I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
2
,
A
pr
il
20
25
, pp.
1410
~
1419
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
2
.pp
1410
-
1419
1410
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
S
q
u
e
e
z
e
-
e
x
c
i
t
at
i
on
h
al
f
U
-
N
e
t
an
d
sy
n
t
h
e
t
i
c
m
i
n
or
i
t
y
ove
r
sam
p
l
i
n
g t
e
c
h
n
i
q
u
e
ove
r
sam
p
l
i
n
g
f
or
p
ap
i
l
l
e
d
e
m
a
i
m
age
c
l
ass
i
f
i
c
at
i
on
Wi
h
ar
t
o
,
A
n
gga E
xc
a P
r
ad
ip
t
a S
yai
f
u
d
d
in
D
e
pa
r
t
m
e
nt
of
I
nf
or
m
a
t
i
c
s
, F
a
c
ul
t
y of
I
n
f
or
m
a
t
i
on
T
e
c
hnol
ogy a
nd D
a
t
a
S
c
i
e
nc
e
, U
ni
ve
r
s
i
t
a
s
S
e
be
l
a
s
M
a
r
e
t
, S
ur
a
ka
r
t
a
, I
ndone
s
i
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
M
a
r
18
,
2024
R
e
vi
s
e
d
O
c
t
30
,
2024
A
c
c
e
pt
e
d
N
ov
14
,
2024
The
emergence
of
various
convolut
ional
neural
networks
(CNN)
architectu
res
indicates
progress
in
the
computer
vision
field.
However
,
most
of
the
architectures
have
large
parameters,
which
tends
to
increa
se
the
computat
ional
cost
of
the
training
process.
Additi
onaly,
imbalance
d
data
sources
are
often
encountered
,
causing
the
model
to
overfit.
The
aim
of
this
study
is
to
evaluate
a
new
method
to
classify
retinal
fundus
image
s
from
imbalance
d
data
into
the
corresp
onding
classes
by
using
fewer
paramete
rs
than
t
he
previous
method.
To
achieve
this,
squeeze
-
excitati
on
h
alf
U
-
Net
(
SEHUNE
T
)
architectu
re,
a
modifi
cation
of
h
alf
U
-
N
et
with
squeeze
-
excite
process
to
provide
attention
mechanism
on
each
feature
maps
channel
of
the
model,
in
combination
with
synthetic
minority
oversampling
tec
hnique
(
SMOTE
)
is
proposed.
The
test
accuracy
of
SEHUNET
is
98.52%
wit
h
area
under
the
curve
of
receiver
operation
ch
aracteristic
(AUROC)
of
0.99
9.
This
result
outperforms
the
previous
study
that
used
CNN
with
Ba
yesian
optimization
,
achieving
accuracy
of
95.89%
and
AUROC
of
0.992.
SEHUNE
T
is
also
able
to
compet
e
with
the
transf
er
learn
ing
methods
used
in
previous
research
such
as
InceptionV3
with
96.35%
accuracy,
visual
geometry
group
(
VGG
)
with
96.8%,
and
ResNet
with
98.63%.
This
performance
can
be
achieved
by
SE
HUNET
with
only
0.268
million
parameters
compared
to
the
architecture
parameters
used
in
pr
evious
research ranging fro
m 11 mi
llion t
o 33 mil
lion.
K
e
y
w
o
r
d
s
:
C
om
put
a
ti
ona
l
e
f
f
ic
ie
nc
y
P
a
pi
ll
e
de
m
a
R
e
ti
na
l
S
que
e
z
e
-
e
xc
it
a
ti
on
h
a
lf
U
-
N
e
t
S
M
O
T
E
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
W
ih
a
r
to
D
e
pa
r
tm
e
nt
of
I
nf
or
m
a
ti
c
s
, F
a
c
ul
ty
of
I
nf
o
r
m
a
ti
on T
e
c
hnol
ogy
&
D
a
ta
S
c
ie
nc
e
,
U
ni
ve
r
s
it
a
s
S
e
be
la
s
M
a
r
e
t
St
.
I
r
S
ut
a
m
i
N
o.
36, Ke
nt
in
ga
n, J
e
br
e
s
, S
ur
a
ka
r
ta
, C
e
nt
r
a
l
J
a
va
, I
ndone
s
ia
E
m
a
il
:
w
ih
a
r
to
@
s
ta
f
f
.uns
.a
c
.i
d
1.
I
N
T
R
O
D
U
C
T
I
O
N
P
a
pi
ll
e
de
m
a
is
a
s
e
r
io
us
e
ye
c
ondi
ti
on
c
ha
r
a
c
te
r
iz
e
d
by
in
c
r
e
a
s
e
d
in
tr
a
c
r
a
ni
a
l
pr
e
s
s
ur
e
th
a
t
c
a
n
c
a
us
e
pe
r
m
a
ne
nt
da
m
a
g
e
to
th
e
opt
ic
ne
r
ve
if
not
de
te
c
te
d
a
nd
tr
e
a
te
d
pr
om
pt
ly
[
1]
.
T
he
de
te
c
ti
on
of
pa
pi
ll
e
de
m
a
r
e
qui
r
e
s
a
vi
s
ua
l
a
na
ly
s
is
of
th
e
f
undus
im
a
ge
s
ta
ke
n
f
r
om
th
e
pa
ti
e
nt
’
s
e
ye
.
T
he
de
te
r
m
in
in
g
a
tt
r
ib
ut
e
s
f
or
th
e
a
na
ly
s
is
a
r
e
m
a
in
ly
c
ons
i
s
te
d
of
th
e
c
la
r
it
y
of
th
e
opt
ic
a
l
di
s
c
a
nd
th
e
bl
ood
ve
s
s
e
l
s
a
r
ound
th
e
c
e
nt
e
r
of
th
e
e
ye
.
H
ow
e
ve
r
,
th
is
de
te
c
ti
on
pr
ove
s
to
be
r
a
th
e
r
di
f
f
ic
ul
t
be
c
a
us
e
s
im
il
a
r
s
ym
pt
om
s
c
a
n
be
s
e
e
n i
n
ot
he
r
c
ondi
ti
ons
, s
u
c
h a
s
p
s
e
udopa
pi
ll
e
de
m
a
w
hi
c
h r
e
a
l
ly
m
im
ic
s
t
he
r
e
a
l
di
s
or
de
r
[
2]
.
T
he
f
unda
m
e
nt
a
l
di
f
f
e
r
e
nc
e
be
tw
e
e
n
pa
pi
ll
e
de
m
a
a
nd
p
s
e
ud
opa
pi
ll
e
de
m
a
li
e
s
in
th
e
c
a
us
e
s
a
nd
e
f
f
e
c
ts
of
bot
h.
P
a
pi
ll
e
de
m
a
is
c
a
us
e
d
by
th
e
pr
e
s
e
nc
e
of
pr
e
s
s
ur
e
f
r
om
w
it
hi
n
th
e
e
ye
th
a
t
c
a
u
s
e
s
th
e
b
a
c
k
of
th
e
e
ye
to
be
pr
e
s
s
e
d
in
to
th
e
br
a
in
,
c
a
us
in
g
th
e
p
a
ti
e
nt
to
p
os
s
ib
ly
e
xp
e
r
ie
nc
e
bl
ur
r
e
d
a
nd
double
vi
s
io
n
a
c
c
om
pa
ni
e
d
by
he
a
d
a
c
he
s
or
na
u
s
e
a
.
O
n
th
e
ot
he
r
ha
nd,
in
p
s
e
udopa
pi
ll
e
de
m
a
,
th
e
r
e
is
no
pr
e
s
s
ur
e
a
s
in
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sque
e
z
e
-
e
x
c
it
at
io
n
h
al
f
U
-
N
e
t
and s
y
nt
he
ti
c
m
in
or
it
y
o
v
e
r
s
am
p
li
ng t
e
c
hni
que
ov
e
r
s
am
pl
in
g
…
(
W
ih
ar
to
)
1411
c
a
s
e
of
p
a
pi
ll
e
de
m
a
,
but
vi
s
ua
ll
y,
th
e
f
undu
s
im
a
ge
pr
ovi
de
d
h
a
s
s
om
e
s
im
il
a
r
it
ie
s
.
I
n
th
e
f
undus
im
a
ge
of
a
pa
pi
ll
e
de
m
a
pa
ti
e
nt
,
th
e
opt
ic
di
s
c
unde
r
goe
s
e
le
va
ti
on
a
nd
s
w
e
ll
in
g,
c
a
us
in
g
th
e
e
dge
of
th
e
di
s
c
to
be
c
om
e
bl
ur
r
e
d
a
c
c
om
pa
ni
e
d
by
di
la
ti
on
of
th
e
bl
ood
ve
s
s
e
ls
a
r
ound
th
e
di
s
c
,
w
hi
c
h
is
c
ha
r
a
c
te
r
iz
e
d
by
a
n
e
f
f
e
c
t
s
uc
h
a
s
be
in
g
pr
e
s
s
e
d
or
pa
le
r
th
a
n
ot
he
r
bl
ood
ve
s
s
e
ls
.
