lo
gists
ℎ
=
−
1
Ω
,
,
(1)
2.
2.
Se
gme
ntati
on
Segm
entat
ion
base
d
on
act
ive
con
t
our
is
co
nducted
for
se
gm
enting
nodu
le
area
and
se
pa
rati
ng
from
it
s
backgroun
d.
The
c
oncept
of
se
gm
entat
ion
pr
ocess
is
e
m
plo
ye
d
by
groupin
g
sim
il
ar
pix
el
s
or
s
ub
-
reg
i
on
into
the
la
r
ge
r
reg
i
on.
It
is
s
ta
rted
by
deter
m
ining
t
he
it
e
rati
on
num
ber
us
e
d
an
d
set
ti
ng
t
he
sta
rtin
g
po
i
nt
wh
ic
h
is
kn
own
al
so
as
seed
po
i
nt.
T
hen
t
he
sta
rting
point
sp
rea
ds
gr
a
du
al
ly
in
su
b
reg
i
on
(r
e
gion
gro
wing)
base
d
on the num
ber
o
f
it
erat
ion
us
ed
. In
thi
s w
ork
, th
e m
a
xim
u
m
n
u
m
ber
o
f
it
erati
on is
set
to f
ifty
it
er
at
ion
s.
Finall
y,
segm
ented
nodule
known
a
s
f
oregro
und
area
is
ob
ta
ine
d;
ot
herwise
it
can
be
cat
e
gori
es
as
backg
rou
nd ar
ea.
In
al
m
os
t
all
c
ases,
nodule
ar
eas
ob
ta
ine
d
from
rad
iolog
ist
s
m
os
tl
y
app
ear
in
the
centre
RoI
im
age.
Ther
e
f
or
e, t
he
sta
r
ti
ng
point i
s d
et
erm
ined
from
the centre o
f
the Ro
I
im
a
ge.
Ho
wev
e
r,
fa
lse
p
os
it
ive ar
ea sti
ll
occurs
i
n
the
segm
ented
im
age
of
act
ive
con
t
our.
Th
us
,
openi
ng
m
or
phologica
l
ope
rati
on
is
ap
plied
t
o
ov
e
rc
om
e this p
r
oble
m
.
2.3.
Fe
ature
ext
r
act
i
on
Feat
ur
e
e
xtract
ion
a
nd
featu
re
sel
ect
ion
are
cond
ucted
to
obta
in
im
po
rtan
t
featur
es
relat
ed
to
s
ha
pe
analy
sis.
Thes
e
featur
e
s
are
us
e
d
f
or
cl
assi
ficat
ion
proces
s.
The
e
xtract
ed
feat
ur
es
a
r
e
Zern
i
ke
m
ome
nts,
inv
a
riant
m
o
m
ents,
r
oundne
s
s
and
sli
m
ness.
Zern
ik
e
m
ome
nts
are
the
ba
sis
of
Zer
nik
e
po
ly
nom
ia
ls
fr
om
x2
+
y2
≤
1
ci
rcle
[18]
as
fo
rm
ulate
d
in
(2).
N
otati
on
r
is
the
rad
i
us
of
the
(
y,
x)
to
the
ce
ntre
of
m
ass,
θ
is
the
ang
le
betwe
en
r
a
nd the
x
-
a
xi
s,
a
nd Rp
q
is
r
adial
orth
ogon
al
p
olyn
om
ia
ls
.
,
=
c
os
,
s
in
=
.
(2)
Seve
n feat
ur
e
s
of in
var
ia
nt m
om
ent
[19]
are
denoted
in
(
3)
–
(9).
∅
1
=
(ŋ2
0
+
ŋ02)
(3)
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
3
, N
o.
2
,
Fe
bru
ary
201
9
:
8
3
7
–
8
4
4
840
Roun
dn
es
s
kn
own
as
c
om
pactness
feat
ur
e
is
a
si
m
ple
and
po
pu
la
r
m
ea
su
re
of
the
e
ffi
ci
ency
of
a
sh
a
pe
co
ntour
.
