Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
3
,
June
2020,
pp. 2
98
6
~
2996
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
3
.
pp2986
-
29
96
2986
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Decomp
osition o
f color w
avelet
with highe
r order
statisti
cal
textur
e an
d convolutional
neural n
etwork f
eatur
es
s
et based
classific
ation o
f c
olorectal
polyps f
ro
m
video end
osc
opy
A.
S. M.
Shafi
, Mohamm
ad
Motiur
R
ah
m
an
Depa
rtment
o
f
C
om
pute
r
Scie
n
ce a
nd
Engi
n
ee
rin
g,
Mawla
n
a
Bh
a
shani
Sci
enc
e
an
d
Technol
og
y
U
nive
rsit
y
,
Bang
l
ade
sh
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
21, 201
9
Re
vised N
ov 25, 2
019
Accepte
d Dec
1,
2019
Gastroint
esti
n
al
ca
nc
er
is
one
of
the
leading
ca
us
es
of
dea
th
a
cro
s
s
the
world.
The
gastro
intestina
l
po
l
y
ps
are
conside
red
as
th
e
pre
cur
sors
of
deve
lop
ing
thi
s
m
al
igna
nt
c
anc
er
.
In
orde
r
t
o
conde
nse
the
proba
bil
i
t
y
of
c
anc
er
,
e
ar
l
y
det
e
ct
ion
and
r
e
m
oval
of
col
ore
ct
a
l
pol
y
ps
ca
n
be
cogi
t
at
ed
.
Th
e
m
ost
used
dia
gnostic
m
odal
ity
for
col
o
r
ec
t
al
pol
y
ps
is
vide
o
endosc
op
y
,
b
ut
the
accurac
y
of
dia
gnosis
m
ost
l
y
de
pends
on
doct
ors'
expe
ri
e
nce
that
is
cru
cial
to
d
et
e
ct
pol
y
ps
in
m
an
y
c
ase
s.
Com
pute
r
-
a
ide
d
pol
y
p
det
e
ct
ion
is
prom
ising
to
red
uce
the
m
iss
d
et
e
ct
ion
ra
te
of
the
pol
y
p
and
t
hus
improve
the
accurac
y
of
dia
gnosis
result
s.
The
proposed
m
et
hod
first
det
ec
ts
po
l
y
p
and
non
-
pol
y
p
t
hen
il
lustrates
a
n
aut
om
at
ic
pol
y
p
class
ifi
c
at
ion
te
chni
qu
e
from
endosc
opic
vide
o
th
rough
col
or
wave
let
with
highe
r
-
ord
er
stat
isti
cal
te
xtur
e
feature
a
nd
Convolut
ional
Neura
l
Ne
twork
(CNN
).
Gra
y
Le
ve
l
Run
Le
ngth
Ma
tri
x
(
GLRLM)
is
use
d
for
highe
r
-
ord
er
statistical
te
xt
ure
fe
at
ur
e
s
of
diffe
ren
t
dir
ec
t
ions
(
Ɵ
=
0
0
,
45
0
,
90
0
,
135
0
).
The
fea
tur
es
are
fed
int
o
a
l
inear
support
vec
tor
m
ac
hine
(SV
M)
to
tra
in
t
he
class
ifi
er
.
The
expe
r
iment
al
result
demo
nstrat
es
that
t
he
proposed
appr
oac
h
is
auspic
ious
and
o
per
ative
with
res
idua
l
ne
twork
ar
chi
t
ec
tur
e,
whi
c
h
tri
um
phs
the
best
p
erf
or
m
anc
e
of
a
cc
ur
acy
,
s
ensit
ivit
y
,
and
spec
i
ficit
y
of
98.
83%
,
97.
87%,
and
99
.
13%
respe
c
ti
v
ely
fo
r
c
la
ss
ifica
t
i
on
of
co
lorect
a
l
pol
y
ps
on
standa
rd
pub
li
c
endosc
opic vi
d
e
o
databa
ses.
Ke
yw
or
d
s
:
Color
ect
al
ca
nc
er
Conv
olu
ti
onal
neural
netw
ork
Gastr
oin
te
sti
na
l l
esi
on
s
Run l
en
gth
m
at
rix
Suppor
t
v
ect
or m
achine
Vide
o
e
ndos
c
opy
Wav
el
et
Copyright
©
202
0
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
:
A.
S.
M.
S
haf
i,
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce a
nd E
ng
i
ne
erin
g,
Ma
wlana B
has
han
i
Scie
nce
a
nd Tec
hnology
Unive
rsity
,
Santosh
, Ta
ngai
l
-
1902, Ba
ngla
desh.
Em
a
il
:
sh
afi.cs
e.m
bs
tu1
1@g
m
ai
l.co
m
1.
INTROD
U
CTION
Color
ect
al
ca
nc
er
or
bowel
i
s
the
sec
ond
m
os
t
pr
om
inent
cause
of
ca
nc
er
in
w
om
en
and
the
t
hir
d
m
os
t
le
ading
c
ause
i
n
m
en
[
1].
Althou
gh
c
olorect
al
poly
ps
a
re
foreru
nners
to
col
or
e
ct
al
cancer
,
f
or
thes
e
po
ly
ps
it
ta
kes
sever
al
ye
ars
to
pote
ntial
ly
t
ran
s
f
or
m
into
cancer
[2
]
,
whic
h
m
a
y
dev
el
op
if
t
hese
pol
yps
ar
e
le
ft
un
treat
e
d.
Ma
j
or
ty
pes
of
color
ect
al
pol
yps
are
ade
no
m
as,
hyperplas
ti
c
and
serated
.
If
early
detect
ion
a
n
d
cl
assifi
cat
ion
of
these
po
ly
ps
is
possible,
the
y
can
be
rem
ov
ed
be
f
or
e
this
transm
issi
on
arises.
Seve
ral
te
sts
are
rec
omm
end
ed
i
n
al
l
colon
ca
ncer
sc
re
enin
g
guideli
ne
s,
pa
rtic
ularly
vid
e
o
en
dosc
op
y
a
nd
fecal
occu
lt
blood
te
sti
ng.
W
it
hin
the
U
ni
te
d
Stat
es,
vide
o
e
ndos
c
opy
is
the
m
os
t
co
m
m
on
ly
util
ized
te
st
a
nd
s
houl
d
be
perform
ed
eve
ry
10
ye
a
rs
in
aver
a
ge
-
risk
pe
rsons.
D
ur
i
ng
an
en
dosco
py,
a
long,
fle
xib
l
e
tub
e
(
col
onosc
ope)
is
inserted
into
the
bo
dy.
A
ti
ny
vi
deo
c
am
e
ra
at
the
ti
p
of
the
tu
be
al
lo
ws
the
doct
or
to
vie
w
the
i
nsi
de
of
the
entire
col
on
to
detect
a
nd
rem
ov
e
a
po
ly
p.
Disti
nguis
hin
g
f
ro
m
low
-
ri
sk
po
ly
ps
wit
h
high
-
risk
col
orect
al
po
ly
ps
is
a
n
i
m
po
rtant
pa
rt
of
c
olorect
al
cance
r
sc
ree
ning
th
r
ough
the
de
te
ct
ion
and
hist
op
at
holog
ic
al
char
act
e
rizat
ion
of
col
or
ect
al
poly
ps
.
T
he
ge
ner
at
io
n
of
s
uc
h
a
ty
pical
en
do
s
co
py
vid
e
o
procee
ds
f
or
a
lo
n
g
per
i
od.
For
ba
ck
-
to
-
back
e
ndos
c
opy,
it
i
s
so
tough
f
or
an
en
do
sc
opist
to
exam
i
ne
it
with
su
ff
ic
ie
nt
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Decom
po
sit
io
n of c
olo
r
wavel
et
wi
th h
ig
her
or
de
r statist
ic
al
text
ur
e
…
(
A.
S.
M.
Shafi
)
2987
at
te
ntiveness
duri
ng
s
uch
a
long
pe
rio
d,
as
it
is
an
op
era
tor
-
de
pende
nt
proce
dure.
T
he
accuracy
of
su
c
h
a
chall
eng
in
g
diag
nosis
dep
e
nd
s
on
a
qual
ifie
d
ph
ysi
ci
an.
Ther
e
is
a
lar
ge
de
gree
of
var
ia
bili
ty
fo
r
ho
w
the
physi
ci
an
char
act
e
rizes
a
nd
dia
gnos
e
t
hese
po
ly
ps,
a
s
not
al
l
po
ly
ps
hav
e
sam
e
m
a
li
gn
ant
in
f
lun
ce.