O
n
th
e
ot
he
r
ha
nd,
ps
e
udopa
pi
ll
e
de
m
a
ha
s
c
ha
r
a
c
te
r
is
ti
c
s
th
a
t
r
e
s
e
m
bl
e
th
e
bl
ur
r
in
g
of
th
e
di
s
c
e
dge
,
but
te
nds
to
be
w
it
hout
a
c
c
om
pa
ni
e
d
di
la
ti
on
of
bl
ood
ve
s
s
e
l
s
.
T
hi
s
c
a
n
oc
c
ur
due
to
th
e
f
a
c
to
r
of
th
e
a
ppe
a
r
a
n
c
e
of
dr
us
e
n,
dot
s
th
a
t
a
pp
e
a
r
a
r
ound
th
e
di
s
c
due
t
o a
na
to
m
ic
a
l
di
f
f
e
r
e
nt
ia
ti
on or
a
gi
ng
[
3]
.
A
s
a
n
e
f
f
or
t
to
he
lp
doc
to
r
s
di
a
gnos
e
pa
pi
ll
e
de
m
a
by
ob
s
e
r
v
in
g
th
e
f
undus
im
a
g
e
s
f
a
s
te
r
s
o
th
a
t
a
c
ti
on
f
or
th
e
pa
ti
e
nt
s
c
a
n
be
done
im
m
e
di
a
te
ly
,
m
a
ny
a
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
r
e
s
e
a
r
c
he
s
ha
ve
b
e
e
n
done
to
de
ve
lo
p
m
e
th
ods
a
nd
to
ol
s
th
a
t
c
a
n
s
uppor
t
doc
to
r
s
in
c
la
s
s
if
yi
ng
pa
pi
ll
e
de
m
a
.
F
or
e
xa
m
pl
e
,
r
e
s
e
a
r
c
h
in
[
4]
,
[
5]
w
he
r
e
bot
h
of
th
e
m
us
e
m
a
nua
l
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
th
a
t
is
th
e
n
pr
oc
e
s
s
e
d
by
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
(
S
V
M
)
[
6]
.
T
he
la
te
s
t
r
e
s
e
a
r
c
h
w
it
h
th
e
s
a
m
e
pur
pos
e
a
ls
o
done
A
hn
e
t
al
.
[
7]
th
a
t
us
e
s
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
)
w
it
h
B
a
ye
s
ia
n
opt
im
iz
a
ti
on
ut
il
iz
in
g
lo
c
a
l
a
ve
r
a
ge
c
ol
or
s
ubt
r
a
c
ti
on
f
or
t
he
pr
e
pr
oc
e
s
s
in
g s
te
p
.
I
n
a
ddi
ti
on
to
th
e
s
ta
nda
r
d
C
N
N
a
r
c
hi
te
c
tu
r
e
[
8]
,
th
e
de
r
iv
a
t
iv
e
a
r
c
hi
te
c
tu
r
e
of
C
N
N
,
e
s
pe
c
ia
ll
y
U
-
N
e
t
[
9]
,
is
a
ls
o
u
s
e
d
to
c
la
s
s
if
y
pa
pi
ll
e
de
m
a
im
a
g
e
s
.
F
or
in
s
ta
nc
e
,
a
s
tu
dy
c
onduc
t
e
d
by
M
il
e
a
e
t
al
.
[
10]
e
m
pl
oye
d t
he
one
ve
r
s
us
r
e
s
t
(
O
V
R
)
t
e
c
hni
que
on t
he
U
-
N
e
t
a
nd D
e
ns
e
N
e
t
a
r
c
hi
te
c
tu
r
e
. T
he
r
e
i
s
a
l
s
o a
s
tu
dy
c
onduc
te
d by S
a
ba
e
t
al
.
[
11
]
th
a
t
us
e
s
a
c
om
bi
na
ti
on o
f
U
-
N
e
t
a
nd D
e
ns
e
N
e
t
a
c
c
om
p
a
ni
e
d by a
G
a
bor
f
il
te
r
.
F
ur
th
e
r
m
or
e
,
th
e
r
e
a
r
e
r
e
s
e
a
r
c
h
c
onduc
te
d
in
[
12]
,
[
13]
w
ho
us
e
d
th
e
s
a
m
e
da
ta
s
e
t
f
or
pa
pi
ll
e
de
m
a
c
la
s
s
if
ic
a
ti
on,
w
he
r
e
th
e
da
ta
us
e
d
c
a
m
e
f
r
om
by
K
im
’
s
E
ye
H
os
pi
ta
l
[
7]
.
K
okul
u
a
nd
G
öke
r
[
12]
us
e
s
M
obi
le
N
e
tV2
w
it
h
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
a
nd
3D
box
f
il
te
r
i
ng
te
c
hni
que
s
to
pr
e
pr
oc
e
s
s
th
e
in
put
im
a
ge
s
.
M
e
a
nw
hi
le
,
A
l
-
A
z
z
a
w
i
e
t
al
.
[
13]
us
e
s
R
e
s
N
e
t
-
50
in
c
om
bi
na
ti
on
w
it
h
s
e
gm
e
nt
a
ti
on
opt
im
iz
a
ti
on
to
p
r
e
pa
r
e
th
e
i
nput
s
.
T
h
e
us
e
of
t
he
U
-
N
e
t
a
r
c
h
it
e
c
tu
r
e
i
s
not
l
im
it
e
d
to
pa
pi
ll
e
d
e
m
a
c
l
a
s
s
if
i
c
a
ti
o
n a
lo
ne
. A
be
da
ll
a
e
t
a
l.
[
14]
us
e
s
t
he
E
ns
4B
-
U
N
e
t
a
r
c
hi
te
c
tu
r
e
t
ha
t
c
om
bi
ne
s
f
our
t
r
a
ns
f
e
r
l
e
a
r
ni
ng mode
ls
a
s
pa
r
t
of
t
he
e
nc
ode
r
f
r
om
t
he
U
-
N
e
t
a
r
c
hi
te
c
tu
r
e
to
pe
r
f
or
m
th
or
a
x
im
a
ge
s
e
gm
e
nt
a
ti
on.
T
hi
s
s
how
s
th
e
pot
e
nt
ia
l
of
f
e
r
e
d
by
U
-
N
e
t
in
th
e
bi
om
e
di
c
a
l
f
ie
ld
.
H
ow
e
ve
r
,
th
e
m
a
in
c
ha
ll
e
nge
of
us
in
g
C
N
N
de
r
iv
a
ti
ve
a
r
c
hi
te
c
tu
r
e
s
s
uc
h
a
s
U
-
N
e
t
a
nd
tr
a
ns
f
e
r
le
a
r
ni
ng
m
ode
ls
i
s
th
e
n
e
e
d
f
or
la
r
ge
c
om
put
a
ti
ona
l
c
o
s
ts
due
to
th
e
la
r
ge
num
be
r
of
pa
r
a
m
e
te
r
s
.
T
he
s
ta
nda
r
d
U
-
N
e
t
ha
s
m
or
e
th
a
n
30
m
il
li
on
pa
r
a
m
e
te
r
s
[
15
]
,
w
hi
l
e
tr
a
ns
f
e
r
le
a
r
ni
ng
m
ode
ls
s
uc
h
a
s
V
G
G
[
16]
,
R
e
s
N
e
t
[
17]
,
a
nd
I
nc
e
pt
io
nV
3
[
18]
ha
ve
m
or
e
th
a
n
24
m
il
li
on
pa
r
a
m
e
te
r
s
[
7]
.
T
he
hi
gh
num
be
r
s
s
ig
na
l
th
e
ne
e
d f
or
pa
r
a
m
e
te
r
e
f
f
ic
ie
nc
y s
o t
ha
t
c
om
put
a
ti
ona
l
c
os
ts
a
nd
m
ode
l
s
iz
e
c
a
n be
m
in
im
iz
e
d. I
n a
ddi
ti
on t
o t
he
num
be
r
of
pa
r
a
m
e
te
r
s
,
a
not
he
r
is
s
ue
th
a
t
a
r
is
e
s
i
s
th
e
pr
oba
bi
li
ty
of
im
ba
la
nc
e
d
da
ta
,
w
hi
c
h
m
a
ke
s
th
e
tr
a
in
in
g
pr
oc
e
s
s
di
f
f
ic
ul
t.
D
a
ta
w
it
h
unba
la
nc
e
d
c
la
s
s
e
s
c
a
n
e
a
s
il
y
c
a
u
s
e
th
e
m
ode
l
to
ove
r
f
it
f
or
ju
s
t
one
c
la
s
s
,
he
nc
e
m
a
ki
ng
it
us
e
le
s
s
f
or
de
te
c
ti
ng
ot
h
e
r
c
la
s
s
e
s
.
T
hi
s
c
a
n
be
a
ddr
e
s
s
e
d
by
s
ynt
he
ti
c
m
in
or
it
y
ove
r
s
a
m
pl
in
g t
e
c
hni
que
(
S
M
O
T
E
)
[
19]
,
w
hi
c
h of
f
e
r
s
a
w
a
y t
o
s
ynt
he
s
iz
e
da
ta
f
r
om
i
m
ba
la
nc
e
d c
om
pos
it
io
n.
T
he
us
e
of
S
M
O
T
E
m
a
ke
s
t
h
e
da
ta
ba
la
n
c
e
d, t
hus
pr
ovi
di
ng be
tt
e
r
de
te
c
ti
on r
e
s
ul
ts
.