Slim
ness
or
a
sp
ect
rati
o
is
ob
ta
ine
d
f
r
om
the
rati
o
bet
w
een
wi
dth
an
d
le
ng
th
of
t
he
sh
a
pe.
Roun
dn
es
s a
nd slim
ness
is f
orm
ula
te
d
by t
he
(
4
)
a
nd
(
5
).
=
4
×
2
(4)
=
ℎ
ℎ
(5)
2.4.
Clas
si
ficat
i
on
A
m
ajo
r
ta
s
k
after
feat
ur
e
e
xtracti
on
is
to
cl
assify
the
obj
ect
i
nto
on
e
of
t
he
se
ver
a
l
cat
ego
ries
.
Mult
i
-
la
ye
r
per
ceptr
on
(MLP
)
is
an
exam
pl
e
of
an
arti
fici
al
neu
ral
net
w
ork
that
is
us
ed
exten
sively
fo
r
th
e
so
luti
on
of
a
num
ber
of
diff
e
ren
t
pro
blem
s.
It
is
a
la
ye
re
d
netw
ork
c
om
pr
isi
ng
i
nput
node
s,
hi
dden
node
s
an
d
ou
t
pu
t
node
s a
s sho
wn in Fi
gure
4.
Figure
4
.
Mult
i
-
la
ye
r
pe
rce
ptr
on
3.
RESU
LT
S
A
ND AN
ALYSIS
To
facil
it
at
e
t
he
processin
g
of
breast
ultra
so
un
d
im
age,
or
i
gin
al
RGB
form
at
was
co
nv
e
rted
i
nt
o
gr
ey
scal
e
as
de
scribe
d
in
Figure
5.
Fil
te
r
process
w
as
cond
ucted
to
overc
om
e
so
m
e
com
m
on
pr
ob
lem
s
in
ultraso
und
i
m
a
ges,
su
c
h
as
pr
esence
s
pec
kle
noise
s
a
nd
la
be
ls.
Sele
ct
ion
of
filt
erin
g
te
c
hniq
ue
was
al
s
o
base
d
on m
easur
em
e
nt s
peck
le
i
nd
e
x (SI) as
sho
w
n
in
Fig
ure
6.
(a)
(b)
(c)
Figure
5. The
c
onve
rsion
resu
l
t (a) o
rigin
al
i
m
age (b)
R
oI
RGB i
m
age (
b)
gr
ey
scal
e
of
RoI
im
age
As
s
how
n
i
n
Figure
6,
the
com
bin
at
ion
of
a
dap
ti
ve
m
e
dian
filt
er
a
nd
SRB
F
ob
ta
in
s
the
l
ow
es
t
sp
ec
kle
ind
e
x
(S
I
).
It
in
dicat
es
that
com
bin
at
ion
of
ada
ptive
m
edian
filter
an
d
SRB
F
is
m
or
e
ap
pro
pri
at
e
to
ov
e
rc
om
e
sp
ec
kle
no
ise
tha
n
the
oth
e
r
filt
ering
te
ch
niques
.
In
ad
diti
on,
th
e
com
bin
at
ion
of
a
dap
ti
ve
m
e
dian
and SRB
F m
eth
od is
capa
ble
to elim
inate
the i
m
age label a
s sho
wn in Fi
gure
7.
Inp
u
t lay
er
Hid
d
en
lay
er
Ou
tp
u
t
lay
er
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
Shape
analysis
for
cl
as
sif
ic
ation of
br
e
as
t
nodu
le
s
on digit
al
u
lt
ra
s
ound im
ag
e
s
(
H
anung
Adi Nu
gro
ho
)
841
Figure
6. The
c
om
par
ison s
pe
ckle in
dex of s
ever
al
filt
ering m
e
tho
ds
(a)
Or
i
gin
al
im
ages
(b)
Me
dian fil
te
red
im
ages
(c)
Ad
a
ptive m
edian fi
lt
ered
im
a
ges
(d)
Com
bin
at
ion
of a
dap
ti
ve
m
edian
an
d SRB
F
filt
ered
im
ages
(e)
SRB
F f
il
te
red im
ages
Figure
7.
