Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
23
,
No.
1
,
Ju
ly
2021
, p
p.
216
~
2
28
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v
23
.i
1
.
pp
21
6
-
228
216
Journ
al h
om
e
page
:
http:
//
ij
eecs.i
aesc
or
e.c
om
ROI
-
b
ased f
ea
tu
res for cl
assificati
on of ski
n di
sea
s
es usin
g a
mu
lti
-
layer
n
eural net
wo
rk
Thanh
-
H
ai
N
guyen,
Ba
-
Vie
t
N
go
Ho Chi
Minh C
ity
Un
ive
rsit
y
of Tec
hno
log
y
a
nd
Educ
a
ti
on,
Vie
tn
am
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Feb
1
7
, 202
1
Re
vi
sed
A
pr
1
4
, 2
021
Accepte
d
Apr
20
, 202
1
Skin
disea
ses
h
ave
a
serious
i
m
pac
t
on
hum
a
n
li
fe
and
h
ea
l
t
h.
Thi
s
art
i
cle
ai
m
s
to
rep
rese
nt
the
class
ifi
cati
on
ac
cur
a
c
y
of
skin
disea
ses
for
supporting
the
ph
y
sic
ia
ns’
c
orre
ct
d
ecision
o
n
pat
i
ent
s
for
e
a
rl
y
treatme
nt
.
In
par
ticular,
100
images
in
e
a
ch
t
y
p
e
of
five
skin
dise
ase
s
fro
m
ISIC
dat
aba
s
e
are
used
for
bal
an
ce
d
da
ta
set
s re
la
te
d
to
the
c
la
ss
ifi
c
at
ion
a
ccuracy
.
In
a
ddit
io
n,
thi
s pa
p
e
r
foc
uses
on
proc
essing
images
f
or
ext
r
ac
t
ing
si
x
opti
m
al
t
y
pes
of
e
le
ve
n
fea
tu
r
es
of
skin
disea
se
imag
e
fo
r
highe
r
cl
assifi
c
at
ion
p
erf
orm
an
ce
and
al
so
thi
s
ta
k
es
le
ss
tim
e
for
tra
in
ing.
The
ref
or
e,
skin
disea
se
imag
es
are
fi
ltere
d
and
segm
ent
ed
for
sepa
rat
ing
r
egi
on
of
int
er
ests
(ROIs
)
bef
or
e
ext
ra
ct
in
g
opti
m
al
fe
at
ure
s
.
First,
the
s
kin
d
isea
se
images
ar
e
proc
essed
b
y
n
orm
al
iz
ing
size
s,
removing
noises,
segm
ent
ing
to
sepa
r
at
e
reg
ion
of
int
er
ests
(ROIs
)
show
ing
skin
disea
se
signs.
Next
,
a
gr
a
y
-
le
v
el
co
-
occ
urre
n
ce
m
at
ri
x
(GLCM)
m
et
hod
is
appl
ie
d
for
te
xture
a
naly
s
is
to
ext
ra
ct
eleve
n
fe
at
ur
es.
W
it
h
th
e
opti
m
al
six
feat
ure
s
chose
n,
th
e
high
cl
assifi
cati
on
ac
cur
acy
of
s
kin
disea
ses
is
about
92%
e
val
ua
te
d
using
a
m
at
rix
conf
us
ion.
The
r
esult
show
ed
to
il
lustrate
the
e
ff
ec
t
ive
ness
of
the
proposed
m
et
ho
d.
Furthermore,
thi
s
m
et
hod
ca
n
be
dev
el
op
ed
for
oth
er
m
edi
c
al
d
at
ase
ts
for
supporting
in
disea
se
dia
gnosis.
Ke
yw
or
d
s
:
GLCM
al
gorit
hm
Ma
trix con
fu
si
on
MLNN
str
uctu
re
ROI feat
ures
Sk
in
d
ise
ase
s
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Tha
nh
-
Hai
N
guye
n
Faculty
of Elec
tric
al
an
d El
ect
ronics E
nginee
rin
g
Ho Chi Mi
nh
Ci
ty
U
niv
er
sit
y of T
ech
nolo
gy
an
d E
du
cat
io
n
01 Vo
Va
n Ngan St
reet, Li
nh Chieu
W
a
rd,
Th
u Du
c
Ci
ty
, H
o C
hi Mi
nh
Ci
ty
, V
ie
t Nam
Em
a
il
:
nth
ai
@h
cm
ute.ed
u.v
n
1.
INTROD
U
CTION
Sk
in
disease
,
wh
ic
h
is
one
of
the
m
os
t
com
m
on
diseases
i
n
hu
m
ans,
ca
n
aff
ect
t
o
al
l
ag
es
an
d
se
xes
.
In
cu
rrent,
the
r
e
are
m
any
patie
nts
with
dif
f
eren
t
sk
i
n
dise
ases
and
a
lot
of
de
at
hs
w
or
l
dw
i
de.
Mo
reover
,
the
sk
in
disease
is
a
glo
bal
pro
ble
m
,
ran
king
18th
in
the
glob
al
ran
king
of
world
wide
hea
lt
h
bu
r
de
ns
.
I
n
a
2017
global
sta
ti
sti
c
,
sk
in
diseases
wer
e
a
bout
1.
79%
of
ot
her
diseases
[
1],
in
w
hich
t
he
s
kin
diseases
i
nclu
de
d
erm
at
itis
(aller
gies,
c
on
ta
ct
,
sebu
m
),
acne
,
and
ur
ti
caria,
ps
ori
asi
s,
vira
l
sk
in
disease,
fu
ngal
sk
i
n
di
sease,
scabies, m
el
ano
m
a, p
yode
rm
at
it
is, cel
luli
ti
s,
carcin
om
a, u
lc
erati
ve by canc
er.
In
rece
nt
ye
ar
s,
enh
a
ncem
ent
of
m
edical
i
m
ages
befor
e
segm
entat
io
n
for
i
m
age
reco
gnit
ion
or
cl
assifi
cat
ion
ha
ve
bee
n
a
ppli
ed
[
2].
I
n
s
kin
disease
i
de
ntific
at
ion
,
al
l
s
ki
n
disease
im
a
ges
wer
e
se
gme
nte
d
us
in
g
en
ha
nce
d
le
vel
[
3
]
,
[
4].
A
gray
le
vel
co
-
occ
urren
ce
m
at
rix
(G
LC
M)
wer
e
a
ppli
ed
to
these
seg
m
ented
i
m
ages
f
or
ext
ra
ct
ing
feat
ur
e
s
ap
plied
to
a
drag
onfly
-
bas
ed
dee
p
neura
l
netw
orks
(
D
NN)
cl
assifi
e
r.
It
is
obvious
that
a
m
edical
i
m
ag
e
is
of
te
n
filt
ered
noise
f
or
segm
enting
reg
io
n
of
inte
r
est
(
ROI
)
,
in
wh
ic
h
enh
a
ncem
ent of the
im
age f
or increasi
ng cla
ssific
at
ion
acc
ur
acy
is
ve
ry im
po
rtant.
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
ROI
-
ba
se
d
fe
atu
res f
or
cl
as
sif
ic
ation
of ski
n disease
s
us
in
g a multi
-
layer
neur
al…
(
Thanh
-
H
ai
Ng
uye
n
)
217
Im
age
s
are
captu
red
from
di
ff
e
ren
t
s
ources
and
al
ways
e
xist
diff
e
re
nt
noise
s.
I
n
pra
c
ti
ce,
there
ar
e
ty
pes
of
filt
ers
fo
r
rem
ov
ing
no
ise
s
in
im
age.
In
order
t
o
el
i
m
inate
unwa
nted
noise
in
a
n
i
m
age
and
s
harpe
n
it
,
the
com
bin
ed
filt
er
of
G
a
ussi
an
an
d
Butt
erwor
t
h
hi
gh
pa
ss
was
ap
plie
d
[
5
]
,
[
6]
in
the
fr
e
qu
e
ncy
do
m
ai
n.
In
[
7
]
-
[
9],
im
ages
of
s
kin
le
sion
s
we
re
segm
en
ta
ti
on
us
in
g
a
gr
owing
m
et
hod,
i
n
wh
ic
h
a
utom
at
ic
init
ia
li
zation
of
seed
points
w
ere
us
e
d
.
The
r
esult
of
segem
entat
ion
t
o
ext
r
act
le
sion
area
s
in
im
ages
was
us
e
d
in
the
fusio
n
of
a
sup
port
vector
m
achine
(SVM)
a
nd
k
-
ne
ur
al
netw
ork
(k
-
NN)
for
cl
as
sific
at
ion
with
61%
of
F
-
m
easur
e.
In
our
researc
h,
the
B
utterw
or
t
h
high
pass
filt
er
will
be
em
plo
ye
d
to
sh
a
rp
e
n
areas
of
sk
i
n
diseas
e
befor
e
se
gm
ent
at
ion
.
In
im
age
proc
essing,
im
age
segm
entat
ion
al
gorithm
[1
0]
is
com
m
on
ly
app
li
ed
t
o
se
par
at
e
obj
ect
from
an
im
age
f
or
analy
sis
or
rec
ogniti
on,
Otsu
m
et
ho
d
i
s
one
of
th
res
holdin
g
m
et
hods
are
oft
en
em
plo
ye
d
for
segm
entat
i
on
th
res
ho
l
ds
[11].
Furthe
r
m
or
e,
the
Otsu
segm
entat
ion
m
et
ho
d
is
util
iz
ed
to
determ
ine
thres
ho
l
d
for
di
sease
reg
io
n
de
te
ct
ion
for
cl
assifi
cat
ion
[
12]
.