As
an
exam
ple,
serr
at
ed
po
ly
ps
can
pote
ntia
ll
y
dev
el
op
m
or
e
aggressi
vely
into
color
ect
al
cancer
as
com
par
e
d
to
oth
e
r
colo
re
ct
al
po
ly
ps
,
be
cause
of
the
serr
at
ed
path
wa
y
in
tu
m
or
igen
esi
s
[3
]
.
The
re
are
on
ly
co
nsi
ste
nt
pr
e
vaili
ng
m
et
hods
f
or
diagnosi
ng
se
rr
at
ed
poly
ps
is
histo
path
ologica
l
char
act
eri
zat
ion
beca
use
oth
e
r
screeni
ng
m
eth
ods
de
sig
ned
to
det
ect
pr
em
al
ign
ant
le
sion
s
(su
c
h
as
fecal
blood,
f
ecal
DN
A
,
or
virtu
al
colo
nosco
py)
and
are
not
well
m
at
ched
f
or
dif
fer
e
nt
ia
ti
ng
se
rr
at
ed
poly
ps
from
othe
r
poly
ps
[
4].
The
chall
en
ging task
f
or
a do
ct
or
is to d
if
fere
ntiat
e b
et
wee
n
a serr
at
e
d
po
ly
ps
an
d hyper
plasti
c p
olyps
. Th
is i
s
because
hype
r
plasti
c
po
ly
ps
of
te
n
la
ck
t
he
dysp
la
sti
c
nu
cl
ear
c
hanges
that
chara
ct
erize
conve
ntion
al
adenom
as
po
ly
p,
an
d
their
histop
at
ho
l
ogic
al
diagnosis
of
hy
perplas
ti
c
po
ly
ps
is
entirel
y
based
on
m
or
phologica
l
featur
es
,
su
c
h
as
s
err
at
io
n,
dilat
at
ion
,
an
d
br
a
nch
i
ng
a
nd
oft
en
la
ck
t
he
dys
plasti
c
nu
cl
ea
r
changes
that
de
pict co
nventio
nal ad
e
nom
as p
olyp
s [5].
In
t
he
fiel
d
of
arti
fici
al
intel
lig
ence
,
c
onvolu
ti
on
al
ne
ural
ne
twork
m
od
el
s
ha
ve
been
pro
po
s
ed
a
nd
app
li
ed
f
or
co
m
pu
te
r
-
ai
de
d
p
olyp
detect
io
n
a
nd
cl
assifi
c
at
ion
syst
em
.
Color
wa
velet
cov
a
riance
(C
WC)
of
diff
e
re
nt
colo
r
bands
base
d
on
t
he
co
va
riance
of
sec
ond
-
orde
r
te
x
t
ual
m
easur
es
hav
e
bee
n
use
d
t
o
the
detect
ion
of
tum
or
s
in
colo
no
sc
opic
vid
e
o
with
a
sp
eci
fici
ty
and
se
ns
it
ivit
y
of
97%
a
nd
90%
resp
ect
ively
[
6]
.
In
te
ll
igent
proces
sin
g
te
ch
niques
of
S
V
Ms
and
c
olor
-
t
extu
re
analy
sis
m
et
ho
dolo
gies
hav
e
been
pr
opos
e
d
in
t
heir
c
ons
ecuti
ve
w
ork
for
a
uto
m
at
ic
detect
ion
of
ga
stroin
te
sti
nal
adenom
as
in
vid
e
o
endosc
opy
ha
vi
ng
the
accu
rac
y
of
9
4
%
[
7].
94.
20%
accu
rac
y
was
obta
ine
d
with
the
c
om
bin
at
ion
o
f
c
ol
or
an
d
sh
a
pe
featu
res
to
recog
nize
intest
inal
po
ly
p
from
capsu
le
endosc
opy
as
a
cl
assifi
er
of
m
ulti
la
ye
r
per
ce
ptr
on
(MLP)
[
8].
A
deep
c
onvoluti
on
al
ne
ural
network
-
base
d
cl
assifi
cat
ion
f
or
dig
es
ti
ve
orga
ns
in
wireless
caps
ule
endosc
opy
im
a
ges
was
co
ns
id
ered
in
Y.
Z
ou
et
al.
[
9].
A
tr
ai
nab
le
featu
re
extract
or
ba
se
d
on
a
c
onvolu
ti
on
al
neural
netw
ork
is
util
iz
ed
f
or
le
sion
detect
ion
f
r
om
end
os
cop
y
im
ages
in
R.
Z
hu
et
al.
[
10]
.
In
pa
pe
r
[
11]
CNN
feat
ur
es
hav
e
been
pr
opose
d
f
or
the
a
uto
m
at
ed
cl
assifi
cat
ion
of
col
on
ic
m
uco
sa
f
or
col
on
poly
p
sta
ging
in
the
co
ntex
t
of
colo
n
ca
ncer
sc
reen
i
ng
with
a
sen
sit
ivit
y
of
95
.
16%
an
d
sp
e
ci
fici
ty
of
74.19%.
The
CN
N
-
de
ri
ved
featu
res
s
how
gr
eat
e
r
inv
a
riance
t
o
viewin
g
a
ngle
s
an
d
im
age
q
ualit
y
facto
rs
wh
e
n
com
par
ed
t
o
t
he
ei
gen
im
age
m
od
el
[12].
Color
wa
velet
(C
W)
featu
re
s
an
d
c
onvolu
ti
on
al
neural
ne
twork
featur
e
s
of
vide
o
f
ram
es
are
extracte
d
an
d
c
om
bin
ed
wh
ic
h
ar
e
use
d
to
tr
ai
n
a
li
nea
r
s
uppo
rt
vect
or
m
achine
,
gaining
accu
ra
cy
of
98.
65%,
sensiti
vity
of
98.79%
an
d
s
pecifici
ty
of
98.
52%
[
13]
.
A
uthors
[
14]
en
han
ce
d
autom
at
ic
po
ly
p
de
te
ct
ion
a
ccur
acy
by
util
iz
ing
fe
at
ur
e
fu
si
on
(
wa
velet
,
local
bin
a
ry
patte
r
n,
an
d
Ga
bo
r
featur
e
s)
a
nd
m
ul
ti
ple
cl
assif
ie
r
te
ch
niques
.
They
ac
hi
ev
ed
80%
tr
ue
posit
ive
rate
by
inco
rpor
at
in
g
local
bin
a
ry
patte
r
n
an
d
wav
el
et
featur
e
s.
A
uthors
[
15
]
reli
es
on
a
faster
r
egio
n
-
base
d
c
onvoluti
onal
ne
ur
al
netw
ork
(F
ast
e
r
R
-
C
NN)
m
odel
for
poly
ps’
detect
ion
i
n
e
ndosc
opic
vi
deos.
Ta
j
ba
khsh
et
.
al.
[
16
]
f
oc
use
d
on
a
novel
vote
accum
ulati
on
schem
e
fo
r
det
ect
ing
col
on
ic
po
ly
ps
that
e
nab
le
poly
p
de
te
ct
ion
from
par
ti
al
ly
identifie
d
bounda
ries
of
po
l
yps.
H
ow
e
ve
r,
sever
al
novel
te
chn
ol
og
ie
s
are
em
erg
ing
within
the
fiel
d
of
endolum
inal
i
m
aging
a
nd
c
onf
ocal
la
ser
endom
ic
ro
sco
py
to
per
f
or
m
t
he
i
m
aging
of
the
color
ect
al
area,
the
nu
m
ber
of
com
pu
te
r
-
ai
de
d
decisi
on
sup
port
syst
e
m
(C
ADSS)
relat
ed
to
color
ect
al
po
ly
ps’
cl
assif
ic
at
ion
is
sti
l
l
lim
it
ed.
Hen
ce
,
the
m
ai
n
con
trib
ut
ion
of
this
pa
per
is
to
effe
ct
ively
us
e
m
achine
le
arni
ng
to
su
ccess
fu
ll
y
de
te
ct
and
cl
assify
color
ect
al
po
ly
ps
pr
ese
nted
in
en
dos
cop
y
vi
deo.
Howe
ver,
the
oth
er
intenti
ons
of
t
hi
s p
a
per
i
nclu
de
the
fo
ll
owin
gs:
A
novel
intel
li
gen
t
a
ppr
oac
h
util
iz
ing
hi
ghe
r
-
order
sta
ti
stical
te
xtu
re
fea
ture
on
the
col
or
wa
velet
an
d
conv
olu
ti
onal
neural
netw
ork
learne
d from
v
ideo
e
ndosc
op
y datase
t.