Com
par
is
on of
or
i
gi
nal i
m
ages and fil
te
red
im
ages u
si
ng v
a
rio
us fil
te
rin
g
m
et
ho
ds
Fil
te
red
i
m
ages
su
bse
que
ntly
underwe
nt
act
ive
co
ntour
-
ba
sed
segm
entat
ion.
The
visu
al
com
par
ison
of
se
gm
entat
ion
res
ult
can
be
seen
in
Fi
gure
8.
T
he
segm
ented
im
ages
of
adap
ti
ve
m
edian
an
d
SRB
F
f
il
te
red
hav
e
the
cl
ose
st
resu
lt
to
t
he
act
ual
nodule
i
m
ages
as
sess
ed
by
rad
i
olog
ist
s.
Segm
ented
no
du
le
was
then
processe
d
to
ta
ke
it
s
featur
e
s.
Feat
ur
e
e
xtr
act
ion
proce
ss
ob
ta
ine
d
ni
ne
featur
es
incl
udin
g
of
rou
ndness,
slim
ness
an
d
s
even
feat
ur
es
of
in
var
ia
nt
m
om
ent.
Furthe
r
,
ext
racted
fea
tures
a
re
cl
ass
ifie
d
base
d
on
MLP
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
3
, N
o.
2
,
Fe
bru
ary
201
9
:
8
3
7
–
8
4
4
842
cl
assifi
er.
T
he
pe
rfor
m
ance
of
trai
ni
ng
an
d
te
sti
ng
featu
res
was
e
valu
at
ed
by
m
easur
in
g
s
om
e
sta
t
ist
ic
al
par
am
et
ers
su
c
h
as
accu
racy,
sensiti
vi
ty
and
sp
eci
fici
ty
u
sin
g (
6
)
-
(
7
).
a
ccur
a
c
y
=
TP
+
TN
TP
+
FP
+
TN
+
FN
×
100%
(6)
sensi
t
iv
ity
=
TP
TP
+
FN
×
100%
(7)
specificit
y
=
TN
TN
+
FP
×
100%
(8)
(a)
O
rigin
al
im
age
(b)
Se
gm
ented im
age b
y m
edian fi
lt
er
(c)
Segm
ented im
age b
y ad
apt
ive m
edian
filt
er
(d)
Se
gm
ented im
age b
y ad
apt
ive m
edian
+ S
RB
F
Figure
8
.
The
c
om
par
ison o
f
i
m
age seg
m
entat
ion
resu
lt
s
Ther
e
are
six
ty
pes
of
cl
as
sific
at
ion
b
ase
d
on
e
xtracted
fea
tures.
T
hey
are
Zer
nik
e
m
o
m
e
nt,
in
va
rian
t
m
o
m
ent,
r
oundne
ss
a
nd
sli
m
ness,
and
th
e
com
bin
at
ion
of
these
th
re
e
kinds
of
fea
tures.
T
he
num
ber
of
featur
e
s
desc
ribes
the
nu
m
ber
of
input
la
ye
rs
in
the
M
LP
cl
assifi
er.
Co
m
par
ison
of
cl
assifi
cat
ion
re
s
ults
of
these feat
ur
es
i
s sho
wn in T
ab
le
1
.
Table
1.
