Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
13
,
No.
1
,
Jan
uar
y
201
9
,
pp.
405
~
410
IS
S
N:
25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
1
.pp
405
-
410
405
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Classific
atio
n
en
hancem
en
t o
f b
reast canc
er histop
atholo
gica
l
image us
ing pena
lized l
og
i
stic r
egressi
on
Moham
med
A
bdulraz
aq K
ahy
a
Depa
rtment
o
f
C
om
pute
r
scie
n
ce,
Educat
ion
Col
lege
for
Pure
Sci
e
nce
,
Univer
si
t
y
of
Mos
ul,
Mos
ul
,
Ir
aq
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
un
1
9
, 2
018
Re
vised
A
ug
2
1,
2018
Accepte
d
Nov
18
, 201
8
Cla
ss
ifi
c
at
ion
of
bre
ast
ca
n
ce
r
hist
opat
hologica
l
im
age
s
pl
a
y
s
a
sign
ifi
c
ant
ro
le
in
computer
-
ai
d
e
d
di
agnosis
s
y
s
tem
.
Feat
ur
es
m
at
r
ix
was
ext
r
acte
d
in
ord
er
to
cl
assif
y
those
images
and
th
e
y
m
a
y
contain
out
li
e
r
val
ues
adv
erse
l
y
th
at
aff
e
c
t
the
class
ifi
c
at
ion
per
form
an
ce
.
S
m
oothi
ng
of
feat
ure
s
m
at
rix
h
as
bee
n
prove
d
to
be
an
eff
ectiv
e
wa
y
t
o
improve
the
class
ifi
c
at
i
on
result
v
ia
el
i
m
ina
ti
ng
o
f
outl
ie
r
va
lue
s.
I
n
thi
s
pap
er,
a
n
ada
p
ti
ve
pen
a
li
z
ed
log
isti
c
re
gre
ss
ion
is
proposed,
with
t
he
ai
m
of
sm
oothi
ng
fe
at
ure
s
an
d
provide
s hi
gh
cl
assifi
ca
t
ion
ac
cur
acy
of
hist
opat
hologica
l
i
m
age
s,
b
y
combin
ing
th
e
pen
alize
d
logi
sti
c
reg
ression
wi
th
t
he
sm
oothe
d
fe
atures
m
at
rix
.
Exp
eri
m
ent
a
l
r
esult
s
base
d
on
a
publi
cly
re
ce
nt
bre
ast
ca
nc
er
hi
stopat
hologica
l
i
m
age
dataset
s
s
how
tha
t
the
proposed
m
et
ho
d
significantl
y
o
utpe
rform
s
penalized
log
isti
c
re
gre
ss
ion
in
te
rm
s
of
cl
assifi
ca
t
ion
a
cc
ura
c
y
and
ar
ea
und
er
t
he
cur
v
e.
Thus,
t
he
propose
d
m
et
hod
ca
n
b
e
useful
for
histo
pat
hologica
l
images
class
ifi
c
at
i
on
and
other
cl
assifi
ca
t
ion
of
disea
ses
t
y
p
es
u
sing
DN
A
gene
expr
ession
da
ta
in
the
re
al
cl
inica
l
pr
ac
t
ic
e
.
Ke
yw
or
d
s
:
Breast
cance
r
Histo
path
ologica
l im
age
L1
-
norm
Pe
naliz
ed
l
ogist
ic
r
egr
e
ssio
n
Sm
oo
thing
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
:
Moh
am
m
ed
A
bdulraza
q Ka
hy
a
,
Dep
a
rtm
ent o
f
Com
pu
te
r
sci
e
nce
,
Ed
ucati
on Coll
ege
for
P
ure Sc
ie
nce
,
Un
i
ver
sit
y o
f M
os
ul,
Mos
ul, Ira
q.
Em
a
il
:
m
oh
a
m
m
edk
ahya
@uo
m
os
ul.ed
u.i
q
1.
INTROD
U
CTION
Nowa
days,
ca
ncer
is
t
he
sec
ond
le
adin
g
ca
us
e
of
death
w
or
l
dw
i
de.
O
n
t
he
oth
e
r
ha
nd,
the
Wo
rl
d
Healt
h
O
rg
a
niz
at
ion
(WH
O)
c
onfirm
ed
that
8
.
2
m
il
l
ion
d
ea
ths w
ere
ca
us
e
d
by
ca
nce
r
i
n
2012
a
nd 8
.
8 m
illi
on
in
20
15.
M
or
e
ov
e
r,
it
ex
pect
ed
27
m
illi
on
of
ne
w
ca
ses
o
f
this
disease
be
fore
2030
[1]
.
I
n
par
ti
cula
r,
br
eas
t
cancer
is
one
of
t
he
le
adin
g
ca
us
es
of
w
om
en'
s
death
in
the
world
.
A
rece
nt
stu
dy
c
onfirm
ed
that
breast
c
ancer
accounts
for 1
8% of al
l t
ypes
of wom
en
can
cers a
nd the
fif
th r
eas
on
of d
e
at
h
in t
he worl
dw
i
de
[2]
.
Howe
ver,
t
he
e
arly
sta
ge
dia
gnos
is
a
nd
the
ra
py
ca
n
i
ncr
eas
e
the
sur
viv
al
r
at
es
to
98%
[3]
.
The
re
are
m
any
noninva
sive
im
aging
te
chn
iq
ues
f
or
br
ea
st
cance
r
su
c
h
as
m
agn
et
ic
res
on
a
nc
e
i
m
aging
(MRI),
m
a
m
m
og
ra
m
s
(
X
-
rays)
,
ultr
aso
nogr
a
phy
a
nd
histo
path
olo
gical
im
age
[4
-
7]
.