In
on
e
a
pp
l
ic
at
ion
,
auth
or
s
app
li
ed
to
fi
nd
key
po
i
nts
of
pix
el
s
con
si
der
e
d
a
s
the
seed
po
i
nt
s
of
disease
d
sk
in
.
The
refore
,
featu
res
co
rresp
onding
to
t
he
key
po
i
nts
ext
racted
after
the
se
gm
entat
ion
we
r
e
app
li
e
d
to
be
input
of
the
cl
assifi
er
[
13]
.
I
n
ou
r
arti
cl
e,
the
Ots
u
m
et
ho
d
is
appl
ie
d
to
segm
ent sk
in
lesi
on are
as
, call
ed
R
OI
.
In
pr
a
c
tic
e
,
the
re
are
dif
fere
nt
filt
er
m
et
ho
ds
fo
r
e
nha
nci
ng
i
m
ages
su
ch
a
s
m
edian
and
bo
tt
om
hat
filt
ers
wh
ic
h
are
of
te
n
ap
plied
f
or
de
-
no
isi
ng
in
m
edical
i
m
ages.
A
u
th
or
s
pro
po
se
d
a
novel
ap
pro
ach
f
or
detect
ing
featu
res
,
in
w
hich
a
Bott
om
hat
fil
te
ring
was
em
plo
ye
d
for
sm
oo
t
hing
im
age
s
[4
]
,
[
14]
.
Mo
reove
r
,
th
e
m
edian
filt
er
is
of
te
n
util
iz
ed
f
or
rem
ove
sal
t
an
d
pe
p
per
noise
in
m
edical
im
ages.
In
this
resear
ch,
we
will
app
ly
m
e
dian
an
d
bo
tt
om
hat
filt
ers
fo
r
el
im
inati
ng
no
ise
s
in
s
kin
diseases
an
d
sh
a
pp
e
n
sk
i
n
disease
areas.
Fo
r
ef
fecti
ve
segm
entat
ion
to
se
par
at
e
R
O
I
in
im
age,
ed
ge
detect
ion
of
ob
j
ect
s
in
th
e
i
m
age
is
necessa
ry.
I
n
[
15
]
,
a
uthor
s
pro
posed
a
Ca
nny
ed
ge
de
te
ct
or
with
a
it
era
ti
ve
m
edian
filt
er
(I
MF
)
f
or
sk
i
n
le
sion
bor
de
r
de
te
ct
ion
,
in
w
hi
ch
the
i
m
pr
ov
ed
it
erati
ve
segm
entat
ion
al
go
rithm
fo
r
bor
der
detect
ion
of
real
sk
in
le
sions
produce
d
th
e
ef
f
ect
ive
res
ult.
The
perf
or
m
ance
of
this
al
go
rithm
is
bette
r
than
the
tra
diti
on
al
segm
entat
ion
al
gorithm
.
In
ou
r
pa
per,
the
Ca
nn
y
m
et
ho
d
is
app
li
ed
f
or
det
ect
ing
ed
ges
i
n
sk
in
disease
a
reas
in im
age f
or
se
par
at
i
ng
R
OI.
M
orpholo
gical
op
e
rati
on
[
16
]
,
[
17]
was
em
plo
ye
d
in
proc
essing
T1
-
w
e
igh
te
d
m
agn
et
ic
resona
nce
i
m
aging
(MRI
)
i
m
ages,
in
w
hi
ch
dilat
ion
is
t
o
en
ha
nce
the
est
i
m
ation
vol
um
e
m
et
ho
d.
I
t
is
ob
vi
ous
th
at
this
m
et
ho
d
is
bett
er
com
par
ed
t
o
Stere
ology
m
et
ho
d
f
or
le
ss
MR
I
sli
ces
a
nd
le
ss
te
st
po
ints.
In
dee
p
le
arn
i
ng
m
et
ho
ds
for
s
kin
le
sio
n
cl
assifi
cat
ion
,
pr
i
m
ary
m
or
phol
og
ic
al
el
em
ent
s
are
a
m
od
ul
e
in
the
auto
m
at
i
c
detect
ion
syst
e
m
.
The
resu
lt
of
this
cl
assifi
cat
ion
was
the
76
.
00%
accu
r
acy
fo
r
5
cl
asses
of
th
e
pr
i
m
ary
m
or
phologica
l
el
e
m
ents
and
the
81.67%
ac
c
ur
acy
for
3
cl
a
sses
[
18]
.
I
n
our
stu
dy,
dilat
i
on
an
d
e
rosio
n
in
the
m
or
phologica
l
appr
oach
are
a
pp
li
ed
for
rem
ov
i
ng
unwa
nted
sm
all
ob
j
ec
t
s
and
co
nnect
in
g
dott
ed
li
nes
in
sk
in
disease a
reas.
In
m
edical
i
m
ag
es,
the
sepa
r
at
ion
of
RO
Is
for
featu
re
e
xtracti
on
is
ver
y
necess
ary.
A
si
m
ple
and
reli
able
ap
proa
ch
for
s
kin
re
gi
on
s
egm
entat
i
on
to
ge
ner
at
e
ROI
ca
n
be
a
ppli
ed.
I
n
on
e
r
esearch
,
a
m
et
hod
i
n
te
rm
s
of
segm
enting
s
olid
s
ki
n
reg
i
on
s
without
ge
ner
at
in
g
m
uch
no
i
sy
segm
ents
was
pro
po
se
d
in
[
19]
.
A
ROI
c
ou
l
d
be
s
epar
at
e
d
base
d
on
a
naly
sis
of
fine
ge
om
et
ric
detai
ls
of
s
kin
le
sion
im
ages.
The
res
ul
t
wa
s
that
the
ave
ra
ge
perform
ance
of
t
his
m
et
ho
d
was
bette
r
the
cu
rrent
sta
te
-
of
-
the
-
art
te
chn
i
qu
es
without
trai
ning
[2
0].
It
is
ob
vious
t
hat
there
ha
ve
been
dif
fer
e
nt
m
et
ho
ds
for
separ
at
in
g
a
ROI
in
im
age.
In
our
researc
h,
after
pr
e
-
processi
ng
im
age,
a
Ca
nn
y
e
dge
detec
ti
on
will
be
a
ppli
ed
to
se
par
a
te
ROI
f
or
e
xtr
act
ing
op
ti
m
al
featur
e
s u
si
ng GLC
M m
e
tho
d.
Fo
r
s
kin
disea
s
e
cl
assifi
cat
io
n,
im
age
featu
re
e
xtracti
on
usi
ng
GLCM
m
et
hod
play
s
an
i
m
po
rta
nt
ro
le
beca
us
e
it
can
con
ta
in
i
m
po
rtant
infor
m
at
ion
of
the
aff
ect
ed
s
kin.
Fo
r
feature
ex
tract
ion
f
or
te
xture
recog
niti
on
,
a
GLCM
m
et
ho
d
w
as
em
ploy
ed
[
21
]
-
[
24]
.
I
n
t
his
re
sea
rch,
th
ree
dif
f
eren
t
data
set
s
we
re
pro
po
se
d,
in
wh
ic
h
the
fir
s
t
set
is
a
sim
ple
m
od
ifie
d
f
eat
ur
es
e
xtract
ed
from
the
tradit
ion
al
GLC
M;
the
seco
nd
set
use
s
two
s
ub
-
se
ts
of
featu
res
extracte
d
f
r
om
two
GLC
M
m
et
ho
ds
c
al
culat
ed
us
in
g
two
disp
la
cem
ent
va
lues;
the
thi
r
d
one
was
pas
sing
t
he
e
xtrac
te
d
set
of
the
GLCM
th
rou
gh
a
n
arti
fici
al
neural
netw
ork
(AN
N)
for
cl
assifi
cat
ion
.
In
cu
rrent,
m
any
m
eth
ods
ha
ve
be
e
n
a
pp
li
ed
f
or
cl
assifi
cat
ion
of
s
kin
le
sion
s
[25
]
,
[
26]
.
The
refo
re,
these
featu
re
c
om
po
ne
nts
we
re
us
e
d
to
be
t
h
e
in
pu
t
of
t
he
cl
assifi
er
to
cl
assify
diff
e
re
nt
sk
i
n
diseases.
O
ur
stud
y
is
that
a
GLCM
is
uti
li
zed
for
e
xtra
ct
ing
feat
ur
es
and
just
6
optim
al
feat
u
re
s
of
ele
ven ones
u
se
d
for
cl
assifi
cat
ion o
f
s
kin dise
ases.
W
it
h
m
any
di
ff
e
ren
t
hum
an
sk
in
diseases
,
it
is
ver
y
i
m
portant
to
cl
assify
and
fin
d
the
disease
autom
at
ic
ally
because
it
help
s
doct
ors
to
ea
rly
diag
no
se
th
e
disease
a
nd
t
he
patie
nt
can
be
treat
ed
ea
rl
y
and
eff
ect
ively
.
In
recent
ye
a
rs,
the
cl
assifi
cat
ion
of
diff
e
ren
t
com
po
ne
nts
us
in
g
a
n
a
rtifi
ci
al
i
nte
ll
igence
(AI
)
m
et
ho
d
has
be
en
ve
ry
po
pu
la
r
an
d
e
ff
ect
ive
[27
]
,
[
28]
.
I
n
par
ti
cula
r,
a
uthors
pro
posed
a
dee
p
co
nvol
ution
al
netw
orks
with
a
su
sta
ined
c
om
bin
at
ion
f
or
the
cl
assifi
cat
ion
of
dam
a
ged
s
kin
dise
ases
[27].