The
pro
posed
appr
oach
is
fa
vora
ble
an
d
e
f
fecti
ve,
w
hich
achiev
es
the
best
perform
a
nce
of
accu
rac
y,
com
par
e to
only
CNN
featu
re
-
base
d
m
et
ho
d.
Dr
ast
ic
al
ly
reducin
g
the
nece
ssary
tim
e
fo
r
a
qu
al
ifie
d
physi
ci
an
to
exam
i
ne
the
entire
e
ndos
c
opy
vid
e
o,
by in
dicat
ing c
olorect
al
poly
ps an
d/or cla
ssif
y t
hem
.
The
rest
of
t
he
pap
e
r
is
orga
nized
as
f
ollows:
sect
io
n
2
descr
i
bes
the
arc
hitec
ture
a
nd
m
et
ho
ds
of
the
propose
d
syst
e
m
.
Sect
io
n
3
d
eal
s
wit
h
the
analy
zi
ng
of
e
xp
e
rim
e
ntal
resu
lt
s.
Finall
y,
discuss
i
on
of
the r
es
ults as
w
el
l as the c
on
cl
us
io
ns
of this
s
tud
y i
s s
umm
a
rized i
n
sect
io
ns 4 a
nd
5 res
pe
ct
ively
.
2.
SY
STE
M A
R
CHI
TE
CT
U
R
E
The
im
ple
m
entat
ion
of
the
pr
opos
e
d
syst
em
is
base
d
on
MATL
AB
20
17b
that
ca
n
acce
pt
sta
nda
r
d
vid
e
o
file
s
of
diff
e
re
nt
f
or
m
a
ts
su
ch
as
A
V
I
,
MP4
,
an
d
W
VM
as
in
put
and
pro
duce
out
pu
ts
with
a
cl
a
ssifie
d
po
ly
p
i
n
the
vi
deo
fram
e
sequ
ence
.
Wh
e
n
e
ndos
c
opy
vid
e
o
is
fe
d
int
o
th
e
pro
po
se
d
sys
tem
,
it
utilizes
colo
r
wav
el
et
with
higher
-
or
der
s
ta
ti
sti
ca
l
te
xtu
re
i
m
age
featur
es
a
nd
c
onvoluti
onal
ne
ural
network
fe
at
ur
es.
The
f
us
io
n
of
al
l
the
featur
e
s
is
incorporat
ed
into
S
VMs
to
achieve
im
pro
ved
detect
ion
a
nd
cl
assifi
cat
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
3
,
J
une
2020 :
29
8
6
-
299
6
2988
accuracy.
The
r
e
are
five
sta
ge
s
of
this
sect
io
n
nam
ely
the
acqu
isi
ti
on
of
vi
deo
en
dosco
py
with
norm
al
lesio
ns
and
d
if
fer
e
nt
ty
pes
of
col
or
ec
ta
l
po
ly
ps
, p
re
-
processi
ng,
fea
ture
e
xtracti
on,
cl
assifi
cat
ion
, an
d
post
-
proce
ssing
m
od
ule as s
hown in Fi
gure
1.
Figure
1
.
A fra
m
ewo
r
k of t
he pr
opos
e
d
a
ppr
oach
2.1.
Ac
quisi
ti
on
of v
i
deo en
do
sc
opy
The
a
vaila
bili
ty
of
en
dosc
op
y
dataset
s
is
an
im
po
rtant
issue
for
this
pur
po
s
e
.
H
ow
e
ver,
the
perf
or
m
a
nce
of
any
Com
pu
te
r
-
ai
de
d
desi
gn
(CAD)
syst
em
dep
e
nds
on
the
trai
ning
da
ta
set
,
this
pro
pose
d
syst
e
m
util
i
zes
m
or
e
tha
n
86
sta
nda
r
d
e
ndos
c
opy
vid
e
os
f
ro
m
dif
fer
e
nt
s
ources
.
Sam
ple
dataset
of
dif
fere
nt
poly
p
cl
as
ses
an
d
nor
m
al
le
sion
a
re
s
how
n
in
Fig
ur
e
2.
Most
of
the
en
do
sc
op
y
vid
eo
em
brace
color
e
ct
al
po
ly
ps
a
nd
on
ly
sm
all
par
t
of
the
e
ndos
c
opy
vide
o
are
associat
ed
with
the
norm
al
colo
n.
Im
po
rta
nt
sourc
es
of
the
data
base
i
nclu
de
Dep
a
rtm
ent
of
Ele
ct
r
on
ic
s
,
Un
i
ver
sit
y
of
Alcal
a
(
http:/
/
www.de
peca.
ua
h.
es/
c
olono
sc
op
y
_d
at
aset
/
)
[17],
E
ndos
c
opic
Visio
n
C
ha
ll
eng
e
(
https:/
/p
olyp.
gran
d
-
c
halle
nge
.org
/
databases/
)
[
18
]
.
Ta
ble
1
sho
ws
th
e
dataset
col
le
ct
ed
from
vid
e
o
endosc
opy.
F
or
our
ex
per
im
e
nt,
we
ha
ve
e
xt
racted
4,2
16
f
ram
es
fr
om
end
osc
opy
vi
deo,
w
hich
c
onsis
ts
of
3,162 c
olorect
al
p
olyp
s,
a
nd
1,054 n
or
m
al
f
ram
es.
(a)
(b)
(c)
(d)
Figure
2. Sam
ple d
at
aset
: (a)
hy
perplast
ic
poly
ps
, (b
)
a
de
nom
as p
olyps
,
(c)
ser
rated
poly
ps
, a
nd (d
) norm
al
les
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Decom
po
sit
io
n of c
olo
r
wavel
et
wi
th h
ig
her
or
de
r statist
ic
al
text
ur
e
…
(
A.
S.
M.
Shafi
)
2989
Table
1.
O
ur
da
ta
set
: t
he
distr
ibu
ti
on
of col
orect
al
poly
p
ty
pes
i
n
the
pre
-
processe
d
im
a
ge fram
e
us
e
d
in
our p
r
opose
d sy
ste
m
Co
lo
recta
l
p
o
ly
p
s’
ty
p
e
Sh
o
rt
-
f
o
r
m
Nu
m
b
e
r
o
f
end
o
sco
p
y
vid
eo
Size of
vid
eo
end
o
sco
p
y
(
MB
)
Hy
p
erplast
ic
po
ly
p
s
Hp
21
4
0
.8
Ad
en
o
m
as
po
ly
p
s
Ad
40
126
Serr
at
ed
po
ly
p
s
Sr
15
3
6
.4
No
r
m
al
lesio
n
-
10
3
2
.6
Total
-
86
2
3
5
.8
8
0
%
(3,3
7
2
v
i
d
eo
f
ra
m
e)
o
f
th
is
d
ataset
was
u
sed
f
o
r
tr
ain
in
g
,
wh
ile
th
e
re
m
ain
in
g
2
0
%
(8
4
4
v
id
e
o
f
ra
m
es,
2
1
1
f
rom
ea
ch
class
)
was
u
sed
f
o
r
v
alid
atio
n
.
2.2.
Pre
-
proc
essing m
odul
e
En
do
sc
opy
vide
o
is
loade
d
in
the
com
pu
te
r
to
fin
d
possi
ble
cat
ego
ries
of
color
ect
al
pol
yps
that
ca
n
pro
du
ce
t
he
r
unni
ng
seq
ue
nc
e
of
sti
ll
i
m
ag
es
cal
le
d
f
ram
e
us
in
g
the
pr
opos
e
d
intel
li
ge
nt
syst
em
.
Or
iginal
vid
e
o
f
ram
e
a
s
sh
ow
n
in
Fi
gure
3(a)
with
su
pe
rf
l
uous
r
egio
ns
are
dis
card
e
d
res
ulti
ng
in
a
pre
-
pro
cessed
fr
am
e
as
sh
own
in
Fig
ur
e
3(
b)
in
order
to
e
xc
lusio
n
of
ir
releva
nt
te
xtua
l
inf
or
m
at
ion
,
suc
h
as
patie
nt’s
nam
e,
date of
birth a
nd ti
m
e o
f
the e
xam
inati
on
.
(a)
(b)
Figure
3. Pr
e
-
proces
sin
g: (
a
) or
i
gin
al
vid
e
o fr
am
e and
(
b)
pr
e
-
pr
ocesse
d
f
ram
e
2.3.