Diag
no
s
is
us
i
ng
histol
og
ic
al
i
m
ages h
as
bec
om
e a p
ower
ful
g
ol
d
sta
nda
rd f
or
dea
dly dis
eases suc
h
a
s breast
and l
ung ca
ncer
s
, whic
h give
s
a sati
sfactor
y
di
agnosis c
om
par
ed
w
it
h ot
her m
e
tho
ds s
uch
as m
a
m
m
og
raphy an
d ult
ras
onog
raphy
[
8]
.
On
the
ot
her
ha
nd,
m
achine
l
earn
i
ng
te
ch
ni
qu
e
s
hav
e
be
e
n
us
e
d
to
e
nh
a
nce
t
he
diag
nosti
c
accuracy
for
br
east
ca
nc
er
th
rou
gh
a
c
o
m
pu
te
r
-
assist
ed
syst
em
[9]
.
I
n
gen
e
ral,
br
ea
st
cancer
is
cl
assifi
ed
i
nto
be
ni
gn
a
nd
m
al
ign
ant ty
pe
s and t
his d
ia
gnos
is
is
ver
y i
m
po
rtant in
dr
ug d
isc
overy a
nd treat
m
ent
[10
-
11]
.
Lo
gisti
c
re
gr
es
sion
(
LR)
is
c
onside
red
on
e
of
the
fam
ou
s
m
a
chine
le
ar
ning
t
echn
i
qu
e
s
of
cl
assifi
cat
io
n
su
c
h
as
sup
port
vect
or
m
achines
(
SV
M
),
r
andom
forests
(RF
),
a
nd
ne
ur
al
netw
orks
(NNet)
[
12
]
.
L
og
ist
ic
regressio
n
is
a
n
exte
ns
ive
cl
a
ssific
at
ion
te
ch
nique
an
d
has
m
any
app
li
ed
f
ie
lds
li
ke
ge
ne
expressi
on
d
at
a
[13]
,
pr
e
dicti
on of t
her
a
py
ou
tc
ome
[
14
]
a
nd pr
otein fu
nction
[15
]
.
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.
13
, N
o.
1
,
Ja
nu
a
ry 20
19
:
4
05
–
4
10
406
The
cl
assi
ficat
ion
pe
r
form
ance
im
pr
ov
em
ent
is
the
co
re
of
t
he
br
ea
st
cance
r
histo
path
olog
ic
al
i
m
age
cl
assifi
cat
ion
to
inc
rease
the
diag
nosti
c
accuracy
th
r
ough
the
feat
ur
es
sel
ect
ion
process
[16
-
17]
,
pre
-
proces
s
e
d
i
m
age
[18
-
19]
,
or
a
ny
oth
er
te
chn
i
qu
e
s.
How
ever,
the
pro
posed
m
et
ho
d
dif
fer
s
from
pr
evi
ou
s
te
ch
niques
in
th
e
pr
e
proc
essi
ng
act
ion
of
the
f
eat
ur
es
m
at
rix
wh
ic
h
ai
m
s
to
el
i
m
inati
ng
the
ou
tl
ie
r
va
lues
in
these
feat
ures
to
increase t
he
cl
assifi
cat
ion
acc
ur
acy
t
hroug
h
sm
oo
thing feat
ur
es
m
at
rix
data proces
s.
2.
THE
PROPO
SED
METHO
D
2.1
.
Pe
na
li
z
ed Lo
gistic
Re
gr
ession
Lo
gisti
c
regres
sion
is
one
of
the
powe
rful
c
la
ssific
at
ion
al
gorithm
s
that
is
com
par
at
ively
easy
and
rob
us
t for class
ific
at
ion
b
et
we
en
tw
o
cl
asses.
I
n
this
pa
per
, l
og
ist
ic
r
e
gr
es
sion
tec
hn
i
qu
e
w
as u
se
d
to il
lus
trat
e
the
relat
ionshi
p
bet
ween
i
ndepende
nt
va
riables
(
br
east
c
ancer
histo
pat
ho
l
og
ic
al
im
a
ge
feat
ur
es
)
a
nd
t
he
var
ia
ble of
res
pons
e
(1
for
th
e b
e
nign class
or 0 f
or the m
al
ign
ant cla
s
s).
Let
we
ha
ve
n
ind
e
pe
nd
e
nt
ob
s
er
vations
ii
y
,
x
;
i
1
,
2
,
.
.
.
,
n
wh
e
re
i
y
1
,
0
are
re
sp
on
s
e
var
ia
bles, a
nd
T
i
i
1
i
p
x
x
,
.
.
.
,
x
is a vect
or of i
m
age f
eat
ures.
Con
se
quently
, t
he
lo
gisti
c re
gressi
on m
od
el
i
s
exp
la
ine
d as
i
i
i
P
r
o
b
(
y
1
:
x
)
(
x
)
,
(
1
)
i
i
i
P
r
o
b
(
y
0
:
x
)
1
(
x
)
,
(
2
)
This
pro
bab
il
it
y ca
n be e
xp
la
i
ned as
fo
ll
ows:
T
i
T
i
ii
T
i
i
e
x
p
x
(
x
)
(
x
)
,
l
o
g
x
,
1
(
x
)
1
e
x
p
x
(
3
)
wh
e
re
T
i
i
1
1
i
2
2
i
p
p
x
x
x
.
.
.
.
x
.