T
he
us
e
of
U
-
N
e
t
a
r
c
hi
te
c
tu
r
e
is
w
id
e
ly
u
s
e
d
f
or
m
e
di
c
a
l
im
a
ge
s
e
gm
e
nt
a
ti
on
but
ha
s
hi
gh
a
r
c
hi
te
c
tu
r
a
l
c
om
pl
e
xi
ty
.
I
n
r
e
s
e
a
r
c
h
c
onduc
t
e
d
by
L
u
e
t
al
.
[
1
5]
pr
opos
e
d
th
e
u
s
e
of
ha
lf
U
-
N
e
t
f
or
m
e
di
c
a
l
im
a
ge
s
e
gm
e
nt
a
ti
on
w
hi
c
h
is
a
de
ve
lo
pm
e
nt
of
U
-
N
e
t,
na
m
e
ly
by
us
in
g
ghos
t
m
odul
e
a
s
a
s
ub
s
ti
tu
te
f
or
c
onvolut
io
na
l
bl
oc
ks
in
U
-
N
e
t.
T
e
s
t
r
e
s
ul
ts
c
onduc
t
e
d
us
in
g
m
a
m
m
ogr
a
phy,
lu
ng
nodule
,
a
nd
le
f
t
ve
nt
r
ic
ul
a
r
m
a
gne
ti
c
r
e
s
ona
nc
e
im
a
gi
ng
(
M
R
I
)
da
ta
s
e
ts
s
how
th
a
t,
w
h
e
n
u
s
in
g
th
e
s
ta
nd
a
r
d
H
a
lf
U
-
N
e
t
a
r
c
hi
te
c
tu
r
e
ha
s
0.21
m
il
li
on
pa
r
a
m
e
te
r
s
,
w
hi
le
th
e
s
ta
nda
r
d
U
-
N
e
t
ha
s
31.04
m
il
li
on
pa
r
a
m
e
te
r
s
.
T
he
s
e
r
e
s
ul
ts
s
how
th
a
t
h
a
lf
U
-
N
e
t
is
a
bl
e
to
r
e
duc
e
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
w
hi
le
m
a
in
ta
in
in
g
it
s
pe
r
f
or
m
a
nc
e
c
om
pa
r
e
d
to
th
e
s
ta
nda
r
d
U
-
N
e
t.
C
N
N
a
r
e
bui
lt
on
c
onvolut
io
na
l
ope
r
a
ti
ons
,
w
h
ic
h
e
xt
r
a
c
t
in
f
or
m
a
ti
ve
f
e
a
tu
r
e
s
by
c
om
bi
ni
ng
s
pa
ti
a
l
a
nd
c
ha
nn
e
l
in
f
or
m
a
ti
on
to
ge
th
e
r
in
a
lo
c
a
l
r
e
c
e
pt
iv
e
f
ie
ld
.
T
o
im
pr
ove
th
e
r
e
pr
e
s
e
nt
a
ti
ona
l
pow
e
r
of
th
e
ne
twor
k,
s
e
ve
r
a
l
r
e
c
e
nt
a
ppr
oa
c
h
e
s
h
a
ve
s
how
n
th
e
be
ne
f
it
s
of
im
pr
ove
d
s
pa
ti
a
l
e
nc
odi
ng.
I
n
th
e
r
e
s
e
a
r
c
h
c
onduc
te
d
by
H
u
e
t
al
.
[
20]
pr
opos
e
d
a
ne
w
a
r
c
hi
te
c
tu
r
a
l
uni
t,
na
m
e
ly
s
que
e
z
e
-
a
nd
-
e
x
c
it
a
ti
on
(
S
E
)
,
w
hi
c
h
a
da
pt
iv
e
ly
r
e
c
a
li
br
a
te
s
c
h
a
nne
l
f
e
a
tu
r
e
r
e
s
pon
s
e
s
e
xpl
ic
it
ly
m
ode
li
ng
th
e
in
te
r
de
pe
nde
nc
ie
s
b
e
twe
e
n
c
ha
nne
ls
. T
he
pe
r
f
or
m
a
nc
e
of
S
E
ha
s
be
e
n
d
e
m
ons
tr
a
te
d
in
a
c
om
pe
ti
ti
on
by
A
be
da
ll
a
e
t
al
.
[
14]
,
w
he
r
e
4
of
th
e
to
p
5
r
a
nki
ng
s
u
s
e
d
S
E
bl
oc
k
s
c
om
bi
ne
d
in
to
th
e
R
e
s
N
e
X
t
[
21]
m
ode
l
to
im
pr
ove
e
nc
ode
r
p
e
r
f
or
m
a
nc
e
on s
ta
nda
r
d U
-
N
e
t.
S
E
ha
s
a
ls
o be
e
n s
how
n t
o pr
ovi
de
i
m
pr
ove
m
e
nt
s
t
o U
-
N
e
t
on U
S
E
-
N
e
t
[
22]
.
R
e
f
e
r
r
in
g
to
a
num
be
r
of
s
tu
di
e
s
th
a
t
h
a
ve
b
e
e
n
c
onduc
te
d,
th
is
s
tu
dy
pr
opos
e
s
a
pa
pi
ll
e
de
m
a
c
la
s
s
if
ic
a
ti
on
m
ode
l
us
in
g
a
c
om
bi
na
ti
on
of
S
que
e
z
e
-
E
xc
it
a
ti
on
a
nd
H
a
lf
U
-
N
e
t,
he
r
e
in
a
f
te
r
r
e
f
e
r
r
e
d
to
a
s
s
que
e
z
e
-
e
x
c
it
a
ti
on
ha
lf
U
-
N
e
t
(
S
E
H
U
N
E
T
)
.
T
he
c
om
bi
na
ti
on
of
S
E
a
im
s
to
pr
ovi
de
w
e
ig
ht
c
a
li
br
a
ti
on
f
or
e
a
c
h
gho
s
t
m
odul
e
c
ont
a
in
e
d
in
th
e
ha
lf
U
-
N
e
t,
s
o
th
a
t
it
s
o
ut
put
ha
s
a
n
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
f
or
c
e
r
ta
in
f
e
a
tu
r
e
m
a
p
c
ha
nne
ls
.
T
o
ove
r
c
om
e
da
ta
im
ba
la
nc
e
,
th
e
S
M
O
T
E
m
e
th
od
is
a
ppl
ie
d.
T
he
s
ys
te
m
m
ode
l
is
bui
lt
us
in
g
a
da
ta
s
e
t
f
r
om
th
e
K
im
E
ye
H
os
pi
ta
l.
M
ode
l
pe
r
f
or
m
a
nc
e
is
m
e
a
s
ur
e
d
u
s
in
g
a
c
c
ur
a
c
y
pa
r
a
m
e
te
r
s
,
a
r
e
a
unde
r
t
he
c
ur
ve
of
r
e
c
e
iv
e
r
ope
r
a
ti
on c
ha
r
a
c
te
r
is
ti
c
(
A
U
R
O
C
)
, a
nd t
he
numbe
r
of
m
ode
l
pa
r
a
m
e
te
r
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
2
,
A
pr
il
20
25
:
1410
-
1419
1412
2.
M
E
T
H
O
D
F
ig
ur
e
1
s
how
s
th
e
f
lo
w
s
um
m
a
r
y
of
th
e
m
e
th
od
us
e
d
in
th
i
s
s
tu
dy
.
T
he
a
f
or
e
m
e
nt
io
ne
d
f
lo
w
is
c
ons
is
te
d
of
da
ta
s
e
t
c
ol
le
c
ti
on,
da
ta
s
e
t
pr
e
pr
oc
e
s
s
in
g,
m
ode
l
bui
ld
in
g,
m
ode
l
tr
a
in
in
g,
a
nd
p
e
r
f
or
m
a
nc
e
e
va
lu
a
ti
on.
T
he
s
e
s
te
ps
a
r
e
done
s
e
que
nt
ia
ll
y,
m
e
a
ni
ng
th
a
t
th
e
f
lo
w
is
not
a
bl
e
to
be
p
r
oc
e
s
s
e
d
f
ur
th
e
r
if
th
e
pr
e
vi
ous
s
te
p i
s
not
done
.
F
ig
ur
e
1. M
e
th
od
f
lo
w
c
ha
r
t
2
.1.
C
ol
le
c
t
in
g d
at
as
e
t
T
he
da
ta
s
e
t
u
s
e
d
in
th
is
s
tu
dy
is
f
undus
r
e
ti
na
l
da
ta
s
e
t
f
r
om
K
im
’
s
E
ye
H
os
pi
ta
l
us
e
d
in
[
7]
w
hi
c
h
c
a
n
be
a
c
c
e
s
s
e
d
on
K
a
ggl
e
.
T
he
a
va
il
a
bl
e
c
la
s
s
e
s
a
r
e
nor
m
a
l
w
it
h
779
im
a
ge
s
,
p
a
pi
ll
e
de
m
a
w
it
h
295
im
a
g
e
s
,
a
nd
p
s
e
udopa
pi
ll
e
de
m
a
w
it
h
295
im
a
ge
s
.
T
he
da
ta
s
e
t
w
a
s
c
o
ll
e
c
te
d
f
r
om
pa
ti
e
nt
s
w
ho
unde
r
w
e
nt
f
undus
phot
ogr
a
phy a
s
pa
r
t
of
t
he
ir
e
ye
e
xa
m
in
a
ti
on.