W
it
h
this
pro
po
se
d
m
et
ho
d,
the
dee
p
le
arn
i
ng
netw
or
k
wa
s
em
plo
ye
d
f
or
te
sti
ng
ov
e
r
900
s
kin
disease
im
ages
and
it
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.
23
, N
o.
1
,
Ju
ly
2021
:
2
1
6
-
2
2
8
218
gav
e
relat
ivel
y
good
re
su
l
ts.
A
rtific
ia
l
intel
li
gen
ce
al
gorithm
s
were
pro
posed
f
or
cl
assi
fyi
ng
sk
in
diseases
[
29
]
,
[
30
]
.
In
pa
rtic
ular,
a
C
NN
m
od
el
com
bin
ed
i
ntra
-
a
rch
it
ect
ur
e
a
nd
i
nt
er
-
a
rch
it
ect
ure
netw
ork
fu
si
on
[31]
or
a
com
pact
deep
le
ar
ning
-
ba
sed
cl
assifica
ti
on
m
od
el
wit
h
a
sepa
ra
ble
conv
olu
ti
onal
neural
netw
ork
[32],
t
he
sta
te
-
of
-
art
DNN
[33]
were
util
iz
ed
.
In
one
researc
h,
au
thors
pro
po
se
d
a
re
gion
-
base
d
C
N
N
(F
RC
N
N)
f
or
trai
ning
cl
inica
l
i
m
ages
of
pig
m
ented
sk
i
n
le
sion
s
,
in
wh
ic
h
there
w
ere
m
a
li
gn
ant
tu
m
or
i
m
ages
[34].
T
he
res
ults
of
t
his
resea
rch
w
as
com
par
ed
t
o
an
oth
e
r
rese
arch
t
o
il
lustra
te
the
bette
r
propose
d
m
et
ho
d.
I
n
res
ea
r
ch
[
35
]
,
a
C
NN
was
a
ppli
ed
in
t
he
cl
assif
ic
at
ion
o
f
Ps
ori
asi
s
sk
in
disea
se
with
the
acc
ur
acy
rate
of
82.
9%
and
72.
4%
f
or
Plaq
ue
an
d
Gu
tt
at
e
Pso
ri
asi
s
sk
in
disea
se,
resp
ect
i
vely
.
It
is
ob
vious
that
cl
assifi
cat
ion
m
od
el
s u
sin
g AI
m
et
hods
for
b
ig
m
edical
i
m
age sets are
very
im
po
rtant.
In
this
arti
cl
e,
we
ha
v
e
pro
posed
an
aut
om
a
ti
c
cl
assifi
cation
m
e
tho
d,
in
wh
ic
h
six
opti
m
al
featur
es
are
e
xtracted
f
ro
m
five
balan
ced
ty
pes
of
s
kin
diseases
a
nd
the
n
they
a
re
c
om
bin
ed
t
o
a
m
ulti
la
ye
r
ne
ur
al
netw
ork
(ML
N
N)
[36
]
,
[
37]
fo
r
hi
gh
cl
assifi
cat
ion
.
All
thes
e
i
m
ages
are
pr
eprocesse
d
to
r
e
m
ov
e
unnece
ssary
com
po
ne
nts,
a
s
well
as
to
en
han
ce
t
hem
bef
ore
se
g
m
entation
for
se
par
at
ing
R
OI
s
befo
re
extracti
ng
optim
al
featur
e
s
us
i
ng
a
GLCM
.
Th
ese
extracte
d
f
eat
ur
es
are
use
d
to
be
t
he
inputs
of
the
MLNN
cl
assif
ie
r.
Th
e
aver
a
ge
acc
ura
cy
o
f
the
high
cl
assifi
cat
ion
i
s ev
al
uated usi
ng a c
onf
us
io
n m
at
rix
.
2.
METHO
DOL
OGY
It
is
ver
y
im
portant
t
o
det
ect
an
d
cl
assi
fy
hum
an
sk
i
n
disease
ea
rl
y
beca
us
e
it
can
help
the
ph
ysi
ci
ans
’
c
or
rect
dia
gnos
is
f
or
treat
m
ent
early
.
I
n
t
his
rese
a
r
ch
,
a
sk
i
n
dis
e
ase
detect
io
n
m
et
ho
d
is
base
d
on
the
extracti
on
of
th
e
surface
te
xtu
re
featu
re
s
of
t
he
diseas
e
sk
in
im
age.
Ther
e
f
or
e,
t
he
featur
e
set
wi
ll
be
sta
ti
sti
cally
analy
zed
to
sel
e
ct
the
opt
i
m
al
featu
res
f
or
trai
ning
a
nd
cl
assifi
cat
ion
process
us
in
g
a
MLN
N
structu
re
.
It
is
obvious
t
hat
th
is
will
ta
ke
m
or
e
ti
m
e
fo
r
pro
cessi
ng
disease
sk
in
im
ages
and
sel
ect
ing
op
tim
a
l
featur
e
s
before
sk
i
n
disease
cl
assifi
cat
ion
with
high
ac
cur
acy
.
W
it
h
sel
ect
ed
opt
i
m
al
featu
res,
a
si
m
pler
MLNN
str
uctu
re ca
n be a
pp
li
ed
, i
n wh
ic
h
it
s
t
rainin
g
ta
kes l
ess tim
e and
c
la
ssific
at
ion
pe
rfor
m
ance is h
i
gh
.
2.1.
S
kin im
age pre
-
pr
ocess
ing
fo
r
det
er
mi
ning
ROI
Im
age
pr
e
-
pro
cessi
ng
is
a
n
im
po
rtant
sta
ge
in
the
extracti
on
of
im
age
fe
at
ur
es
f
or
cl
assify
ing
s
kin
diseases.
T
herefo
re,
im
age
dataset
s
are
pro
cessed
to
e
nh
a
nce
the
im
age
qu
al
it
y
befor
e
extracti
ng
struc
tural
featur
e
s
f
or
the
sk
in
disease
cl
assifi
cat
ion
with
higher
accuracy
us
in
g
neural
net
w
orks.
To
anal
yz
e
the
diseased
s
kin
obj
ect
s
in
the
i
m
age,
we
nee
d
to
disti
nguish
the
obj
ect
s
of
i
nterest
f
ro
m
the
rest
of
the
i
m
age,
cal
le
d
the
bac
kgr
ound.
The
s
e
obj
ect
s
can
be
identifie
d
us
ing
a
se
gm
entat
ion
m
e
t
hod
to
sep
ara
te
the
foregr
ound
f
r
om
the
i
m
age
f
or
c
ollec
ti
ng
ROI.
The
RO
I
par
t
m
ay
con
t
ai
n
di
ff
e
ren
t
f
eat
ur
es
su
c
h
a
s
colo
r,
un
i
form
i
ty
,
te
xtu
re,
gra
y
le
ve
l,
fr
e
quency
or
m
o
m
ent.
The
se
featu
res
c
a
n
form
a
featu
re
vector
ass
oc
ia
te
d
with the
RO
I
a
nd this
helps
to
d
ist
in
gu
is
h dif
fer
e
nt s
kin
lesi
on
s
usi
ng a
ne
ur
al
netw
ork.
In
this
pa
per
,
the
process
of
RO
I
se
parat
ion
will
be
carrie
d
out
thr
ough
se
ve
n
ste
ps
.
T
he
pre
-
pr
ocessin
g
ste
ps
will
m
a
ke
the
bette
r
s
kin
disease
im
age
befor
e
pe
r
form
ing
the
se
gm
entat
ion
to
separ
at
e
the
ROI
f
ro
m
the
sk
in
diseas
e
i
m
age
fo
r
fe
at
ur
e
extra
ct
io
n.
I
n
pa
rtic
ular,
the
ste
ps
f
o
r
pr
oc
essin
g
of
sk
in
i
m
ages ar
e,
Step
1:
T
he
ori
gin
al
im
age
f
(
x
,y
)
will
be
nor
m
al
iz
ed
fo
r
the
sa
m
e
siz
e.
Then
the
im
age
e
nh
a
ncem
ent
g
(
x,y
)
i
s
cal
culat
ed by
c
onvoluti
ng
a
ke
rn
el
k
1
(
s
,
t
)
w
it
h
the
im
age
f
(
x,y
),
(
,
)
=
∑
∑
1
(
,
)
(
−
,
−
)
=
−
=
−
(1)
Step
2:
T
he
i
m
age
after
e
nhance
m
ent
will
be
trans
f
or
m
ed
to
pro
du
ce
the
Fou
rier
im
ages
(
,
)
in
th
e
fr
e
qu
e
ncy
dom
ai
n
befor
e
filt
erin
g
out
the
low
-
fr
e
quency
c
om
po
ne
nt.
I
n
this
stud
y,
the
Butt
erwor
t
h
hi
gh
pass
filt
er
was
a
ppli
ed
to
s
ha
rp
e
n
for
the
pur
pos
e
of
en
ha
ncin
g
detai
ls,
hi
gh
li
gh
ti
ng
par
ti
cl
e
s
in
the
RO
I
r
egio
n,
wh
ic
h
are
c
on
sidere
d
the
le
sion
s
of
sk
i
n
diseases.
T
he
i
m
age
̃
(
,
)
us
ing
t
he
Butt
erwo
rth
filt
er
is
descr
i
bed
,
{
̃
(
,
)
=
(
,
)
∗
(
,
)
(
,
)
=
1
1
+
[
0
/
(
,
)
]
2
(2)
wh
e
re
D(
u
,
v
)
i
s
the
Eu
cl
idea
n Dist
ance
from
an
y
po
i
nt (
u
,
v
)
in
the im
age G
(
u
,
v
)
to
the
ori
gin
of the
fre
qu
e
ncy
plane,
(
,
)
=
√
2
+
2
and
D
0
desc
ribes
the
cu
t
-
off fre
que
ncy.
