I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
43
,
No
.
2
,
A
u
g
u
s
t
2
0
2
6
,
p
p
.
4
9
5
~
5
0
6
I
SS
N:
2
5
0
2
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
43
.i
2
.
pp
495
-
5
0
6
495
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
O
CTM
o
dNe
t:
a d
eep learning
-
ba
sed framewo
rk
f
o
r
o
ptical
co
herence
tom
o
g
ra
phy
ima
g
e multi
-
cla
ss
clas
sifica
tio
n
Sa
j
a
At
a
a
lla
h M
uh
a
m
m
ed
1
,
Abba
s
M
.
Al
i
2
,
Dler
Sa
lih
H
a
s
a
n
1
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
I
n
f
o
r
ma
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
C
o
l
l
e
g
e
o
f
S
c
i
e
n
c
e
,
S
a
l
a
h
a
d
d
i
n
U
n
i
v
e
r
si
t
y
-
Er
b
i
l
,
K
u
r
d
i
st
a
n
R
e
g
i
o
n
,
I
r
a
q
2
D
e
p
a
r
t
me
n
t
o
f
S
o
f
t
w
a
r
e
En
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
,
S
a
l
a
h
a
d
d
i
n
U
n
i
v
e
r
si
t
y
-
Er
b
i
l
,
K
u
r
d
i
s
t
a
n
R
e
g
i
o
n
,
I
r
a
q
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Dec
3
0
,
2
0
2
5
R
ev
is
ed
J
u
n
1
6
,
2
0
2
6
Acc
ep
ted
J
u
n
2
7
,
2
0
2
6
S
ig
h
t
is o
n
e
o
f
t
h
e
m
o
st
si
g
n
ifi
c
a
n
t
se
n
se
s
in
h
u
m
a
n
b
e
i
n
g
s.
Lo
si
n
g
sig
h
t
c
a
n
c
h
a
n
g
e
a
h
u
m
a
n
’s
l
ife
d
ra
m
a
ti
c
a
ll
y
.
Early
d
iag
n
o
sis
a
n
d
d
e
tec
ti
o
n
o
f
re
ti
n
a
l
d
ise
a
se
s
will
p
re
v
e
n
t
sig
h
t
lo
s
s.
Op
ti
c
a
l
c
o
h
e
re
n
c
e
to
m
o
g
ra
p
h
y
(OCT)
ima
g
in
g
is
h
e
lp
f
u
l
in
o
c
u
lar
i
m
a
g
in
g
with
it
s
h
i
g
h
re
so
l
u
ti
o
n
a
n
d
n
o
n
-
in
v
a
siv
e
ima
g
i
n
g
fo
r
d
iag
n
o
si
n
g
re
ti
n
a
l
d
ise
a
se
s.
Th
is
stu
d
y
p
ro
p
o
se
s
OCTM
o
d
Ne
t,
a
n
o
v
e
l
m
u
lt
i
-
c
las
s
h
y
b
rid
c
las
sifica
ti
o
n
m
o
d
e
l
t
h
a
t
i
s
c
a
p
a
b
le
o
f
c
las
sify
in
g
se
v
e
n
d
iffere
n
t
re
ti
n
a
l
d
ise
a
se
s
u
sin
g
a
c
o
m
b
i
n
a
ti
o
n
o
f
pu
b
li
c
ly
a
v
a
i
lab
le
d
a
tas
e
ts,
Ka
g
g
le
OCT
-
C8
,
a
n
d
OCTDL
d
a
tas
e
ts.
F
o
r
e
ffe
c
ti
v
e
e
x
trac
ti
o
n
o
f
t
h
e
sig
n
ifi
c
a
n
t
fe
a
tu
re
s,
th
e
c
o
l
lec
ted
d
a
ta
is
re
siz
e
d
a
n
d
n
o
rm
a
li
z
e
d
.
A
u
g
m
e
n
tati
o
n
t
e
c
h
n
iq
u
e
s
a
re
u
se
d
t
o
imp
r
o
v
e
t
h
e
train
i
n
g
a
n
d
lea
rn
in
g
p
ro
c
e
ss
.
F
o
r
m
o
d
e
l
d
e
v
e
l
o
p
m
e
n
t,
we
a
p
p
l
ied
t
ra
n
sfe
r
lea
rn
in
g
a
n
d
fi
n
e
-
tu
n
e
d
se
v
e
ra
l
sta
te
-
of
-
th
e
-
a
rt
d
e
e
p
lea
rn
in
g
m
o
d
e
ls,
Re
sN
e
t5
0
,
In
c
e
p
ti
o
n
V3
,
De
n
se
Ne
t1
2
1
,
VG
G
1
6
,
a
n
d
Ef
ficie
n
tNe
tV2
S
.
Tw
o
m
o
d
e
ls
t
h
a
t
p
e
rfo
rm
e
d
b
e
st,
VG
G
1
6
a
n
d
Eff
icie
n
tNe
tV2
S
,
we
re
se
lec
ted
a
s
t
h
e
m
o
d
e
l
b
a
c
k
b
o
n
e
s
with
in
teg
ra
ti
o
n
o
f
a
sq
u
e
e
z
e
a
n
d
e
x
c
it
a
ti
o
n
b
lo
c
k
.
Th
e
re
su
lt
s
sh
o
we
d
th
a
t
th
e
OCT
M
o
d
Ne
t
a
c
h
iev
e
d
a
n
o
u
tstan
d
i
n
g
p
e
rfo
rm
a
n
c
e
with
9
9
%
tes
t
a
c
c
u
ra
c
y
,
9
9
%
p
re
c
isio
n
,
9
9
%
re
c
a
ll
,
9
9
%
F
1
-
sc
o
re
,
a
n
d
a
n
a
re
a
u
n
d
e
r
th
e
c
u
rv
e
(
AUC
)
a
s
9
9
%
a
c
ro
ss
th
e
u
n
se
e
n
d
a
t
a
.
Th
e
se
re
su
lt
s
p
o
i
n
t
o
u
t
th
e
r
o
b
u
stn
e
ss
a
n
d
re
li
a
b
il
it
y
o
f
t
h
e
p
r
o
p
o
se
d
m
o
d
e
l
t
o
c
la
ss
ify
OCT
ima
g
e
s
a
n
d
h
a
v
e
t
h
e
p
o
te
n
ti
a
l
t
o
e
n
h
a
n
c
e
c
li
n
ica
l
d
e
c
isio
n
-
m
a
k
in
g
a
n
d
a
ss
ist
o
p
h
th
a
lmo
l
o
g
ists
in
th
e
e
a
rly
d
e
tec
ti
o
n
o
f
d
ise
a
se
s.
K
ey
w
o
r
d
s
:
Dee
p
lear
n
in
g
Me
d
ic
al
im
ag
e
an
aly
s
is
Mu
lti
-
class
clas
s
if
icatio
n
OC
T
im
ag
e
class
if
icatio
n
R
etin
al
d
is
ea
s
e
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Saja
Ataa
llah
Mu
h
am
m
ed
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
,
C
o
lleg
e
o
f
Scien
ce
Salah
ad
d
in
Un
iv
er
s
ity
-
E
r
b
il
Ku
r
d
is
tan
R
eg
io
n
,
I
r
aq
E
m
ail: saja.
m
u
h
am
m
ed
@
s
u
.
ed
u
.
k
r
d
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
r
etin
a
is
a
th
i
n
,
lay
e
r
ed
s
tr
u
ctu
r
e
o
f
ce
lls
at
th
e
p
o
s
ter
io
r
p
ar
t
o
f
th
e
ey
eb
all.
T
h
is
th
in
lay
er
ac
ts
as
th
e
b
r
id
g
e
b
etwe
en
th
e
lig
h
t
r
ec
eiv
ed
b
y
th
e
ey
es
an
d
th
e
im
ag
es
we
s
ee
.
T
h
e
l
ig
h
t
r
ec
eiv
ed
is
co
n
v
er
ted
in
to
s
ig
n
als
an
d
p
ass
ed
to
th
e
b
r
ain
th
r
o
u
g
h
t
h
e
o
p
tic
n
er
v
e
to
b
e
p
r
o
ce
s
s
ed
.
An
y
d
am
ag
e
to
th
ese
lay
er
ce
lls
co
u
ld
af
f
ec
t
th
e
v
is
io
n
q
u
ality
o
r
co
u
ld
ca
u
s
e
p
e
r
m
an
e
n
t
v
is
io
n
lo
s
s
.
Ma
n
y
tech
n
iq
u
es
we
r
e
in
tr
o
d
u
ce
d
o
v
er
th
e
y
ea
r
s
t
o
ca
p
tu
r
e
th
e
la
y
er
ed
s
tr
u
ctu
r
e
o
f
th
e
r
etin
a
to
d
eter
m
in
e
wh
eth
e
r
i
t'
s
n
o
r
m
al
o
r
n
o
t.
O
n
e
o
f
th
ese
h
elp
f
u
l
tech
n
iq
u
es
is
o
p
tical
co
h
er
en
ce
to
m
o
g
r
ap
h
y
(
OC
T
)
im
ag
in
g
,
wh
ic
h
is
a
v
ital
n
o
n
-
in
v
asiv
e
im
a
g
in
g
m
o
d
ality
in
d
iag
n
o
s
in
g
,
tr
ea
tin
g
,
an
d
f
o
llo
win
g
m
an
y
r
e
tin
al
d
is
ea
s
es
(
e.
g
.
,
d
iab
etic
r
etin
o
p
ath
y
(
DR
)
,
ep
ir
etin
al
m
em
b
r
an
e
(
E
R
M)
,
d
r
u
s
en
,
d
iab
etic
m
ac
u
lar
ed
e
m
a
(
DM
E
)
,
m
ac
u
lar
h
o
le
(
M
H)
,
an
d
ce
n
tr
al
s
er
o
u
s
r
etin
o
p
ath
y
(
C
SR
)
)
.
E
ar
ly
d
ia
g
n
o
s
is
an
d
tr
ac
k
i
n
g
o
f
o
cu
la
r
d
is
ea
s
es a
r
e
r
ea
lized
b
y
OC
T
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
43
,
No
.
2
,
Au
g
u
s
t
20
26
:
495
-
5
0
6
496
No
n
eth
eless
,
class
if
y
in
g
OC
T
im
ag
e
m
o
d
alities
is
s
till
ch
alle
n
g
in
g
b
ec
a
u
s
e
o
f
d
iv
er
s
e
im
ag
e
q
u
ality
,
ac
q
u
is
itio
n
p
r
o
to
c
o
l,
an
d
im
b
alan
ce
d
d
atasets
.
Alth
o
u
g
h
m
ac
h
in
e
an
d
d
ee
p
lear
n
in
g
m
o
d
els
h
av
e
s
h
o
wn
s
ig
n
if
ican
t
p
r
o
m
is
e
in
au
to
m
atica
lly
class
if
y
in
g
r
etin
al
d
is
ea
s
es,
th
ey
h
av
e
lar
g
ely
f
ailed
to
en
ab
le
d
if
f
er
en
tiatio
n
b
etw
ee
n
OC
T
m
o
d
alities
,
th
er
eb
y
lim
itin
g
th
eir
p
o
ten
tial
f
o
r
au
to
m
ated
d
ia
g
n
o
s
is
an
d
clin
ical
d
ec
is
io
n
-
m
ak
in
g
[
1
],
[2
]
.
De
ep
lear
n
in
g
-
b
ased
ap
p
r
o
ac
h
es
f
o
r
OC
T
im
ag
e
an
aly
s
is
h
av
e
g
ain
ed
in
ter
est
r
ec
en
tly
a
n
d
h
av
e
b
ee
n
s
tu
d
ie
d
in
a
f
ew
r
ec
e
n
t
p
u
b
licatio
n
s
[
3
]
.
W
ith
im
p
r
o
v
ed
class
if
icatio
n
p
e
r
f
o
r
m
an
ce
in
OC
T
im
ag
es,
co
n
tr
asti
v
e
lea
r
n
in
g
was
u
s
ed
to
e
n
h
an
ce
r
ep
r
esen
tatio
n
lear
n
in
g
[
4
]
.
T
o
o
v
er
co
m
e
th
e
lim
itatio
n
o
f
t
h
e
s
m
all
n
u
m
b
e
r
o
f
lab
elled
im
a
g
es
co
r
r
esp
o
n
d
in
g
t
o
r
a
r
e
r
etin
al
d
is
ea
s
es
[
5]
,
f
ew
-
s
h
o
t
lear
n
in
g
me
th
o
d
s
h
av
e
b
ee
n
p
r
o
p
o
s
ed
.