2
.
2
.
P
r
e
p
r
oc
e
s
s
in
g d
at
as
e
t
T
hi
s
s
tu
dy
a
dopt
s
S
M
O
T
E
a
s
a
c
r
uc
ia
l
te
c
hni
qu
e
to
ta
c
k
le
th
e
im
ba
la
nc
e
d
c
la
s
s
c
ha
ll
e
ng
e
.
I
t
ope
r
a
te
s
by
f
la
tt
e
ni
ng
th
e
in
put
im
a
ge
s
in
to
a
one
-
di
m
e
ns
io
na
l
num
e
r
ic
a
l
da
ta
,
a
nd
th
e
n
s
ynt
he
s
iz
in
g
a
r
ti
f
ic
ia
l
in
s
ta
nc
e
s
f
or
th
e
m
in
or
it
y
c
la
s
s
by
in
te
r
pol
a
ti
ng
b
e
tw
e
e
n
it
s
c
lo
s
e
s
t
n
e
ig
hbor
s
in
th
e
f
e
a
tu
r
e
s
pa
c
e
[
19]
.
T
hus
,
u
s
in
g
th
e
S
M
O
T
E
te
c
hni
que
is
a
bl
e
to
e
qua
li
z
e
th
e
c
la
s
s
di
s
tr
ib
ut
io
n
a
nd
e
nr
ic
h
th
e
in
f
or
m
a
ti
on
f
or
t
he
m
ode
l
to
l
e
a
r
n f
r
om
.
T
he
pr
e
pr
oc
e
s
s
in
g
s
te
p
is
il
lu
s
tr
a
te
d
in
F
ig
ur
e
2
w
he
r
e
20%
of
t
he
da
ta
w
il
l
be
ta
ke
n
f
r
om
th
e
in
it
ia
l
da
ta
to
be
th
e
te
s
t
in
g
s
e
t
.
T
he
r
e
s
t
th
e
n
w
il
l
go
th
r
ough
th
e
S
M
O
T
E
ove
r
s
a
m
pl
in
g
pr
oc
e
s
s
f
o
r
s
ynt
he
s
iz
in
g,
c
r
e
a
ti
ng
ne
w
s
im
il
a
r
in
s
ta
nc
e
s
on
e
ve
r
y
c
la
s
s
ba
s
e
d
on
th
e
in
put
im
a
ge
s
,
he
nc
e
f
or
m
in
g
a
ba
la
nc
e
d
c
om
pos
it
io
n
be
twe
e
n
e
a
c
h
c
l
a
s
s
.
W
he
n
th
e
s
ynt
he
s
i
z
in
g
pr
oc
e
s
s
is
done
,
th
e
s
ynt
he
ti
c
d
a
ta
w
il
l
e
nt
e
r
th
e
hor
iz
ont
a
l
f
li
p
a
ugm
e
nt
a
ti
on
s
ta
ge
w
hi
c
h
w
il
l
f
li
p
th
e
im
a
ge
hor
iz
ont
a
ll
y
a
nd
a
dd
it
to
th
e
tr
a
in
in
g
pr
oc
e
s
s
e
v
e
r
yt
im
e
th
e
da
t
a
ge
ne
r
a
to
r
i
s
c
a
ll
e
d.
T
he
r
e
s
ul
t
of
t
he
la
s
t
pr
oc
e
s
s
w
il
l
b
e
di
vi
de
d
in
to
20%
a
s
va
li
da
ti
on da
ta
a
nd 80%
a
s
t
he
tr
a
in
da
ta
f
or
t
he
S
E
H
U
N
E
T
m
ode
l.
F
ig
ur
e
2. D
a
ta
s
e
t
pr
e
pr
oc
e
s
s
in
g
(
N
:
nor
m
a
l
, P
E
:
pa
pi
ll
e
de
m
a
, P
P
E
:
ps
e
udopa
pi
ll
e
de
m
a
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sque
e
z
e
-
e
x
c
it
at
io
n
h
al
f
U
-
N
e
t
and s
y
nt
he
ti
c
m
in
or
it
y
o
v
e
r
s
am
p
li
ng t
e
c
hni
que
ov
e
r
s
am
pl
in
g
…
(
W
ih
ar
to
)
1413
2
.
3
.
B
u
il
d
in
g m
od
e
l
T
he
s
tr
uc
tu
r
e
of
S
E
H
U
N
E
T
a
r
c
hi
te
c
tu
r
e
pr
opos
e
d
in
th
is
s
tu
dy
is
il
lu
s
tr
a
te
d
in
F
ig
u
r
e
3.
T
he
S
E
H
U
N
E
T
a
r
c
hi
te
c
tu
r
a
l
de
s
ig
n
is
c
ons
is
te
d
of
ghos
t
m
odul
e
a
nd
S
E
bl
oc
k
c
om
bi
na
ti
on
w
hi
c
h
f
or
m
s
th
e
ghos
t
-
S
E
m
odul
e
,
s
ta
c
ke
d
to
m
im
ic
th
e
s
tr
uc
tu
r
e
f
r
om
ha
lf
U
-
N
e
t
a
r
c
hi
te
c
tu
r
e
.
T
he
gho
s
t
m
odul
e
[
23]
f
unc
ti
ons
a
s
ge
ne
r
a
to
r
to
ge
ne
r
a
te
m
or
e
f
e
a
tu
r
e
m
a
ps
f
r
om
th
e
in
put
w
it
h
li
ght
e
r
ope
r
a
ti
ons
th
a
n
th
e
r
e
gul
a
r
c
onvolut
io
na
l
bl
oc
k.
T
hi
s
m
a
k
e
s
th
e
gho
s
t
m
odul
e
pr
oduc
e
s
t
he
s
a
m
e
num
be
r
of
out
put
s
c
om
p
a
r
e
d
to
th
e
r
e
gul
a
r
c
onvolut
io
na
l
bl
oc
k
w
i
th
m
uc
h
le
s
s
c
om
put
a
ti
ona
l
c
os
t
a
nd
he
nc
e
w
il
l
c
ont
r
ib
u
te
to
r
e
duc
e
th
e
f
in
a
l
m
ode
l
s
iz
e
.
H
a
lf
U
-
N
e
t
f
r
o
m
L
u
e
t
al
.
[
15]
us
e
s
t
he
ghos
t
m
odul
e
to
r
e
duc
e
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
a
nd a
voi
d
a
ddi
ng
m
or
e
pa
r
a
m
e
te
r
s
f
r
om
th
e
s
ta
nda
r
d
U
-
N
e
t.
I
n
a
ddi
ti
on,
ha
lf
U
-
N
e
t
a
r
c
hi
te
c
tu
r
e
s
ta
nda
r
di
z
e
s
th
e
num
be
r
of
a
ll
c
ha
nne
ls
t
o be
c
om
bi
ne
d i
n t
he
de
c
ode
r
pa
r
t
of
th
e
a
r
c
hi
te
c
tu
r
e
by pe
r
f
or
m
in
g f
ul
l
-
s
c
a
le
f
e
a
tu
r
e
f
us
io
n.
T
hi
s
w
il
l
f
ur
th
e
r
r
e
duc
e
th
e
m
ode
l
s
iz
e
a
nd
th
e
to
ta
l
c
om
put
a
ti
ona
l
c
os
t
of
it
s
tr
a
in
in
g
w
hi
le
s
ti
ll
m
a
in
ta
in
in
g t
he
pe
r
f
or
m
a
nc
e
.
F
ig
ur
e
3. S
E
H
U
N
E
T
m
ode
l
a
r
c
hi
te
c
tu
r
e
T
he
de
ve
lo
pm
e
nt
c
a
r
r
ie
d
out
in
th
is
s
tu
dy
in
vol
ve
s
th
e
in
s
e
r
t
io
n
of
a
n
S
E
bl
oc
k
a
f
te
r
e
a
c
h
ghos
t
m
odul
e
in
ha
lf
U
-
N
e
t
a
s
gho
s
t
-
S
E
m
odul
e
.
T
he
m
a
in
pur
po
s
e
of
th
e
S
E
bl
oc
k
a
ddi
ti
on
i
s
to
pr
ovi
de
a
n
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
a
nd
th
e
a
bi
li
ty
to
r
e
c
a
li
br
a
te
e
a
c
h
f
e
a
t
ur
e
m
a
ps
by
c
a
lc
ul
a
ti
ng
th
e
ir
c
or
r
e
s
ponding
w
e
ig
ht
s
[
20]
.
T
hi
s
e
n
a
bl
e
s
th
e
S
E
bl
oc
k
to
e
nha
n
c
e
th
e
r
e
pr
e
s
e
nt
a
ti
ona
l
pow
e
r
of
th
e
f
e
a
tu
r
e
m
a
ps
by
le
a
r
ni
ng
th
e
in
te
r
de
pe
nde
nc
ie
s
be
twe
e
n
th
e
m
.