T
he
Cl
assifi
cat
ion
R
esults o
f
t
he
Fe
at
ur
es
of Ze
rn
i
ke
M
om
ents,
I
nv
a
riant M
om
e
nts,
R
oundne
ss
and
Slim
ness
Para
m
et
ers
Extracted f
eatu
res
Nu
m
b
e
r
o
f
f
eatu
re
s
Accurac
y
(
%)
Sen
sitiv
ity
(
%
)
Sp
ecif
icity
(
%
)
Zer
n
ik
e
m
o
m
en
t
28
8
5
.70
7
9
.20
9
1
.20
Inv
ariant
m
o
m
en
t
7
9
0
.50
8
7
.50
9
3
.00
Ro
u
n
d
n
ess
and
sli
m
n
ess
2
9
5
.20
9
7
.90
9
3
.00
Zer
n
ik
e and
inv
ari
an
t
m
o
m
en
ts
35
8
5
.70
8
1
.30
8
9
.50
Inv
ariant
m
o
m
en
t,
rou
n
d
n
ess
,
sli
m
n
es
s
9
9
6
.20
9
7
.90
9
4
.70
Inv
ariant
m
o
m
en
t,
Zer
n
ik
e,
rou
n
d
n
ess
,
sli
m
n
ess
37
9
5
.20
9
3
.80
9
6
.50
As
de
picte
d
in
Table
1,
t
he
co
m
bin
at
ion
of
ni
ne
extracte
d
f
eat
ur
es
c
on
sist
ing
of
r
ound
ne
ss,
slim
ness
and
se
ve
n
feat
ur
es
of
in
var
ia
nt
m
o
m
ent
ob
t
ai
ns
the
best
c
la
ssific
at
ion
re
su
lt
s
with
accu
racy,
sensiti
vity
and
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
Shape
analysis
for
cl
as
sif
ic
ation of
br
e
as
t
nodu
le
s
on digit
al
u
lt
ra
s
ound im
ag
e
s
(
H
anung
Adi Nu
gro
ho
)
843
sp
eci
fici
ty
96
.
20%,
97.
90%
and
94.70%.
More
ov
e
r,
Ze
r
nik
e
m
o
m
ents
do
not
hav
e
sign
ific
a
nt
effe
ct
of
cl
assifi
cat
ion
ei
ther
in
sel
f
or
com
bin
at
ion
with
othe
r
feat
ur
es
.
The
c
ombinati
on
of
Ze
rn
i
ke
m
o
m
ents
with
inv
a
riant
m
ome
nts,
rou
ndnes
s
an
d
slim
ness
eve
n
re
du
ces
the
re
su
lt
s
of
cl
assifi
cat
ion
c
om
par
ed
t
o
t
ha
t
of
without
Zer
ni
ke
m
o
m
ents.
More
ov
e
r,
a
c
om
par
ison
of
our
res
ults
to
that
of
oth
e
r
publishe
d
m
eth
ods
is
pr
ese
nted
in
T
able
2.
As
sho
wn
i
n
Ta
ble
2,
the
pro
pose
d
m
et
ho
d
is
c
ompara
ble
to
oth
e
r
existi
ng
m
et
ho
ds
by
m
ai
ntaining
hi
gh accu
racy, se
ns
it
ivit
y a
nd spec
ific
it
y rate
s.
Table
2.
C
om
par
iso
n of
Re
su
l
ts t
o Other
Met
hods
Accurac
y
(
%)
Sen
sitiv
ity
(
%
)
Sp
ecif
icity
(
%
)
Hu
an
g
et
al.
[
1
6
]
8
2
.2
9
4
.10
N/A
Tah
m
asb
i et
al.
[
1
3
]
N/A
9
7
.60
9
7
.50
Ro
u
h
i et
al.
[
1
2
]
9
6
.47
9
6
.87
9
5
.94
Prop
o
sed
app
roach
9
6
.20
9
7
.90
9
4
.70
4.
CONCL
US
I
O
N
This
pap
e
r
pro
po
s
es
a
sc
hem
e
to
cl
assify
nodule
s
ha
pe
of
breast
ultras
ound
im
ages.
T
he
sc
hem
e
consi
sts
of
f
our
m
ai
n
sta
ges,
i.e.
pre
-
proces
sing,
segm
entat
ion
,
featu
re
e
xtracti
on
an
d
cl
assifi
cat
ion
.
At
th
e
i
m
age
enh
a
nce
m
ent
sta
ge,
th
e
com
bin
at
ion
m
et
ho
d
of
a
da
ptive
m
edian
f
il
te
r
and
SRB
F
is
able
to
el
im
inate
the
la
bels
of
the
ultraso
und
br
east
nodule
i
m
ages.