The
l
og
-
li
kelih
ood f
un
ct
io
n f
or r
es
pons
e
v
a
r
ia
bles
i
y
can
be writt
en
as:
n
TT
i
i
i
i1
l
(
)
=
y
x
l
o
g
1
e
x
p
x
,
(
4
)
The
pen
al
iz
ed
l
og
ist
ic
re
gr
es
sion
m
od
el
(
PLR
)
ad
ds
a
nonne
gative
pe
nalty
te
rm
to
Equ
at
io
n
(
4
)
,
an
d
is
de
fine
d
as
f
ollows:
p
n
TT
i
i
i
j
i
1
j
1
l
(
)
=
y
x
l
o
g
1
e
x
p
x
,
(
5
)
Mi
ni
m
iz
ing
th
e PLR
functi
on
g
ive
s
us
the
param
et
ers
and
[
20
-
21]
.
2.2
.
His
topat
ho
lo
gical
Ima
ges
Fe
atures
Extr
act
i
on
In
t
his
pap
e
r,
discrete
wav
el
et
trans
form
was
us
e
d
t
o
dec
om
po
se
hist
opat
ho
l
ogic
al
im
ages
of
breast
cancer
[
22]
.
P
r
eci
sel
y,
each
im
age
was
dec
om
po
sed
to
le
vel
VII
base
d
on
Haar
disc
re
te
wa
velet
tra
nsfo
r
m
to
extract
t
he
feat
ur
es
[23]
.
Le
vel
I
of
im
age
dec
om
po
sit
io
n
giv
e
s
fou
r
e
qu
al
siz
e
of
s
ub
-
im
ages,
nam
el
y
A
1
(appro
xim
a
ti
on
coe
ff
ic
ie
nts
)
,
H
1
(
horizo
nt
al
coef
fici
ent
s),
V
1
(v
e
rtic
al
coef
fici
ent
)
and
D1
(d
ia
gonal
coeffic
ie
nt).
T
hen,
the
ne
xt
le
vel
dec
om
po
sit
ion
is
base
d
onl
y
on
t
he
pre
vi
ous
A
of
the
pre
vious
deco
m
posit
ion
.
Ther
e
f
or
e,
le
ve
l
II
of
im
age
deco
m
po
sit
io
n
gi
ves
an
oth
e
r
four
e
qu
al
siz
e
of
sub
-
im
ages,
nam
el
y
A2
,
H2,
V2
and
D
2
res
ult
of
dec
om
po
sit
ion
A
1.
T
he
de
com
po
sit
ion
c
on
ti
nues
unti
l
r
eachin
g
th
e
le
ve
l
VII.
Th
us
,
twenty
-
ei
gh
t
s
ub
-
im
ages
are
deco
m
po
se
d.
Ne
xt,
thr
ee
ne
w
s
ub
-
im
ages
are
ge
ne
r
at
ed
from
the
colo
r
c
ha
nn
el
s
(
red,
gr
ee
n,
bl
ue)
of each
s
ub
-
im
age.
T
hus,
t
he
or
i
gin
al
im
age
is deco
m
po
se
d
t
o
the 28 x
3
f
rom
su
b
-
im
ages.
The
n,
nin
e
of
t
he
tra
di
ti
on
al
sta
ti
sti
c
al
sta
nd
a
rd
s
(m
ean
,
m
ean
abs
ol
ute
de
viati
on
,
m
edian
abs
olu
t
e
dev
ia
ti
on,
sta
nd
a
r
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
Cl
as
sif
ic
ation
enha
ncem
e
nt
of
b
re
as
t ca
ncer
h
ist
op
atholo
gical im
ag
e
…
(
Mo
hamm
e
d Ab
du
lr
azaq K
ahy
a
)
407
dev
ia
ti
on,
entr
op
y,
ene
r
gy,
s
kewness
,
kurto
sis,
r
oo
t
m
ean
sq
ua
re
)
are
ext
racted
from
ever
y
sub
-
im
age.
As
a
resu
lt
, 7
56 f
eat
ur
es
h
a
ve bee
n o
btained
fro
m
each hist
opat
holo
gical
im
age
.
2.3
.
Sm
oot
hing D
ata of Fe
at
u
res
Matrix
Suppose
a
seq
uen
ce
of
data
po
i
nts
(i.e.:
a
f
eat
ur
e
in
patte
rn
rec
ogniti
on,
a
ge
ne
i
n
ge
ne
ex
pressi
on
data
or
a
var
ia
ble
in
sta
ti
sti
cs)
are
giv
e
n
whic
h
re
pr
ese
nt
t
he
c
har
act
erist
i
c
featu
res
f
or
kinds
of
cl
asse
s.
This
data
is
often
c
onta
ined
outl
ie
r (extrem
e)
valu
es as
no
ise
fo
r t
he
cl
assifi
cat
i
on
pro
blem
in data
m
ining fie
ld.
To
reduce
this
no
i
se
data
(R
ough
data),
we
ca
n
consi
der
t
he
se
qu
e
nce
of
data
po
i
nts
as
a
dis
crete
sign
al
in
tim
e
do
m
ai
n
us
i
ng the
dig
it
al
f
il
te
r
s in
s
i
gn
al
proc
essing.
Digital
filt
ers
t
echn
i
qu
e
s
a
re
us
e
d
t
o
e
xtract
be
ne
fici
al
pa
rts
of
the
sig
nal
or
to
cl
ear
ou
t
unwelcom
e
par
ts
of the
sig
nal
[24
-
26]
.
Fi
gure
1
cl
ari
fies
the
basic idea
of the
filt
er.
Row
sig
nal
→
FILT
ER
→ F
il
te
re
d
sig
nal
Figure
1
.