T
he
c
om
bi
na
ti
on
of
ghos
t
m
odul
e
a
nd
S
E
bl
oc
k
a
ll
ow
s
S
E
H
U
N
E
T
to
ge
ne
r
a
te
f
e
a
tu
r
e
m
a
ps
e
f
f
ic
ie
nt
ly
w
hi
le
ha
vi
ng
t
he
a
bi
li
ty
to
pr
ovi
de
a
tt
e
nt
io
n
f
oc
us
on
e
a
c
h
of
th
e
s
e
f
e
a
tu
r
e
m
a
ps
s
o
th
a
t
th
e
m
ode
li
s
a
bl
e
to
a
c
hi
e
ve
be
tt
e
r
pe
r
f
or
m
a
nc
e
th
a
n
th
e
r
e
gul
a
r
ha
lf
U
-
N
e
t.
T
he
ghos
t
-
S
E
m
odul
e
s
tr
uc
tu
r
e
c
a
n be
s
e
e
n i
n F
ig
ur
e
4.
2
.
4
.
T
r
ai
n
in
g m
od
e
l
I
n
th
is
s
tu
dy
,
a
s
e
t
of
tr
a
in
in
g
e
xpe
r
im
e
nt
s
w
a
s
c
onduc
te
d
to
de
te
r
m
in
e
w
hi
c
h
one
is
be
tt
e
r
be
twe
e
n
ha
lf
U
-
N
e
t
a
nd
S
E
H
U
N
E
T
us
in
g
th
e
or
ig
in
a
l
da
ta
,
th
e
n
th
e
be
s
t
m
ode
l
f
r
om
th
e
pr
e
vi
ous
tr
ia
l
w
il
l
be
tr
a
in
e
d
us
in
g t
he
S
M
O
T
E
s
ynt
he
s
iz
e
d da
ta
.
T
he
t
r
a
in
in
g
pr
oc
e
s
s
w
a
s
r
un on G
oogl
e
C
ol
a
bor
a
to
r
y
w
it
h t
he
c
ode
f
ul
l
y
w
r
it
te
n
in
P
yt
hon
pr
ogr
a
m
m
in
g
la
ngua
ge
ut
il
iz
in
g
T
e
ns
or
f
lo
w
f
r
a
m
e
w
or
k
.
I
t
r
a
n
w
it
h
G
P
U
r
unt
im
e
ty
pe
s
o
th
a
t
th
e
tr
a
in
in
g
t
im
e
c
a
n
be
m
in
im
iz
e
d
.
T
he
tr
a
in
in
g
us
e
d
th
e
A
da
m
opt
im
iz
e
r
[
24]
w
it
h
th
e
s
ta
r
ti
ng
le
a
r
ni
n
g
r
a
te
va
lu
e
of
10
-
3
.
T
he
b
a
tc
h
s
i
z
e
u
s
e
d
f
or
tr
a
in
in
g
w
a
s
32
w
it
h
a
pr
e
s
e
t
of
300
e
poc
hs
.
T
h
e
tr
a
in
in
g
s
y
s
te
m
w
il
l
r
e
duc
e
th
e
le
a
r
ni
ng
r
a
te
to
1/
10
of
th
e
pr
e
vi
ou
s
va
lu
e
if
th
e
va
li
da
ti
on
a
c
c
ur
a
c
y
doe
s
not
in
c
r
e
a
s
e
f
or
20
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
2
,
A
pr
il
20
25
:
1410
-
1419
1414
e
poc
hs
w
it
h
th
e
m
in
im
um
le
a
r
ni
ng
r
a
te
va
lu
e
be
in
g
10
-
6
.
I
n
a
ddi
ti
on,
th
e
tr
a
in
in
g
w
il
l
be
s
to
ppe
d
a
nd
th
e
m
ode
l
w
il
l
r
e
s
to
r
e
t
he
be
s
t
w
e
ig
ht
s
i
f
t
he
r
e
i
s
no i
nc
r
e
a
s
e
i
n va
li
da
ti
on a
c
c
ur
a
c
y f
or
50 e
poc
hs
.
F
ig
ur
e
4. G
hos
t
-
S
E
m
odul
e
s
tr
uc
tu
r
e
2
.
5
.
P
e
r
f
or
m
an
c
e
e
val
u
at
io
n
T
he
pe
r
f
or
m
a
nc
e
of
th
e
tr
a
in
in
g
e
xpe
r
im
e
nt
s
w
il
l
be
e
va
lu
a
te
d
us
in
g
te
s
t
in
g
s
e
t
w
hi
c
h
c
ons
i
s
ts
of
20%
f
r
om
th
e
or
ig
in
a
l
da
ta
.
T
he
m
a
in
de
te
r
m
in
in
g
m
e
tr
ic
us
e
d
is
a
c
c
ur
a
c
y
a
nd
f
ol
lo
w
e
d
by
bot
h
c
a
te
gor
ic
a
l
c
r
os
s
e
nt
r
opy
(
C
C
E
)
lo
s
s
a
nd
A
U
R
O
C
[
25]
a
s
th
e
s
e
c
onda
r
y
m
e
tr
ic
s
.
T
he
m
a
in
f
a
c
to
r
of
us
in
g
a
c
c
ur
a
c
y
a
s
th
e
pr
im
a
r
y
m
e
tr
ic
in
vol
ve
s
th
e
pr
e
vi
ous
s
tu
di
e
s
s
ho
w
n
in
[
7]
,
[
12]
,
[
13]
th
a
t
c
ol
le
c
ti
ve
ly
us
e
a
c
c
ur
a
c
y
a
s
th
e
m
a
in
de
te
r
m
in
is
ti
c
f
a
c
to
r
.
3.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
T
he
e
xpe
r
im
e
nt
s
c
onduc
te
d
in
th
i
s
r
e
s
e
a
r
c
h
w
il
l
ha
ve
th
r
e
e
m
a
in
s
te
ps
.
F
ir
s
t,
th
e
e
xpe
r
im
e
nt
of
tr
a
in
in
g
H
a
lf
U
-
N
e
t
w
it
hout
S
M
O
T
E
.
S
e
c
ond,
th
e
e
xpe
r
im
e
nt
of
t
r
a
in
in
g
S
E
H
U
N
E
T
w
hi
c
h
a
ls
o
w
it
hout
S
M
O
T
E
.
L
a
s
tl
y,
th
e
e
xpe
r
im
e
nt
f
or
th
e
be
s
t
a
r
c
hi
te
c
tu
r
e
de
f
in
e
d
by
c
om
pa
r
is
on
of
th
e
pr
e
vi
ous
e
xp
e
r
im
e
nt
s
,
tr
a
in
e
d us
in
g S
M
O
T
E
.
3.1.
E
xp
e
r
im
e
n
t
s
T
he
f
ir
s
t
e
xpe
r
im
e
nt
c
onduc
te
d
w
a
s
th
e
tr
a
in
in
g
of
H
a
lf
U
-
N
e
t
us
in
g
th
e
or
ig
in
a
l
da
ta
s
e
t
,
in
th
e
ot
he
r
w
or
d
w
it
hout
S
M
O
T
E
ove
r
s
a
m
pl
in
g.
T
he
r
e
s
ul
ts
of
th
e
te
s
ts
c
onduc
t
e
d
c
a
n
be
s
ho
w
n
in
T
a
bl
e
1.
I
t
pe
r
f
or
m
e
d
w
e
ll
w
it
h
th
e
te
s
t
a
c
c
ur
a
c
y
s
c
or
e
r
e
a
c
hi
ng
93.7
2%
.
T
hi
s
e
xpe
r
im
e
nt
la
s
te
d
f
or
130
e
poc
h
s
,
m
e
a
ni
ng t
ha
t
it
r
e
a
c
he
d i
ts
pe
a
k on
e
poc
h 80
.
T
h
e
c
or
r
e
s
pondin
g t
e
s
t
lo
s
s
gr
a
ph
is
s
how
n i
n F
ig
ur
e
5
.
T
a
bl
e
1. H
a
lf
U
-
N
e
t
pe
r
f
or
m
a
nc
e
w
it
hout
S
M
O
T
E
M
e
t
r
i
c
s
S
c
or
e
(%)
V
a
l
i
da
t
i
on
a
c
c
ur
a
c
y
96.78
V
a
l
i
da
t
i
on
l
os
s
T
e
s
t
a
c
c
ur
a
c
y
T
e
s
t
l
os
s
A
U
R
O
C
0.15
93.72
0.16
0.994
T
he
s
e
c
ond
e
xpe
r
im
e
nt
c
onduc
te
d
w
a
s
th
e
tr
a
in
in
g
of
S
E
H
U
N
E
T
us
in
g
or
ig
in
a
l
da
ta
s
e
t
w
it
hout
S
M
O
T
E
ove
r
s
a
m
pl
in
g.
T
he
r
e
s
ul
t
f
r
om
th
is
e
xp
e
r
im
e
nt
w
il
l
be
c
om
pa
r
e
d
to
th
e
r
e
s
ul
t
f
r
om
th
e
f
ir
s
t
e
xpe
r
im
e
nt
t
o de
te
r
m
in
e
w
hi
c
h one
pe
r
f
o
r
m
s
t
he
be
s
t
on t
he
s
a
m
e
s
c
e
na
r
io
.
T
he
r
e
s
ul
ts
of
t
he
t
e
s
ts
c
onduc
t
e
d
c
a
n
be
s
how
n
in
T
a
bl
e
2.
I
t
pe
r
f
or
m
e
d
w
e
ll
w
it
h
th
e
te
s
t
a
c
c
ur
a
c
y
s
c
or
e
r
e
a
c
hi
ng
95.94%
.