Mo
r
eo
v
er
,
d
ee
p
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
lik
e
R
es
Net,
I
n
ce
p
tio
n
V3
,
an
d
Den
s
eNe
t
h
av
e
b
ee
n
a
p
p
r
o
ac
h
ed
f
o
r
r
et
in
al
d
is
ea
s
e
clas
s
if
icatio
n
[
6
]
,
[
7
]
.
On
t
h
e
o
th
er
h
an
d
,
th
ese
m
o
d
els
h
av
e
a
s
in
g
le
f
ea
tu
r
e
ex
tr
ac
tio
n
m
ec
h
a
n
is
m
with
lim
itat
io
n
s
to
th
e
h
ig
h
er
h
ier
ar
c
h
ical
r
ep
r
esen
tatio
n
o
f
OC
T
im
ag
es.
No
t
f
ea
tu
r
in
g
ad
a
p
tiv
e
lear
n
i
n
g
r
ate
s
tr
ateg
ies
also
m
ea
n
s
p
o
o
r
an
d
less
s
tab
le
co
n
v
er
g
en
ce
.
T
h
ese
d
r
awb
ac
k
s
em
p
h
asize
th
e
u
r
g
en
c
y
o
f
p
r
o
p
o
s
in
g
a
d
ee
p
lear
n
in
g
f
r
am
ewo
r
k
u
s
in
g
an
i
n
teg
r
ated
h
y
b
r
id
-
m
o
d
el
f
ea
tu
r
e
ex
tr
ac
to
r
,
a
d
v
an
ce
d
f
ea
tu
r
e
f
u
s
es,
an
d
o
p
tim
ized
tr
ain
in
g
s
t
r
ateg
ies.
T
h
is
r
esear
ch
p
r
esen
ts
a
n
o
v
el
h
y
b
r
id
d
ee
p
lear
n
in
g
-
b
ased
f
r
am
ewo
r
k
ca
lled
OC
T
Mo
d
Net
f
o
r
OC
T
im
ag
e
m
o
d
ality
class
if
icatio
n
.
Fu
s
io
n
o
f
E
f
f
icien
tNetV2
S
a
n
d
VGG1
6
f
o
r
m
u
lti
-
s
ca
le
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
th
e
u
s
e
o
f
atten
tio
n
m
ec
h
an
is
m
,
s
q
u
ee
ze
an
d
ex
citatio
n
b
l
o
ck
(
SE)
to
im
p
r
o
v
e
class
if
icatio
n
ac
cu
r
ac
y
an
d
r
o
b
u
s
tn
ess
.
I
t
u
s
es
a
f
ea
t
u
r
e
f
u
s
io
n
s
tr
ateg
y
to
co
m
b
in
e
t
h
e
ex
tr
ac
te
d
f
ea
t
u
r
es,
im
p
r
o
v
in
g
d
is
cr
im
in
ativ
e
ab
ilit
y
.
B
esid
es,
th
e
lear
n
in
g
r
ate
s
ch
ed
u
le
was
im
p
lem
e
n
ted
,
as
th
e
c
o
s
in
e
d
ec
ay
r
estar
t
was
ap
p
lied
to
th
e
o
p
tim
izer
to
s
tab
ilize
co
n
v
er
g
en
ce
b
etter
an
d
p
r
ev
en
t o
v
er
f
it
tin
g
.
B
y
co
m
b
in
in
g
d
if
f
er
e
n
t state
-
of
-
th
e
-
ar
t
d
ee
p
lear
n
in
g
a
r
ch
itectu
r
es
with
a
n
ad
ap
tiv
e
o
p
tim
izatio
n
s
tr
ateg
y
an
d
au
g
m
en
ted
f
ea
tu
r
e
p
r
o
c
ess
in
g
m
eth
o
d
s
,
we
d
em
o
n
s
tr
ate
im
p
r
o
v
ed
class
if
icatio
n
p
er
f
o
r
m
an
ce
c
o
m
p
a
r
e
d
to
s
tate
-
of
-
th
e
-
ar
t
ar
ch
itect
u
r
es,
wh
ich
is
th
e
n
o
v
elty
o
f
th
is
wo
r
k
.
T
h
is
wo
r
k
p
r
esen
ts
a
s
tr
u
ctu
r
ed
an
d
s
y
s
tem
atic
f
r
am
ewo
r
k
ch
ar
ac
ter
i
ze
d
b
y
s
ev
er
al
k
e
y
co
n
tr
ib
u
tio
n
s
:
a)
Hy
b
r
id
m
u
lti
-
b
ac
k
b
o
n
e
i
n
teg
r
atio
n
:
a
d
u
al
-
b
ac
k
b
o
n
e
ar
ch
it
ec
tu
r
e
co
m
b
in
in
g
two
E
f
f
icie
n
tNetV2
S
an
d
VGG1
6
to
ex
am
i
n
e
co
m
p
lem
e
n
tar
y
m
u
lti
-
s
ca
le
f
ea
tu
r
e
ex
t
r
a
ctio
n
an
d
im
p
r
o
v
e
d
is
cr
im
in
ativ
e
ca
p
ab
ilit
y
in
OC
T
r
etin
al
im
ag
in
g
.
b)
Atten
tio
n
-
en
h
an
ce
d
f
ea
tu
r
e
r
e
f
in
em
en
t
:
t
o
e
n
h
an
ce
f
ea
tu
r
e
r
ec
alib
r
ati
o
n
an
d
r
o
b
u
s
tn
ess
to
s
u
b
tle
r
etin
al
s
tr
u
ctu
r
al
v
ar
iatio
n
s
.
SE
ch
a
n
n
el
atten
tio
n
m
ec
h
an
is
m
was in
teg
r
ated
an
d
ass
ess
ed
.
c)
C
o
m
p
r
eh
en
s
iv
e
c
o
m
p
ar
ativ
e
e
v
alu
atio
n
:
a
s
y
s
tem
atic
ev
alu
a
tio
n
o
f
wid
ely
u
s
ed
b
ac
k
b
o
n
e
m
o
d
els
s
u
ch
as
R
esNet5
0
,
I
n
ce
p
tio
n
V3
,
Den
s
eNe
t1
2
1
,
VGG1
6
,
a
n
d
E
f
f
icien
tNetV2
S,
to
p
o
s
itio
n
th
e
p
r
o
p
o
s
e
d
f
r
am
ewo
r
k
with
in
ex
is
tin
g
OC
T
class
if
icatio
n
m
eth
o
d
o
lo
g
ies.
d)
Op
tim
ized
tr
ain
in
g
s
tr
ateg
y
f
o
r
r
o
b
u
s
t
co
n
v
er
g
e
n
ce
:
t
h
e
d
ep
lo
y
m
e
n
t
an
d
em
p
ir
ical
v
alid
atio
n
o
f
ad
ap
tiv
e
o
p
tim
izatio
n
u
s
in
g
Nad
am
with
co
s
in
e
d
ec
ay
r
estar
ts
to
im
p
r
o
v
e
co
n
v
er
g
e
n
c
e
s
tab
ilit
y
an
d
g
en
er
aliza
tio
n
ac
r
o
s
s
h
eter
o
g
e
n
eo
u
s
OC
T
d
atasets
.
T
h
e
clin
ical
s
ig
n
if
ican
ce
a
n
d
th
e
u
n
iq
u
e
co
n
tr
ib
u
tio
n
o
f
th
is
s
tu
d
y
ca
n
b
e
s
u
m
m
ar
iz
ed
as
th
e
in
tr
o
d
u
ctio
n
o
f
a
n
o
v
el
h
y
b
r
id
f
r
am
ewo
r
k
th
at
f
u
s
es
E
f
f
icien
tN
etV2
S
an
d
VGG1
6
b
ac
k
b
o
n
es
in
teg
r
ated
with
SE
atten
tio
n
f
o
r
th
e
f
ir
s
t
tim
e
f
o
r
th
e
class
if
icatio
n
o
f
s
ev
en
d
if
f
er
en
t
r
etin
al
d
is
e
ases
.
B
y
co
m
b
in
in
g
co
m
p
lem
en
tar
y
f
ea
tu
r
e
e
x
tr
ac
tio
n
m
ec
h
a
n
is
m
s
with
atten
tio
n
-
d
r
i
v
en
c
h
an
n
el
r
ec
alib
r
ati
o
n
an
d
a
s
tr
u
ctu
r
ed
f
ea
t
u
r
e
f
u
s
io
n
s
tr
ateg
y
,
th
e
p
r
o
p
o
s
ed
OC
T
Mo
d
Net
en
h
a
n
c
es
d
is
cr
im
in
ativ
e
p
er
f
o
r
m
a
n
c
e
f
o
r
s
u
b
tle
r
etin
al
ab
n
o
r
m
alities
.
T
h
e
in
teg
r
atio
n
o
f
ad
ap
tiv
e
o
p
tim
izatio
n
th
r
o
u
g
h
co
s
in
e
d
ec
ay
r
estar
t
s
ch
ed
u
li
n
g
an
d
Nad
am
f
u
r
th
er
im
p
r
o
v
es
tr
ain
in
g
s
tab
ilit
y
an
d
g
en
er
aliza
t
io
n
u
n
d
er
h
eter
o
g
en
eo
u
s
an
d
p
ar
tiall
y
im
b
alan
ce
d
m
u
lti
-
d
ataset
co
n
d
itio
n
s
.
C
lin
ically
,
th
e
f
r
am
ewo
r
k
s
u
p
p
o
r
ts
co
m
p
r
eh
en
s
iv
e
au
to
m
ated
s
cr
ee
n
in
g
o
f
m
u
ltip
le
r
etin
al
d
is
ea
s
es
with
in
a
u
n
if
ied
s
y
s
tem
,
o
f
f
e
r
in
g
p
o
ten
tial
as
a
r
o
b
u
s
t
d
ec
is
io
n
-
s
u
p
p
o
r
t
t
o
o
l
f
o
r
o
p
h
th
alm
ic
d
iag
n
o
s
tics
an
d
r
ea
l
-
w
o
r
ld
OC
T
d
ep
lo
y
m
en
t.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
r
e
m
ain
d
e
r
o
f
th
is
p
ap
e
r
is
as
f
o
llo
ws.
Sectio
n
2
is
a
tab
u
lated
liter
atu
r
e
r
ev
iew.
T
h
e
th
ir
d
s
ec
tio
n
c
o
n
tain
s
th
e
o
v
er
v
iew
o
f
th
e
p
r
o
p
o
s
ed
m
e
th
o
d
,
w
h
ich
c
o
m
p
r
is
es
th
e
d
at
aset
p
r
ep
r
o
ce
s
s
in
g
,
d
ee
p
lear
n
in
g
ar
ch
itectu
r
e
,
an
d
tr
ain
in
g
s
tr
ateg
ies
.
T
h
e
ex
p
er
im
en
tal
r
esu
lts
ar
e
r
ep
o
r
ted
in
th
e
4
th
s
ec
tio
n
.
T
h
e
5
th
Sectio
n
s
u
m
m
a
r
izes a
n
d
co
n
clu
d
es th
e
wo
r
k
with
a
d
i
s
cu
s
s
io
n
o
f
th
e
im
p
licatio
n
s
o
f
th
e
f
in
d
in
g
s
.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
OC
T
h
as
em
er
g
ed
as
a
k
e
y
i
m
ag
in
g
m
o
d
ality
in
ass
ess
in
g
r
etin
al
d
is
ea
s
e
d
u
e
to
its
n
o
n
-
in
v
asiv
e,
h
ig
h
-
r
eso
lu
tio
n
cr
o
s
s
-
s
ec
tio
n
a
l
r
etin
a
im
ag
in
g
.
L
ately
,
d
ee
p
lear
n
in
g
h
as
witn
ess
ed
a
r
ap
id
d
ev
elo
p
m
en
t;
au
to
m
ated
an
aly
s
is
o
f
OC
T
i
m
ag
es
is
m
ak
in
g
f
u
r
th
er
s
tr
i
d
es
in
ef
f
icien
cy
an
d
ac
cu
r
ac
y
in
clin
ical
d
iag
n
o
s
is
.
OC
T
im
ag
e
class
if
icatio
n
h
as
b
ee
n
ap
p
r
o
ac
h
ed
with
d
i
f
f
er
en
t
m
eth
o
d
o
lo
g
ies,
s
u
ch
as
C
NNs,
tr
an
s
f
er
lear
n
in
g
,
GANs,
an
d
e
n
s
em
b
l
e
lear
n
in
g
.