Step
3:
T
he
filt
ered
im
age
us
ing
the
Butt
er
w
or
t
h
filt
er
will
be
c
onve
rted
i
nto
t
he
im
age
̃
(
,
)
in
the
sp
at
ia
l
do
m
ai
n
be
fore
us
in
g
t
he
Ots
u
[
4]
m
et
ho
d
for
fi
ndin
g
th
e
thres
hold
f
or
im
age
segm
entat
ion
a
nd
f
or
t
he
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
ROI
-
ba
se
d
fe
atu
res f
or
cl
as
sif
ic
ation
of ski
n disease
s
us
in
g a multi
-
layer
neur
al…
(
Thanh
-
H
ai
Ng
uye
n
)
219
conve
rsion
to b
ina
ry
i
m
ages
̃
(
,
)
.
The
segm
entat
ion
is
to r
em
ov
e
the b
ac
kgr
ound o
f
t
he
nor
m
al
sk
in
to b
e
black, an
d kee
p
the
d
ise
ase
d skin
to be
w
hite i
n
the
b
i
nar
y
i
m
age.
Step
4: T
he bin
ary im
age w
il
l be
rem
ov
ed
sal
t and pe
pper
noise
by the m
edian fi
lt
er
̃
(
,
)
.
Step
5:
A
bott
om
-
hat
filt
er
will
be
ap
plied
to
the
im
age
̃
(
,
)
after
rem
ov
i
ng
so
m
e
el
e
m
ent
s
not
to
be
i
n
the
disease
sk
i
n
im
age
area,
su
c
h
as
hairs
or
oth
e
rs.
The
i
m
age
̃
(
,
)
is
cal
culat
ed
us
in
g
t
he
bo
tt
om
-
hat
filt
er
m
ulti
plied
with the
ke
r
ne
l
k
2
(
s
,
t
)
,
̃
(
,
)
=
̃
(
,
)
+
∑
∑
2
(
,
)
̃
(
−
,
−
)
=
−
=
−
(3)
Step
6: Th
e
filt
ered
im
age
̃
(
,
)
is
erode
d
to
rem
ov
e the
un
know
n
sm
al
l areas and the
n detec
te
d
the
edge
of
the
disease
s
ki
n
area u
si
ng
a Cann
y
m
et
ho
d
[15].
T
her
e
fore
,
the
im
age
with
the d
ise
ase
s
kin
e
dg
e
is
dilat
ed
to
connect
the
do
t
te
d
li
nes.
Step
7:
Finall
y,
fin
ding
the
re
gion
with
the
l
arg
est
area
t
o
s
epar
at
e
the
RO
I
is
perform
ed.
In
pa
rtic
ular
,
to
fi
nd
this
la
rg
est
regi
on
,
it
is
necessary
to
identif
y
the
reg
io
ns
f
il
le
d
in
the
bin
ary
i
m
age.
Ther
ef
or
e
,
the
co
nnect
e
d
obj
ect
com
pone
nts
are
fou
n
d
and
eac
h
fill
ed
ob
j
ect
is
ass
ign
e
d
a
la
bel
f
o
r
i
den
ti
ficat
io
n.
F
ro
m
the
la
be
le
d
obj
ect
s,
we
ca
n
dete
rm
ine
the
area
of
the
obj
ect
s
base
d
on
the
nu
m
ber
of
pix
el
s
a
nd
t
he
in
dex
f
or
e
xtracti
ng
the ob
j
ect
w
it
h t
he
la
r
gest a
re
a.
The
res
ult
obta
ined
a
fter
pre
-
proce
ssin
g
is
the
RO
I
a
rea
of
the
s
kin
di
sease
to
ser
ve
f
or
featu
re
extracti
on.
It
m
eans
that
pr
e
-
proce
ssin
g
for
sepa
rati
ng
the
ROI
area
ca
n
ta
ke
m
or
e
ti
m
e
,
but
it
will
ta
ke
le
ss
tim
e fo
r
trai
ning
process
u
si
ng the
MLN
N f
or cla
ssifyi
ng s
kin
disease im
ages.
2.2
.
Fe
ature
ext
r
act
i
on
of
ROI
Af
te
r
e
xtracti
ng
the
R
OI
regi
on
from
the
sk
in
disease
im
age,
a
GLCM
m
et
ho
d
is
em
plo
ye
d
f
or
featur
e
ext
racti
on
.
W
it
h
di
fferent
s
kin
col
or
im
ages,
ther
e
is
a
diff
e
rence
in
the
s
kin
le
sion
a
rea
relat
ed
t
o
structu
re,
fr
e
quency
,
a
nd
ot
her
par
am
et
ers.
W
it
h
the
te
xture
distri
bu
t
ion
a
nd
par
a
m
et
ers
in
the
GLCM
al
gorithm
of
t
he
s
kin
diseas
e
im
age,
we
can
ob
ta
i
n
dif
fer
e
nt
feat
ur
es
by
t
he
a
naly
sis
of
s
kin
te
xture
,
rou
ghness,
un
i
form
i
ty
o
f
the
sk
in
, a
nd sk
i
n condit
ion.
The
GLCM
al
gorith
m
is
on
e
of
m
et
ho
ds
use
d
for
e
xtract
ing
im
po
rta
nt
featur
e
s
relat
e
d
to
im
age
te
xtu
re
a
naly
sis.
I
n
pa
rtic
ular
,
each
el
em
ent
in
the
im
age
represe
nts
the
pro
bab
il
it
y
of
occurri
ng
t
he
sam
e
intensit
y
at
the
ty
pical
distan
ce
d
and
t
he
a
ng
le
θ
.
The
refor
e
,
the
re
can
be
m
any
diff
e
r
ent
GLCM
m
at
rices
dep
e
ndin
g
on
the p
ai
r
of
d
a
nd
θ
.
I
n
this
st
udy, with
da
m
a
ged
sk
i
n
diseas
e,
only
so
m
e
im
po
rtant
featu
res
s
uch
as
con
tra
st,
en
erg
y,
ho
m
og
e
ne
it
y,
m
ean,
stan
da
r
d
de
viati
on
,
ent
ropy
,
w
hi
ch
are
sync
hrono
us
an
d
re
pe
ti
ti
ve,
are
co
ns
ide
re
d.
From
these
sel
ect
ed
featu
re
s,
their
vecto
rs
are
us
e
d
in
t
he
MLNN
cl
ass
ifie
r.
I
n
the
G
LCM
al
gorithm
,
feat
ur
es
s
uc
h
as
c
on
t
rast,
ene
r
gy,
hom
og
eneit
y,
m
ean,
sta
nd
a
rd
dev
ia
ti
on,
e
ntr
op
y
can
gre
at
ly
diff
e
r
f
ro
m
othe
r
gro
ups
an
d
so
they
can
be
dataset
s
ch
os
e
n
f
or
trai
ning
and
cl
assify
in
g
in
the
MLN
N
.
Af
te
r
pre
-
pr
ocessin
g
sk
in
im
age
and
sepa
rati
ng
th
e
ROI
of
the
i
m
age
G
(
i
,
j
)
,
th
e
op
ti
m
al
fo
ll
ow
in
g
feat
ur
es
will
be
cal
culat
ed usin
g
the
GLC
M.
Con
tra
st
feat
ure
is
to
m
easur
e
the
s
patia
l
f
r
equ
e
ncy
of
s
kin
im
age
w
hich
is
the
dif
fer
e
nc
e
bet
wee
n
the
hi
gh
e
st
an
d
lo
west
value
s
of
a
c
onti
guou
s/a
djacent
s
et
of
pi
xels.
In
pa
rtic
ular,
the
co
ntrast
ca
n
m
easure
the
am
ou
nt
of
local
va
riat
ions
present
t
he
s
kin
im
age.
I
n
a
dd
it
io
n,
t
he
c
ontrast
descr
i
b
e
s
the
dep
t
h
of
"t
extil
e
gro
ov
es"
of
t
he
im
age.
The
re
fore,
if
t
he
c
ontrast
val
ue
is
hi
gh
e
r,
the
“g
roo
ves”
is
dee
pe
r.
The
c
on
tra
st
f
eat
ur
e
cab
be
cal
c
ula
te
d usin
g
t
he
f
ol
lowing
form
ul
a
,
=
∑
∑
(
−
)
2
(
,
)
−
1
−
1
(4)
in
w
hich
|i
-
j
|
i
s
the
gray
scal
e
diff
e
re
nce
bet
ween
a
djacent
pix
el
s
,
P
(i,
j
)
is
the
el
e
m
ent
(i,j)
of
the
norm
al
iz
ed
sy
m
m
e
tric
al
GLCM,
cal
le
d
the
distri
bu
ti
on
pro
bab
il
it
y
of
the
dif
fer
e
nt
gr
ay
scal
e
le
vels
bet
wee
n
the
ad
j
acent
pix
el
s.
L is t
he nu
m
ber
of gra
y l
evels in t
he skin
d
ise
ase
im
age.
Entr
op
y
is
an
i
m
po
rtant
featu
re
w
hich
al
lo
w
s
to
m
easur
e
inf
or
m
at
ion
of
the
disor
der
or
com
plexity
of
s
kin
im
age.