T
he
cl
assifi
cat
ion
of
nodule
sh
a
pes
was
div
ide
d
i
nt
o
tw
o
cl
asses,
i.e.
re
gu
la
r
and
ir
regular
cl
asses.
T
he
best
pe
rform
ance
of
cl
assifi
cat
ion
is
ob
t
ai
ned
by
us
i
ng
nin
e
extracte
d
feat
ures
wh
ic
h
c
om
pr
ise
of
se
ve
n
feat
ur
e
s
of
inv
a
riant
m
ome
nts,
r
oundne
ss
an
d
sli
m
ness.
T
he
resu
lt
s
of
cl
ass
ific
at
ion
proce
ss
achie
ve
the
le
vel
of
acc
ur
a
cy
,
sensiti
vity
and
spe
ci
fici
ty
at
96.
20%,
97
.90%
and
94.70%,
r
especti
vely
.
Th
ese
res
ults
ind
ic
at
e
that
the
pr
op
os
e
d
has
a
po
te
ntial
to
be
i
m
ple
m
ente
d
in
a
com
pu
te
rised
a
ided diag
nosis
syst
e
m
.
In
fu
t
ur
e
resea
rch,
the
oth
e
r
m
al
ign
ancy
pa
ram
et
er
su
ch
a
s
poste
rio
r
ac
ousti
c
pa
ram
et
e
r
can
be
us
e
d
to
analy
se
the
br
east
no
du
le
s f
or
furthe
r
i
m
pr
ovem
ent
of
the
rad
iol
og
ist
di
agnosis.
I
n
ad
di
ti
on
,
to
i
m
pr
ove
the
perform
ance o
f
this r
esea
rc
h,
t
he other
selec
ti
on m
et
ho
ds
ca
n be d
e
vel
op
e
d wit
h
the
o
t
her
featur
e
s for a
be
tt
er
cl
assifi
cat
ion
r
esult an
d
acce
l
erate t
he
c
om
pu
ta
ti
on
proce
ss
.
ACKN
OWLE
DGE
MENTS
This
researc
h
work
is
fun
de
d
by
Di
rectora
te
Gen
e
ral
of
Higher
Ed
ucat
ion
,
Mi
nistry
of
Re
searc
h,
Tech
no
l
og
y
a
nd
Hi
gh
e
r
E
du
c
at
ion
,
Re
public
of
Ind
on
e
sia
.
The
a
uthors
w
ou
l
d
li
ke
to
tha
nk
ra
dio
lo
gists
from
the
Dep
a
rtm
e
nt
of
Ra
di
ology,
RSUP
Sa
rdjito
an
d
Ha
rdj
olukit
o
H
ospit
al
fo
r
c
oope
rati
ng
a
nd
s
har
i
ng
exp
e
riences
.
We
w
ould
al
so
li
ke
to
than
k
t
he
I
ntell
igent
Syst
e
m
s
resear
ch
gro
up
m
e
mb
ers
i
n
our
De
par
tm
ent
for
s
har
i
ng all
o
f
m
eaning
f
ul
knowle
dge.
REFERE
NCE
S
[1]
C.
-
M.
Chen,
Y.
-
H.
Chou,
K.
-
C.
Han,
G.
-
S.
Hung,
C.
-
M.
Ti
u
,
H.
-
J.
Chiou
,
e
t
al.
,
"Brea
st
le
sions
on
sonograms:
computer
-
ai
d
ed dia
gnosis with
n
ea
rl
y
se
tt
ing
-
ind
epe
nden
t
fe
at
ure
s a
nd
art
if
ic
i
al
n
eur
al
n
et
works
1
,
"
Radi
olog
y,
vo
l
.
226,
pp
.
504
-
51
4,
2003
.
[2]
Y.
L.
Huang,
D.
R.
Chen,
Y.
R
.
Jiang
,
S.
J
.
K
uo,
H.
K
.
W
u,
and
W
.