B
ab
a
et
a
l.
[
8
]
in
tr
o
d
u
ce
d
a
cu
s
to
m
co
n
v
o
lu
tio
n
al
m
o
d
el
f
o
r
m
a
cu
lar
d
is
o
r
d
er
clas
s
if
icatio
n
,
ac
h
iev
in
g
9
7
%
tr
ain
in
g
ac
cu
r
ac
y
,
9
3
%
v
alid
atio
n
ac
c
u
r
ac
y
,
an
d
a
t
est
ac
cu
r
ac
y
r
ea
ch
ed
u
p
to
9
8
%
o
n
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
OC
T
Mo
d
N
et:
a
d
ee
p
lea
r
n
in
g
-
b
a
s
ed
fr
a
mewo
r
k
f
o
r
o
p
tica
l
co
h
eren
ce
… (
S
a
ja
A
ta
a
lla
h
Mu
h
a
mme
d
)
497
Ker
m
an
y
d
ataset.
T
alaa
t
et
a
l.
[
9
]
d
em
o
n
s
tr
ated
th
at
att
en
tio
n
m
o
d
els
d
o
n
o
t
alwa
y
s
r
ea
ch
n
ea
r
-
p
e
r
f
ec
t
p
er
f
o
r
m
an
ce
[
9
]
.
T
h
e
y
p
r
o
p
o
s
ed
an
atten
tio
n
-
au
g
m
en
ted
Den
s
Net
f
o
r
r
etin
al
d
is
ea
s
e
d
iag
n
o
s
is
;
th
e
s
tu
d
y
r
ep
o
r
ted
v
alid
atio
n
ac
cu
r
ac
y
o
f
0
.
9
1
6
7
[
9
]
.
Milad
in
o
v
i
ć
et
a
l.
[
1
0
]
h
i
g
h
lig
h
ted
c
h
allen
g
es
in
OC
T
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
.
T
h
ey
s
h
o
wed
th
at
d
if
f
er
e
n
t
cl
ass
if
ier
s
h
av
e
p
er
f
o
r
m
ed
d
if
f
er
en
tly
u
n
d
er
d
ata
s
ca
r
city
an
d
lab
el
n
o
is
e
[
1
0
]
.
I
n
th
is
s
tu
d
y
,
e
x
p
er
im
e
n
ts
p
r
o
v
ed
th
at
I
n
ce
p
tio
n
V3
was
th
e
b
est
p
er
f
o
r
m
er
,
r
eg
is
ter
in
g
p
er
f
o
r
m
a
n
ce
g
r
ea
t
er
th
an
9
0
%
[
1
0
]
.
I
n
an
o
th
er
s
tu
d
y
o
n
r
etin
o
p
ath
y
d
etec
tio
n
u
s
in
g
VGG1
6
an
d
s
em
i
-
s
u
p
er
v
is
ed
lear
n
in
g
[
1
1
]
,
p
r
o
p
o
s
ed
a
n
etwo
r
k
th
at
ac
h
iev
ed
ar
o
u
n
d
9
4
%
an
d
9
3
%
o
n
s
m
all
OC
T
d
atasets
,
s
tatin
g
th
at
p
er
f
o
r
m
a
n
ce
ca
n
b
e
c
o
n
s
id
er
ab
l
y
b
elo
w
9
8
% in
lim
ited
d
ata
s
ettin
g
s
[
1
1
]
.
A
r
ec
e
n
t stu
d
y
h
as
p
er
f
o
r
m
ed
Xce
p
tio
n
a
n
d
I
n
ce
p
tio
n
V3
with
s
p
ec
if
ic
au
g
m
en
tatio
n
tech
n
iq
u
es
o
n
th
e
OC
T
-
C
8
d
ataset,
an
d
b
o
th
n
etwo
r
k
s
r
eg
is
ter
ed
a
g
o
o
d
class
if
icatio
n
ac
cu
r
ac
y
.
T
h
e
I
n
ce
p
tio
n
m
o
d
el
ac
h
iev
e
d
9
4
.
8
2
%,
wh
ile
th
e
Xce
p
tio
n
m
o
d
el
p
er
f
o
r
m
e
d
b
etter
b
y
r
e
g
is
ter
in
g
9
5
.
2
5
ac
r
o
s
s
th
e
8
class
es
[
1
2
]
.
So
m
e
o
f
th
e
s
tu
d
ies
th
at
ca
r
r
ied
o
u
t
r
esear
c
h
o
n
d
if
f
er
en
t
OC
T
d
atasets
f
o
r
d
if
f
er
e
n
t
d
is
ea
s
e
d
etec
tio
n
a
n
d
class
if
icatio
n
h
a
v
e
b
ee
n
r
ev
iewe
d
an
d
s
u
m
m
ar
ized
i
n
T
ab
le
1
.
T
ab
le
1
.
Su
m
m
a
r
y
o
f
liter
atu
r
e
r
ev
iew
f
in
d
i
n
g
s
R
e
f
.
Y
e
a
r
M
e
t
h
o
d
o
l
o
g
y
D
a
t
a
s
e
t
s
c
o
p
e
I
d
e
n
t
i
f
i
e
d
l
i
m
i
t
a
t
i
o
n
s
P
a
r
r
a
-
M
o
r
a
e
t
a
l
.
[
3
]
2
0
2
1
D
e
e
p
C
N
N
f
o
r
ER
M
d
e
t
e
c
t
i
o
n
Lo
c
a
l
D
a
t
a
se
t
(
ER
M
o
n
l
y
)
B
i
n
a
r
y
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
si
n
g
l
e
d
i
se
a
se
d
e
t
e
c
t
i
o
n
,
n
o
f
e
a
t
u
r
e
f
u
si
o
n
o
r
a
t
t
e
n
t
i
o
n
mo
d
u
l
e
s fo
r
b
r
o
a
d
e
r
p
a
t
h
o
l
o
g
y
r
e
p
r
e
s
e
n
t
a
t
i
o
n
.
Y
o
o
e
t
a
l
.
[
5
]
2
0
2
1
F
e
w
-
sh
o
t
l
e
a
r
n
i
n
g
f
r
a
mew
o
r
k
K
e
r
man
y
2
0
1
8
S
i
n
g
l
e
f
e
a
t
u
r
e
e
x
t
r
a
c
t
o
r
w
h
i
c
h
r
e
s
t
r
i
c
t
s rep
r
e
se
n
t
a
t
i
o
n
a
l
d
i
v
e
r
s
i
t
y
;
f
e
w
-
sh
o
t
l
e
a
r
n
i
n
g
r
a
t
h
e
r
t
h
a
n
l
a
r
g
e
-
sca
l
e
mu
l
t
i
-
c
l
a
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
;
n
o
f
u
s
i
o
n
o
r
a
t
t
e
n
t
i
o
n
mo
d
u
l
e
s.
A
l
t
a
n
[
7
]
2
0
2
2
Ex
p
l
a
i
n
a
b
l
e
C
N
N
D
i
sea
s
e
-
sp
e
c
i
f
i
c
O
C
T
d
a
t
a
se
t
S
p
e
c
i
f
i
c
d
i
se
a
se
c
a
t
e
g
o
r
y
;
l
a
c
k
s m
u
l
t
i
-
b
a
c
k
b
o
n
e
r
e
p
r
e
se
n
t
a
t
i
o
n
l
e
a
r
n
i
n
g
;
s
t
a
t
i
c
t
r
a
i
n
i
n
g
p
i
p
e
l
i
n
e
w
i
t
h
o
u
t
a
d
a
p
t
i
v
e
l
e
a
r
n
i
n
g
r
a
t
e
st
r
a
t
e
g
i
e
s
o
r
f
e
a
t
u
r
e
f
u
s
i
o
n
.
B
a
b
a
e
t
a
l
.
[
8
]
2
0
2
4
C
u
s
t
o
m C
N
N
A
r
c
h
i
t
e
c
t
u
r
e
K
e
r
man
y
2
0
1
8
S
i
n
g
l
e
-
b
a
c
k
b
o
n
e
a
r
c
h
i
t
e
c
t
u
r
e
t
h
a
t
l
i
m
i
t
s m
u
l
t
i
-
sc
a
l
e
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
;
l
a
c
k
s
a
t
t
e
n
t
i
o
n
me
c
h
a
n
i
sm
s t
o
e
m
p
h
a
s
i
z
e
p
a
t
h
o
l
o
g
y
-
r
e
l
e
v
a
n
t
r
e
t
i
n
a
l
l
a
y
e
r
s;
f
i
x
e
d
l
e
a
r
n
i
n
g
r
a
t
e
t
r
a
i
n
i
n
g
st
r
a
t
e
g
y
r
e
d
u
c
e
s c
o
n
v
e
r
g
e
n
c
e
e
f
f
i
c
i
e
n
c
y
a
n
d
g
e
n
e
r
a
l
i
z
a
t
i
o
n
.
Ta
l
a
a
t
e
t
a
l
.
[
9
]
2
0
2
4
A
t
t
e
n
t
i
o
n
-
b
a
se
d
D
e
n
seN
e
t
K
e
r
man
y
2
0
1
8
V
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
~
9
1
.
7
%
;
A
l
t
h
o
u
g
h
a
t
t
e
n
t
i
o
n
i
s
u
se
d
,
t
h
e
a
r
c
h
i
t
e
c
t
u
r
e
r
e
l
i
e
s
o
n
a
s
i
n
g
l
e
b
a
c
k
b
o
n
e
,
r
e
st
r
i
c
t
e
d
c
o
m
p
l
e
me
n
t
a
r
y
f
e
a
t
u
r
e
r
e
p
r
e
se
n
t
a
t
i
o
n
,
c
o
n
v
e
n
t
i
o
n
a
l
LR
sch
e
d
u
l
i
n
g
w
i
t
h
o
u
t
a
d
a
p
t
i
v
e
o
p
t
i
mi
z
a
t
i
o
n
.
M
i
l
a
d
i
n
o
v
i
ć
e
t
a
l
.
[
1
0
]
2
0
2
4
C
o
m
p
a
r
a
t
i
v
e
d
e
e
p
l
e
a
r
n
i
n
g
e
v
a
l
u
a
t
i
o
n
H
e
t
e
r
o
g
e
n
e
o
u
s
O
C
T
d
a
t
a
se
t
s
F
o
c
u
ses
o
n
e
v
a
l
u
a
t
i
o
n
r
a
t
h
e
r
t
h
a
n
a
r
c
h
i
t
e
c
t
u
r
a
l
i
n
n
o
v
a
t
i
o
n
;
n
o
f
e
a
t
u
r
e
f
u
si
o
n
,
a
t
t
e
n
t
i
o
n
i
n
t
e
g
r
a
t
i
o
n
,
o
r
a
d
a
p
t
i
v
e
o
p
t
i
m
i
z
a
t
i
o
n
t
o
a
d
d
r
e
s
s
d
o
m
a
i
n
s
h
i
f
t
a
c
r
o
ss
d
a
t
a
se
t
s.
Lu
o
e
t
a
l
.
[
1
1
]
2
0
2
1
V
G
G
1
6
w
i
t
h
semi
-
s
u
p
e
r
v
i
se
d
l
e
a
r
n
i
n
g
K
e
r
man
y
2
0
1
8
~
9
4
%
a
c
c
u
r
a
c
y
;
s
i
n
g
l
e
a
r
c
h
i
t
e
c
t
u
r
e
;
l
a
c
k
s
a
t
t
e
n
t
i
o
n
mo
d
u
l
e
s
f
o
r
d
i
scri
mi
n
a
t
i
v
e
f
e
a
t
u
r
e
w
e
i
g
h
t
i
n
g
;
n
o
mu
l
t
i
-
sc
a
l
e
o
r
mu
l
t
i
-
b
a
c
k
b
o
n
e
f
e
a
t
u
r
e
f
u
si
o
n
st
r
a
t
e
g
y
.
N
o
o
r
e
t
a
l
.