Fil
te
r
Bl
ock D
ia
gr
a
m
In
gen
e
ral,
the
re
are
se
ver
al
dig
it
al
filt
ers
te
chn
i
qu
e
s
to
s
m
oo
th
data
suc
h
as
m
ov
ing
aver
a
ge,
l
ocal
regressio
n
(lo
wess
a
nd
l
oess
),
a
nd
r
obus
t
local
re
gr
es
sio
n
(
rlo
wess
an
d
rloess
)
an
d
Sa
vitzky
-
Go
la
y
[
24
-
28]
.
This
pa
per
us
e
s
the
m
ov
ing
a
ver
a
ge
te
ch
nique
w
hich
c
onsidere
d
as
the
m
os
t
com
m
on
dig
it
al
filt
er
in
sign
a
l
processi
ng
to
e
ase
the
cal
cula
ti
on
an
d
underst
and
i
ng.
I
n
a
s
i
m
ple
way
t
he
work
of
m
ov
in
g
ave
rage
te
ch
nique
can
be
su
m
m
arized
as
f
ollo
w,
if
we
ha
ve
an
arr
ay
of
raw
(
Roug
h)
data
x
(
1
)
,
x
(
2
)
,
.
.
.
.
,
x
(
N
)
,
it
can
be
ref
i
ne
d
to a
ne
w
a
rr
ay
of sm
oo
the
d d
at
a
x
(
1
)
,
x
(
2
)
,
.
.
.
.
,
x
(
N
)
. T
he
sm
oo
thed
point
x
(
k
)
equal
the a
ver
a
ge
nu
m
ber
of
an
odd
neig
hb
or
points
f
or
t
he
c
urren
t
poi
nt.
T
he
fo
ll
owing
f
or
m
ula
re
pr
ese
nts
t
he
e
quat
ion
of
t
he
m
ov
ing
aver
a
ge fil
te
r.
m
im
1
x
(
k
)
x
(
k
i
)
,
m
1
,
2
,
3
,
.
.
.
2m
1
(
6
)
The
odd
nu
m
ber
2
m
1
is
al
ways
na
m
ed
filt
er
s
pa
n.
S
ubseq
ue
ntly
,
the
sm
oo
thed
f
eat
ur
es
m
at
rix
dat
a
(
Fig
ur
e
2
)
is
use
d for classi
fic
at
ion
pro
blem
.
11
1
p
11
1
p
n
1
n
p
n
1
n
p
S
m
oo
t
h
Ev
e
r
y
F
e
a
t
ur
e
s
C
ol
um
n
x
x
x
x
S
a
m
pl
e
s
.
x
x
x
x
D
e
f
a
ul
t
Mov
i
ng
S
m
oo
t
he
d
F
e
a
t
ur
e
s
Ma
t
r
i
a
v
e
r
a
x
F
e
a
t
ur
e
s
Ma
t
r
e
ix
g
F
IL
T
E
R
Figure
2
.
Feat
ures Ma
trix
D
at
a
Sm
oo
thin
g
P
ro
ces
s
2.4
.
Br
e
as
t
Cancer
Classi
fi
cat
i
on
Af
te
r
the
prep
r
ocessin
g
f
or
t
he
featu
res
m
at
rix,
the
PLR
with
sm
oo
th
featu
res
m
at
rix
is
ut
il
iz
ed
to
get
hi
gh
cl
assifi
cat
ion
a
ccur
acy
.
The
de
ta
il
ed
of the
Ad
a
ptive
PLR
(A
P
LR) c
om
pu
ta
ti
on
is
d
es
cr
ibed
i
n
Algorit
hm
1
.
Algori
th
m
1
: T
he co
m
pu
tat
ion
of A
PL
R
Step
1: E
xtract
f
eat
ures m
at
rix via wa
velet
t
ran
s
f
or
m
.
Step
2: Sm
oo
th
featur
e
s m
at
rix
via m
ov
i
ng
aver
a
ge
te
c
hn
i
qu
e
(
Fig
ur
e
2
).
Step
3: S
olv
e t
he APLR,
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.
13
, N
o.
1
,
Ja
nu
a
ry 20
19
:
4
05
–
4
10
408
p
n
TT
i
i
i
j
i
1
j
1
A
L
R
=
a
r
g
M
i
n
l
o
g
1
e
x
p
x
y
x
.
(
7
)
3.
RESEA
R
CH MET
HO
D
3.1
.
Datasets
Descripti
on
The
database
that
ha
s
be
en
use
d
is
t
he
Bre
aKH
is
(T
he
B
reast
Ca
nce
r
Histo
path
ologica
l
Im
ages),
BreaK
His
data
base
s
upplies
us
with
7,9
09
of
m
ic
ro
sco
pic
biopsy
i
m
ages
wh
i
c
h
ha
ve
i
nclu
ded
t
wo
ty
pes
of
ben
i
gn
an
d
m
a
li
gn
ant
t
um
or
s
that
ha
d
colle
ct
ed
f
ro
m
82
patie
nts
usi
ng
diff
e
re
nt
m
agn
ify
ing
f
act
ors:
40X
,
100X,
200X,
a
nd
400X
[
4]
.
T
he
avail
able
his
top
at
holo
gical
i
m
age
s
of
tr
ue
colo
rs
in
P
or
ta
ble
Netw
ork
G
raphics
(P
N
G
)
f
or
m
at
with
700
×
460
pi
xels’
r
e
so
luti
on
are
th
e
ra
w
im
ages
of
neither
nor
m
al
iz
ation
nor
col
or
sta
nd
a
rd
iz
at
io
n.
The
se
im
ages
are
ac
qu
i
red
in
RGB
cha
nne
ls.