T
hi
s
e
xpe
r
im
e
nt
la
s
te
d
f
or
159
e
poc
h
s
,
m
e
a
ni
ng
th
a
t
it
r
e
a
c
h
e
d
it
s
p
e
a
k
on
e
p
oc
h
109.
T
h
e
c
or
r
e
s
ponding
te
s
t
lo
s
s
gr
a
ph
i
s
s
how
n i
n F
ig
ur
e
6.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sque
e
z
e
-
e
x
c
it
at
io
n
h
al
f
U
-
N
e
t
and s
y
nt
he
ti
c
m
in
or
it
y
o
v
e
r
s
am
p
li
ng t
e
c
hni
que
ov
e
r
s
am
pl
in
g
…
(
W
ih
ar
to
)
1415
F
ig
ur
e
5. H
a
lf
U
-
N
e
t
lo
s
s
gr
a
ph w
it
hout
S
M
O
T
E
T
a
bl
e
2. S
E
H
U
N
E
T
pe
r
f
or
m
a
nc
e
w
it
hout
S
M
O
T
E
M
e
t
r
i
c
s
S
c
or
e
(%)
V
a
l
i
da
t
i
on
a
c
c
ur
a
c
y
97.24
V
a
l
i
da
t
i
on
l
os
s
T
e
s
t
a
c
c
ur
a
c
y
T
e
s
t
l
os
s
A
U
R
O
C
0.17
95.94
0.17
0.993
F
ig
ur
e
6. S
E
H
U
N
E
T
lo
s
s
gr
a
ph
w
it
hout
S
M
O
T
E
F
r
om
th
e
r
e
s
ul
ts
of
two
e
xpe
r
im
e
nt
s
a
bove
,
S
E
H
U
N
E
T
m
od
e
l
pe
r
f
or
m
e
d
be
tt
e
r
th
a
n
th
e
or
ig
in
a
l
h
a
lf
U
-
N
e
t
m
ode
l
in
th
e
m
a
in
de
te
r
m
in
in
g
m
e
tr
ic
c
ol
le
c
ti
ve
ly
u
s
e
d
in
th
e
pr
e
vi
ous
s
tu
di
e
s
,
a
c
c
ur
a
c
y,
f
or
bot
h
va
li
da
ti
on
a
nd
te
s
t
da
ta
.
C
on
s
e
que
nt
ly
,
it
c
om
e
to
th
e
c
onc
lu
s
io
n
th
a
t
S
E
H
U
N
E
T
ha
s
b
e
tt
e
r
pe
r
f
or
m
a
nc
e
th
a
n
th
e
s
ta
nda
r
d
h
a
lf
U
-
N
e
t
a
nd
w
il
l
go
th
r
ough
th
e
ne
xt
e
xpe
r
i
m
e
nt
us
in
g
th
e
S
M
O
T
E
s
ynt
he
s
iz
e
d
da
ta
s
e
t.
T
he
S
M
O
T
E
w
il
l
c
r
e
a
te
a
s
ynt
he
ti
c
im
a
ge
f
r
om
th
e
or
ig
in
a
l
im
a
ge
,
in
or
de
r
to
ba
la
nc
e
th
e
da
ta
.
T
he
s
a
m
pl
e
of
th
e
S
M
O
T
E
r
e
s
ul
ts
on
a
r
e
ti
na
l
im
a
ge
c
a
n
be
s
how
n
in
F
ig
ur
e
7.
I
f
th
e
or
ig
in
a
l
im
a
ge
is
a
s
s
how
n
in
F
ig
ur
e
7(
a
)
,
th
e
n
th
e
s
ynt
he
s
iz
e
d
im
a
g
e
is
s
how
n
in
F
ig
ur
e
7(
b)
.
T
he
r
e
s
ul
ts
of
th
e
S
E
H
U
N
E
T
e
xpe
r
im
e
nt
w
it
h S
M
O
T
E
a
r
e
s
how
n i
n T
a
bl
e
3.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
2
,
A
pr
il
20
25
:
1410
-
1419
1416
(
a
)
(
b)
F
ig
ur
e
7. T
he
s
a
m
pl
e
of
S
M
O
T
E
r
e
s
ul
t
im
a
ge
:
(
a
)
o
r
ig
in
a
l
pa
pi
ll
e
de
m
a
a
nd
(
b)
s
ynt
he
ti
c
pa
pi
ll
e
de
m
a
T
a
bl
e
3. S
E
H
U
N
E
T
pe
r
f
or
m
a
nc
e
w
it
h S
M
O
T
E
M
e
t
r
i
c
s
S
c
or
e
(%)
V
a
l
i
da
t
i
on
a
c
c
ur
a
c
y
99.19
V
a
l
i
da
t
i
on
l
os
s
T
e
s
t
a
c
c
ur
a
c
y
T
e
s
t
l
os
s
A
U
R
O
C
0.04
98.52
0.07
0.997
T
hi
s
e
xpe
r
im
e
nt
la
s
te
d
f
or
1
17
e
poc
hs
,
m
e
a
ni
ng
th
a
t
it
r
e
a
c
he
d
it
s
pe
a
k
on
e
poc
h
67.
T
he
c
or
r
e
s
ponding
te
s
t
lo
s
s
gr
a
ph
is
s
how
n
in
F
ig
ur
e
8.
S
E
H
U
N
E
T
w
it
h
S
M
O
T
E
ga
ve
th
e
be
s
t
r
e
s
ul
t
c
om
pa
r
e
d t
o t
he
f
ir
s
t
two e
xpe
r
im
e
nt
s
, r
e
a
c
hi
ng t
he
t
e
s
t
a
c
c
ur
a
c
y s
c
or
e
of
98.52%
.
F
ig
ur
e
8.
S
E
H
U
N
E
T
lo
s
s
gr
a
ph
w
it
h S
M
O
T
E
3.2.
C
om
p
ar
is
on
w
it
h
p
r
e
vi
ou
s
s
t
u
d
y
T
hi
s
s
tu
dy
in
ve
s
ti
g
a
te
d
th
e
e
f
f
e
c
ts
of
S
M
O
T
E
a
nd
a
n
e
nha
n
c
e
d
h
a
lf
U
-
N
e
t
a
r
c
hi
te
c
tu
r
e
to
im
a
ge
c
la
s
s
if
ic
a
ti
on
f
ie
ld
in
pa
pi
ll
e
de
m
a
w
it
h
a
p
r
e
m
is
e
to
a
c
hi
e
ve
th
e
ti
ni
e
s
t
m
ode
l
s
iz
e
a
nd
a
c
ons
id
e
r
a
bl
e
pe
r
f
or
m
a
nc
e
.
W
hi
le
e
a
r
li
e
r
s
tu
di
e
s
ha
ve
e
xpl
or
e
d
th
e
im
pa
c
t
of
di
f
f
e
r
e
nt
tr
a
ns
f
e
r
le
a
r
ni
ng
m
ode
l
on
th
e
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
ba
s
e
d
on
K
im
’
s
E
ye
H
os
pi
ta
l
da
ta
,
th
e
y
ha
ve
not
a
ddr
e
s
s
e
d
th
e
pa
r
a
m
e
te
r
s
iz
e
m
in
im
iz
a
ti
on
th
or
oughly.
T
o
ga
in
s
om
e
in
s
ig
ht
r
e
ga
r
di
ng
th
a
t
m
a
tt
e
r
,
t
he
S
E
H
U
N
E
T
m
ode
l
in
c
om
bi
na
ti
on
w
it
h
S
M
O
T
E
ove
r
s
a
m
pl
in
g
w
il
l
be
c
om
pa
r
e
d
in
te
r
m
s
of
pe
r
f
or
m
a
nc
e
a
nd
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
w
it
h
pr
e
vi
ous
s
tu
di
e
s
to
pr
ove
th
a
t
th
is
a
r
c
hi
te
c
tu
r
e
c
a
n
a
c
hi
e
ve
c
om
put
a
ti
ona
l
c
os
t
e
f
f
ic
ie
nc
y
w
hi
le
s
ti
ll
e
xc
e
ll
in
g
in
pe
r
f
or
m
a
nc
e
.
T
he
m
e
tr
ic
s
us
e
d
a
s
e
va
lu
a
to
r
s
a
r
e
te
s
t
a
c
c
ur
a
c
y,
A
U
R
O
C
,
a
nd
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
.
T
he
r
e
s
e
a
r
c
h
e
s
th
a
t
w
il
l
be
c
om
pa
r
e
d
to
S
E
H
U
N
E
T
a
r
e
r
e
s
e
a
r
c
h
A
hn
e
t
al
.
[
7
]
w
hi
c
h
us
e
s
th
e
C
N
N
B
a
ye
s
i
a
n
opt
im
iz
a
ti
on
,
V
G
G
,
R
e
s
N
e
t,
a
nd
I
nc
e
pt
io
nV
3
a
r
c
hi
te
c
tu
r
e
s
;
r
e
s
e
a
r
c
h
K
okul
u
a
nd
G
öke
r
[
12]
w
hi
c
h
us
e
s
M
obi
le
N
e
tV2;
a
nd r
e
s
e
a
r
c
h A
l
-
A
z
z
a
w
i
e
t
al
.