[
1
2
]
2
0
2
6
Tr
a
n
sf
e
r
l
e
a
r
n
i
n
g
(
X
c
e
p
t
i
o
n
,
I
n
c
e
p
t
i
o
n
V
3
)
M
u
l
t
i
-
c
l
a
ss
O
C
T
-
C
8
d
a
t
a
se
t
~
9
4
–
9
5
%
a
c
c
u
r
a
c
y
;
s
i
n
g
l
e
-
b
a
c
k
b
o
n
e
mo
d
e
l
s
w
i
t
h
n
o
f
e
a
t
u
r
e
f
u
si
o
n
;
a
b
s
e
n
c
e
o
f
a
t
t
e
n
t
i
o
n
m
e
c
h
a
n
i
s
ms r
e
d
u
c
e
s
t
h
e
a
b
i
l
i
t
y
t
o
e
mp
h
a
s
i
z
e
d
i
s
e
a
se
-
s
p
e
c
i
f
i
c
r
e
g
i
o
n
s;
e
mp
l
o
y
s s
t
a
t
i
c
o
p
t
i
m
i
z
a
t
i
o
n
s
t
r
a
t
e
g
i
e
s wi
t
h
o
u
t
d
y
n
a
mi
c
l
e
a
r
n
i
n
g
r
a
t
e
a
d
a
p
t
a
t
i
o
n
.
Pre
v
io
u
s
ly
r
ev
iewe
d
OC
T
cla
s
s
if
icatio
n
s
tu
d
ies
h
av
e
d
em
o
n
s
tr
ated
g
o
o
d
p
er
f
o
r
m
a
n
ce
u
s
in
g
s
in
g
le
b
a
c
k
b
o
n
e
C
N
N
s
(
e
.
g
.
,
V
G
G
1
6
,
I
n
c
e
p
t
i
o
n
,
X
c
e
p
t
i
o
n
,
a
n
d
D
e
n
s
N
e
t
)
[
3
]
-
[1
2
]
.
F
o
r
i
n
s
t
a
n
c
e
,
P
a
r
r
a
-
M
o
r
a
e
t
a
l
.
[
3
]
p
r
o
p
o
s
ed
a
d
ee
p
C
NN
ar
ch
ite
ctu
r
e
f
o
r
E
R
M
d
etec
tio
n
u
s
in
g
OC
T
im
a
g
es;
d
esp
ite
th
eir
a
p
p
r
o
ac
h
f
o
cu
s
ed
o
n
s
in
g
le
-
d
is
ea
s
e
class
if
icatio
n
u
s
in
g
a
f
ea
tu
r
e
ex
tr
ac
to
r
.
Simil
ar
ly
,
Yo
o
et
a
l.
[
5
]
in
v
esti
g
ated
f
ew
-
s
h
o
t
lear
n
i
n
g
s
tr
ateg
ies
to
im
p
r
o
v
e
OC
T
d
ia
g
n
o
s
is
o
f
r
ar
e
r
etin
al
d
is
ea
s
es,
y
et
th
eir
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
r
elied
o
n
a
s
in
g
le
b
ac
k
b
o
n
e
ar
ch
itectu
r
e
with
o
u
t
m
u
lti
-
m
o
d
el
f
ea
tu
r
e
f
u
s
io
n
.
W
h
ile
Altan
[
7
]
in
tr
o
d
u
c
ed
Dee
p
OC
T
,
an
ex
p
lain
ab
le
C
NN
m
o
d
el
f
o
r
m
ac
u
lar
ed
em
a
an
aly
s
is
,
th
eir
s
tu
d
y
r
em
ain
s
d
is
ea
s
e
-
s
p
ec
if
ic
an
d
d
id
n
o
t
in
co
r
p
o
r
ate
co
m
p
lem
en
tar
y
m
u
lti
-
s
ca
le
f
ea
tu
r
e
ex
tr
ac
t
io
n
o
r
ad
ap
tiv
e
lear
n
in
g
r
ate
s
ch
e
d
u
li
n
g
.
Ho
wev
er
,
m
o
s
t
ex
is
tin
g
f
r
am
e
wo
r
k
s
r
ely
o
n
a
s
in
g
le
f
ea
tu
r
e
ex
tr
ac
to
r
,
wh
ic
h
ca
n
lim
it
th
e
ca
p
tu
r
e
o
f
co
m
p
lem
en
tar
y
r
etin
al
p
atter
n
s
ac
r
o
s
s
s
ca
le
an
d
tex
tu
r
es,
p
ar
ticu
lar
ly
wh
en
class
b
o
u
n
d
ar
ies
ar
e
s
u
b
tle.
I
n
ad
d
it
io
n
,
tr
ain
in
g
o
n
s
tatic
lea
r
n
in
g
r
ates
o
r
c
o
n
v
o
lu
tio
n
al
s
ch
ed
u
les
m
a
y
r
esu
lt
in
u
n
s
tab
le
co
n
v
er
g
en
ce
an
d
s
u
b
o
p
tim
al
g
en
er
aliza
tio
n
u
n
d
er
h
eter
o
g
en
e
o
u
s
OC
T
ac
q
u
is
itio
n
s
.
C
o
n
v
er
s
ely
,
OC
T
Mo
d
Net
ad
d
r
ess
es
th
es
e
lim
itatio
n
s
th
r
o
u
g
h
d
u
al
-
b
ac
k
b
o
n
e
f
u
s
io
n
(
E
f
f
icie
n
tNetV2
S
an
d
VGG1
6
)
to
lear
n
co
m
p
le
m
en
tar
y
m
u
lti
-
s
ca
le
f
ea
tu
r
es.
Mo
r
eo
v
e
r
,
to
r
ec
alib
r
ate
in
f
o
r
m
ativ
e
r
etin
al
r
esp
o
n
s
es,
an
SE
-
b
lo
ck
ch
an
n
el
atten
t
io
n
was
u
s
ed
,
an
d
an
ad
ap
tiv
e
lear
n
in
g
r
ate
s
tr
ateg
y
was
u
s
ed
(
co
s
in
e
d
ec
ay
r
estar
ts
with
Nad
am
)
to
en
h
an
ce
co
n
v
er
g
en
c
e
s
tab
ilit
y
an
d
r
o
b
u
s
tn
ess
.
T
h
is
co
m
b
in
atio
n
a
d
v
an
ce
s
b
e
y
o
n
d
p
r
io
r
s
in
g
le
-
s
tr
ea
m
OC
T
class
if
ier
s
b
y
ex
p
licitly
in
teg
r
atin
g
m
u
lti
-
m
o
d
el
f
ea
t
u
r
e
f
u
s
io
n
an
d
a
d
ap
tiv
e
o
p
tim
i
za
tio
n
with
in
a
u
n
if
ied
s
ev
en
-
class
OC
T
d
is
ea
s
e
class
if
icatio
n
f
r
am
ew
o
r
k
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
43
,
No
.
2
,
Au
g
u
s
t
20
26
:
495
-
5
0
6
498
3.
M
AT
E
R
I
AL
S AN
D
M
E
T
H
O
DS
3
.
1
.
Da
t
a
s
et
des
cr
iptio
n
A
m
ix
o
f
two
h
ig
h
-
q
u
ality
r
et
in
al
OC
T
d
atasets
f
r
o
m
th
e
Kag
g
le
OC
T
-
C
8
d
ataset
an
d
th
e
OC
T
DL
h
as
b
ee
n
u
s
ed
to
d
ev
elo
p
th
e
m
o
d
el
an
d
test
it
f
o
r
th
e
d
ia
g
n
o
s
is
o
f
r
etin
al
d
is
ea
s
es
[
1
3
]
,
[
1
4
]
.
C
o
m
b
in
i
n
g
s
am
p
les
f
r
o
m
th
ese
two
d
atasets
in
tr
o
d
u
ce
s
i
n
ter
-
d
ataset
v
a
r
iab
ilit
y
,
wh
ich
n
o
r
m
ally
h
as
d
if
f
er
en
t
ac
q
u
is
itio
n
p
r
o
to
co
ls
,
d
if
f
e
r
en
t
s
ca
n
n
er
s
,
an
d
d
em
o
g
r
ap
h
ic
d
iv
er
s
ity
.
T
h
is
m
ix
will
im
p
r
o
v
e
th
e
m
o
d
el's
r
o
b
u
s
tn
ess
an
d
g
en
er
aliza
tio
n
ca
p
ac
ity
.
Six
d
i
f
f
er
en
t
class
es
f
r
o
m
th
e
OC
T
-
C
8
d
ataset
h
av
e
b
ee
n
ch
o
s
en
;
ea
ch
class
h
as
DR
,
C
S
R
,
DM
E
,
DR
U
SEN,
MH
,
an
d
NORMAL
.
T
h
e
d
ataset
is
s
p
lit
in
to
t
r
ain
,
v
alid
atio
n
,
an
d
test
s
ets
[
1
3
]
.
I
m
ag
es
f
r
o
m
t
h
is
d
ataset
ar
e
n
o
t
eq
u
al
in
s
ize;
th
e
s
ize
o
f
th
ese
im
ag
es
d
if
f
er
s
.
T
h
e
E
R
M
class
was
ch
o
s
en
f
r
o
m
th
e
OC
T
DL
d
ataset
to
ad
d
d
iv
er
s
ity
;
th
e
s
izes
o
f
i
m
ag
es
in
th
is
d
ataset
ar
e
n
o
t
eq
u
al
eith
er
[
1
4
]
.
Dif
f
er
en
t
s
am
p
les
ar
e
s
h
o
wn
in
Fig
u
r
e
1
.
Data
d
is
tr
ib
u
tio
n
ac
r
o
s
s
th
e
s
ev
en
d
if
f
er
en
t
class
es
i
s
s
h
o
wn
in
T
ab
le
2
.
Fig
u
r
e
1.
Sam
p
les f
r
o
m
7
-
class
d
ataset
T
ab
le
2
.
Sam
p
le
d
is
tr
ib
u
tio
n
a
cr
o
s
s
th
e
7
-
class
d
ataset
C
l
a
s
s
Ta
r
i
n
V
a
l
i
d
a
t
i
o
n
Te
st
To
t
a
l
p
e
r
c
l
a
ss
C
S
R
2
,
3
0
0
3
5
0
3
5
0
3
,
0
0
0
D
M
E
2
,
3
0
0
3
5
0
3
5
0
3
,
0
0
0
DR
2
,
3
0
0
3
5
0
3
5
0
3
,
0
0
0
D
R
U
S
EN
2
,
3
0
0
3
5
0
3
5
0
3
,
0
0
0
M
a
c
u
l
a
H
o
l
e
2
,
3
0
0
3
5
0
3
5
0
3
,
0
0
0
N
o
r
mal
2
,
3
0
0
3
5
0
3
5
0
3
,
0
0
0
ER
M
2
,
3
5
3
4
9
0
4
8
3
3
,
3
2
6
To
t
a
l
1
6
,
1
5
3
2
,
5
9
0
2
,
5
8
3
2
1
,
3
2
6
T
o
ass
ess
th
e
g
en
er
aliza
tio
n
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
an
d
m
iti
g
ate
d
ataset
-
s
p
ec
if
ic
b
ias,
th
e
m
o
d
el
was
ev
alu
ated
o
n
two
d
e
p
en
d
e
n
t
d
atasets
:
Ker
m
ay
2
0
1
8
[
1
5
]
an
d
OC
T
-
C
8
[
1
3
]
d
atasets
.
T
h
e
Ker
m
a
n
y
2
0
1
8
co
n
s
is
ts
o
f
ap
p
r
o
x
im
ately
2
0
7
,
1
3
0
im
ag
es,
co
llected
f
r
o
m
d
if
f
er
en
t
p
atien
ts
an
d
d
e
v
ices
b
etwe
en
2
0
1
3
an
d
2
0
1
7
.
T
h
e
d
is
tr
ib
u
tio
n
o
f
s
am
p
les
ac
r
o
s
s
th
e
f
o
u
r
class
es
is
(
C
NV(
n
=3
7
.
2
k
)
,
DM
E
(
n
=1
1
,
3
k
)
,
DR
USEN
(
n
=8
,
6
1
6
)
,
an
d
NO
R
MA
L
(
n
=2
6
,
3
k
)
)
i
n
th
e
tr
ain
in
g
s
u
b
s
et,
wh
ile
th
e
v
alid
ati
o
n
s
et
c
o
n
tain
s
2
5
0
im
ag
es p
er
class
,
an
d
th
e
test
s
et
2
4
2
s
am
p
les p
er
ea
ch
o
f
th
e
f
o
u
r
class
es.
E
x
p
er
im
en
ts
wer
e
ex
te
n
d
ed
t
o
th
e
r
etin
al
OC
T
-
C
8
d
ataset.
I
t
co
n
s
is
ts
o
f
8
class
es,
ea
ch
with
4
,
6
0
0
im
ag
es:
Ag
e
-
R
elate
d
Ma
cu
la
r
Oed
em
a,
C
NV,
C
SR
,
DM
E
,
DR
,
DR
USEN
,
MH
,
an
d
NORMAL
.