In
par
ti
cula
r,
t
he
ent
ropy
is
la
rg
e
w
hen
t
he
sk
in
im
age
is
no
t
te
xt
ur
al
ly
un
i
form
,
in
whic
h
the
entr
op
y
is
hi
gh,
po
te
ntial
ly
t
he
pa
rt
of
sk
i
n
im
age
has
com
plex
te
xtu
r
es.
I
n
ad
diti
on,
w
hen
t
he
ent
ropy
is
strongly, it
m
ay
inv
e
rsely
cor
relat
e to e
nerg
y. The
en
tr
opy
value
is
calc
ul
at
ed
usi
ng the
foll
ow
i
ng for
m
ula:
=
∑
∑
(
,
)
(
,
)
−
1
−
1
(5)
Inverse
Diff
e
r
ence
M
om
ent
is
cal
le
d
hom
og
e
neity
of
s
ki
n
im
age
and
can
s
how
la
ge
r
va
lues
f
or
sm
a
ll
er
gr
ay
ton
e
dif
fer
e
nce
s
in
pair
el
em
ents.
In
a
ddit
ion,
it
is
m
or
e
sensiti
ve
t
o
the
pr
ese
nce
of
ne
a
r
diag
on
al
el
em
e
nts
in
the
GLC
M.
Th
us
,
it
has
m
axi
m
u
m
val
ue
w
he
n
al
l
el
em
ents
in
the
ski
n
i
m
age
are
sa
m
e.
In
a
ddit
ion,
hom
og
eneit
y
is
us
e
d
to
desc
r
ibe
the
rou
ghness
of
the
i
m
age
struct
ur
e
.
It
m
eans
tha
t
if
th
e
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.
23
, N
o.
1
,
Ju
ly
2021
:
2
1
6
-
2
2
8
220
ho
m
og
e
neity
Hg
value
is
great
er,
the
rou
ghne
ss
o
f
the
i
m
a
ge
structu
re
at
that
da
m
ag
ed
area
will
be
gr
eat
er.
The u
niform
v
al
ue
of s
kin dis
eases can
b
e
d
e
scribe
d
,
=
∑
∑
1
1
+
(
−
)
2
(
,
)
−
1
−
1
(6)
W
it
h
the
featu
re
base
d
on
e
ne
rg
y,
the
im
ag
e
with
le
sion
sk
in
pa
rt
can
produce
a
m
axim
u
m
value.
More
ov
e
r,
this
energy
m
et
hod
is
to
m
easur
e
the
te
xtu
re
unif
or
m
ity
and
detect
s
disorde
rs
in
te
xtures
and
is
us
e
d
to
de
scrib
e
the
t
hick
ness
of
the
str
uctu
r
e
of
a
sk
i
n
dise
ase
im
a
ge.
The
ene
rg
y
of
this
sk
in
disease
im
age
can
be descri
be
d
us
in
g
t
he
f
ol
lowing e
xpres
sion,
=
∑
∑
(
,
)
2
−
1
−
1
(7)
Me
an
is
the
aver
a
ge
gray
le
vel
of
the
im
a
ge
of
s
kin
dis
ease
wh
e
re
G
(
x
,
y
)
is
the
i
m
a
ge
after
p
r
e
-
processi
ng
with
the
siz
e
m
×
n
.
In
ad
diti
on,
the
gray
intensi
ty
of
pix
el
s
is
norm
al
iz
ed
in
t
he
ra
ng
e
[
0,1]
before
cal
culat
ing
t
he
aver
a
ge of
s
kin disease i
m
ages,
=
∑
∑
(
,
)
−
1
=
0
−
1
=
0
255
×
×
(8)
Feat
ur
e
relat
e
d
to
cal
culat
io
n
is
sta
ndar
d
dev
ia
ti
on.
Thi
s
feat
ur
e
re
pr
e
sents
a
com
par
iso
n
of
th
e
m
ean
sta
ndar
d
de
viati
on
val
ues
of
dif
fe
re
nt
sk
i
n
diseas
es.
It
ca
n
be
s
een
that
t
he
diff
e
ren
ce
bet
we
en
the
disease cl
asses
can be
desc
rib
ed wh
e
n
a
naly
zi
ng
t
he
sta
nd
a
rd d
e
viati
on of
the gray
level
of the s
kin i
m
a
g
e,
=
√
∑
∑
(
(
,
)
255
−
)
2
−
1
=
0
−
1
=
0
×
(9)
An
al
ysi
s
of
th
e
featu
res
of
s
kin
diseases
is
i
m
po
rtant
f
or
c
la
ssific
at
ion
.
H
ow
e
ve
r,
t
her
e
is
an
exce
ss
betwee
n
pa
ram
e
te
rs
or
s
om
e
f
eat
ur
es
with
out
the
dif
fer
e
nce
betwee
n
s
kin
diseases.
It
is
obvi
ous
that
thi
s
will
ta
ke
tim
e
fo
r
cal
culat
ion
with
ou
t
ben
e
fit
for
trai
nin
g
proce
ss
of
s
kin
dise
ase
cl
assifi
cat
i
on.
I
n
this
stu
dy
,
the
cal
culat
ion
an
d
sta
ti
st
ic
s
of
the
featur
es
of
th
e
sk
in
diseases
will
be
the
basis
fo
r
the
sel
e
ct
ion
of
featu
r
es
that
can
re
flect
the
sk
in
featu
res
at
diff
e
re
nt
le
vels.
Af
te
r
a
naly
zi
ng
datas
et
s
from
el
even
ty
pes
of
dif
fer
e
nt
featur
e
s
us
in
g
the
GLCM
al
go
rithm
,
the
resu
lt
s
sh
owe
d
th
at
there
are
six
op
ti
m
a
l
featur
es
cho
se
n:
co
nt
rast,
energy,
ho
m
ogeneit
y,
m
ean,
sta
nd
a
rd
de
viati
on,
e
ntropy
.
T
hese
feature
ty
pes
s
how
t
he
l
arg
e
c
ha
nge
be
tween
diff
e
re
nt sk
i
n d
ise
ases an
d
a
re
u
se
d
t
o
ef
fecti
vely
train
for
t
he
s
kin disease
classi
ficat
ion
.
2.3.
Multila
ye
r neur
al
ne
t
w
ork
s
truc
tu
r
e
f
or
skin
dise
as
e cl
as
sific
ati
on
In
this
resea
rc
h,
as
the
num
ber
of
sk
i
n
di
seases
increas
es,
it
beco
m
es
m
or
e
diff
ic
ul
t
to
directl
y
cl
assify
sk
in
di
seases
with
hig
h
acc
uracy
w
it
ho
ut
pr
e
-
proc
essing
im
ages.
Ther
e
f
or
e
,
it
is
necessa
ry
to
pr
e
-
process
the
im
age
befor
e
s
egm
entat
ion
of
separ
at
i
ng
R
OI
s
a
nd
this
will
op
tim
iz
e
trai
ning
proce
ss
and
cl
assifi
cat
ion
with
high
acc
uracy
us
i
ng
the
MLNN
.
P
re
vious
w
orks
ha
ve
sho
wn
that
th
e
ne
ur
al
netw
ork
can
be
well
app
li
e
d
in
m
edical
diagnostic
syst
e
m
s
[2
9
]
,
[
30]
.
Ther
e
f
or
e,
in
our
arti
cl
e,
the
cl
assifi
cat
ion
of
sk
in
diseases
will
be
perform
ed
usi
ng
t
he
MLN
N,
this
ca
n
re
su
lt
in
a
highe
r
accu
racy
cl
assifi
cat
ion
us
i
ng
six
op
ti
m
al
f
eat
ur
e
s of s
kin
diseas
e i
m
ages.
The
ML
N
N
is
the
a
bili
ty
of
bette
r
processi
ng
com
p
le
x
re
la
ti
on
sh
i
ps
between
di
ff
e
ren
t
pa
ram
et
ers
and
t
hen
e
ffe
ct
ively
cl
assify
ing
ba
sed
on
le
arn
in
g
f
r
om
the
process
ed
trai
ni
ng
da
ta
.
The
s
ucce
ss
of
a
cl
assifi
cat
ion
s
yst
e
m
based
on
the
MLN
N
de
pends
on
t
he
m
od
el
arch
it
ect
ur
e
of
the
net
work
a
nd
the
t
rainin
g
a
lgorit
hm
.
Fu
r
therm
or
e,
the
nu
m
ber
of
hi
dden
la
ye
rs
as
well
as
the
nu
m
ber
of
node
s
in
the
net
wor
k
a
re
determ
ined
us
i
ng
the
tria
l
an
d
erro
r
m
e
tho
d
du
ri
ng
the
cl
assifi
cat
ion
pro
cess
rep
eat
e
d.
In
pa
rtic
ular
,
the
loss
functi
on
MSE
an
d
t
he
act
iv
at
ing
functi
on
Lo
g
-
sigm
oid
can
be
sel
ect
ed
t
o
be
s
uitable
f
or
trai
ning
a
nd
cl
assifi
cat
ion
of s
kin
diseases
eff
ect
ively
.
In
t
he
MLN
N
with
back
-
prop
a
gatio
n
,
m
ean
s
qu
a
re
e
rror
(MSE
)
is
the
m
os
t
com
m
on
ly
us
ed
regressio
n
lo
ss
functi
on.
T
he
MSE
is
the
sum
of
squa
red
di
sta
nces
be
tw
e
en
the
ta
r
get
va
riable
an
d
pr
e
dicte
d
values
and calc
ulate
d,
=
1
∑
(
−
̂
)
=
1
2
(10)
wh
e
re
y
i
is
the
desire
d
ne
ur
al
netw
ork
outp
ut,
a
nd
i
y
ˆ
is
the
neural
netw
ork
outp
ut
an
d
n
is
the
nu
m
ber
of
ou
t
pu
t
no
des.