A
s
umm
ary
of
this
databas
e
is
li
ste
d
in
T
able
1
and sam
ples o
f
these im
ages in
Fi
gure
3
.
Table
1
.
Su
m
m
ary o
f
the
Brea
KH
is
data
base
Magn
if
icatio
n
Ben
ig
n
Malign
an
t
Total
40X
625
1370
1995
100X
644
1437
2081
200X
623
1390
2013
400X
588
1232
1820
Total
2480
5429
7909
#
Patients
24
58
82
Figure
3
.
Sam
ples Breast C
an
cer
Histop
at
ho
l
og
ic
al
Im
ages
3.2
.
Per
fo
r
m
an
ce E
valua
tion
In
order
t
o
e
valuate
the
propose
d
m
et
ho
d,
tw
o
pe
rfo
r
m
ance
m
et
ric
s
wer
e
us
ed:
the
patie
nt
cl
assifi
cat
ion
r
at
e
(P
CR
)
a
nd
the
ov
e
rall
cl
as
sific
at
ion
accu
r
acy
(
OCA
)
[
4]
.
The
sta
ndar
d
P
CR
is
the
rate
of
the
nu
m
ber
of im
a
ges
cl
assifi
e
d
c
orrectl
y t
o
the
nu
m
ber
of all
im
ages for each
p
at
ie
nt,
t
he P
CR
can
e
xp
la
i
n
as:
c
o
r
r
e
c
t
a
l
l
n
P
C
R
x
1
0
0
%
,
n
(
8
)
wh
e
re
c
o
r
r
e
c
t
n
is t
he
num
ber
o
f
h
ist
opat
holo
gical
im
ages classi
fied
co
rr
ect
ly
for
the p
at
ie
nt
i
an
d
a
l
l
n
is t
he
nu
m
ber
of h
ist
op
at
ho
l
og
ic
al
i
m
ages of
the
pa
ti
ent
i
.
The OCA
can
be
e
xp
la
in
ed
a
s:
pa
t
i
e
nt
s
n
i
i1
p
a
t
i
e
n
t
s
PCR
OC
A
,
n
(
9
)
wh
e
re
p
a
t
i
e
n
t
s
n
is t
he n
um
ber
of
patie
nts.
40X
100X
200X
400X
Be
nign
Ma
li
gn
ant
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
Cl
as
sif
ic
ation
enha
ncem
e
nt
of
b
re
as
t ca
ncer
h
ist
op
atholo
gical im
ag
e
…
(
Mo
hamm
e
d Ab
du
lr
azaq K
ahy
a
)
409
3.3
.
E
xp
eri
m
ent
al Settin
g
To
co
nf
i
rm
the
us
e
f
uln
ess
of
the
pro
posed
m
et
hod,
com
pr
e
hensi
ve
ex
per
i
m
ent
with
LR
i
s
co
nducte
d.
To
do so,
t
he
s
m
oo
th
feat
ur
es
m
at
rix
data
is
p
arti
ti
on
e
d
int
o
t
he
t
rainin
g
s
et
an
d
t
he
te
st
set,
w
he
re 70
%
of
the
sam
ples
are
sel
ect
ed ra
ndom
l
y
for t
he
t
raini
ng
set
an
d t
he r
est
30%
are
sel
ect
ed f
or
te
sti
ng
set
. T
o m
it
igate
the
eff
ect
s
of
the
f
eat
ur
es
m
at
rix
data
par
ti
ti
on,
al
l
the
resu
lt
s
obta
ined
were
th
e
aver
a
ge
of
fi
ve
tria
ls
for
pa
r
ti
ti
on
s
.
4.
RESU
LT
S
AND
DI
SCUS
S
ION
4.1
.
Clas
si
ficat
i
on
Per
f
orm
an
ce
Table
2
re
ports
,
on
a
ver
a
ge,
t
he
OC
A
for
th
e
trai
ni
ng
a
nd
t
est
ing
dataset
s
of
a
pp
ly
in
g
t
he
A
PLR
a
nd
PLR
m
et
ho
ds
.
The
num
ber
in
par
e
nth
esi
s
is
t
he c
orres
ponding
sta
nda
rd
de
viati
on
.
I
n
ad
di
ti
on
,
the
la
st
co
lum
n
represe
nts the f
il
te
r
sp
an
v
al
ue
.
At th
e
be
ginning
with
the
m
a
gn
i
ficat
ion 4
0X,
r
e
gardi
ng th
e ove
rall
classi
ficat
ion acc
ur
a
cy
and
b
as
e
d
on
the
t
rainin
g
dataset
,
t
he
propose
d
m
et
ho
d,
A
PLR,
ac
hie
ves
100.0
0%,
def
eat
in
g
PLR,
by
3.098%
w
hethe
r
the
fil
te
r
sp
a
n
va
lue
eq
ual
five
or
th
ree.
Dep
e
ndin
g
on
the
te
st
ing
dataset
,
the
AP
LR
w
hich
de
pends
on
t
he
f
il
te
r
sp
a
n
five
is
be
tt
er
tha
n
P
LR
of
ove
rall
cl
assifi
cat
ion
acc
uracy
bec
ause
it
achie
ved
91.922
%,
wh
ic
h
is
6.9
62
bette
r
tha
n PL
R.
Wh
il
e
the
m
agn
ific
at
io
n
100X,
the
APLR
al
so
prov
i
des
e
nhancem
ent
ove
r
the
PLR
by
6.
608%
for
the
trai
ning
datase
t
reg
ar
dless
of
the
filt
er
sp
a
n
value.