T
h
e
v
alid
atio
n
a
n
d
tes
t
s
u
b
s
ets
h
av
e
7
0
0
im
a
g
es/clas
s
.
T
h
e
s
tan
d
ar
d
OC
T
class
if
icatio
n
s
ettin
g
h
as
b
ee
n
ex
p
an
d
e
d
in
th
is
d
ataset
to
co
m
p
r
is
e
eig
h
t
r
etin
al
d
is
o
r
d
er
ca
teg
o
r
ies,
wr
ap
p
in
g
b
o
th
p
ath
o
l
o
g
ical
an
d
n
o
r
m
al
co
n
d
itio
n
s
.
T
h
e
in
ter
-
class
v
is
u
al
s
im
ilar
ity
an
d
th
e
elev
a
ted
n
u
m
b
e
r
o
f
class
es
m
ak
e
th
is
d
ataset
m
o
r
e
d
if
f
icu
lt a
n
d
ch
allen
g
in
g
th
an
th
e
Ker
m
an
y
2
0
1
8
d
ataset.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
OC
T
Mo
d
N
et:
a
d
ee
p
lea
r
n
in
g
-
b
a
s
ed
fr
a
mewo
r
k
f
o
r
o
p
tica
l
co
h
eren
ce
… (
S
a
ja
A
ta
a
lla
h
Mu
h
a
mme
d
)
499
3
.
2
.
Da
t
a
prepro
ce
s
s
ing
3
.
2
.
1
.
I
m
a
g
e
re
s
izing
All
th
e
im
ag
es
in
th
e
m
ix
ed
d
ataset
wer
e
r
esized
to
h
av
e
th
e
ex
ac
t
d
im
en
s
io
n
s
an
d
,
h
en
ce
,
ar
e
u
n
if
o
r
m
an
d
r
ea
d
y
to
b
e
u
tili
ze
d
to
f
ee
d
in
to
th
e
d
ee
p
lear
n
i
n
g
m
o
d
els.
I
m
ag
es
a
r
e
r
esized
to
2
56
×2
56
p
i
x
els,
k
ee
p
in
g
th
e
asp
ec
t
r
atio
d
u
e
to
d
is
to
r
tio
n
.
R
esizin
g
im
ag
es
t
o
a
s
tan
d
a
r
d
r
eso
lu
tio
n
‘
2
56
×2
56
’
is
c
r
u
cial
in
th
e
d
ev
elo
p
m
e
n
t
o
f
th
e
m
o
d
el,
allo
win
g
an
ef
f
ec
tiv
e
an
d
u
n
if
o
r
m
p
r
o
ce
s
s
o
f
im
a
g
es
b
y
th
e
co
n
v
o
l
u
tio
n
n
e
u
r
al
n
etwo
r
k
s
.
3
.
2
.
2
.
No
rma
liza
t
io
n
I
t
is
th
e
p
r
o
ce
s
s
o
f
co
n
v
er
tin
g
p
ix
el
v
alu
es
(
in
ten
s
ities
)
in
t
o
a
co
n
s
is
ten
t
r
a
n
g
e.
No
r
m
ali
zin
g
p
i
x
el
in
ten
s
ity
v
alu
es
h
elp
s
th
e
n
etwo
r
k
to
lear
n
im
ag
e
p
atter
n
s
m
o
r
e
ef
f
icien
tly
a
n
d
ac
c
u
r
atel
y
.
I
n
th
is
s
tu
d
y
,
all
in
ten
s
ity
v
alu
es o
f
im
ag
e
p
ix
el
s
wer
e
s
ca
le
d
f
r
o
m
[
0
,
2
5
5
]
to
[
0
,
1
]
u
s
in
g
(
1
)
.
=
0
255
⁄
(
1
)
3
.
2
.
3
.
Da
t
a
a
ug
m
ent
a
t
io
n
Du
e
to
th
e
f
ac
t
th
at
h
u
g
e
am
o
u
n
ts
o
f
d
ata
ar
e
r
eq
u
ir
ed
b
y
d
ee
p
lear
n
in
g
m
o
d
els
to
b
e
ab
le
to
p
r
ed
ict
an
d
in
f
er
co
r
r
ec
tl
y
,
wo
r
k
in
g
o
n
m
ed
ical
im
a
g
es
p
o
s
es
a
ch
allen
g
e
b
ec
au
s
e
o
f
d
ata
s
ca
r
ci
ty
an
d
av
aila
b
ilit
y
.
R
esear
ch
er
s
h
ad
to
u
s
e
Au
g
m
en
tatio
n
tech
n
i
q
u
es
to
b
o
o
s
t
th
e
am
o
u
n
t
o
f
d
ata
an
d
im
p
r
o
v
e
t
h
e
lear
n
i
n
g
p
r
o
ce
s
s
an
d
m
o
d
el
g
en
er
aliza
tio
n
ab
ilit
y
to
r
ec
o
g
n
ize
n
ew
u
n
s
ee
n
d
ata.
T
h
e
u
s
ef
u
ln
ess
o
f
th
ese
tech
n
iq
u
es
d
ep
en
d
s
o
n
th
e
task
at
h
a
n
d
o
r
th
e
p
r
o
b
lem
to
b
e
s
o
lv
ed
.
I
n
th
is
wo
r
k
,
s
ev
e
r
al
au
g
m
en
tatio
n
tech
n
i
q
u
es
wer
e
u
s
ed
,
as lis
ted
in
T
ab
le
3
.
T
ab
le
3
.
Au
g
m
en
tatio
n
tec
h
n
i
q
u
es u
s
ed
in
th
e
s
tu
d
y
A
u
g
m
e
n
t
a
t
i
o
n
me
t
h
o
d
D
e
scri
p
t
i
o
n
R
o
t
a
t
i
o
n
R
o
t
a
t
e
s
t
h
e
i
n
p
u
t
i
m
a
g
e
r
a
n
d
o
m
l
y
b
y
a
n
g
l
e
[
-
3
0
,
3
0
]
d
e
g
r
e
e
s
H
o
r
i
z
o
n
t
a
l
f
l
i
p
p
i
n
g
F
l
i
p
s t
h
e
i
n
p
u
t
i
ma
g
e
a
r
o
u
n
d
t
h
e
h
o
r
i
z
o
n
t
a
l
a
x
i
s,
p
e
r
f
o
r
mi
n
g
a
mi
r
r
o
r
i
n
v
e
r
s
i
o
n
Zo
o
mi
n
g
R
a
n
d
o
m
l
y
z
o
o
ms
t
h
e
i
n
p
u
t
i
ma
g
e
i
n
o
r
o
u
t
.
S
h
e
a
r
r
a
n
g
e
Th
e
i
n
p
u
t
i
ma
g
e
w
i
l
l
b
e
d
i
st
o
r
t
e
d
a
l
o
n
g
t
h
e
h
o
r
i
z
o
n
t
a
l
a
x
i
s.
F
i
l
l
m
ode
F
i
l
l
s
t
h
e
e
m
p
t
y
a
r
e
a
s
(
b
o
r
d
e
r
s)
c
a
u
s
e
d
b
y
f
l
i
p
p
i
n
g
o
r
r
o
t
a
t
i
o
n
w
i
t
h
t
h
e
n
e
a
r
e
st
v
a
l
i
d
v
a
l
u
e
B
r
i
g
h
t
n
e
ss
a
n
d
c
o
n
t
r
a
st
a
d
j
u
st
m
e
n
t
U
p
d
a
t
e
s t
h
e
c
o
n
t
r
a
s
t
a
n
d
t
h
e
b
r
i
g
h
t
n
e
s
s o
f
t
h
e
i
n
p
u
t
i
ma
g
e
3
.
3
.
P
r
o
po
s
ed
a
rc
hite
ct
ure
T
h
is
s
tu
d
y
p
r
o
p
o
s
es
OC
T
M
o
d
Net,
a
h
y
b
r
id
d
ee
p
lear
n
i
n
g
f
r
am
ew
o
r
k
d
esig
n
ed
f
o
r
m
u
lti
-
class
r
etin
al
OC
T
im
ag
e
class
if
icati
o
n
.
T
h
e
f
r
am
ewo
r
k
b
r
in
g
s
to
g
eth
er
two
C
NN
b
ac
k
b
o
n
es,
E
f
f
icien
tNetV2
S
an
d
VGG1
6
,
o
p
e
r
atin
g
i
n
p
a
r
allel
to
ex
tr
ac
t
co
m
p
lem
en
tar
y
f
e
atu
r
e
r
e
p
r
esen
tatio
n
s
f
r
o
m
th
e
s
am
e
in
p
u
t
im
ag
e
.
An
SE
atten
tio
n
b
lo
c
k
is
in
t
eg
r
ated
with
in
th
e
E
f
f
icien
t
NetV2
S
b
r
an
ch
to
en
h
a
n
ce
c
h
an
n
el
-
wis
e
f
ea
tu
r
e
im
p
o
r
tan
ce
.
T
h
e
ex
tr
ac
ted
f
ea
t
u
r
es
f
r
o
m
t
h
e
two
b
r
an
ch
es
ar
e
th
en
f
u
s
ed
th
r
o
u
g
h
c
o
n
ca
ten
atio
n
an
d
p
ass
ed
to
a
f
u
l
ly
co
n
n
ec
ted
class
if
icatio
n
h
ea
d
.
T
h
e
o
v
er
all
ar
ch
itectu
r
e
o
f
th
e
m
o
d
el
is
illu
s
t
r
ated
in
Fig
u
r
e
2.
3.
4
.
B
a
c
k
bo
ne
m
o
del selec
t
i
o
n
T
o
id
en
tif
y
s
u
itab
le
b
ac
k
b
o
n
e
ar
ch
itectu
r
es,
f
iv
e
d
if
f
er
en
t,
wid
ely
u
s
ed
C
NN
ar
ch
itectu
r
es
wer
e
f
ir
s
t
ev
alu
ated
o
n
th
e
d
atas
et:
R
es
Net5
0
,
I
n
ce
p
tio
n
V3
,
Den
s
eNe
t1
2
1
,
VGG1
6
,
an
d
E
f
f
icien
tN
etV2
S.
Pre
lim
in
ar
y
ex
p
er
im
en
ts
s
h
o
wed
th
at
E
f
f
i
cien
tNetV2
S
p
er
f
o
r
m
e
d
b
etter
th
an
th
e
o
th
e
r
m
o
d
els
b
y
ac
h
i
ev
in
g
th
e
s
tr
o
n
g
est
class
if
icatio
n
p
er
f
o
r
m
an
ce
,
as
s
h
o
wn
in
T
ab
le
4
.
Alth
o
u
g
h
Den
s
eNe
t1
2
1
was
th
e
s
ec
o
n
d
b
est
p
er
f
o
r
m
e
r
,
b
u
t
VGG1
6
was
s
elec
ted
in
s
tead
f
o
r
its
co
m
p
lem
en
tar
y
ar
c
h
itectu
r
al
ch
ar
ac
ter
is
tics
.
VGG1
6
em
p
lo
y
s
a
s
im
p
le
an
d
s
eq
u
en
tial
co
n
v
o
lu
tio
n
al
d
esig
n
th
at
ef
f
ec
tiv
ely
ca
p
tu
r
es
f
in
e
-
g
r
ain
e
d
s
p
atial
p
atter
n
s
an
d
tex
tu
r
e
in
f
o
r
m
atio
n
in
OC
T
im
ag
es.
I
n
co
n
tr
ast,
E
f
f
icien
tNetV2
S
f
o
cu
s
es
o
n
ef
f
icien
t
h
ier
ar
c
h
ical
f
ea
tu
r
e
ex
tr
ac
tio
n
th
r
o
u
g
h
o
p
tim
ized
co
n
v
o
lu
tio
n
al
b
lo
ck
s
an
d
co
m
p
o
u
n
d
s
ca
l
in
g
.
C
o
m
b
in
in
g
t
h
ese
two
ar
ch
it
ec
tu
r
es
in
tr
o
d
u
ce
s
r
ep
r
esen
tatio
n
al
d
iv
er
s
ity
,
en
ab
lin
g
th
e
p
r
o
p
o
s
ed
mode
l
to
lear
n
b
o
th
d
etailed
lo
ca
l
s
tr
u
ctu
r
e
an
d
g
lo
b
ally
o
p
tim
ized
f
ea
t
u
r
e
r
ep
r
esen
tatio
n
s
.