Moon,
"Com
pute
r
ai
de
d
dia
gnosis
usin
g
m
orphologi
ca
l
f
ea
tur
es
for
c
la
ss
if
y
ing
bre
ast
l
esi
ons
on
ult
rasoun
d,
"
Ultrasound
i
n
Obs
te
tric
s
&
Gyne
col
og
y,
vo
l
.
32,
pp
.
565
-
572
,
2008.
[3]
S.
W
u,
Q.
Zhu,
and
Y.
Xie,
"Eva
luation
of
var
iou
s
spec
kle
red
uct
ion
filters
on
m
edi
ca
l
ult
r
asound
images,
"
in
Annual
In
te
rnati
onal
Conf
ere
nce of
the IE
E
E Eng
ine
ering
in Medi
ci
ne
and
Bi
olog
y
Society
,
2013
,
p
p.
1148
-
1151
.
[4]
D.
A.
Khus
na,
H
.
A.
Nugroho
,
an
d
I.
Soesan
ti
,
"P
erf
orm
anc
e
anal
y
sis
of
edge
and
det
a
iled
pre
serv
ed
spec
k
le
no
ise
red
uction
fil
t
ers
for
bre
ast
ult
ras
ound
images,
"
in
Inte
rnational
Confe
renc
e
on
Information
Technol
ogy,
Comput
er,
and
Elec
tri
cal E
ngine
ering
,
201
5,
pp
.
76
-
80
.
[5]
H.
K.
N.
Yus
ufi
y
ah
,
H.
A.
Nugroho,
T.
B.
Adji,
and
A.
Nugroh
o,
"F
ea
ture
ext
ra
ction
for
cl
assif
y
i
ng
le
sion'
s
shape
of
bre
ast
ult
r
asound
images,
"
in
2015
2nd
Int
ernati
onal
Con
fe
r
enc
e
on
Informa
ti
on
Techno
logy,
Computer,
and
El
e
ct
rica
l
Eng
in
ee
ring (
ICITACEE
)
,
2015,
pp.
1
02
-
106.
[6]
A.
K.
Jum
aa
t,
W
.
E.
Z.
W
.
A.
Rahman,
A.
I
bra
him,
and
R.
Mahm
ud,
"S
egmenta
ti
on
of
m
asses
from
bre
a
st
ult
rasound
imag
es
using pa
ramet
ric
activ
e
cont
ou
r
al
gorit
hm
,
"
Pr
oce
dia
-
So
ci
al
an
d
Be
havi
oral
S
cienc
es,
vo
l.
8,
pp
.
640
-
647,
2010
.
[7]
A.
Nugroho,
H.
A.
Nugroho,
an
d
L.
Chorid
ah,
"A
ct
ive
cont
ou
r
bil
atera
l
filter
fo
r
bre
ast
l
esions
segm
ent
at
ion
o
n
ult
rasound im
age
s,"
in
Inte
rnat
io
nal
Conf
ere
nce
on
Scienc
e
in
Inf
orm
ati
on
Techn
ology
,
2015,
pp.
36
-
40.
[8]
M.
Savel
onas
,
D.
Marouli
s,
D
.
Iak
ovidi
s,
S.
Karka
nis,
and
N.
Dim
it
ropoul
os,
"A
var
ia
bl
e
bac
kgro
und
active
cont
our
m
odel
f
or
aut
om
atic
d
etec
t
ion
of
th
y
roid
nodule
s
in
ul
tras
ound
images,
"
i
n
IEEE
Int
ernational
Confe
r
ence
on
Image Proce
s
sing
,
2005,
pp.
I
-
17
-
20.
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
3
, N
o.
2
,
Fe
bru
ary
201
9
:
8
3
7
–
8
4
4
844
[9]
T.
F.
Chan
and
L.
A.
Vese,
"A
ctive
cont
ours
wit
hout
edge
s,"
IE
EE
Tr
ansacti
ons
on
image
p
roce
ss
ing,
vol.
10,
pp.
266
-
277,
2001
.
[10]
S.
La
nkton
an
d
A.