M
or
e
over
,
the
propo
sed
m
et
ho
d
be
at
s
PLR
in
te
r
m
s
of
ov
e
rall
classi
fi
cat
ion
acc
ur
ac
y
base
d on the
te
sti
ng
dataset
.
Lo
ok
i
ng
at
t
he
m
agn
ific
at
ion
200X,
the
OC
A
of
t
he
pro
posed
m
et
ho
d
pe
rfor
m
ance
is
be
tt
er
tha
n
t
he
non
-
sm
oo
thed
data
of
PLR
.
I
n
te
rm
s
of
OCA,
th
e
OC
A
obta
ined
f
r
om
the
pro
po
se
d
m
eth
od
was
100.0
0%
f
or
the
trai
ning d
at
aset
and
92.46%
that
de
pends
on
t
he
filt
er
s
pa
n
fi
ve
as
well
93.49
6%
that d
epends on
the fi
lt
er
sp
a
n
th
ree fo
r
t
he
te
sti
ng d
at
a
set
.
This i
nd
ic
a
te
s the supe
rio
r
it
y of
the
pro
pose
d
m
et
ho
d.
Eve
ntu
al
ly
, r
e
ga
rd
i
ng
the
m
agn
ific
at
ion
400X,
the
AP
LR
s
hows
a
consi
der
a
bl
e
do
m
inance
a
gainst
non
-
sm
oo
t
hed
data
P
LR.
It
achie
ve
d
t
he high
e
r o
ver
al
l cl
assifi
cat
ion
acc
ur
acy
for b
oth
the traini
ng and test
in
g datas
et
s.
Table
2
.
Cl
assi
ficat
ion
pe
rform
ance of th
e
APLR
and
PLR
Metho
d
s
Tr
ain
in
g
datas
et
Testin
g
datas
et
Filter
Sp
an
OCA %
OCA %
40X
APLR
1
0
0
.00
(
0
.000)
9
1
.92
2
(
4
.413)
5
APLR
1
0
0
.00
(
0
.000)
9
1
.54
4
(
4
.830)
3
PLR
9
6
.90
2
(
0
.949)
8
4
.96
0
(
4
.602)
No
n
S
m
o
o
th
ed
100X
APLR
1
0
0
.00
(
0
.000)
9
2
.82
2
(
1
.431)
5
APLR
1
0
0
.00
(
0
.000)
9
0
.99
6
(
0
.898)
3
PLR
9
6
.01
8
(
1
.010)
8
6
.21
4
(
0
.544)
No
n
S
m
o
o
th
ed
200X
APLR
1
0
0
.00
(
0
.000)
9
2
.46
0
(
3
.169)
5
APLR
1
0
0
.00
(
0
.000)
9
3
.49
6
(
2
.532)
3
PLR
9
6
.86
2
(
1
.147)
8
6
.85
8
(
2
.323)
No
n
S
m
o
o
th
ed
400X
APLR
1
0
0
.00
(
0
.000)
8
8
.36
4
(
1
.985)
5
APLR
1
0
0
.00
(
0
.000)
8
7
.24
6
(
2
.076)
3
PLR
9
4
.87
0
(
1
.035)
8
2
.57
2
(
1
.919)
No
n
S
m
o
o
th
ed
4.2
.
S
tatistic
al Signi
fica
nc
e Test
To
c
onfirm
the
util
it
y
of
th
e
pro
posed
m
et
hod
in
hi
gh
cl
assifi
cat
ion
perf
or
m
ance,
a
pair
wise
com
par
ison
bet
ween
the
pro
posed
m
et
ho
d
a
nd
eac
h
c
om
petito
r
m
et
ho
d
was
us
e
d
us
in
g
Ma
nn
–
Wh
it
ney
U
te
st.
The
area
unde
r
t
he
c
urve
(AUC)
f
or
t
he
t
r
ai
nin
g
dataset
was
us
e
d
f
or
t
his
te
st.
Ta
ble
3
s
hows
t
he
Ma
nn
–
Wh
it
ney
U
te
s
t
resu
lt
s
at
si
gnific
ance
le
vel
0
.
0
5
.
As
highli
ghte
d
in
Ta
ble
3
,
the
A
UC
of
th
e
pro
pose
d
m
et
ho
d
is
stat
ist
ic
al
ly
sign
ific
antly
bette
r
tha
n
P
LR.
Table
3
.
P
-
valu
es for
the Ma
nn
–
Wh
it
ney U t
est
o
f
the
pro
pose
d
m
et
ho
d re
su
lt
s w
it
h com
petit
or
m
et
ho
d.
(*)
m
eans th
at
the
two
m
et
ho
ds
ha
ve
si
gn
ific
a
nt
diff
e
re
nces
Dataset
APLR vs
P
LR
40X
0
.00
6
8
(*)
100X
0
.00
5
4
(*)
200X
0
.00
6
0
(*)
400X
0
.00
1
1
(*)
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.
13
, N
o.
1
,
Ja
nu
a
ry 20
19
:
4
05
–
4
10
410
5.
CONCL
US
I
O
N
This
pa
per
pr
es
ents
a
n
ada
ptiv
e
pe
naliz
ed
lo
gi
sti
c
regressio
n
by
m
eans
of
s
m
oo
thing
of
fe
at
ur
es
m
at
rix
to
increase
ove
rall
cl
assifi
cat
i
on
acc
uracy
of
br
east
ca
ncer
hi
stop
at
holo
gica
l
i
m
ages.
The
s
up
e
rio
r
cl
assifi
cat
ion
perform
ance
of
the
pro
pose
d
m
et
ho
d
was
s
how
n
thr
ough
two
as
pects:
hig
h
ov
e
rall
cl
as
sific
at
ion
acc
ur
acy
an
d
the
Ma
nn
–
Wh
i
tney
U
te
st
for
the
A
UC.