T
h
is
co
m
p
lem
en
tar
y
co
m
b
in
atio
n
im
p
r
o
v
es
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
f
ea
tu
r
e
f
u
s
io
n
s
tag
e
an
d
co
n
tr
ib
u
tes
to
th
e
o
v
er
all
p
er
f
o
r
m
an
ce
im
p
r
o
v
e
m
en
t
o
f
th
e
OC
T
Mo
d
Net
m
o
d
el.
Fu
r
th
er
m
o
r
e,
co
m
b
in
in
g
d
if
f
er
en
t
d
esig
n
ar
ch
itectu
r
es
h
as
b
ee
n
s
h
o
wn
to
im
p
r
o
v
e
th
e
p
e
r
f
o
r
m
an
ce
o
f
h
y
b
r
id
m
o
d
els
b
y
r
e
d
u
cin
g
f
ea
tu
r
e
r
ed
u
n
d
a
n
cy
an
d
in
cr
ea
s
in
g
r
ep
r
esen
tatio
n
a
l d
iv
er
s
ity
in
m
u
lti
-
b
a
ck
b
o
n
e
n
etwo
r
k
s
[
16
].
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
43
,
No
.
2
,
Au
g
u
s
t
20
26
:
495
-
5
0
6
500
Fig
u
r
e
2.
OC
T
Mo
d
Net
ar
c
h
itectu
r
e
3
.
5
.
E
f
f
icient
Net
V2
S
A
co
n
v
o
l
u
tio
n
n
e
u
r
al
n
etwo
r
k
,
an
im
p
r
o
v
ed
,
s
m
aller
v
e
r
s
io
n
o
f
th
e
E
f
f
icien
tNet
f
am
ily
.
T
h
e
m
ain
ar
ch
itectu
r
e
b
lo
ck
o
f
th
is
n
etwo
r
k
is
MB
C
o
n
v
an
d
Fu
s
ed
-
MBC
o
n
v
.
T
h
e
m
ain
id
ea
b
e
h
in
d
u
s
in
g
Fu
s
ed
-
MBC
o
n
v
is
to
im
p
r
o
v
e
th
e
tr
ain
in
g
s
p
ee
d
[
1
7
]
.
T
h
e
m
a
in
m
o
d
if
icatio
n
m
a
d
e
was
r
em
o
v
in
g
th
e
f
u
lly
co
n
n
ec
ted
lay
er
(
FC
)
to
b
e
u
s
ed
o
n
ly
as
a
f
ea
tu
r
e
ex
tr
ac
to
r
,
allo
win
g
th
e
ar
ch
itectu
r
e
to
ex
tr
ac
t
h
ig
h
-
lev
el
v
is
u
al
f
ea
tu
r
es
s
u
ch
as
tex
tu
r
e
s
,
ed
g
es,
an
d
s
tr
u
ctu
r
al
p
atter
n
s
r
elev
an
t
to
r
etin
al
OC
T
im
ag
es.
T
h
e
m
o
d
el
was
f
in
e
-
tu
n
e
d
to
ad
a
p
t its
lear
n
ed
r
ep
r
esen
tatio
n
s
to
d
o
m
ain
-
s
p
e
cif
ic
r
etin
al
s
tr
u
ctu
r
es
[1
7
].
3.
6
.
Sq
ueez
e
-
a
nd
-
ex
cit
a
t
i
o
n
blo
ck
T
h
e
m
ain
g
o
al
b
e
h
in
d
u
s
in
g
SE
-
b
lo
ck
was
to
im
p
r
o
v
e
th
e
r
ep
r
esen
tatio
n
q
u
ality
th
at
a
n
etwo
r
k
in
tr
o
d
u
ce
s
.
T
h
is
b
l
o
ck
allo
ws
f
ea
tu
r
e
r
ec
alib
r
atio
n
,
wh
ich
e
n
ab
les
th
e
lear
n
in
g
p
r
o
ce
s
s
to
ca
r
ef
u
lly
h
ig
h
lig
h
t
f
ea
tu
r
es
th
at
ar
e
in
f
o
r
m
ativ
e
a
n
d
f
ilter
o
u
t
th
e
less
in
f
o
r
m
ativ
e
o
r
less
u
s
ef
u
l
o
n
es
[
1
8
]
.
T
h
e
SE
-
b
lo
ck
h
as
two
m
ain
p
h
ases
:
th
e
s
q
u
ee
ze
p
h
ase
an
d
th
e
ex
citatio
n
p
h
ase.
I
n
th
e
s
q
u
ee
ze
p
h
ase,
th
e
m
o
d
el
g
en
er
ates
ch
an
n
el
wis
e
s
tatis
t
ics
th
r
o
u
g
h
g
lo
b
all
y
p
o
o
lin
g
f
ea
tu
r
es
ac
r
o
s
s
th
e
s
p
atial
d
im
en
s
io
n
s
to
ca
p
tu
r
e
th
e
m
o
s
t
s
ig
n
if
ican
t
in
f
o
r
m
atio
n
f
o
r
ea
ch
ch
a
n
n
el.
I
n
th
e
s
ec
o
n
d
p
h
ase,
th
e
ex
c
ita
tio
n
p
h
ase,
th
e
lear
n
ed
ch
a
n
n
el
wis
e
s
tat
is
tic
s
f
r
o
m
th
e
f
ir
s
t
p
h
ase
will
b
e
u
t
ilized
to
p
r
o
d
u
ce
atten
tio
n
we
ig
h
ts
[
1
9
]
.
T
h
ese
weig
h
ts
r
ep
r
esen
t
th
e
in
f
lu
en
ce
o
f
ea
ch
ch
an
n
el
o
n
th
e
f
in
al
d
ec
is
io
n
o
r
r
ep
r
esen
tatio
n
.
T
h
r
o
u
g
h
th
is
p
r
o
ce
s
s
,
th
e
n
etwo
r
k
will
b
e
e
n
ab
led
to
g
iv
e
m
o
r
e
atten
tio
n
t
o
th
e
m
o
s
t
r
elev
an
t
o
r
m
o
s
t
in
f
o
r
m
at
iv
e
ch
an
n
els
an
d
b
lo
c
k
o
u
t
t
h
e
less
in
f
o
r
m
ativ
e
ch
an
n
els [
1
9
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
OC
T
Mo
d
N
et:
a
d
ee
p
lea
r
n
in
g
-
b
a
s
ed
fr
a
mewo
r
k
f
o
r
o
p
tica
l
co
h
eren
ce
… (
S
a
ja
A
ta
a
lla
h
Mu
h
a
mme
d
)
501
3
.
7
.
VG
G
1
6
I
t
is
a
ty
p
e
o
f
C
NN
in
tr
o
d
u
ce
d
b
y
th
e
Vis
u
al
Geo
m
etr
y
Gr
o
u
p
in
2
0
1
4
.
I
t
b
ec
am
e
o
n
e
o
f
th
e
ar
ch
itectu
r
es
th
at
g
ain
ed
s
i
g
n
if
ican
t
p
o
p
u
lar
ity
a
n
d
wa
s
d
ep
lo
y
ed
in
m
an
y
v
is
io
n
task
s
d
u
e
to
its
ef
f
ec
tiv
en
ess
,
s
im
p
licity
,
an
d
tr
an
s
f
er
ab
ilit
y
.
T
h
e
m
ain
g
o
al
o
f
th
e
p
r
o
p
o
s
ed
ar
c
h
itectu
r
e
was
to
d
em
o
n
s
tr
ate
th
e
cr
itical
r
o
le
o
f
th
e
d
e
p
th
o
f
th
e
n
etwo
r
k
in
ac
h
iev
i
n
g
h
ig
h
p
er
f
o
r
m
a
n
ce
in
v
is
io
n
ta
s
k
s
.
T
h
e
to
tal
lay
er
n
u
m
b
er
o
f
th
is
m
o
d
el
is
1
6
,
co
n
s
is
tin
g
o
f
1
3
co
n
v
o
lu
tio
n
s
an
d
th
r
ee
FC
lay
er
s
.
T
h
e
d
e
s
ig
n
o
f
th
e
n
etwo
r
k
im
p
r
o
v
es its
ca
p
ac
ity
to
e
x
tr
ac
t c
o
m
p
lex
f
ea
t
u
r
es f
r
o
m
th
e
in
p
u
t im
ag
e
[
2
0
]
.
I
n
th
e
cu
r
r
en
t
wo
r
k
,
a
p
r
etr
a
in
ed
VGG1
6
m
o
d
el
o
n
I
m
a
g
eNe
t
is
u
s
ed
as
th
e
s
ec
o
n
d
b
ac
k
b
o
n
e
n
etwo
r
k
.
T
h
e
m
ai
n
m
o
d
if
icat
io
n
m
ad
e
to
th
is
m
o
d
el
was
r
em
o
v
in
g
th
e
to
p
class
if
ier
lay
er
o
f
th
e
m
o
d
el,
allo
win
g
th
e
m
o
d
el
to
o
p
e
r
ate
as
a
f
ea
tu
r
e
ex
tr
ac
to
r
.
T
h
e
e
x
tr
ac
ted
f
ea
tu
r
e
m
a
p
s
ar
e
th
en
p
r
o
ce
s
s
ed
u
s
in
g
a
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
(
GAP
)
lay
er
,
f
o
llo
wed
b
y
b
atc
h
n
o
r
m
aliza
tio
n
(
B
N)
to
p
r
o
d
u
ce
a
co
m
p
ac
t
f
ea
tu
r
e
r
ep
r
esen
tatio
n
b
ef
o
r
e
th
e
f
ea
tu
r
e
f
u
s
io
n
s
tag
e.
T
h
e
m
ain
g
o
al
o
f
u
s
in
g
th
is
ar
ch
itectu
r
e
is
to
ca
p
tu
r
e
f
in
e
-
g
r
ain
ed
s
p
atial
p
atter
n
s
in
OC
T
im
ag
es,
th
u
s
co
m
p
lem
en
tin
g
th
e
h
ier
a
r
ch
ical
f
ea
tu
r
e
r
e
p
r
esen
tatio
n
s
lear
n
ed
b
y
E
f
f
icien
tNetV2
S in
th
e
o
th
er
b
r
an
c
h
o
f
th
e
f
r
a
m
ewo
r
k
[
2
0
]
.
3
.
8
.
F
e
a
t
ure
f
us
io
n
T
h
e
r
esu
ltin
g
te
n
s
o
r
s
ex
tr
ac
te
d
f
r
o
m
th
e
two
d
if
f
e
r
en
t
b
ac
k
b
o
n
es
wer
e
c
o
n
ca
ten
ated
.
E
ac
h
o
f
th
ese
two
f
ea
tu
r
e
v
ec
to
r
s
was
p
r
o
ce
s
s
ed
in
d
ep
en
d
en
tly
an
d
p
a
s
s
ed
th
r
o
u
g
h
t
h
e
d
if
f
er
en
t
la
y
er
s
o
f
th
ese
two
d
if
f
er
en
t
b
ase
m
o
d
els.
T
h
e
k
n
o
wled
g
e
o
b
tai
n
ed
f
r
o
m
th
es
e
b
r
an
c
h
es
was
co
m
b
in
ed
.
E
ac
h
o
f
t
h
ese
two
m
o
d
els
h
as
lea
r
n
ed
a
d
if
f
er
e
n
t
s
et
o
f
p
atter
n
s
f
r
o
m
t
h
e
s
am
e
in
p
u
t.
T
h
e
f
u
s
io
n
o
f
th
e
s
e
f
ea
tu
r
e
m
ap
s
b
y
co
n
ca
ten
atio
n
will
allo
w
th
e
n
ex
t
FC
lay
er
to
lear
n
f
r
o
m
t
h
ese
co
m
b
in
ed
f
ea
tu
r
e
r
e
p
r
es
en
tatio
n
s
to
ass
is
t
i
n
th
e
p
r
o
ce
s
s
o
f
th
e
f
in
al
d
ec
is
io
n
-
m
ak
in
g
b
y
th
e
class
if
ier
lay
er
.
3
.
9
.
Cla
s
s
if
ier
hea
d
T
h
e
f
in
al
class
if
icatio
n
d
ec
is
io
n
is
m
ad
e
b
y
th
e
class
if
icatio
n
h
ea
d
th
at
r
ec
eiv
es
th
e
f
u
s
ed
f
ea
tu
r
e
m
ap
s
f
r
o
m
th
e
two
b
ase
m
o
d
els
an
d
d
ec
id
es
to
wh
ich
cl
ass
th
e
in
p
u
t
R
etin
al
O
C
T
i
m
ag
e
b
elo
n
g
s
.