Fea
tu
re e
xt
r
act
i
on
mo
d
ule
A
sli
ding
wind
ow
of
us
e
r
-
de
fined
siz
e
a
nd
sli
din
g
ste
p
is
sli
ded
over
t
he
pr
e
-
proce
ssed
vi
deo
fr
am
es
to
ge
ner
at
e
s
m
al
l
i
m
ages
cal
le
d
window
as
show
n
in
F
igure
4.
De
pe
nd
i
ng
on
the
siz
e,
dim
ension
s,
a
nd
sli
din
g st
ep, ea
ch win
dow p
rodu
ce
s a
num
ber
of
featu
re
vec
tors.
Figure
4. Feat
ure e
xtracti
on te
chn
i
qu
e
2.3.1
.
Higher
order
st
at
is
tics o
n
the w
avel
et dom
ain
as
gr
ay sc
ale t
e
xture fe
atures
Mult
iresolutio
n
analy
sis
of
an
i
m
age
is
achieve
d
by
usi
ng
a
disc
rete
wav
el
et
trans
form
fo
r
ou
r
pro
po
se
d
f
ram
ewor
k
as
the
s
iz
e
of
the
poly
p
va
ries.
T
he
m
os
t
relevan
t
t
extu
re
inf
or
m
at
ion
oft
en
a
ppears
i
n
m
idd
le
-
f
reque
ncy
cha
nn
el
s
[
19
]
.
Te
xture
pro
vid
es
i
nfor
m
at
ion
ab
out
th
e
sp
at
ia
l
arr
an
gem
ent
of
col
or
s
or
intensit
ie
s
in
a
n
im
age
that
he
lps
im
age
seg
m
entat
ion
a
nd
cl
assifi
cat
ion
.
Wav
el
et
s
pe
rfor
m
well
for
t
extu
re
analy
sis.
The
de
com
po
sit
ion
of
the
wav
el
et
trans
form
pr
ovides
sp
at
ia
l/
fr
e
qu
e
ncy
re
pr
ese
ntati
on
f
ro
m
ori
gin
al
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
3
,
J
une
2020 :
29
8
6
-
299
6
2990
i
m
ag
es
w
her
e
eve
ry
sub
-
im
age
preser
ve
s
both
local
an
d
global
inf
or
m
at
ion
of
a
s
pecific
s
cal
e
an
d
or
ie
ntati
on
.
W
hen
deco
m
po
si
ti
on
le
vel
decre
ases
in
the
spa
ti
al
do
m
ai
n,
it
increases
in
th
e
fr
eq
ue
ncy
dom
ai
n
pro
vid
in
g
z
oo
m
ing
ca
pab
il
it
ie
s
an
d
l
ocal
c
har
act
erizat
i
on
of
the
im
age
[6
]
.
Since
m
aj
or
te
xture
i
nfo
rm
ation
pro
du
ce
d
by
this
tra
ns
f
or
m
at
ion
do
es
not
con
ta
in
in
the
low
-
f
reque
ncy
i
m
age
and
m
os
t
of
the
sub
sta
ntial
te
xtu
re
i
nfor
m
at
ion
of
te
n
lo
ok
s
as
if
i
n
the
m
idd
le
fr
e
qu
e
ncy
c
hann
el
,
so
our
pro
po
s
ed
a
ppr
oac
h
use
s
a d
isc
rete
wa
ve
le
t t
ran
sf
or
m
(
D
WT
)
f
or the
deco
m
po
sit
io
n of t
he fre
quenc
y do
m
ai
n
of th
e i
m
age.
A
tw
o
-
dim
ension
al
(2
-
D
)
di
screte
w
avelet
trans
f
or
m
at
ion
is
obta
ined
by
ap
plyi
ng
t
he
filt
ering
consecuti
vely
al
ong
with
ver
t
ic
al
and
horiz
onta
l
directi
ons
(sep
a
ra
ble
filt
er
ba
nk).
T
his
fi
lt
ering
pr
ocedu
re
is
gr
as
pe
d
by
c
onvolvi
ng
the
im
age
with
a
high
pass
filt
er
H
a
nd
a
l
ow
pa
ss
filt
er
L
wh
ic
h
pro
du
c
e
s
a
low
-
res
olu
ti
on
im
age
B
0
(k
)
at
scal
e
k
a
nd
detai
l
im
ag
es
D
j
(
k),
j
=
1,
2,
3,
at
sca
le
k
as
descr
i
bed
by
the foll
owin
g
r
ecur
si
ve [1
9]:
B
0
(k
)
= {
L
x
*
[
L
y
*
B
0
(k
-
1)]
↓
2,
1
}
↓
1,2.
D
1
(
k)
=
{
H
x
* [ L
y
*
B
0
(k
-
1)
]
↓
2,
1
}
↓
1,2.
D
2
(
k)
=
{
H
x
* [ H
y
*
B
0
(k
-
1)]
↓
2,
1
}
↓
1,2.
D
3
(
k)
=
{ L
x
*
[
H
y
*
B
0
(k
-
1)]
↓
2,
1
}
↓
1,2
(1)
wh
e
re
ar
row
(
↓)
denotes
the
s
ub
-
sam
pling
pr
ocedu
re,
the
as
te
risk
(
*)
is
the
convo
l
ution
operat
or
an
d
H
and
L
are the
tw
o
filt
ers fo
r
al
l k
=
1,
2, 3, .
. . . .
. , n.
In
this
pa
per
,
the
gray
le
vel
run
le
ng
t
h
m
atr
ix
has
bee
n
c
on
si
der
e
d
f
or
the
desc
riptio
n
of
hi
gher
-
order
sta
ti
sti
cal
te
xtu
re
feat
ures
i
m
ple
m
ente
d
withi
n
the
de
com
po
sed
s
ub
-
i
m
ages.
A
r
un
-
le
ng
th
m
at
rix
P
(i,
j)
for
a
giv
e
n
i
m
age
is
def
ine
d
by
the
sp
eci
fyi
ng
directi
on
of
0
0
,
45
0
,
90
0
,
135
0
a
nd
the
n
co
un
t
the
occ
urre
nce
of
a
run
for
eac
h
gr
ay
le
vels
i
a
nd
r
un
-
le
ngth
j
in
t
his
di
re
ct
ion
.
We
c
onsider
only
fou
r
r
un
-
le
ngth
m
at
rix
featur
e
s
nam
ely
Sh
ort
Ru
n
Lo
w
G
ray
-
Le
vel
Em
ph
asi
s
(S
RLG
E)
,
S
hort
Run
High
Gr
ay
-
Level
E
m
ph
asi
s
(S
RH
GE
),
Lo
ng
Ru
n
Lo
w
Gr
ay
-
Le
vel
Em
ph
asi
s
(LRLGE),
Lo
ng
Run
High
Gr
ay
-
Level
E
m
ph
asi
s
(
LRH
GE
)
[
20]
.
SRLGE
=
1
n
r
∑
∑
P
(
i
,
j
)
i
2
.
j
2
N
j
=
1
M
i
=
1
(2)
SRHGE =
1
n
r
∑
∑
P
(
i
,
j
)
.
i
2
j
2
N
j
=
1
M
i
=
1
.
(3)
LRLGE =
1
n
r
∑
∑
P
(
i
,
j
)
.
j
2
i
2
N
j
=
1
M
i
=
1
.
(4)
LRHG
E =
1
n
r
∑
∑
P
(
i
,
j
)
.
i
2
.
j
2
N
j
=
1
M
i
=
1
.
(5)
wh
e
re
P(
i,
j
)
is
the ru
n
-
le
ngt
h m
at
rix
an
d n
r
i
s the t
otal n
umber
of
runs.
2.3.2
.
Higher
order c
olo
r
w
av
el
e
t
c
ovaria
nce fe
atures
In
our
pr
opos
e
d
ap
proac
h
col
or
te
xt
ur
e
featur
es
e
xtracte
d
by
us
in
g
D
W
T
for
the
dec
ompo
sit
io
n
of
the
f
reque
ncy
do
m
ai
n
of
the
colo
r
im
age
are
est
i
m
at
ed
over
the
GLRL
M.
The
i
nput
i
m
age
is
dec
om
po
sed
into th
ree c
olor
ch
a
nn
el
s
:
I
C
i
, i =
1, 2,
3.
(6)
I
n
this
w
ork,
t
he
fr
am
es
are
deco
m
po
se
d
at
le
vel
3
us
in
g
Daubec
hies
2
(db
2)
wa
velet
fam
i
ly
.