Co
nse
quently
,
the
resu
lt
s
c
onfirm
that
AP
LR
is
a
prom
isi
ng
m
et
hod
for
m
edical
i
m
age
cl
assifi
cat
ion
,
m
edical
diag
nosis
of
tum
or
s
an
d
ve
ry
us
e
f
ul
in
oth
e
r
ty
pes
of
high
-
dim
ension
al
classi
ficat
ion
dat
a relat
ed
t
o
the
b
iol
og
ic
al
fiel
d.
REFERE
NCE
S
[
1]
Bo
y
l
e, P.
and
B
.
Le
vin
,
World
ca
nce
r report 2008
.
2008:
IARC Pr
ess,
Internat
ional
Agenc
y
for
R
ese
arc
h
on
C
ancer.
[2]
Zha
ng,
Y.
,
B
.
Z
hang,
and
W
.
Lu
,
Breast
canc
er
histol
ogic
a
l
image
c
lassifi
ca
ti
on
wit
h
mult
ipl
e
f
e
atures
and
rand
om
subs
pace
cl
ass
ifier
ense
mble
,
in
Knowle
dge
-
Based
Syste
ms
in
Biom
edi
ci
n
e
and
Computati
onal
Life
Scienc
e
.
20
13,
Springer.
p
.
27
-
4
2.
[3]
Cent
er
,
M.,
R.
Si
ege
l
,
and
A
.
Jem
al
,
Globa
l
can
cer
fac
ts
&
fi
gures.
Atla
nt
a:
Am
eri
c
an
Cancer
Soci
ety
,
2011:
p
.
1
-
52
.
[4]
Spanhol,
F.A.
,
et
a
l.,
A
dataset
for
breast
can
ce
r
histopat
ho
lo
gic
al
image
cl
a
ss
if
ic
ati
on
.
I
EEE
Tr
ansa
ctions
on
Biom
edi
cal
Eng
i
nee
ring
,
2016
.
6
3(7):
p.
1455
-
14
62.
[5]
Vu,
T
.
H.
,
et
al
.
,
Histopatho
logical
image
c
lassifi
cation
using
d
i
scriminati
v
e
fe
a
ture
-
oriented
d
i
ct
ionary
l
earnin
g.
IEE
E
Tr
ansa
ctio
ns on
Medical
I
m
agi
ng,
2016
.
3
5(3):
p.
738
-
751
.
[6]
Acha
r
y
a
,
U.R
.
,
et
al.,
Aut
omat
e
d
charac
t
erization
of
fatty
liver
disease
and
ci
rr
hosis
using
curv
el
e
t
transfor
m
a
nd
ent
ropy
fe
atures
ex
trac
te
d
from u
lt
ras
ound
image
s.
Com
pute
rs i
n
Biol
og
y
and
Me
dic
in
e, 2016. 79: p. 250
-
258.
[7]
U
y
un,
S.
and
L
.
Chorida
h,
F
eat
u
re
Selec
ti
on
Ma
mm
ogram
based
on
Breast
Can
c
er
Mini
ng
.
Inte
r
nat
ion
al
Journ
al
of
El
e
ct
ri
ca
l
and
C
om
pute
r
Engi
n
e
eri
ng
(IJEC
E), 2
018.
8(1):
p.
60
-
69.
[8]
Huang,
C.
-
H.
,
et
a
l.,
Time
-
effi
ci
e
nt
spar
se
analy
s
is
of
histopa
thological
whol
e
sli
de
images.
Com
pute
ri
ze
d
m
edi
c
a
l
imaging
and
gra
phic
s,
2011
.
35(
7
-
8):
p.
579
-
591
.
[9]
Bel
sare
,
A.D.
,
e
t
al
.
Classifi
cat
io
n
of
breast
can
cer
histopat
hology
images
using
tex
ture
fe
a
ture
anal
ysis
.
in
TENCO
N
2015
-
2015
IEEE
R
egi
on
10
Co
nfe
renc
e
.
2015.
[10]
Korkm
az
,
S.,
G.
Za
rar
si
z,
and
D.
Goks
uluk,
Dr
ug/nondrug
cl
ass
if
i
cat
ion
u
sing
sup
port
vect
or
mach
ine
s
wit
h
various
fe
ature
select
ion
strategies.
Com
pute
r
Me
thods and Programs
in Biom
edi
ci
n
e, 20
14.
117(2):
p.
51
-
60.
[11]
Zha
ng,
L.,
et
al.,
Simil
arit
y
-
bala
nce
d
dis
crimina
nt
nei
ghbor
embe
dding
and
it
s
a
ppli
cation
to
ca
nce
r
c
lassifi
ca
tion
based
on
ge
ne e
xpre
ss
ion
data.
Com
pute
rs i
n
bi
olog
y
and
m
edi
c
ine
,
2015.
64:
p.
236
-
245.
[12]
Abu
-
Nim
eh,
S.,
et
al
.
A
comparison
of
mach
ine
lear
ning
techniques
for
phishing
d
e
te
c
ti
on
.
in
Proceedi
ngs
of
the
ant
i
-
phishing
workin
g
groups
2nd
an
nual
eCri
me
researche
rs
sum
mit
.
2007
.
ACM
.
[13]
Algamal,
Z.,
C
la
ss
if
ic
ati
on
of
gen
e
e
xpre
ss
ion
aut
i
sm
data
based
o
n
adaptive
penalized
log
isti
c
regress
ion.