T
h
e
class
if
icatio
n
h
ea
d
co
n
s
is
ts
o
f
two
d
en
s
e
lay
er
s
with
1
0
2
4
an
d
5
1
2
u
n
its
with
s
wi
s
h
ac
ti
v
atio
n
.
E
ac
h
d
en
s
e
lay
er
is
f
o
llo
wed
b
y
a
d
r
o
p
o
u
t
lay
er
b
y
0
.
6
,
an
d
0
.
4
,
r
esp
ec
tiv
ely
to
p
r
ev
en
t
m
o
d
el
o
v
er
f
itti
n
g
.
T
h
e
f
in
a
l
m
u
lti
-
class
clas
s
if
icatio
n
was
m
ad
e
b
y
t
h
e
So
f
tMa
x
lay
er
.
T
h
is
lay
er
co
n
v
e
r
ts
th
e
o
u
tp
u
t
f
r
o
m
th
e
last
m
o
d
el
lay
er
in
to
a
p
r
o
b
ab
ilit
y
d
is
t
r
ib
u
tio
n
.
W
h
er
e
ea
ch
v
alu
e
in
th
is
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
r
ep
r
esen
ts
th
e
p
r
o
b
a
b
ilit
y
o
f
b
elo
n
g
in
g
th
e
in
p
u
t
R
etin
a
l
OC
T
im
ag
e
in
to
o
n
e
o
f
th
e
s
ev
en
class
es.
T
h
e
h
ig
h
er
v
alu
e
r
ep
r
esen
ts
th
e
tr
u
e
class
o
f
th
e
in
p
u
t im
ag
e
.
3
.
1
0
.
H
a
nd
lin
g
cla
s
s
im
ba
la
nce
T
o
h
an
d
le
th
e
class
im
b
alan
ce
in
th
e
(
7
-
class
an
d
Ker
m
an
y
2
0
1
8
)
d
atasets
,
class
weig
h
tin
g
was
ap
p
lied
d
u
r
in
g
tr
ai
n
in
g
o
f
t
h
e
f
r
am
ewo
r
k
.
E
v
er
y
class
r
ec
eiv
ed
a
weig
h
t
i
n
v
er
s
ely
p
r
o
p
o
r
tio
n
al
to
its
f
r
eq
u
e
n
cy
i
n
th
e
tr
ain
i
n
g
s
et.
B
y
en
s
u
r
in
g
th
at
u
n
d
er
r
e
p
r
esen
ted
class
es
co
n
tr
ib
u
te
p
r
o
p
o
r
tio
n
ately
to
th
e
lo
s
s
f
u
n
ctio
n
,
th
is
m
eth
o
d
k
ee
p
s
th
e
m
o
d
el
f
r
o
m
b
ein
g
s
k
ewe
d
i
n
f
av
o
u
r
o
f
m
ajo
r
ity
class
es.
T
h
e
o
v
er
all
wo
r
k
f
lo
w
o
f
th
e
p
r
o
p
o
s
ed
f
r
a
m
ewo
r
k
is
s
h
o
wn
in
Alg
o
r
ith
m
1
.
Alg
o
r
ith
m
1.
OC
T
Mo
d
Net
p
ip
elin
e
f
o
r
OC
T
im
ag
e
class
if
icatio
n
Input:
OCT image dataset
, number of classes
, learning rate
, dropout
.
Output:
Predicted class
̂
.
Steps:
1.
Preprocessing:
−
Resize images to
256
×
256
, normalize pixel values, and apply augmentation
2.
Feature Extraction:
−
Compute feature map
=
ℱ
(
)
using EfficientNetV2S backbone.
−
Ap
pl
y
SE
-
bl
oc
k
to
re
ca
li
br
at
e
,
Th
en
us
e
Gl
ob
al
Av
er
ag
e
Po
ol
in
g
(G
AP
)
an
d
Batch Normalization to obtain
−
Compute
feature
map
=
ℱ
(
)
us
in
g
VG
G1
6,
fo
ll
ow
e
d
by
GA
P
an
d
Ba
tc
h
Normalization to obtain
3.
Feature Fusion
−
Concatenate the extracted feature vectors:
=
[
;
]
.
4.
Classification:
−
Pass
through
the
Dense(1024,
swish)
+
Dropout(0.6),
then
Dense(512
,
swish) + Dropout(0.4).
−
Apply SoftMax to produce class probabilities and assign the final
label.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
43
,
No
.
2
,
Au
g
u
s
t
20
26
:
495
-
5
0
6
502
5.
Training:
−
Tr
ai
n
th
e
mo
de
l
us
in
g
c
at
eg
or
ic
al
cr
os
s
-
en
tr
op
y
lo
ss
,
cl
as
s
we
ig
ht
in
g,
an
d
update weights using the Nadam optimizer.
−
Apply Cosine Decay Restart for learning rate.
−
Train for
50
epochs with batch size
16
.
6.
Evaluation:
Validate on
, t
est on
, save the trained model
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
4
.
1
.
E
x
perim
ent
a
l
s
et
up
T
h
e
m
o
d
el
was
tr
ain
ed
o
n
a
s
y
s
tem
with
an
I
n
tel
C
o
r
e
i9
-
1
2
9
0
0
K
C
PU,
6
4
GB
R
AM
,
an
d
an
NVI
DI
A
R
T
X
3
0
9
0
(
2
4
GB
VR
AM
)
,
wh
ich
was
u
s
ed
f
o
r
all
o
f
th
e
tr
ai
n
in
g
a
n
d
v
alid
atio
n
p
r
o
ce
s
s
es.
Py
th
o
n
3
.
8
was
u
s
ed
f
o
r
tr
ain
in
g
,
w
h
ile
T
en
s
o
r
Flo
w
2
.
9
with
th
e
Ker
as
b
ac
k
en
d
was
u
s
ed
f
o
r
th
e
d
ee
p
lear
n
in
g
f
r
am
ewo
r
k
im
p
lem
e
n
tatio
n
.
T
h
e
d
ataset
was
p
r
o
ce
s
s
ed
u
s
in
g
Op
e
n
C
V
an
d
N
u
m
Py
,
an
d
p
er
f
o
r
m
a
n
ce
was
v
is
u
alis
ed
u
s
in
g
Ma
tp
lo
tlib
.
Pix
el
in
ten
s
ities
wer
e
n
o
r
m
alis
ed
with
in
th
e
in
ter
v
al
[
0
,
1
]
,
a
n
d
all
OC
T
p
ictu
r
es
wer
e
co
n
v
e
r
ted
to
2
5
6
×
2
5
6
×
3
R
GB
f
o
r
m
at.
T
h
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
was
tr
ain
e
d
f
o
r
5
0
ep
o
ch
s
,
b
atch
s
ize
eq
u
al
to
3
2
,
u
s
in
g
th
e
Nad
am
Op
tim
izatio
n
tech
n
iq
u
e
with
a
n
in
itial lea
r
n
in
g
r
ate
o
f
0
.
0
0
0
1
.
4
.
2
.
E
v
a
lua
t
i
o
n
m
et
rics
T
o
ev
alu
ate
th
e
p
e
r
f
o
r
m
an
ce
o
f
OC
T
Mo
d
Net
f
o
r
OC
T
i
m
ag
e
class
if
icatio
n
,
s
ev
er
al
p
er
f
o
r
m
a
n
ce
m
etr
ics
wer
e
u
s
ed
to
ass
es
s
b
o
th
o
v
er
all
an
d
class
-
wis
e
class
if
icatio
n
b
eh
av
io
r
.
Fo
r
a
m
u
lti
-
class
class
if
icatio
n
p
r
o
b
lem
with
N
s
am
p
les
an
d
C
class
es,
s
tan
d
ar
d
class
if
icatio
n
m
etr
ics
wer
e
ad
o
p
ted
[
21
],
[
2
2
]
in
(
2
)
-
(
5
)
e
x
p
r
ess
th
e
m
etr
ics
u
s
ed
f
o
r
m
o
d
el
e
v
alu
atio
n
:
T
h
e
o
v
e
r
all
class
if
icatio
n
ac
cu
r
ac
y
is
d
ef
in
ed
as:
=
∑
1
(
̂
=
)
=
1
(
2
)
Pre
cisi
o
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
wer
e
ca
lcu
lated
p
er
class
as f
o
llo
ws [
2
1
]
,
[
2
3
]
.
=
+
(
3
)
=
+
(
4
)
1
−
=
2
×
×
+
(
5
)
Fo
r
m
u
lti
-
class
ev
alu
atio
n
,
m
ac
r
o
-
av
e
r
ag
in
g
was
ad
o
p
te
d
to
en
s
u
r
e
eq
u
al
co
n
tr
ib
u
t
io
n
o
f
ea
ch
class
ir
r
esp
ec
tiv
e
o
f
class
im
b
alan
c
e
[
2
3
]
.
Fo
r
f
u
r
th
e
r
ass
ess
in
g
d
is
cr
im
in
ativ
e
ca
p
ab
ilit
y
,
th
e
ar
ea
u
n
d
er
th
e
r
ec
ei
v
er
o
p
er
atin
g
c
h
ar
ac
ter
is
ti
c
cu
r
v
e
(
AUC
-
R
OC
)
was
co
m
p
u
ted
u
s
in
g
t
h
e
On
e
-
vs
-
R
est
s
t
r
ateg
y
[
2
4
]
,
[
2
5
]
.
Fo
r
ea
ch
cla
s
s
c,
a
b
in
ar
y
R
OC
cu
r
v
e
was
co
n
s
tr
u
cted
b
y
t
r
e
atin
g
class
c
as
p
o
s
itiv
e
an
d
all
o
th
er
class
es
as
n
eg
ativ
e.
T
h
e
class
-
wis
e
ar
ea
u
n
d
er
t
h
e
cu
r
v
e
(
AUC
)
v
alu
es
wer
e
th
en
ag
g
r
e
g
ate
d
u
s
in
g
m
ac
r
o
av
er
a
g
in
g
(
1
3
)
an
d
(
1
4
)
.
=
∫
(
)
(
)
1
0
=
1
∑
=
1
All
ev
alu
atio
n
m
etr
ics
wer
e
c
o
m
p
u
ted
o
n
th
e
in
d
ep
en
d
en
t
t
est
s
et
th
at
h
ad
n
o
t
b
ee
n
u
s
ed
d
u
r
in
g
tr
ain
in
g
th
e
m
o
d
el
o
r
h
y
p
er
p
ar
a
m
eter
tu
n
i
n
g
.
4
.
3
.
Resul
t
s
OC
T
Mo
d
Net
was
ev
alu
ated
o
n
an
i
n
d
ep
e
n
d
en
t
test
s
et
co
n
s
is
tin
g
o
f
s
ev
en
d
if
f
e
r
en
t
r
etin
al
d
is
o
r
d
er
s
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
e
l a
ch
iev
ed
an
o
v
er
all
test
ac
cu
r
ac
y
o
f
9
9
% a
n
d
a
m
ac
r
o
-
av
er
ag
ed
AUC o
f
9
9
%,
o
u
tp
er
f
o
r
m
in
g
th
e
ev
al
u
ated
s
in
g
le
-
b
ac
k
b
o
n
e
m
o
d
els
(
R
E
SNet5
0
,
I
n
ce
p
tio
n
V
3
,
De
n
s
eNe
t1
2
1
,
VGG1
6
,
an
d
E
f
f
icien
tNetV2
S)
u
n
d
e
r
id
en
ti
ca
l tr
ain
in
g
an
d
v
alid
atio
n
co
n
d
itio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
OC
T
Mo
d
N
et:
a
d
ee
p
lea
r
n
in
g
-
b
a
s
ed
fr
a
mewo
r
k
f
o
r
o
p
tica
l
co
h
eren
ce
… (
S
a
ja
A
ta
a
lla
h
Mu
h
a
mme
d
)
503
C
l
as
s
-
w
is
e
p
r
e
c
i
s
i
o
n
,
r
e
c
a
l
l
,
an
d
F
1
-
s
c
o
r
e
s
(
s
e
e
F
i
g
u
r
e
3
)
s
h
o
w
c
o
n
s
is
t
e
n
t
p
e
r
f
o
r
m
a
n
c
e
a
c
r
o
s
s
a
ll
c
a
t
e
g
o
r
i
e
s
,
i
n
d
i
c
a
ti
n
g
b
a
l
a
n
c
ed
d
i
s
c
r
i
m
i
n
a
ti
v
e
c
a
p
a
b
i
li
t
y
d
e
s
p
i
t
e
m
o
d
e
r
a
te
c
l
as
s
i
m
b
al
an
c
e
.