In
a
dd
it
io
n,
in
this
MLNN,
log
si
g
is
a
tran
sfer
f
unct
ion
f
or
ca
lc
ulati
ng
a
la
ye
r’
s
the
outp
ut
y
from
it
s n
et
input x
:
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
ROI
-
ba
se
d
fe
atu
res f
or
cl
as
sif
ic
ation
of ski
n disease
s
us
in
g a multi
-
layer
neur
al…
(
Thanh
-
H
ai
Ng
uye
n
)
221
(
)
=
1
1
+
−
(11)
In
t
his
ML
NN,
the
i
nput
dataset
is
a
vect
or
of
six
opti
m
al
featur
es
e
xtracted
f
ro
m
the
G
LCM
al
gorithm
.
In
add
it
io
n,
eac
h
sk
in
diseas
e
in
five
ty
pes
is
con
st
ru
ct
e
d
wit
h
a
set
of
10
0
i
m
ages,
wh
ic
h
woul
d
gr
eat
ly
enha
nc
e the classi
fica
ti
on
acc
ur
acy
.
3.
E
X
PERI
MEN
TAL RES
UL
TS
3.1.
Skin
dise
as
e d
atase
ts
Dataset
s,
i
n
th
is
stud
y,
are
f
ro
m
IS
IC
data
base,
inclu
ding
10,00
0
im
ages
f
or
se
ven
ty
pes
of
ski
n
diseases.
For t
he
e
valuati
on
of cla
ssific
at
ion
eff
ect
ive
ness,
500
im
ages of skin
diseases
wer
e c
hose
n
to
b
e
100
i
m
ages
for
eac
h
ty
pe,
i
n
wh
i
ch
eac
h
dataset
was
div
ide
d
80
im
ages
fo
r
trai
ning
a
nd
20
on
es
f
or
te
st
ing
a
s
descr
i
bed in
T
able 1.
Table
1.
Re
pr
e
sentat
ion o
f da
ta
set
s f
or trai
nin
g
an
d
te
sti
ng
of f
ive
ty
pes of ski
n disease
Tr
ain
in
g
and
testin
g
i
m
ag
es
Bas
al cell
c
arcino
m
a
Ben
ig
n
keratos
is
Der
m
ato
f
ib
ro
m
a
Melano
cy
tic
nev
u
s
Melano
m
a
Tr
ain
in
g
100
100
100
100
100
Testin
g
20
20
20
20
20
3.2.
Sep
ar
at
i
on o
f ROI
In
this
a
rtic
le
,
the
sk
in
diseas
e
i
m
ages
wer
e
resized
to
the
512x512
sam
e
siz
e
and
the
n
la
beled
as
sh
ow
n
in
Fig
ur
e
1.
Be
f
or
e
separ
at
in
g
RO
I
from
a
sk
in
disease
i
m
age,
i
m
age
proc
essing
m
et
ho
ds
were
app
li
ed
.
I
n
pa
rtic
ular,
the
i
m
age
with
the
ROI
was
e
nh
anced
us
in
g
t
he
ke
rn
el
K
1
in
(
1)
as
descri
bed
in
Figure
2.
I
n
ad
diti
on
, u
nneces
sary
no
ise
s
in
the
i
m
age
wer
e
rem
ov
ed
us
i
ng
the
Butt
erw
ort
h
high
-
pass
filt
er
i
n
the
fr
e
que
ncy
do
m
ai
n
as
sho
wn
i
n
Fig
ure
3
.
T
his
im
age
was
c
onver
te
d
to
the
bin
a
ry
i
m
age
with
the
blac
k
ROI
a
nd
the
w
hite
bac
kg
r
ound
as
sho
wn
in
Figure
4
.
Ne
xt
,
the
bin
a
ry
im
age
was
proce
ssed
t
o
c
on
ti
nu
ou
sly
el
i
m
inate
unne
cessary
d
et
ai
ls
us
in
g
the
m
ed
ia
n
filt
er
as
sh
own
in
Fig
ure
5
,
the
n
it
re
m
ov
ed
s
kin
hair
detai
l
us
in
g
th
e
bott
om
-
hat
filt
er
with
kernel
K
2
in
(3)
as
desc
ribe
d
in
Fig
ur
e
6.
The
e
rosio
n
al
gorithm
was
a
pp
li
ed
to f
il
l t
he
s
ki
n disease
a
reas a
s sho
wn in
Fig
ur
e
7.
Fo
r
the
RO
I
s
epar
at
io
n,
the
Ca
nn
y
e
dge
de
te
ct
ion
was
ut
il
iz
ed
in
the
bin
a
ry
im
age
as
sho
wn
in
Figure
8.
The
ero
si
on
m
et
ho
d
was
a
pp
li
ed
for
li
nkin
g
e
dg
es
of
obj
ect
s
a
s
sho
wn
i
n
Fi
gure
9,
the
n
the
y
wer
e
fill
ed
al
l
obj
ect
reg
i
ons
a
nd
la
beled
a
s
s
how
n
in
Fi
gure
10.
The
R
OI
area
i
n
fill
ed
bi
nar
y
i
m
age
was
e
xtr
act
ed
base
d
on
t
he
la
rg
est
R
OI
a
rea
as
show
n
i
n
Fi
gure
11
,
the
n
it
was
m
ulti
plied
to
t
he
e
nh
a
nc
ed
or
igi
nal
im
a
ge
to
pro
du
ce
t
he
e
nhance
d
or
i
gin
a
l
ROI
as
sho
w
n
in
Fig
ur
e
12
befor
e
extr
act
ing
featu
res
for
cl
assifi
cat
ion
of
s
ki
n
diseases.
Figure
1. Im
age r
esi
zed
512x51
2
Figure
2. Im
age af
te
r
enh
a
ncem
ent
Figure
3. Im
age af
te
r
bu
tt
er
w
or
th
hi
g
h
-
pass
Figure
4. Bi
nary
i
m
age
0
0
0
0
5
.
1
0
0
0
0
1
K
0
0
1
1
1
0
0
0
1
1
1
1
1
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
1
1
1
1
0
0
0
1
1
1
0
0
2
K
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.
23
, N
o.
1
,
Ju
ly
2021
:
2
1
6
-
2
2
8
222
Figure
5. Im
age af
te
r
t
he
m
edian
filt
er
Figure
6. Im
age
el
i
m
inate
d
s
kin hairs
Figure
7. Im
age af
te
r
t
he
ero
si
on
Figure
8. Im
age w
it
h
t
he
Ca
nn
y e
dge
de
te
ct
ion
Figure
9. Ed
ge
i
m
age
erode
d
Figu
re
10. Im
a
ge wit
h
the f
il
le
d o
bj
ec
ts
Figure
11. Im
a
ge wit
h
the lar
gest area
Figure
12. Ori
gin
al
R
OI
i
m
age
All
sk
in
diseas
e
i
m
ages
hav
e
been
processe
d
us
i
ng
im
age
processin
g
m
et
hods
.
Fi
gure
13
s
how
e
d
ROI
im
ages
separ
at
e
d
f
r
om
the
proce
ssed
i
m
age
of
5
t
y
pes
of
sk
i
n
di
seases
s
uch
a
s
basal
cel
l
car
ci
no
m
a
(F
ig
ur
e
a0
-
a1
),
ben
i
gn
ke
r
at
os
is
(F
i
gure
b0
-
b1),
de
r
m
at
of
ibr
om
a
(F
ig
ur
e
c0
-
c1
),
m
el
ano
cy
ti
c
ne
vus
(F
ig
ur
e
d0
-
d1
)
,
m
el
ano
m
a
(
Figure
e
0
-
e
1).
The
RO
I
im
age
retai
ns
m
os
t
of
t
he
dis
eased
s
kin
are
a
an
d
unnecess
a
ry
ar
eas
are
rem
ov
ed
f
or
cl
assifi
cat
ion
.
T
he
ac
cur
at
e
R
OI
se
par
at
io
n
is
im
portant,
beca
use
th
e
a
m
ou
nt
of
t
he
i
m
po
rtant
feat
ur
e
i
nfor
m
at
ion
in
the
R
OI
m
akes
it
cal
culat
e
featur
es
fas
te
r
an
d
m
or
e
accurate
for
cl
assifi
cat
ion.
T
her
e
f
or
e,
featur
e
s
in
the
ROI
i
m
ages
w
ere
e
xtracted
usi
ng
t
he
GLC
M
m
et
ho
d,
in
wh
ic
h
6
op
ti
m
al
feature
s
of
11
ones
wer
e
sel
ect
ed
for
trai
ning
an
d
cl
assify
sk
in
diseases.
W
it
h
the
or
i
gin
al
ROIs
separ
at
e
d
f
r
om
5
s
kin
disease
s,
it
is
obvious
that
there
is
t
he
str
uctu
ral
di
ff
e
ren
ce
am
ong
s
ha
pes
,
c
olo
r
s
an
d
oth
e
rs.
From
t
hese
diff
e
re
nt
factors,
6
op
i
t
m
al
featur
es
e
xtracted
us
i
ng
the
G
LCM
possibly
e
nh
a
nc
e
the
cl
assifi
cat
ion
a
ccur
acy
.