E
lectr
oni
c
Journal
of
App
lied
Sta
ti
sti
ca
l
An
aly
s
is,
2017
.
10(
2):
p.
561
-
571.
[14]
Park,
H.
,
e
t
a
l.,
A
nov
el
a
dap
ti
v
e
penal
ized
logi
st
ic
regress
ion
for
uncovering
bio
mar
ke
r
ass
oci
ate
d
wit
h
anti
-
can
ce
r
drug sensit
ivity.
IEE
E
/ACM T
ra
nsac
ti
ons on
Co
m
puta
ti
onal Bi
o
l
og
y
and
B
ioi
nfo
rm
at
ic
s,
2016.
[15]
Le
e
,
H.
,
et
al.,
Diff
usion
ke
rne
l
-
b
ased
logi
sti
c
reg
ress
ion
models
f
o
r
protei
n
fun
cti
on
prediction.
Om
ic
s:
a
journa
l
o
f
int
egr
at
iv
e
bio
lo
g
y
,
2006
.
10(1):
p.
40
-
55
.
[16]
Kah
y
a,
M.A
.
,
W
.
Al
-
Ha
y
ani,
and
Z
.
Y.
Algam
al
,
C
lassifi
ca
ti
o
n
of
breast
can
c
er
histopat
holog
y
images
based
on
adapti
v
e
spar
se
suppor
t
ve
c
tor m
achi
ne
.
Journa
l
of
Applie
d
Math
emati
cs
and
B
ioinform
at
ic
s,
2017
.
7(1):
p.
49.
[17]
Hlai
ng,
T
.
,
Fe
a
ture
se
le
c
ti
on
a
nd
fuzzy
decisio
n
tre
e
for
n
et
w
ork
in
trusion
d
et
e
ct
ion
.
Inte
rn
a
ti
onal
Journa
l
o
f
Inform
at
ic
s
and Com
m
unic
at
ion Te
chno
log
y
(IJ
-
I
CT),
2012
.
1(2
): p.
109
-
118
.
[1
8]
Spanhol,
F.A.
,
e
t
al.
Breast
can
c
er
histopa
thol
og
ic
al
image
cl
ass
if
ic
a
ti
on
using
C
onvol
uti
ona
l
Ne
ural
Ne
tworks
.
i
n
Inte
rnational
Jo
i
nt
Conf
ere
nce o
n
Neural
N
et
wor
ks
(
IJCNN)
.
2016.
Van
couve
r
,
C
ana
da
:
I
EE
E
.
[19]
Motla
gh,
N.H.
,
et
al.,
Breast
Ca
nce
r
Histop
a
thological
Image
Cl
ass
if
ic
ati
on:
A
Dee
p
Learning
Approach.
bioR
xiv
,
2018:
p.
242818.
[20]
Algamal,
Z.
Y.
a
nd
M.H.
L
ee
,
A
nove
l
mol
ec
u
lar
desc
riptor
sel
ection
method
in
q
sar
cl
ass
if
ic
at
io
n
model
based
o
n
wei
ghte
d
pena
li
z
ed
log
isti
c
regre
ss
ion.
Journal
of
Chemom
et
ric
s,
2017.
31(10)
.
[21]
Algamal,
Z
.
,
An
efficien
t
g
ene
select
ion
method
for
h
igh
-
dimen
sional
microarr
ay
data
based
o
n
spar
se
logi
st
i
c
regress
ion.
Elec
t
ronic
Journa
l
of
Applie
d
St
at
ist
ical
Ana
l
y
s
is,
201
7.
10(1):
p.
242
-
256.
[22]
Gonza
le
z,
R
.
and
R.
W
oods,
Digital
Image
Proce
s
sing
.
2008
,
Uppe
r
Saddl
e
Rive
r
,
New
Jerse
y
,
US
A:
Pear
son
Prent
ic
e
Hall
.
[23]
El
A
y
ac
h
i,
R
.
,
B.
Bouikhalen
e,
and
M.
Faki
r,
New
Image
Com
press
ion
Al
gorit
hm
using
Haar
Wav
elet
Tr
ansform
.
Inte
rna
ti
ona
l
Jou
rna
l
of
Inform
a
tics a
nd
Com
m
unic
ation
T
ec
hno
lo
g
y
(IJ
-
ICT), 201
7.
6(1):
p.
43
-
48
.
[24]
Antoniou,
A
.
,
Di
git
al
signa
l
proc
essing
.
2016:
McGraw
-
Hill
.
[25]
Rabi
ner
,
L
.
R.
an
d
B.
Gold,
Theor
y and
applicati
o
n
of digi
tal
signa
l pr
oce
ss
ing.
En
gle
wood
Cli
ffs,
NJ
,
Prenti
ce
-
Hal
l,
Inc
.
,
1975.
777
p
.
,
1975.
[26]
Quinquis,
A.
,
Di
git
al
signa
l
proc
essing
using M
A
TL
AB
.
2008:
W
i
ley
Onl
ine L
ibr
a
r
y
.
[27]
Champagne,
B.
and
F.
L
abeau,
Discrete
ti
me
si
gnal
proc
essing.
Cla
ss
Not
es
for
the
Cours
e
ECSE
-
412,
Depa
r
tment
of
Elec
tr
ical & Com
pute
r
Engi
n
e
eri
ng
,
McGi
ll
Univer
sit
y
,
2004
:
p.
190
-
194.
[28]
Te
rre
ll
,
T
.
J.,
Intr
oduct
ion
to
d
igi
t
al
filt
ers
.
1988:
Springer.
Evaluation Warning : The document was created with Spire.PDF for Python.