Ta
nnenbaum
,
"Loc
al
i
zi
n
g
reg
ion
-
base
d
ac
ti
v
e
cont
our
s,"
IEE
E
transacti
ons
on
imag
e
proce
ss
ing,
vol
.
17,
pp
.
2029
-
20
39,
2008
.
[11]
A.
Nugroho,
H
.
A.
Nugroho,
N
.
A.
Setiawan,
an
d
L.
Chorid
ah,
"
Inte
rna
l
con
te
n
t
cl
assifi
ca
t
ion
of
ult
rasound
th
y
ro
id
nodule
s ba
sed
o
n
te
x
tura
l
fe
at
ur
es,
"
Comm
unic
a
ti
ons i
n
Scienc
e and Tec
hnolog
y,
vol. 1, 2016.
[12]
R.
Rouhi,
M.
Jafa
ri,
S.
Kasae
i,
a
nd
P.
Keshava
rzi
an,
"Beni
gn
an
d
m
al
igna
nt
bre
ast
tumors
cl
assific
a
ti
on
base
d
o
n
reg
ion
growing
and
CNN
segm
e
nta
ti
on
,
"
Ex
pert
Syste
ms
wit
h
Ap
pli
cations,
vo
l. 4
2,
pp
.
990
-
1002
,
2015.
[13]
A.
Ta
hm
asbi,
F.
Saki,
and
S.
B.
Shokouhi,
"Classificat
ion
of
beni
gn
and
m
al
igna
nt
m
asses
base
d
on
Ze
rni
ke
m
om
ent
s,"
Com
pute
rs
in Bi
olog
y
and
M
edi
ci
ne
,
vol.
41
,
pp
.
726
-
735,
2011
.
[14]
M.
Yang,
K.
Kp
al
m
a,
and
J.
Ron
sin,
"A
surve
y
of
shape
fe
at
ur
e
ex
tra
c
ti
on
t
ec
hn
iqu
es,
"
Pattern
rec
o
gnit
ion,
pp.
43
-
90,
2008
.
[15]
H.
A.
Nugroho,
H.
K.
N.
Yus
ufi
y
ah,
T.
B
.
Adji,
and
A.
N
ugroho,
"Ze
rn
ik
e
m
om
ent
fea
tu
re
e
xtr
ac
t
ion
fo
r
cl
assif
y
ing
le
sio
n'
s
shape
of
br
eas
t
ult
rasound
images,
"
in
Int
ernati
onal
Con
fe
ren
c
e
on
In
formation
Technol
og
y
and
El
e
ct
rica
l
Eng
in
ee
ring
,
2015,
pp.
458
-
463.
[16]
Y.
-
L.
Huang
,
"Com
pute
r
-
ai
de
d
dia
gnosis
using
neur
al
n
et
w
orks
and
support
vec
t
or
m
achine
s
for
bre
as
t
ult
rasonogra
ph
y
,
"
Journal
of
Me
dic
al
Ultrasound,
vol
.
17
,
pp
.
17
-
24,
2009
.
[17]
W.
-
J.
W
u
and
W
.
K.
Moon,
"U
lt
rasound
b
rea
st
tumor
im
age
comput
er
-
a
ide
d
di
agn
osis
with
te
x
ture
an
d
m
orphologi
ca
l
f
ea
tur
es,
"
A
cadem
ic
radiolog
y,
v
ol.
15
,
pp
.
873
-
8
80,
2008
.
[18]
M.
R.
Teague
,
"I
m
age
an
aly
s
is via
th
e
g
ene
r
al
theor
y
of
m
om
ent
s,
"
JOSA
,
vol
.
70
,
pp.
920
-
930
,
19
80.
[19]
M.
-
K.
Hu,
"V
is
ual
p
at
t
ern
r
ec
o
gnit
ion
b
y
m
om
ent
inv
ariant
s,"
I
RE
transacti
ons
on
inf
orm
ati
on
t
heory
,
vo
l.
8,
pp
.
179
-
187,
1962
.
Evaluation Warning : The document was created with Spire.PDF for Python.