A
th
re
e
-
le
vel
two
dim
ension
al
discrete
wav
el
et
tra
ns
f
or
m
at
ion
is
c
on
s
eq
ue
ntly
app
li
ed
on
ea
c
h
col
or
c
hann
el
(
I
C
i
),
pro
du
ci
ng
a
l
ow
res
olu
ti
on
i
m
age
B
i
(
k
)
,
at
scal
e
k
an
d
t
he
de
ta
il
i
m
ages
D
i
(
k
)
,
wh
e
re
i
=
1,
2,
3
an
d
k
=
1, 2, .
. . . .
. ,9 acc
ordin
g
t
o
the
w
a
velet
de
com
po
sit
ion
of (1
).
Ther
e
f
or
e
, we
have
:
I
C
i
= {
B
i
(
k
)
,
D
i
(
k
)
}, i = 1,
2,
3
a
nd k =
1,
2,
3,
. . . . .
. ,9
.
(7)
wh
e
re
k
is t
he decom
po
sit
io
n l
evel.
As
it
has
al
r
ea
dy
bee
n
no
te
d,
the
m
os
t
sign
ific
ant
te
xt
ua
l
inform
at
ion
is
pr
e
sente
d
in
the
m
idd
le
wav
el
et
detai
le
d
c
hannels.
S
o,
we
c
onside
r
only
the
detai
l
i
m
ages
for
k
=
4,
5,
6.
S
o,
t
he
nin
e
dif
fer
e
nt
su
b
-
im
ages p
r
oduce
d from
(
7) for
t
he values
k
=
4, 5,
6.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Decom
po
sit
io
n of c
olo
r
wavel
et
wi
th h
ig
her
or
de
r statist
ic
al
text
ur
e
…
(
A.
S.
M.
Shafi
)
2991
D
i
(
k
)
,
i = 1, 2
, 3 a
nd
k
=
4, 5,
6.
(8)
Fo
r
extracti
ng
the
hi
gh
e
r
-
order
sta
ti
sti
cal
te
xtu
al
inf
orm
at
ion
,
we
c
onsider
run
-
le
ng
t
h
m
at
rices
cal
culat
ed
ove
r
the
ab
ov
e
nin
e
dif
fer
e
nt
sub
-
im
ages.
The
se
m
at
rices
ref
le
ct
the
sp
at
ia
l
relat
ion
sh
i
p
be
tween
m
or
e
than
two
pix
el
s
in
a
de
finite
directi
on
.
Run
-
le
ngth
m
at
rices
gen
er
at
e
36
m
at
rice
s
cal
culat
ed
from
fou
r
diff
e
re
nt d
ir
ect
ion
s
of inte
ns
it
ie
s r
el
at
ion 0
0
,
45
0
, 90
0
,
and
135
0
.
R
α
{
D
i
(
k
)
}
, i =
1,
2, 3.
k
=
4, 5,
6.
a
nd α
= 0
0
, 45
0
,
90
0
,
135
0
(9)
Finall
y,
the
four
run
-
le
ng
t
h
m
at
rix
featur
e
s
nam
ely
Sh
or
t
Run
L
ow
Gray
-
Level
Em
ph
asi
s,
S
hort
Run
Hi
gh
Gr
a
y
-
Level
Em
ph
asi
s,
Long
Ru
n
Lo
w
Gr
ay
-
L
evel
Em
ph
asi
s,
and
L
ong
Run
High
Gr
ay
-
Leve
l
Em
ph
asi
s ar
e c
al
culat
ed
f
or e
ach
36 m
at
rice
s r
es
ulti
ng in 1
44 w
a
velet
f
eat
ur
es
.
F
m
[
R
α
{
D
i
(
k
)
}
]
,
i =
1, 2,
3.
k
= 4
,
5,
6,
α =
0
0
, 4
5
0
, 90
0
, 135
0
, a
nd m
= 1
, 2, 3,
4.
(10)
wh
e
re m
is the
resp
ect
ive
ru
n
-
le
ng
th
m
at
rix
f
eat
ur
e.
2.3.3
.
C
onvo
lu
tional ne
ural
network
f
e
atu
res
Conv
olu
ti
onal
neural
netw
ork
f
eat
ur
e
s
ar
e
extracte
d
from
each
wind
ow
of
siz
e
227
*
227.
A
co
nvol
utional
neural
net
work,
a
hie
rarc
hical
neural
netw
ork
is
th
e
br
a
nc
h
of
de
ep
le
ar
ning,
can
be
com
po
sed
of
c
onvoluti
on
la
ye
rs,
poolin
g
la
ye
rs,
recti
fied
li
near
u
nit
(Re
LU)
la
ye
rs
,
f
ully
con
necte
d
la
ye
rs
,
and
lo
ss
la
ye
rs.
In
a
sim
ple
C
NN
ar
chite
ct
ur
e,
a
Re
ct
ifie
d
Linear
U
nit
la
ye
r
fo
ll
ows
ea
ch
co
nvolu
ti
on
la
ye
r.
Af
te
r
eac
h
co
nvol
ution
lay
er, t
her
e is a
m
ax
-
poolin
g
la
ye
r.
Finall
y, on
e or
m
or
e fu
ll
y con
nected lay
ers,
wh
ic
h
can
be
at
ta
ined
after
on
e
or
m
or
e
con
volut
ion
la
ye
rs.
A
n
i
m
po
rtant
cha
r
act
erist
ic
that
disti
nguish
e
s
CNN
to
tradit
ion
al
m
ul
ti
la
ye
r
per
ceptron
(MLP
)
m
od
el
s
is
ta
king
into
acco
unt
the
structur
e
of
the
im
ages
wh
il
e
processi
ng
the
m
.
Du
e
to
f
ul
l
connecti
vity
betwe
en
the
la
ye
rs,
ML
P
m
od
el
s
suffe
r
f
r
om
the
cu
rse
of
dim
ension
al
it
y
thu
s
do
no
t
sc
al
e
well
to
hi
gh
-
res
olu
ti
on
im
ages
an
d
le
ss
sensiti
ve
to
po
sit
ion
al
cha
nge
s
.
[
21
]
In
s
pire
the
str
uctu
re
of
CNN
us
ed
in
t
his
pa
per
c
on
ta
in
s
the
f
ollow
i
ng
r
epr
ese
ntati
o
n
as
sh
ow
n
in
Fig
ur
e
5
.
Table
2
ca
n de
scribe a
bout t
he
n
e
ur
al
netw
ork
arc
hitec
ture im
ple
m
entat
io
n.
Figure
5.
A
n
il
lustrati
on of t
he
prop
os
ed
CN
N
f
eat
ur
e
ex
t
ra
ct
ion
m
od
ule
Table
2.
T
he
stru
ct
ur
e
of the
neural
netw
ork use
d
in
this
pa
per
Lay
e
r
Ty
p
e
Di
m
en
sio
n
s
Co
n
v
o
l
u
tio
n
1
1
x 1
1
-
Pad
d
in
g
:
0
–
Stride
-
4
Max
-
p
o
o
lin
g
3
x 3
–
Stride: 2
Co
n
v
o
l
u
tio
n
5
x 5
–
Pad
d
in
g
: 2
–
Stride: 2
Max
-
p
o
o
lin
g
3
x 3
-
Stride:
2
Co
n
v
o
l
u
tio
n
3
x 3
Co
n
v
o
l
u
tio
n
3
x 3
Co
n
v
o
l
u
tio
n
3
x 3
Max
-
p
o
o
lin
g
3
x 3
-
Stride:
2
Fu
lly
Co
n
n
ected
4096
Fu
lly
Co
n
n
ected
4096
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
3
,
J
une
2020 :
29
8
6
-
299
6
2992
2.4
.
Cl
as
sific
ati
on mo
dule
This
m
od
ule
ha
nd
le
s
t
he
cl
as
sific
at
ion
of
t
he
featu
re
vecto
rs
int
o
one
of
four
cl
asses:
hy
perplast
ic
,
adenom
as, ser
r
at
ed
an
d
norm
al
. F
or
c
om
pu
te
r
-
ai
de
d hist
opat
ho
lo
gical
cl
assifi
cat
ion
syst
e
m
s,
m
any classifier
s
hav
e
bee
n
de
ve
lop
e
d
s
uc
h
as
li
near
discrim
inant
a
naly
sis
(LDA)
[
6],
ne
ur
al
netw
orks
[8
]
,
a
da
ptive
ne
uro
-
fu
zzy
in
fer
e
nc
e
s
yst
e
m
[2
2],
bin
a
ry
cl
assifi
er
[
23
]
,
an
d
s
upport
vector
m
achine
(
SV
M)
[
13
,
24,
25
]
.