[
13]
w
hi
c
h us
e
s
R
e
s
N
e
t
-
50
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sque
e
z
e
-
e
x
c
it
at
io
n
h
al
f
U
-
N
e
t
and s
y
nt
he
ti
c
m
in
or
it
y
o
v
e
r
s
am
p
li
ng t
e
c
hni
que
ov
e
r
s
am
pl
in
g
…
(
W
ih
ar
to
)
1417
T
a
bl
e
4
s
how
s
th
e
c
om
pa
r
is
on
be
twe
e
n
e
a
c
h
s
tu
dy.
C
N
N
B
a
ye
s
ia
n
opt
im
iz
a
ti
on
f
r
om
A
hn
e
t
al
.
[
7
]
a
nd
S
E
H
U
N
E
T
a
r
e
th
e
onl
y
a
r
c
hi
te
c
tu
r
e
s
tr
a
in
e
d
f
r
om
s
c
r
a
tc
h
w
it
hout
ut
il
iz
in
g
a
ny
tr
a
ns
f
e
r
le
a
r
ni
ng
m
ode
l.
I
n
th
e
s
tu
dy
of
A
hn
e
t
al
.
[
7]
a
c
hi
e
ve
d
th
e
be
s
t
a
c
c
ur
a
c
y
of
98.63%
us
in
g
th
e
tr
a
ns
f
e
r
le
a
r
ni
ng
R
e
s
N
e
t
m
ode
l
w
it
h
0.999
A
U
R
O
C
a
nd
24.2
m
il
li
on
pa
r
a
m
e
te
r
s
.
I
n
th
e
s
tu
dy
of
K
okul
u
a
nd
G
öke
r
[
12]
a
c
hi
e
ve
d
96.83%
a
c
c
ur
a
c
y
us
in
g
th
e
tr
a
ns
f
e
r
le
a
r
ni
ng
M
obi
le
N
e
tV2
m
ode
l.
I
n
th
e
s
tu
dy
of
A
l
-
A
z
z
a
w
i
e
t
al
.
[
13]
a
c
hi
e
ve
d
97.5%
a
c
c
ur
a
c
y
us
in
g
th
e
tr
a
ns
f
e
r
le
a
r
ni
ng
R
e
s
N
e
t
-
50
m
ode
l.
T
he
pr
opos
e
d
S
E
H
U
N
E
T
m
ode
l
a
c
hi
e
ve
d
98.52%
w
it
h 0.997 AU
R
O
C
a
nd 0.268 mi
ll
io
n pa
r
a
m
e
te
r
s
.
T
a
bl
e
4. P
e
r
f
or
m
a
nc
e
c
om
pa
r
is
on
A
r
c
h
L
e
a
r
ni
ng
t
ype
P
r
e
pr
oc
e
s
s
i
ng
A
c
c
(%)
A
U
R
O
C
P
a
r
a
m
s
C
N
N
B
a
ye
s
i
a
n
opt
i
m
. [
7]
S
c
r
a
t
c
h
L
oc
a
l
A
vg
c
ol
or
s
ubt
r
a
c
t
i
on
95.89
0.992
11.6 m
i
l
l
i
on
V
G
G
[
7]
T
r
a
ns
f
e
r
L
oc
a
l
A
vg
c
ol
or
s
ubt
r
a
c
t
i
on
96.8
0.999
33.3 m
i
l
l
i
on
I
nc
e
pt
i
onV
3 [
7]
T
r
a
ns
f
e
r
L
oc
a
l
A
vg
c
ol
or
s
ubt
r
a
c
t
i
on
96.35
0.997
24.8 m
i
l
l
i
on
R
e
s
N
e
t
[
7
]
T
r
a
ns
f
e
r
L
oc
a
l
A
vg
c
ol
or
s
ubt
r
a
c
t
i
on
98.63
0.999
24.2 m
i
l
l
i
on
M
obi
l
e
N
e
t
V
2 [
12]
T
r
a
ns
f
e
r
H
i
s
t
ogr
a
m
e
qua
l
i
z
a
t
i
on
3D
box f
i
l
t
e
r
i
ng
96.83
-
-
R
e
s
N
e
t
-
50 [
13]
T
r
a
ns
f
e
r
S
e
gm
e
nt
a
t
i
on
opt
i
m
.
97.5
-
-
S
E
H
U
N
E
T
(
pr
opos
e
d)
S
c
r
a
t
c
h
S
M
O
T
E
ove
r
s
a
m
pl
i
ng
98.52
0.997
0.268 m
i
l
l
i
on
B
a
s
e
d
on
th
e
c
om
p
a
r
is
on
r
e
s
ul
ts
a
bove
,
it
c
a
n
b
e
s
e
e
n
th
a
t
S
E
H
U
N
E
T
be
c
om
e
s
th
e
s
e
c
ond
to
p
pe
r
f
or
m
a
nc
e
a
r
c
hi
te
c
tu
r
e
a
f
te
r
a
lm
os
t
m
a
tc
hi
ng
th
e
a
c
c
ur
a
c
y
of
R
e
s
N
e
t
tr
a
ns
f
e
r
le
a
r
ni
ng
m
ode
l
f
r
om
A
hn
e
t
al
.
[
7]
w
it
h
m
uc
h
s
m
a
ll
e
r
pa
r
a
m
e
te
r
s
.
T
hu
s
,
S
E
H
U
N
E
T
s
uc
c
e
s
s
f
ul
ly
a
c
hi
e
ve
d
c
om
put
a
ti
ona
l
c
os
t
e
f
f
ic
ie
nc
y
w
it
h
to
ta
l
p
e
r
c
e
nt
a
ge
of
pa
r
a
m
e
te
r
s
r
e
duc
ti
on
f
r
o
m
R
e
s
N
e
t
[
7]
by
98%
.
T
he
pe
r
f
or
m
a
nc
e
of
S
E
H
U
N
E
T
i
n t
he
f
or
m
of
a
c
onf
us
io
n m
a
tr
ix
[
26]
f
r
o
m
t
he
t
e
s
t
s
e
t
is
s
how
n i
n F
ig
ur
e
9
.
F
ig
ur
e
9. S
E
H
U
N
E
T
pe
r
f
or
m
a
nc
e
on c
onf
us
io
n m
a
tr
ix
3.
3
.
M
od
e
l
ou
t
p
u
t
T
he
f
ol
lo
w
in
g
im
a
ge
s
c
ont
a
in
th
e
e
xa
m
pl
e
of
s
uc
c
e
s
s
f
ul
pr
e
di
c
ti
on r
e
s
ul
ts
f
or
e
a
c
h c
la
s
s
c
ol
le
c
te
d
in
F
ig
ur
e
10
.
T
he
s
e
im
a
g
e
s
s
how
th
a
t
th
e
S
E
H
U
N
E
T
m
ode
l
c
a
n
a
c
c
ur
a
te
ly
di
s
ti
ngui
s
h
b
e
twe
e
n
nor
m
a
l,
pa
pi
ll
e
de
m
a
,
a
nd
ps
e
udopa
pi
ll
e
de
m
a
in
f
or
th
e
m
os
t
c
a
s
e
s
.
F
ig
ur
e
11
how
e
ve
r
,
s
how
s
th
e
e
xa
m
pl
e
of
f
a
il
e
d
pr
e
di
c
ti
on
a
tt
e
m
pt
s
by
th
e
m
ode
l.
T
he
f
ir
s
t
c
or
r
e
la
te
d
out
put
s
a
r
e
th
e
pr
e
di
c
t
io
n
of
no
r
m
a
l
im
a
ge
a
s
ps
e
udopa
pi
ll
e
de
m
a
a
nd
th
e
oppos
it
e
, w
hi
c
h
is
s
u
s
pe
c
te
d
to
be
c
a
us
e
d
by
th
e
la
c
k
s
e
gm
e
nt
a
ti
on
of
dr
us
e
n
th
a
t
m
a
ke
th
e
di
s
ks
a
pp
e
a
r
to
be
bl
ur
r
y.
T
he
ne
xt
pr
e
di
c
ti
on
is
a
f
a
il
e
d
a
tt
e
m
pt
to
pr
e
di
c
t
ps
e
udopa
pi
ll
e
de
m
a
im
a
ge
a
s
pa
pi
ll
e
de
m
a
w
hi
c
h
i
s
s
us
pe
c
t
e
d
to
b
e
c
a
u
s
e
d
by
th
e
h
ig
h
le
ve
l
of
bl
ur
r
in
e
s
s
of
th
e
opt
ic
di
s
c
w
hi
c
h
ve
r
y c
lo
s
e
ly
m
im
ic
s
t
he
pa
pi
ll
e
de
m
a
i
m
a
ge
f
r
om
F
ig
ur
e
10
,
a
c
c
om
pa
ni
e
d
by t
he
t
hi
nne
s
s
of
t
he
bl
ood ve
s
s
e
ls
a
r
ound the
di
s
c
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
2
,
A
pr
il
20
25
:
1410
-
1419
1418
F
ig
ur
e
10.
C
or
r
e
c
t
pr
e
di
c
ti
ons
f
r
om
S
E
H
U
N
E
T
F
ig
ur
e
11. F
a
il
e
d
pr
e
di
c
ti
ons
f
r
om
S
E
H
U
N
E
T
4.