(a0)
(a1
)
(b0)
(b1)
(c0)
(c1)
(d0)
(d1)
(e0)
(e1)
Figure
13. Re
presentat
io
n of t
he
or
i
gin
al
a
nd the R
OI
im
ages s
epa
rated
aft
er e
nh
a
ncem
ent an
d se
gen
ta
ti
on
;
(a0)
basal
cel
l
carci
-
nom
a
, (
a
1
)
ROI o
f
basa
l
cel
l carci
-
no
m
a i
m
age
,
(b0
)
be
nign
ker
at
osi
s
,
(
b1)
R
OI of
ben
i
gn
ke
ratosi
s im
age
,
(c0)
de
rm
at
o
-
fibrom
a
,
(c
1) RO
I of
der
m
at
o
-
fib
rom
a
i
m
ag
e
,
(d0
)
m
e
la
no
-
cy
ti
c
nevus
,
(d1) RO
I of
m
el
ano
-
cy
ti
c n
ev
us
im
age
,
(e
0)
m
el
ano
-
ma
, and
(e1) R
OI
of
m
el
ano
-
m
a i
m
age
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
ROI
-
ba
se
d
fe
atu
res f
or
cl
as
sif
ic
ation
of ski
n disease
s
us
in
g a multi
-
layer
neur
al…
(
Thanh
-
H
ai
Ng
uye
n
)
223
3.3.
Fe
ature
e
xt
r
act
i
on
fr
om
ROIs
u
sing
a
GL
C
M alg
or
ithm
Figure
14
s
ho
wed
t
he
val
ue
s
of
11
feat
ures
cal
culat
ed
us
in
g
the
GL
CM
al
go
rithm
,
in
w
hich
10
i
m
ages
of
Ba
sa
l
cel
l
carcino
m
a
(cell
carci
nom
a)
wer
e
use
d.
Ba
sed
on
in
f
orm
ation
in
Fig
ur
e
14,
we
can
easi
l
y
evaluate
that
e
ach
of
11 f
eat
ures
f
or
t
he
10
s
kin
d
ise
ase
im
ages
is
nea
rly
si
m
i
la
r.
This
is
the
basis o
f
us
i
ng
t
he
GLMC al
gorit
hm
f
or
e
xtrac
ti
ng f
eat
ur
es
.
Table
2
an
d
T
able
3
showe
d
the
Mi
n
-
Ma
x
threshold
values
of
each
fe
at
ur
e
in
five
t
ypes
of
s
kin
diseases,
i
nclu
ding:
bas
al
ce
ll
carcino
m
a
(No.
1),
be
nig
n
ke
rato
sis
(
No.
2)
,
de
rm
at
of
ib
ro
m
a
(No.
3),
m
el
ano
cy
ti
c
nev
us
(num
ber
4),
m
el
ano
m
a
(
No.
5).
I
n
ad
diti
on
,
f
ro
m
Table
2,
two
feature
s
of
sm
oo
th
ne
ss
an
d
ID
M
hav
e
t
he
Mi
n
(1.0)
an
d
Ma
x
(
1.0)
valu
es
corres
ponding
t
o
each
dise
ase
with
ou
t
c
ha
ng
e
,
res
pecti
ve
ly
,
so
it
was
no
t
sel
ect
ed.
I
n
ad
diti
on
to
tw
o
Sm
oo
t
hn
e
ss
an
d
I
DM
featu
res,
correla
ti
on
fea
ture
has
the
to
o
sm
all
diff
e
re
nce
bet
ween
Mi
n
an
d
Ma
x
value
s
,
j
us
t
0.0
5,
it
was
no
t
ch
ose
n.
Wh
il
e
two
feat
ur
e
pa
irs
of
m
ean
-
va
ri
a
nce
and
RM
S
-
co
nt
rast
are
sim
il
a
r,
s
o
we
just
cho
os
e
one
pair
of
m
ean
an
d
c
on
t
rast
.
T
he
gr
oup
of
six
featu
res
sel
ect
ed
inclu
ding:
con
t
rast,
en
erg
y,
ho
m
og
e
ne
it
y,
m
ean,
stan
da
r
d
de
viati
on
,
an
d
ent
ropy
diff
e
r
gr
eat
ly
whe
n
c
om
par
ing t
he m
in an
d m
ax
va
lues fo
r disea
ses and ca
n be
us
e
d
f
or traini
ng in
the clas
si
fier.
Figure
14. Re
presentat
io
n of
11 f
eat
ur
es
of
10 Basal
cell
c
arcin
om
a i
m
ages
Table
2.
Ele
ve
n
a
ver
a
ge feat
ures e
xtracted
fr
om
f
ive ty
pes of ski
n disease
us
in
g
t
he GLC
M
Sk
in
dis
ease class
Co
n
trast
Ho
m
o
g
en
eit
y
Entro
p
y
Mean
RMS
Energy
m
in
m
a
x
m
in
m
a
x
m
in
m
a
x
m
in
m
a
x
m
in
m
a
x
m
in
m
a
x
1
0
.03
7
0
.21
1
0
.95
8
0
.99
5
0
.12
7
0
.73
3
0.
018
0
.20
5
0
.04
2
0
.37
4
0
.34
6
0
.93
9
2
0
.05
8
0
.82
5
0
.94
6
0
.99
3
0
.23
8
0
.97
8
0
.03
9
0
.41
4
0
.09
6
0
.55
2
0
.33
5
0
.82
5
3
0
.02
2
0
.15
8
0
.96
9
0
.99
8
0
.03
3
0
.35
5
0
.00
3
0
.06
7
0
.02
2
0
.15
3
0
.46
3
0
.97
4
4
0
.03
6
0
.32
4
0
.95
3
0
.99
7
0
.13
0
0
.97
3
0
.01
8
0
.99
7
0
.05
2
0
.56
3
0
.33
4
0
.9
47
5
0
.03
6
0
.42
0
0
.91
5
0
.99
5
0
.07
6
0
.80
6
0
.00
9
0
.24
7
0
.05
0
0
.41
7
0
.25
2
0
.94
1
Table
3.
Ele
ve
n
a
ver
a
ge feat
ures e
xtracted
fr
om
f
ive ty
pes of ski
n disease
us
in
g
t
he GLC
M
Sk
in
d
iseas
e class
Stan
d
ard
Dev
iatio
n
Varience
S
m
o
o
th
n
ess
Co
rr
elatio
n
IDM
m
in
m
a
x
m
in
m
a
x
m
in
m
a
x
m
in
m
a
x
m
in
m
a
x
1
0
.13
1
0
.40
4
0
.01
4
0
.12
3
1
.00
0
1
.00
0
0
.93
8
0
.98
8
1
.00
0
1
.00
0
2
0
.19
4
0
.49
3
0
.02
2
0
.16
0
1
.00
0
1
.00
0
0
.96
5
0
.98
9
1
.00
0
1
.00
0
3
0
.05
9
0
.25
0
0
.00
3
0
.04
6
1
.00
0
1
.00
0
0
.86
1
0
.97
3
1
.00
0
1
.00
0
4
0
.13
3
0
.49
1
0
.01
5
0
.17
7
1
.00
0
1
.0
00
0
.95
3
0
.98
9
1
.00
0
1
.00
0
5
0
.09
6
0
.43
1
0
.00
9
0
.14
1
1
.00
0
1
.00
0
0
.83
5
0
.98
4
1
.00
0
1
.00
0
Figure
15
repr
esented
si
x
fea
ture
sta
ti
sti
cs
fo
r
five
im
age
dataset
s
of
sk
i
n
disease
s
to
il
lustrate
th
e
diff
e
re
nce.
T
hi
s
is
the
basis
for
c
hoos
i
ng
these
six
fea
tur
es
for
trai
ni
ng
i
n
the
cl
a
ssifie
r.
In
pa
rtic
ular
,
Figure
15
(
a
)
presented
th
e
co
ntrast
pa
ram
eter
s
of
five
cl
as
ses
of
s
kin
dis
eases
aver
a
ge
d
from
10
0
im
a
ges
f
or
each
cl
ass.
It
is
si
m
il
ar
to
c
al
culat
ion
of
the
rem
ai
nin
g
featu
re
s
su
c
h
as
the
entr
op
y
values
a
s
sho
wn
i
n
Figure
15
(
b
);
the
unifo
rm
ity
value
as
s
how
n
in
Fi
gure
15
(c)
;
t
he
value
of
ene
r
gy
as
s
how
n
in
Fi
gur
e
15
(
d);
the
m
ean
value
as sh
own
in
Figure 15
(
e)
; and
sta
nd
ar
d
de
via
ti
on
value
as s
how
n
in
Figure
1
5
(
f). From
Fi
gure
15,
it
can
be
s
een
that
the
cl
ass
of
de
rm
at
o
fibrom
a
(N
o.
3)
has
t
he
lo
wes
t
featur
e
value
s,
inclu
ding
c
ontrast
,
entr
op
y,
m
ean
and
sta
ndar
d
de
viati
on
,
w
hile
the
two
featu
r
es
of
hom
og
en
ei
ty
and
ene
rgy
are
the
high
est
at
0.996
a
nd
0.9
26,
res
pecti
vely
.
I
n
a
dd
it
io
n,
f
r
om
the
data
in
Fi
gure
15,
m
elan
ocyt
ic
nevus
(
No.
4)
has
th
e
2nd
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.
23
, N
o.
1
,
Ju
ly
2021
:
2
1
6
-
2
2
8
224
highest
featu
re
values
of
al
l
five
diseases,
fo
ll
owe
d
by
ba
sal
cel
l
carcino
m
a
disease
(No.
1).
Th
r
ough
t
he
analy
sis
of
the
m
ean
values
of
the
s
kin
di
sease
featur
e
s,
it
can
be
seen
that
the
dif
f
eren
ce
betwee
n
f
ive
disease im
age d
at
aset
s is
qu
it
e cle
ar fo
r
a
pp
l
yi
ng
to
d
ise
ase
classi
ficat
ion
.