In
this
pro
pose
d
syst
e
m
,
SV
M
has
been
use
d
for
bette
r
pe
rfor
m
ance
in
t
he
case
of
a
la
r
ge
num
ber
of
f
eat
ur
es
,
trai
ning,
a
nd
s
par
se
data.
Dif
fer
e
nt
ty
pes
of
colo
rectal
po
l
yps’
f
ram
es
are
extracte
d
fro
m
vid
eo
e
ndosc
opy
that
operates
i
n
tw
o
m
od
es:
trai
ning
a
nd
te
sti
ng
m
od
es.
I
n
trai
ni
ng
m
od
e,
dec
om
po
sit
ion
of
c
olor
w
avelet
with
G
LRLM
and
C
N
N
feat
ur
es
a
re
e
xtrac
te
d
from
a
hyperplast
ic
,
ade
no
m
as,
serr
at
e
d
an
d
norm
al
vid
e
o
fr
am
es.
The
in
pu
t
featu
re
ve
ct
or
c
onsist
s
of
144
c
olor
wa
velet
featu
res
and
40
96
CN
N
featu
res,
w
hi
ch
are
com
bin
ed
for
SV
M
trai
ning.
I
n
the
te
sti
ng
m
od
e
of
ope
ra
ti
on
,
t
he
cl
assi
ficat
ion
of
ne
w
sam
ples
is
e
xtracted
from
un
know
n
vi
deo
fr
am
es
base
d
on
th
e
util
iz
at
ion
of
knowle
dge
ga
ine
d
f
ro
m
t
he
trai
ni
ng
sa
m
ples.
If
t
he
unknow
n
sam
ple
is
cl
assifi
ed
as
hy
perplast
ic
,
ad
eno
m
as
or
ser
rated
poly
ps
i
t
go
e
s
to
the
po
st
-
processi
ng
m
odule
ot
herwise
a
new
s
ubse
qu
e
nt
unid
entifi
ed
vi
de
o
fr
am
e
co
m
e
s
under
t
he
featu
r
e
extracti
on m
odule.
2.5
.
P
os
t
-
pr
oc
essing
m
odul
e
The
outc
om
e
of
the
cl
assifi
cat
ion
m
od
ule
is
us
ed
to
a
pro
du
ce
new
vid
e
o
fr
am
e
on
whic
h
po
s
sible
ty
pes
of c
olore
ct
al
p
olyps
are
appr
opriat
el
y l
abeled.
An ill
ust
rati
on
of this
te
chn
iq
ue sh
own
in Fi
gure
6.
No
r
mal
Hp
Sr
Ad
(a)
(b)
(c)
(d)
Figure
6. P
os
t
-
processi
ng m
od
ule: Sy
ste
m
’s
outp
ut af
te
r
post
-
proces
sin
g: (a)
norm
al
lesi
o
n,
(b) hype
rp
la
sti
c poly
ps
,
(
c
)
se
rr
at
ed
poly
ps,
and (
d) ade
no
m
as p
olyps
3.
RESU
LT
S
To
e
valuate
t
he
perform
ance
of
the
pro
pose
d
m
et
ho
d,
we
e
xam
ined
th
ree
m
et
ri
cs
pe
r
cl
ass:
sensiti
vity
(S
en),
sp
eci
fici
ty
(S
pe
),
a
nd
acc
ur
acy
(
Acc).
Sens
it
ivit
y
al
so
cal
le
d
the
true
po
sit
ive
rate
re
fer
s
t
o
the
abili
ty
to
m
easur
e
the
pro
portio
n
of
act
ual
posit
ives
that
are
co
rr
e
ct
ly
identifie
d
as
su
c
h.
Sp
eci
fici
ty
r
efers
to
the
tru
e
negat
i
ve
rate,
m
ea
su
ri
ng
the
pro
portio
n
of
act
ua
l
ne
gatives
t
ha
t
are
c
orrectl
y
identifie
d
as
su
c
h.
Accuracy
is
der
ive
d
from
se
ns
it
ivit
y
and
sp
eci
fici
ty
and
is
def
ine
d
as
the
su
m
of
tru
e
po
sit
ives
an
d
false
po
sit
ives
di
vide
d
by
t
he
tota
l
nu
m
ber
of
e
valuated
cases
(tru
e
posit
ive
s
(TP),
tr
ue
ne
gatives
(TN),
false
po
sit
ives
(F
P
),
and
false
ne
ga
ti
ves
(F
N
).
T
welve
di
ff
e
rent
per
f
or
m
ance
m
et
rics
wer
e
use
d
f
or
cl
assifi
cat
ion
pur
po
ses
(S
e
n,
Sp
e,
an
d
Ac
c
for
each
one
of
the
4
cl
asses
unde
rstudy).
T
he
final
co
nfus
ion
m
at
rice
s
c
an
be
seen
i
n
Ta
ble
3
a
nd
the
pe
rfor
m
ance
analy
sis
with
res
pe
ct
to
Se
n,
S
pe
,
an
d
Acc
sco
res
ac
ro
s
s
the
f
our
diff
e
re
nt
cl
asses
is
sh
own
in
Table
4.
I
n
com
par
ison,
as
is
m
anifested
in
Table
4,
th
e
ou
tc
om
es
ind
ic
at
e
that
the
pro
posed
m
et
ho
d
(
Color
wa
velet
+
CNN
+
G
LRLM
+
SVM
)
achieves
c
om
petit
ive
per
form
ance
no
t
only
f
or
de
te
ct
ing
po
ly
p
an
d
non
-
poly
p
bu
t
al
so
f
or
cl
assify
ing
c
olorect
al
poly
ps
c
om
par
e
to
C
N
N
featur
e
-
based
m
et
ho
d.
Table
3.
C
onf
usi
on m
at
rix
of
our pro
po
s
ed
s
yst
e
m
f
or
dif
fe
ren
t
kind
of po
ly
p
cl
assifi
cat
ion
Actu
al Class
Predicted
Class
Ad
Hp
No
r
m
al
L
esio
n
Sr
Ad
197
6
0
8
Hp
2
209
0
0
No
r
m
al
L
esio
n
0
0
211
0
Sr
2
0
0
209
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Decom
po
sit
io
n of c
olo
r
wavel
et
wi
th h
ig
her
or
de
r statist
ic
al
text
ur
e
…
(
A.
S.
M.
Shafi
)
2993
Table
4.
Per
for
m
ance an
al
ysi
s
of
our pro
po
se
d
m
et
ho
d
i
n
t
he
4
-
cl
ass
prob
l
e
m
Po
ly
p
T
y
p
e
Metho
d
o
lo
g
y
Sen
sitiv
ity
Sp
ecif
icity
Accurac
y
Ad
CNN
8
9
.57
9
8
.73
9
6
.44
Prop
o
sed
M
eth
o
d
9
3
.36
9
9
.37
9
7
.87
Hp
CNN
9
8
.10
9
8
.89
9
8
.69
Prop
o
sed
M
eth
o
d
9
9
.05
9
9
.05
9
9
.05
No
r
m
al
L
esio
n
CNN
9
9
.53
9
9
.52
9
9
.52
Prop
o
sed
M
eth
o
d
100
100
100
Sr
CNN
9
7
.63
9
7
.78
9
7
.74
Prop
o
sed
M
eth
o
d
9
9
.05
9
8
.10
9
8
.41
3.1
.
C
ompari
so
n wi
t
h e
xisti
ng
m
eth
od
s
Howe
ver,
the
com
par
ison
with
existi
ng
a
ppro
ac
hes
ha
s
be
en
extrem
el
y
d
ifficult
for
sev
eral
reason
s
.
First,
the
wide
ly
acce
ptable
r
esearch
pa
pers
of
relat
ed
to
pi
cs
at
tem
pt
to
exp
la
in
dif
fer
e
nt
pro
blem
s:
po
ly
p
detect
ion
[10,
12,
13]
com
m
on
ly
us
ing
c
onvoluti
on
ne
ur
al
net
w
ork
f
eat
ur
es.
Seco
nd,
e
ven
if
t
he
sa
m
e
pro
blem
was
so
lve
d
by
th
ose
m
et
ho
ds
,
t
her
e
is
an
ent
ire
la
ck
of
op
enly
acce
ssibl
e
co
des.