T
h
e
c
o
n
f
u
s
i
o
n
m
a
t
r
i
x
,
F
i
g
u
r
e
3
,
f
u
r
t
h
e
r
c
o
n
f
i
r
m
e
d
l
i
m
i
te
d
i
n
t
e
r
-
c
la
s
s
m
is
c
l
a
s
s
i
f
ic
a
t
i
o
n
,
es
p
e
c
i
al
l
y
a
m
o
n
g
v
i
s
u
a
l
l
y
s
i
m
i
la
r
c
l
a
s
s
es
.
M
u
lt
i
-
cl
a
s
s
R
OC
a
n
al
y
s
is
s
h
o
w
e
d
h
i
g
h
s
p
e
r
a
b
il
i
t
y
a
c
r
o
s
s
al
l
c
l
as
s
e
s
,
w
it
h
m
a
cr
o
-
a
v
e
r
a
g
e
d
A
UC
r
e
f
l
e
c
t
i
n
g
r
o
b
u
s
t
g
e
n
e
r
a
li
z
a
tio
n
w
i
t
h
i
n
t
h
e
t
es
t
e
d
d
a
t
as
e
t
.
O
n
t
h
e
o
t
h
e
r
h
a
n
d
,
a
s
m
e
n
t
i
o
n
e
d
e
a
r
l
i
e
r
,
t
o
i
m
p
r
o
v
e
g
e
n
e
r
a
l
i
z
a
ti
o
n
,
t
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l
w
a
s
t
r
a
i
n
e
d
o
n
K
e
r
m
a
n
y
2
0
1
8
a
n
d
r
e
g
i
s
t
e
r
e
d
an
o
v
e
r
a
l
l
a
c
c
u
r
a
c
y
o
f
9
9
%
,
a
n
d
m
a
c
r
o
-
a
v
e
r
a
g
e
d
p
r
e
c
i
s
i
o
n
,
r
e
c
a
l
l
,
a
n
d
F
1
-
s
c
o
r
e
o
f
0
.
9
9
,
p
r
o
v
i
n
g
b
a
l
a
n
c
e
d
c
l
a
s
s
-
w
i
s
e
d
i
s
c
r
i
m
i
n
a
t
i
o
n
Fig
u
r
e
4.
O
n
t
h
e
OC
T
-
C
8
d
a
t
as
e
t
,
w
h
i
c
h
c
o
n
t
a
i
n
s
a
m
o
r
e
c
o
m
p
l
e
x
8
-
cla
s
s
d
i
s
t
r
i
b
u
t
i
o
n
,
t
h
e
m
o
d
e
l
a
c
h
ie
v
e
d
9
8
%
t
e
s
t
a
cc
u
r
a
c
y
o
n
t
h
e
t
e
s
t
s
e
t
,
wit
h
a
m
a
c
r
o
-
a
v
e
r
a
g
e
d
F
1
-
s
c
o
r
e
o
f
0
.
9
8
.
m
i
n
o
r
v
a
r
i
a
t
i
o
n
s
o
b
s
e
r
v
e
d
a
m
o
n
g
v
i
s
u
a
l
l
y
o
v
e
r
l
a
p
p
i
n
g
c
l
a
s
s
e
s
,
w
h
i
l
e
t
h
e
o
v
e
r
a
l
l
p
e
r
f
o
r
m
a
n
c
e
r
e
m
a
i
n
e
d
s
t
a
b
l
e
,
as
c
a
n
b
e
n
o
t
i
c
e
d
i
n
F
i
g
u
r
e
5
.
T
h
e
s
e
f
i
n
d
i
n
g
s
s
u
g
g
e
s
t
t
h
at
t
h
e
i
n
t
e
g
r
a
t
i
o
n
o
f
c
o
m
p
l
e
m
e
n
t
a
r
y
f
e
at
u
r
e
e
x
t
r
a
c
t
o
r
s
a
n
d
a
d
a
p
t
i
v
e
o
p
t
im
i
z
a
t
i
o
n
i
m
p
r
o
v
es
c
l
a
s
s
i
f
ic
a
t
i
o
n
s
t
a
b
il
i
t
y
r
e
la
t
i
v
e
t
o
s
i
n
g
l
e
-
s
t
r
e
a
m
a
r
c
h
i
t
e
ct
u
r
e
s
,
a
n
d
t
h
a
t
t
h
e
d
u
a
l
-
b
a
c
k
b
o
n
e
a
r
ch
i
t
e
c
t
u
r
e
m
ai
n
t
a
i
n
s
c
o
n
s
i
s
t
e
n
t
d
is
c
r
i
m
i
n
at
i
v
e
c
a
p
a
b
i
l
i
t
y
a
c
r
o
s
s
d
at
a
s
e
ts
w
it
h
d
i
f
f
e
r
e
n
t
c
l
as
s
s
t
r
u
c
t
u
r
es
a
n
d
d
i
s
t
r
i
b
u
t
i
o
n
s
.
Fig
u
r
e
3.
OC
T
Mo
d
Net
co
n
f
u
s
io
n
m
atr
ix
a
n
d
class
if
icatio
n
r
ep
o
r
t
o
n
7
-
class
d
ataset
Fig
u
r
e
4.
C
o
n
f
u
s
io
n
m
atr
i
x
an
d
class
if
icatio
n
r
ep
o
r
t a
cr
o
s
s
Ker
m
an
y
2
0
1
8
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
43
,
No
.
2
,
Au
g
u
s
t
20
26
:
495
-
5
0
6
504
Fig
u
r
e
5.
Mo
d
el'
s
R
O
C
cu
r
v
e
an
d
co
n
f
u
s
io
n
m
atr
ix
ac
r
o
s
s
th
e
OC
T
-
C
8
d
ataset
4
.
4
.
Co
m
pa
riso
n wit
h
ba
s
eline
m
o
dels
T
h
e
co
m
p
a
r
ativ
e
an
aly
s
is
ass
ess
es
O
C
T
Mo
d
Net
b
ased
o
n
s
tate
-
of
-
th
e
-
ar
t,
s
in
g
le
b
ac
k
b
o
n
e
d
ee
p
lear
n
in
g
ar
c
h
itectu
r
es
s
u
ch
as
R
es
Net5
0
,
I
n
ce
p
tio
n
V3
,
Den
s
eNe
t1
2
1
,
VGG1
6
,
an
d
E
f
f
icien
tNetV2
S
(
s
ee
T
ab
le
4
)
.
T
h
e
p
e
r
f
o
r
m
an
ce
s
in
d
ica
te
th
at
OC
T
Mo
d
Net’
s
test
ac
cu
r
ac
y
a
n
d
m
ac
r
o
-
a
v
er
ag
e
d
AUC
o
u
tp
er
f
o
r
m
th
e
p
r
ev
io
u
s
m
eth
o
d
s
,
d
ep
ic
tin
g
th
e
b
en
ef
its
o
f
o
p
tim
izin
g
m
u
lti
-
m
o
d
el
f
ea
tu
r
e
e
x
tr
a
ctio
n
,
f
u
s
io
n
,
a
n
d
tr
ain
in
g
s
tr
ateg
y
.
T
h
is
u
n
d
er
li
n
es
its
b
etter
clas
s
if
icatio
n
p
er
f
o
r
m
a
n
ce
co
m
p
ar
e
d
to
tr
ad
itio
n
al
s
in
g
le
-
m
o
d
el
ar
ch
itectu
r
es.
I
n
co
m
p
ar
is
o
n
,
th
e
s
tr
o
n
g
est
s
in
g
le
-
b
ac
k
b
o
n
e
m
o
d
el
(
E
f
f
icien
tNetV2
S)
ac
h
iev
ed
9
6
.
0
5
%
test
ac
cu
r
ac
y
an
d
9
7
%
AUC.
I
m
p
r
o
v
em
en
ts
wer
e
also
o
b
s
er
v
ed
ac
r
o
s
s
m
ac
r
o
-
a
v
er
ag
e
d
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
m
etr
ics,
in
d
icatin
g
b
alan
ce
d
class
-
wis
e
p
er
f
o
r
m
an
ce
r
a
th
er
th
an
g
ain
s
lim
ited
to
s
p
ec
if
ic
ca
teg
o
r
ies.
T
ab
le
4.
C
o
m
p
a
r
is
o
n
with
b
as
elin
e
m
o
d
els
M
o
d
e
l
Te
st
A
c
c
.
(
%
)
P
r
e
c
i
s
i
o
n
(
%
)
R
e
c
a
l
l
(
%
)
F1
-
S
c
o
r
e
(
%
)
AU
C
(
%
)
R
e
s
N
e
t
5
0
9
2
.
8
5
9
1
.
6
0
9
2
.
3
0
9
1
.
9
5
9
4
.
1
0
I
n
c
e
p
t
i
o
n
V
3
9
4
.
1
2
9
3
.
8
0
9
3
.
9
5
9
3
.
8
7
9
5
.
2
5
D
e
n
s
e
N
e
t
1
2
1
9
5
.
3
7
9
5
.
0
0
9
5
.
1
5
9
5
.
0
7
9
6
.
4
5
V
GG
1
6
9
3
.
2
5
9
2
.
9
0
9
3
.
1
0
9
3
.
0
0
9
4
.
3
5
E
f
f
i
c
i
e
n
t
N
e
t
V
2
S
9
6
.
0
5
9
5
.
8
0
9
6
.
0
0
9
5
.
9
0
9
7
.
0
0
O
C
T
M
o
d
N
e
t
9
9
.0
9
9
.0
9
9
.0
9
9
.0
9
9
.0
4
.
5
.
Abla
t
io
n
s
t
ud
y
T
ab
le
5
r
ev
ea
ls
th
at
th
e
f
r
am
e
wo
r
k
'
s
f
u
ll
co
n
f
ig
u
r
atio
n
h
as
r
eg
is
ter
ed
a
p
ea
k
p
er
f
o
r
m
a
n
ce
(
9
9
%)
in
ter
m
s
o
f
test
ac
cu
r
ac
y
a
n
d
0
.
9
9
AUC.
E
x
clu
d
in
g
in
d
iv
i
d
u
a
l
co
m
p
o
n
en
ts
h
as
d
eg
r
a
d
ed
th
e
p
er
f
o
r
m
an
ce
a
n
d
ac
cu
r
ac
y
o
f
th
e
m
o
d
el.
T
h
e
s
y
n
er
g
y
b
etwe
en
co
m
p
lem
en
ta
r
y
f
ea
tu
r
e
in
teg
r
atio
n
an
d
th
e
o
p
tim
ized
lear
n
in
g
s
ch
ed
u
les is
ess
en
tial f
o
r
ac
h
iev
in
g
r
e
f
in
ed
,
h
ig
h
-
le
v
el
class
if
icatio
n
r
esu
lts
.
T
ab
le
5.
Ab
latio
n
s
tu
d
y
M
o
d
e
l
c
o
n
f
i
g
u
r
a
t
i
o
n
Ef
f
i
c
i
e
n
t
N
e
t
V
2
S
V
G
G
1
6
F
e
a
t
u
r
e
f
u
si
o
n
C
o
s
i
n
e
d
e
c
a
y
r
e
st
a
r
t
A
d
a
p
t
i
v
e
f
e
a
t
u
r
e
p
r
o
c
e
ss
i
n
g
Te
st
A
c
c
.
(
%)
A
U
C
(
%)
O
n
l
y
V
G
G
1
6
✗
✓
✓
✓
✓
9
1
.
8
4
9
4
.
2
O
n
l
y
Ef
f
i
c
i
e
n
t
N
e
t
V
2
S
✓
✗
✓
✓
✓
9
3
.
6
2
9
5
.
8
W
i
t
h
o
u
t
f
e
a
t
u
r
e
f
u
si
o
n
✓
✓
✗
✓
✓
9
4
.
2
7
9
6
.
1
W
i
t
h
o
u
t
c
o
si
n
e
d
e
c
a
y
✓
✓
✓
✗
✓
9
5
.
3
1
9
6
.
9
W
i
t
h
o
u
t
a
d
a
p
t
i
v
e
f
e
a
t
u
r
e
p
r
o
c
e
ss
i
n
g
✓
✓
✓
✓
✗
9
6
.
4
2
9
7
.
5
F
u
l
l
m
o
d
e
l
O
C
TM
o
d
N
e
t
✓
✓
✓
✓
✓
9
8
.
8
99
Evaluation Warning : The document was created with Spire.PDF for Python.