(a)
(b)
(c
)
(d
)
(e)
(f)
Figure
15.
Re
presentat
io
n of t
he
a
ver
a
ge
sta
t
ist
ic
s o
f
6 feat
ures
of 5 ski
n disease
d
at
aset
s
;
(a)
re
pr
ese
ntati
on
of co
ntrast
,
(
b)
. r
e
pr
e
sentat
io
n of entr
opy
,
(
c). rep
rese
ntati
on of
hom
og
en
ei
ty
,
(d
)
. rep
res
entat
ion
of e
ne
rg
y
,
(e)
.
r
e
pr
e
sentat
ion
of m
ean
, a
nd
(
f). r
e
presen
ta
ti
on
of sta
nd
ard de
viati
on
3.4.
Clas
si
ficat
i
on
accur
ac
y
usin
g
a ML
NN st
r
uctu
r
e
The
ML
NNs
m
od
el
e
m
plo
ye
d
in
t
his
st
udy
include
s
an
input
la
ye
r
with
6
nodes
co
rr
e
sp
on
ding
to
6
featur
e
vectors;
3
hidde
n
l
ay
ers
wit
h
100
nodes
f
or
each
la
ye
r;
a
nd
an
ou
t
pu
t
la
ye
r
with
5
node
s
corres
pondin
g
5
cl
asses
of
ski
n
disease
needed
for
cl
assi
ficat
ion
as
s
how
n
in
Fig
ur
e
16.
I
n
a
ddit
ion
,
m
or
e
or
le
ss
hidde
n
la
y
ers
c
ou
l
d
be
c
ho
s
en
to
possi
bly
ens
ur
e
the
best
cl
assi
ficat
ion
.
The
ML
NN
was
em
plo
ye
d
t
o
perform
trai
ni
ng
with
the
le
arn
i
ng
s
pee
d
of
10
-
4
with
un
c
ha
ng
e
duri
ng
t
he
le
arn
i
ng
proce
ss,
the
desire
d
m
od
el
err
or
was
7.10
-
3.
F
igure
17
s
ho
wed
t
he
trai
nin
g
er
ror
c
urv
e,
in
wh
ic
h
t
he
er
r
or
of
th
e
m
od
el
con
ti
nu
ously
decr
eased
f
ollo
wing
the
curve
and
it
achieved
the
best
valu
e
of
0.0
068
aft
er
449
ep
oc
hs
.
This
sh
ows
that the
m
od
el
achieve
d
c
onve
rg
e
nce
with
fast trai
ni
ng tim
e.
Af
te
r
t
rainin
g
400
im
ages
of
5
cl
asses
,
the
MLN
N
was
a
pp
li
ed
to
cl
assify
sk
in
diseases
.
Cl
assifi
cat
ion
resu
lt
s
wer
e
t
est
ed
on
100
i
m
ages
corre
sp
on
ding
to
5
disease
cl
ass
es.
To
e
valua
te
the
cl
assifi
cat
ion
accuracy,
a
c
onfu
si
on
m
at
rix
in
Figure
18
was
em
plo
ye
d
to
sh
ow
th
e
a
ver
a
ge
cl
assifi
cat
ion
accuracy
of
92%,
in
wh
ic
h
the
accuracy
of
the
sk
in
dise
ases
is
85
%
ba
sal
cel
l
carci
no
m
a
(N
o.
1),
95%
Be
nign
ker
at
osi
s
(No.
2),
10
0%
de
rm
at
of
ibr
om
a
(N
o.
3),
85%
m
el
ano
cy
ti
c
nevus
(
No.
4),
95%
m
el
a
no
m
a
(No.
5),
res
pe
ct
ively
.
In
the
c
la
ssific
at
ion
resu
lt
of
dis
eases,
de
rm
ato
-
fib
ro
m
a
disease
has
the
highest
accuracy
of
95
%
due
to
it
s
f
e
at
ur
e
bein
g
ve
ry
dif
fer
e
nt
co
m
par
ed
to
4
re
m
ai
nin
g
diseas
es.
Wh
il
e
basa
l
cel
l
carcin
om
a
and
m
el
ano
cy
ti
c
nevus
diseases
hav
e
the
l
owest
accuracies
of
85
%
.
I
n
th
e
case
of
basa
l
cel
l
carcin
om
a
cl
as
sific
at
ion
,
the
m
ino
r
er
ror
cl
assifi
cat
ion
is
du
e
t
o
it
s
featur
e
m
ai
nly
con
f
us
e
d
with
th
at
of
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
ROI
-
ba
se
d
fe
atu
res f
or
cl
as
sif
ic
ation
of ski
n disease
s
us
in
g a multi
-
layer
neur
al…
(
Thanh
-
H
ai
Ng
uye
n
)
225
ben
i
gn
ke
ratos
is.
I
n
par
ti
cula
r
,
w
he
n
cl
assify
ing
m
el
ano
cy
ti
c
ne
vus
diseas
e,
2
of
20
im
a
ges
(10%
of
th
e
total
nu
m
ber
)
is
er
r
or due t
o
c
onfusing wit
h B
asa
l cel
l carci
no
m
a d
ise
ase.
Figure
16. Clas
sific
at
ion
m
odel
o
f
the ML
N
N
str
uctu
re
for 6 in
put feat
ur
e
s and
5 o
utput
cl
asses
Figure
17. T
rai
ning
resu
lt
us
i
ng the ML
N
N
structu
re
Figure.
18. C
onf
us
io
n
m
at
rix for e
va
luati
on
of 20
te
sti
ng
im
ages each
class
Table
4
pr
e
sen
te
d
the
com
pari
so
n
of
t
he
acc
ur
at
e
res
ults
cl
assify
ing
5
cl
a
sses
of
sk
i
n
dis
eases
based
on
3
gro
up
s
of
diff
ere
nt
featu
res.
The
a
ver
a
ge
res
ults
sh
owed
that
the
tr
ai
nin
g
m
od
el
us
in
g
only
6
featur
e
s
(
co
ntrast,
ene
r
gy,
ho
m
og
e
neity
,
m
ean,
sta
nd
ard
de
viati
on,
entr
op
y
)
has
th
e
92%
highest
accuracy;
the
lowest
accuracy
of
71%
us
i
ng
11
fe
at
ur
es,
a
nd
usi
ng
on
ly
3
featur
es
producin
g
78%.
It
is
obvi
ou
s
that
t
he
s
el
ect
e
d
gro
up of
6 feat
ur
es
u
si
ng the
GLCM re
prese
n
ts t
he
ef
fecti
ve
ness o
f
cl
assif
yi
ng
the
f
i
ve
s
kin
diseases
.
Table
5
s
howe
d
that
the
res
ul
t
us
ing
the
pro
po
s
ed
m
et
ho
d
has
the
92%
cl
assifi
cat
ion
ac
cur
acy
f
or
5
cl
asses
of
sk
i
n
diseases,
it
is
2%
higher
t
ha
n
the
best
m
eth
od
[38]
of
th
e
pr
e
vious
res
earches
an
d
a
bout
6%
higher
t
ha
n
th
e
lowest
acc
ur
acy
[34].
W
it
h
the
high
cl
ass
ific
at
ion
acc
uracy
than
ks
to
i
m
age
processi
ng
to
ext
ract
the
R
OI
with
a
ppr
opriat
e
m
e
thods,
i
n
w
hich
alm
os
t
inform
at
ion
is
possi
bly
kep
t
i
n
the
RO
I
.
More
ov
e
r,
the
GLCM
has
be
en
ap
plied
f
or
m
any
pr
evi
ous
stud
ie
s
[
38]
and
res
ults
has
sh
ow
n
ve
ry
po
sit
ive.
Howe
ver,
in
this
stud
y,
we
chose
on
ly
6
f
eat
ur
es
of
11
f
eat
ur
es
that
can
highly
con
ta
in
a
lot
of
i
m
p
or
ta
nt
inf
or
m
at
ion
relat
ed
to
sk
in
dis
ease.
In
a
dd
it
ion,
the
6
featu
res
ap
plied
for
trai
ning
to
be
a
ble
to
conde
nse
and
le
ss tim
e, so
it
can in
c
rease
th
e cla
ssific
at
ion acc
uracy
.
In
ad
diti
on
to
the
sel
ect
ion
of
i
m
age
pr
oce
s
sing
m
et
ho
ds
and
the
sel
ect
ion
of
6
f
eat
ures,
Table
5
sh
owe
d
that
t
he
MLN
N
wa
s
pro
po
se
d
for
th
e
ap
propriat
e
num
ber
of
no
de
s
an
d
la
ye
rs
t
o
achieve
a
res
ul
t
with
higher
acc
ur
ac
y
com
par
ed
to
oth
er
m
od
el
s
as
CNN
[
31]
,
FRC
NN
[34],
Dep
t
hw
ise
se
pa
rab
le
CN
N
[
32]
and
SV
M
[38].
I
n
par
ti
cula
r,
a
uthors
re
presente
d
com
bin
in
g
th
e
GLCM
an
d
t
he
SV
M
f
or
cl
a
ssifyi
ng
3
cl
as
ses
of
sk
in
diseases
a
nd
ac
hieve
d
th
e
accuracy
of
90%.
I
n
a
dd
it
io
n,
ou
r
pro
pose
d
m
et
ho
d
he
re
has
le
ss
trai
ning
tim
e
Evaluation Warning : The document was created with Spire.PDF for Python.