We
com
par
e
the
m
et
ho
ds
f
ocused
on
the
cl
assifi
cat
ion
of
gastr
oin
te
sti
na
l
le
sion
s
in
vid
eo
e
ndos
c
opy
as
sh
own
Ta
bl
e
5.
In ad
diti
on, th
e
prop
os
ed
syst
e
m
is appraise
d
al
on
gs
ide
w
it
h
sta
nd
a
rd d
at
a
set
.
Table
5.
C
om
par
iso
n
am
on
g
diff
e
re
nt m
et
h
od
s
Au
th
o
r
Used
M
eth
o
d
o
lo
g
y
Used
Dataset
Res
u
lt
Sen
sitiv
ity
Sp
ecif
icity
Accurac
y
Mus
tain
Billah
et
a
l.
[
1
3
]
Co
lo
r
wav
elet +
C
NN +
SVM
1
0
0
vid
eo
s
9
8
.65
%
9
8
.79
%
9
8
.52
%
Xi M
o
et
.
al
.
[
1
5
]
Faster R
-
CN
N
2
9
vid
eo
s
9
8
.1%
Rib
eiro et al. [
1
1
]
CNN
1
0
0
i
m
ag
es
9
5
.16
%
7
4
.19
%
9
0
.96
%
Alaa E
l
Khatib
et
.
al.
[
1
4
]
LBP +
W
av
ele
t f
eatu
res
1
3
vid
eo
s
80%
Karkan
is et al
.
[
6
]
CWC +
L
D
A
60
v
id
eo
s
97%
90%
Ni
m
a
T
ajb
ak
h
sh
et.
al.
[
1
6
]
Vo
te acc
u
m
u
latio
n
sch
e
m
e
3
0
0
i
m
ag
es
86%
Prop
o
sed
m
e
th
o
d
Co
lo
r
wav
elet +
C
NN +
GLR
LM
+ S
VM
8
6
vid
eo
s
9
7
.87
%
9
9
.13
%
9
8
.83
%
3.2
.
C
ompari
so
n wi
t
h hum
an
e
xp
er
ts
We
ha
ve
co
ns
i
der
e
d
the
diag
no
sti
c
ef
ficacy
of
hum
an
experts
to
com
pare
the
perform
a
nce
with
our
pro
po
se
d
a
ppr
oach.
Alth
ough
the
ra
nge
of
ye
ars
of
e
xperi
ence
pe
rfo
rm
in
g
en
dosco
pies
go
e
s
from
8
to
40
for
the
hum
an
ex
per
t’
s
cat
eg
ory
,
seve
ral
hu
m
an
issues
le
ad
to
poly
p
m
isdete
ct
ion
a
nd
m
isc
la
ssifi
cat
ion
.
The
le
sio
ns
wrongly
classi
fied
by all
the
hum
ans
are s
how
n
in Fi
gure
7.
The
le
sio
ns
c
orrectl
y cl
assifi
ed
by all
hu
m
ans
an
d
t
he
best
m
achine
le
arn
i
ng
m
od
el
,
the
ones
wrongly
cl
assifi
ed
by
the
be
st
m
od
el
an
d
c
or
rectl
y
by
al
l
hu
m
ans,
and
t
he
le
sio
ns
correct
ly
cl
assifi
ed
by
the
best
m
od
el
an
d
w
r
ongly
cl
assifi
ed
by
al
l
hum
ans
are
dis
play
ed
in
Fig
ur
es
8,
9
an
d
10,
res
pecti
vely
.
Tex
t
include
d
in
the
fig
ur
e
s
shows
t
he
le
sio
n
nam
e
(‘Hp’, ‘
Ad’,
‘Sr’, ‘
N’ re
fer
s
to
h
yp
e
r
plasti
c, a
denom
as, ser
ra
te
d,
a
nd no
rm
a
l l
esi
on
res
pecti
vely
).
N
-
35
Hp
-
2
Hp
-
21
Figure
7. Lesi
ons
wro
ng
ly
cla
ssifie
d by al
l h
um
ans
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
3
,
J
une
2020 :
29
8
6
-
299
6
2994
Ad
-
9
Hp
-
14
Sr
-
23
Hp
-
25
Hp
-
4
N
-
19
Ad
-
42
Ad
-
66
Sr
-
70
Sr
-
67
Figure
8. Lesi
ons
c
orrectl
y cl
assifi
ed by t
he
prop
os
ed
m
od
el
and all
hum
ans
Ad
-
84
Hp
-
21
Figure
9. Lesi
ons
wro
ng
ly
cla
ssifie
d by
the pr
opos
e
d m
od
el
b
ut c
orr
ect
ly
classi
fied
by
al
l hu
m
ans
Figure
10. Le
sion
s
c
orrectl
y cl
assifi
ed by
the pr
opos
e
d m
od
el
b
ut
wro
ng
ly
classi
fie
d by
al
l hu
m
ans
4.
DISCU
SSI
ON
In
this
pap
e
r,
we
presente
d
an
d
dev
el
ope
d
a
set
of
al
gorithm
s
fo
r
com
pu
te
r
-
ai
de
d
a
uto
m
at
i
c
colo
rectal
poly
p
detect
ion
an
d
cl
assifi
cat
io
n
syst
e
m
from
vid
eo
e
ndosc
op
y.
Alth
ough
t
he
poly
p
detect
ion
an
d
cl
assifi
cat
ion
is
a
disputi
ng
t
ask
beca
us
e
of
it
s
le
gio
n
fact
or
s
s
uc
h
as
th
e
pr
ese
nce
of
trash
a
nd
li
qu
i
ds
an
d
bubble
s,
vi
gnet
ti
ng
an
d
dif
fer
e
nt
ty
pes
of
po
ly
p
s
ha
pe,
we
at
te
m
pt
to
over
whelm
this
pr
oblem
by
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Decom
po
sit
io
n of c
olo
r
wavel
et
wi
th h
ig
her
or
de
r statist
ic
al
text
ur
e
…
(
A.
S.
M.
Shafi
)
2995
inco
rpor
at
in
g
bo
t
h
highe
r
-
order
sta
ti
sti
cal
t
extu
re
an
d
co
nvol
ution
al
neural
netw
ork
fe
at
ur
es.
We
ha
ve
us
ed
a
li
near
SV
M
for
cl
assifi
cat
ion
s
olu
ti
on
f
or
it
s
sta
ble
resu
lt
s
and
faste
r
trai
nin
g
inste
ad
of
a
co
nve
ntion
al
so
ftm
ax
loss
l
ay
er.
Her
e,
th
e
input
to
the
fu
ll
y
c
onnect
ed
la
ye
r
has
be
e
n
us
e
d
as
t
he
input
to
the
SV
M
cl
assifi
er.
As
there
are
se
ver
a
l
kinds
of
im
portant
lo
w
-
le
vel
featur
e
s,
we
co
ns
ide
r
t
hat
the
te
xtu
re
feat
ur
e
su
it
s
this
pa
pe
r
best.
W
e
hav
e
use
d
higher
-
or
der
sta
ti
sti
cal
te
xt
ur
e
desc
ript
or
s
f
or
featu
re
e
xt
racti
on
i
ns
te
a
d
of
s
eco
nd
-
or
der
s
ta
ti
sti
cs
becau
s
e
hig
he
r
-
orde
r
sta
ti
sti
cs
determ
ine
the
relat
i
on
s
hi
p
betwee
n
three o
r
m
or
e
pix
el
s
wh
e
reas
sec
ond
-
orde
r
sta
ti
stics
exam
ine
the
relat
ion
sh
i
p
be
tween
tw
o
pi
xels.
The
refo
re
,
we
ha
ve
co
nsi
der
e
d
D
WT
an
d
GL
RLM
featu
res
for
te
xt
ur
e
rec
ogniti
on.
We
consi
der
that
t
he
CN
N
featu
r
e
is
an
oth
e
r
s
ui
ta
ble
cho
ic
e
f
or
f
eat
ur
e
ext
racti
on
as
CNN
is
an
end
-
to
-
en
d
le
ar
ning
process
.
On
CN
N,
we
pro
vid
e
the
dataset
an
d
their
c
orres
pondin
g
la
bels,
t
he
entire
proce
ss
of
feat
ur
e
en
gi
neer
i
ng
is
done
.
It
has
bot
h
f
e
at
ur
e
e
xtracti
on
a
nd
m
achine
le
ar
nin
g
i
ns
ide
it
.
T
he
propose
d
sy
stem
fo
r
c
olor
ect
al
po
ly
p’s
c
la
ssific
at
ion
s
howe
d
t
hat
the
us
e
of
colo
r
wa
velet
with
G
LRL
M
and
C
NN
featur
e
s
achie
ves
al
m
os
t
hig
h
sensiti
vity
(97.8
7%)
a
nd,
equ
al
ly
i
m
po
rtantl
y,
disp
la
ys
hi
gh s
pe
ci
fici
ty
(
99.13
%).
5.
CONCL
US
I
O
NS
A
ND FUT
UR
E
WO
RKS
In
t
his
pap
e
r,
we
ta
c
kled
a
com
plete
pipe
li
ne
f
or
c
ompu
te
r
-
ai
de
d
a
ut
om
a
ti
c
color
e
ct
al
po
ly
ps
’
detect
ion
a
nd
cl
assifi
cat
ion
s
yst
e
m
capab
le
of
s
upportin
g
the
decisi
on
of
a
m
edical
per
s
on.
Its
ai
m
s
to
enh
a
nce
the
a
bi
li
t
y
of
an
e
ndos
c
op
ist
to
loc
at
e
color
ect
al
po
ly
ps
m
or
e
accuratel
y
w
hich
m
ay
go
unde
te
ct
ed
and
e
volve
int
o
m
align
ant
tu
m
or
s.
The
syst
e
m
exp
loit
s
th
e
higher
-
orde
r
sta
ti
sti
cal
te
xtu
re
featu
re
cal
culat
ed
ov
e
r
th
e
stre
ngth
of
wav
el
et
f
ram
e
trans
for
m
at
ion
of
di
ff
e
ren
t
c
olor
band
s
an
d
c
onvolut
ion
al
neural
ne
twork
.
The
cl
assifi
cat
i
on w
as
per
for
m
ed
us
i
ng a lin
ear S
VM cla
ssi
fier.
The
re
su
lt
s
of
the
exte
ns
ive
e
xp
e
rim
ental
st
ud
y
hav
e
t
o
le
ad
us
to
t
he
use
of
highe
r
-
ord
er
sta
ti
sti
cal
te
xtu
re
feat
ur
e
s
on
t
he
wav
el
et
do
m
ai
n
res
ul
ts
in
hi
gh
e
r
cl
assifi
cat
ion
ac
cur
acy
i
n
te
rm
s
of
s
pecifici
ty
an
d
sensiti
vity
.
Th
e
pro
po
se
d
G
LRLM
perfor
m
s
s
ign
ific
antl
y
bette
r
than
Gr
ay
Le
vel
Co
-
Occ
urren
ce
Ma
trix
(G
LCM
)
f
or
poly
p
detect
io
n
as
the
Gr
ay
Le
vel
Run
L
en
gt
h
Ma
trix
pr
ov
i
des
in
form
at
io
n
ab
out
the
co
nn
ect
e
d
le
ng
th
of
a
part
ic
ular
pi
xel
in
a
def
init
e
dire
ct
ion
.
Co
nvolut
ion
al
neural
ne
tworks
are
a
s
pecial
arc
hitec
ture
of
arti
fici
al
neu
ra
l
networ
ks
that
hav
e
bee
n
wi
dely
us
e
d
in
a
uto
m
at
ic
i
m
age
cl
assifi
cat
ion
syst
e
m
s
by
red
u
ci
ng
le
arn
in
g
c
om
plexity
with
s
ha
rin
g
the
wei
ghts
in
di
ff
e
ren
t
la
ye
rs.
T
he
use
of
li
near
S
V
M
kernel
po
sit
ively
aff
ect
s t
he disc
rim
inati
on
of nor
m
al
an
d
a
bn
or
m
al
sa
m
ples.
Finall
y,
we
m
ay
con
cl
ude
t
hat
the
ex
pe
rim
ental
resu
lt
s
sh
owe
d
the
enh
a
nce
d
pe
rfor
m
ance
of
the
propose
d
de
te
ct
or
an
d
cl
assifi
er,
com
pared
to
oth
e
r
sta
te
-
of
-
t
he
-
a
rt
detect
or
s.
I
n
f
utur
e,
we
woul
d
lik
e
to
determ
ine
a
m
or
e
robu
st
c
la
ssific
at
ion
s
chem
e
based
on
vi
deo
en
dosc
op
y
fr
am
es
an
d
wireless
caps
ule
endosc
opy
(
W
CE)
with
a
bor
der
ra
nge
of
i
m
ages
con
t
ai
nin
g
col
or
ec
ta
l
po
ly
ps
of
dif
fer
e
nt
qua
li
ti
es.
We
w
ou
l
d
li
ke
to
add
an
ot
her
f
unct
ion
al
it
y
that
facil
itate
s
the
i
m
pr
ov
em
ent
(enh
ancem
ent)
of
vid
e
o
endosc
opy by
app
ly
in
g
th
e s
uper
-
res
olu
ti
on t
echn
i
qu
e
.
ACKN
OWLE
DGE
MENTS
We
w
ou
l
d
li
ke
to
ackn
owle
dge
Prof
.
Dr.
Md.
Zul
fikar
Ali,
Pr
ofess
or
a
nd
Head
,
De
pt.
of
Me
dicine
,
Khwa
j
a
Yun
us
Ali
Me
dical
C
ollege
a
nd
Ho
s
pital
,
Enayet
pu
r,
Sirajg
onj
,
B
ang
la
des
h
f
or
his
valua
ble
s
uppo
rt,
su
ggest
io
ns, a
nd c
on
t
rib
ution
to the e
val
uation o
f
t
he
re
su
lt
s.
REFERE
NCE
S
[1]
L.
A.
Alex
andr
e
,
N.
Nobre,
J.
Caste
l
ei
ro
,
“
Pol
y
p
Det
ec
t
ion
in
Endoscope
Vi
deos
Us
ing
SVMs
,
”
European
Confe
renc
e
on
P
rinci
ple
s
and
Pr
act
i
ce
o
f
Know
ledge
Discov
ery
in
Databases,
pp
.
358
-
36,
2007
.
[2]
W
ong
NA
,
Hunt
LP,
Nove
ll
i
MR,
Shepher
ed
,
W
a
rre
n
BF
,
“
Obs
er
ver
agr
ee
m
ent
in
the
dia
gnosis
of
serra
t
ed
pol
y
ps
of
the l
arg
e
bowel
,
”
Histopatho
l
ogy
,
vo
l.
55
,
no.
1,
pp
.
63
-
6
,
200
9.
[3]
Le
gge
tt
B
,
W
hit
ehall
V,
“
Rol
e
of
the
ser
rated
pat
hwa
y
in
co
l
ore
ctal
ca
nc
er
p
at
hogen
esis
,
”
G
astroente
rology
,
vol.
138
,
no
.
6
,
p
p.
2088
-
100
,
20
10.
[4]
Kahi
CJ,
“
How
does
the
serra
t
e
d
pol
y
p
pa
thwa
y
a
lt
er
CRC
scr
ee
ning
and
surv
ei
llanc
e
?
,
”
Dige
stive
Diseases
and
Sci
en
ce
s
,
vol.
60
,
no
.
3
,
pp
.
773
-
80,
2015
.
[5]
Korbar
B,
Olofson AM,
Miraf
lor
AP
,
Nicka
CM,
Suriawina
t
a
MA
,
Torr
esa
n
i
L,
“
Dee
p
le
a
rning
fo
r
cl
assificat
ion
o
f
col
ore
ct
a
l
po
l
y
p
s on
whole
-
slid
e im
age
s
,
”
J Pat
h
ol
Inform
,
vol. 1, no.
30,
pp.
1
-
17
,
2017
.
[6]
S.
A.
Karka
n
is,
et
al
.
,
“
Com
pute
r
-
ai
ded
tumor
d
et
e
ct
ion
in
endo
scopic
vid
eo
usi
ng
col
or
w
ave
l
e
t
fe
at
ure
s
,
”
I
EEE
Tr
ansacti
ons on I
nformation
Tec
hnology
in
B
iomedi
c
ine
,
vol
.
7
,
n
o.
3
,
pp
.
141
–
15
2,
2003
.
[7]
D.
K.
Iak
ovidi
s
,
et
al
.
,
“
An
int
el
l
ige
nt
s
y
s
te
m
for
aut
om
at
ic
det
e
ct
ion
of
gas
troi
nte
st
ina
l
adenom
as
in
vide
o
endosc
op
y
,
” Co
mputers i
n
B
iol
o
gy
and
M
edicine
,
vol
.
36
,
no
.
10
,
pp.
1084
–
1
103
,
200
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.