C
O
N
C
L
U
S
I
O
N
B
a
s
e
d
on
th
e
r
e
s
ul
t
s
of
th
is
s
tu
dy,
it
c
a
n
be
c
on
c
lu
de
d
th
a
t
S
E
H
U
N
E
T
is
e
f
f
e
c
ti
ve
f
or
pa
pi
ll
e
de
m
a
im
a
ge
c
la
s
s
if
ic
a
ti
on
.
T
he
pe
r
f
or
m
a
nc
e
of
f
e
r
e
d
by
S
E
H
U
N
E
T
m
a
tc
he
s
a
nd
e
ve
n
e
xc
e
e
d
s
th
e
m
e
th
ods
us
e
d
pr
e
vi
ous
ly
in
ot
he
r
s
tu
di
e
s
w
it
h
m
uc
h
lo
w
e
r
c
om
put
a
ti
ona
l
c
os
t
in
te
r
m
s
of
th
e
num
be
r
of
m
ode
l
pa
r
a
m
e
te
r
s
.
T
hi
s
pe
r
f
or
m
a
nc
e
c
a
n
be
a
c
hi
e
v
e
d
us
in
g
th
e
c
om
bi
na
ti
on
of
g
hos
t
m
odul
e
,
ha
lf
U
-
N
e
t
,
a
nd
SE
bl
oc
k
a
s
th
e
m
a
in
c
om
pone
nt
s
of
th
e
pr
opos
e
d
a
r
c
hi
te
c
tu
r
e
.
A
s
f
or
im
p
r
o
ve
m
e
nt
s
ugge
s
ti
on
in
th
e
f
ut
ur
e
r
e
s
e
a
r
c
h
,
th
e
im
a
ge
s
c
a
n
be
f
ur
th
e
r
s
ubj
e
c
te
d
to
a
ddi
ti
ona
l
a
ugm
e
nt
a
ti
on
s
uc
h
a
s
c
ol
or
c
ha
nne
l
c
r
oppi
ng
a
nd
c
ont
r
a
s
t
li
m
it
e
d
a
da
pt
iv
e
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
(
C
L
A
H
E
)
.
A
n
a
ddi
ti
ona
l
s
e
gm
e
nt
a
ti
on
pr
oc
e
s
s
to
he
lp
th
e
m
ode
l
unde
r
s
ta
nd t
he
di
s
k be
tt
e
r
c
a
n a
ls
o be
a
ppl
ie
d, s
uc
h a
s
s
e
gm
e
nt
a
ti
on f
or
t
he
d
r
us
e
n, t
he
bl
ood ve
s
s
e
ls
, or
e
ve
n
th
e
di
s
k i
ts
e
lf
.
A
C
K
N
O
WL
E
D
G
E
M
E
N
T
S
We
w
oul
d
li
ke
to
th
a
nk
U
ni
ve
r
s
it
a
s
S
e
be
l
a
s
M
a
r
e
t
f
or
pr
ovi
di
n
g
r
e
s
e
a
r
c
h
f
undi
ng
th
r
ough
th
e
H
ib
a
h
R
is
e
t
G
r
oup
s
c
he
m
e
a
s
s
ti
pul
a
te
d
in
c
ont
r
a
c
t
N
o.
194.2/UN27.22/P
T
.01.03/2024
.
W
e
w
oul
d
a
ls
o
li
ke
t
o
th
a
nk
th
e
F
a
c
ul
ty
of
I
nf
or
m
a
ti
on
T
e
c
hnol
ogy
a
nd
D
a
ta
S
c
ie
nc
e
f
or
f
a
c
il
it
a
ti
ng
th
e
f
a
c
il
it
ie
s
a
nd
in
f
r
a
s
tr
uc
tu
r
e
th
a
t
s
uppor
t
th
e
c
om
pl
e
ti
on of
t
hi
s
r
e
s
e
a
r
c
h.
R
E
F
E
R
E
N
C
E
S
[
1]
J
.
S
.
X
i
e
,
L
.
D
on
a
l
ds
on,
a
nd
E
.
M
a
r
gol
i
n,
“
P
a
pi
l
l
e
de
m
a
:
A
r
e
vi
e
w
of
e
t
i
ol
ogy,
pa
t
hophys
i
ol
ogy,
di
a
gno
s
i
s
,
a
nd
m
a
na
ge
m
e
nt
,
”
Sur
v
e
y
of
O
pht
hal
m
ol
ogy
, vol
. 67, no. 4, pp. 1135
–
1159, 2022, doi
:
10.1016/
j
.s
ur
vopht
ha
l
.2021.11.007.
[
2]
J
.
J
.
C
he
n
a
nd
M
.
T
.
B
ha
t
t
i
,
“
P
a
pi
l
l
e
de
m
a
,”
I
nt
e
r
nat
i
onal
O
pht
hal
m
ol
og
y
C
l
i
ni
c
s
,
vol
.
59,
no.
3,
pp.
3
–
22,
2019,
doi
:
10.1097/
I
I
O
.0000000000000274.
[
3]
G
.
L
.
T
r
i
c
k,
S
.
S
.
B
ha
t
t
,
D
.
D
a
hl
,
a
nd
B
.
S
ka
r
f
,
“
O
pt
i
c
di
s
c
t
opogr
a
phy
i
n
ps
e
udopa
pi
l
l
e
de
m
a
:
A
c
om
pa
r
i
s
on
t
o
ps
e
udot
um
or
c
e
r
e
br
i
,”
J
our
nal
of
N
e
ur
o
-
O
pht
hal
m
ol
ogy
, vol
. 21, no. 4, pp. 240
–
244, 2001, d
oi
:
10.1097/
00041327
-
200112000
-
00002.
[
4]
S
.
A
kba
r
,
M
.
U
.
A
kr
a
m
,
M
.
S
ha
r
i
f
,
A
.
T
a
r
i
q,
a
nd
U
.
U.
Y
a
s
i
n,
“
D
e
c
i
s
i
on
s
uppor
t
s
ys
t
e
m
f
or
de
t
e
c
t
i
on
o
f
P
a
pi
l
l
e
de
m
a
t
h
r
ough
f
undus
r
e
t
i
na
l
i
m
a
ge
s
,”
J
our
nal
of
M
e
di
c
al
Sy
s
t
e
m
s
, vol
. 41, no. 4, A
pr
. 2017, d
oi
:
10.1007/
s
10916
-
017
-
0712
-
9.
[
5]
K
.
N
.
F
a
t
i
m
a
,
T
.
H
a
s
s
a
n,
M
.
U
.
A
kr
a
m
,
M
.
A
kht
a
r
,
a
nd
W
.
H
.
B
ut
t
,
“
F
ul
l
y
a
ut
om
a
t
e
d
di
a
gnos
i
s
of
pa
pi
l
l
e
de
m
a
t
hr
ough
r
obus
t
e
xt
r
a
c
t
i
on
of
va
s
c
ul
a
r
pa
t
t
e
r
ns
a
nd
oc
ul
a
r
p
a
t
hol
ogy
f
r
om
f
undus
phot
ogr
a
phs
,”
B
i
om
e
di
c
al
O
pt
i
c
s
E
x
pr
e
s
s
,
vol
.
8,
no.
2,
2017
,
doi
:
10.1364/
boe
.8.001005.
[
6]
R
. G
. B
r
e
r
e
t
on
a
nd G
.
R
.
L
l
oyd, “
S
uppor
t
ve
c
t
or
m
a
c
hi
ne
s
f
or
c
l
a
s
s
i
f
i
c
a
t
i
on a
n
d r
e
gr
e
s
s
i
on,”
A
nal
y
s
t
,
vol
. 135, no.
2, pp. 230
–
267
,
2010, doi
:
10.1039/
b918972f
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Sque
e
z
e
-
e
x
c
it
at
io
n
h
al
f
U
-
N
e
t
and s
y
nt
he
ti
c
m
in
or
it
y
o
v
e
r
s
am
p
li
ng t
e
c
hni
que
ov
e
r
s
am
pl
in
g
…
(
W
ih
ar
to
)
1419
[
7]
J
.
M
.
A
hn,
S
.
K
i
m
,
K
.
S
.
A
hn,
S
.
H
.
C
ho,
a
nd
U
.
S
.
K
i
m
,
“
A
c
c
ur
a
c
y
of
m
a
c
hi
ne
l
e
a
r
ni
ng
f
or
d
i
f
f
e
r
e
nt
i
a
t
i
on
be
t
w
e
e
n
opt
i
c
ne
ur
opa
t
hi
e
s
a
nd ps
e
udop
a
pi
l
l
e
de
m
a
,”
B
M
C
O
pht
hal
m
ol
ogy
, vol
. 19, no. 1, 2019, doi
:
10.1186/
s
12886
-
019
-
1184
-
0.
[
8]
S
.
A
l
ba
w
i
,
T
.
A
.
M
oha
m
m
e
d,
a
nd
S
.
A
l
-
Z
a
w
i
,
“
U
nde
r
s
t
a
ndi
ng
of
a
c
on
vol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k,”
P
r
oc
e
e
di
ngs
of
2017
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on E
ngi
ne
e
r
i
ng and T
e
c
hnol
ogy
, I
C
E
T
2017
, pp. 1
–
6, 2017, doi
:
10.1109/
I
C
E
ngT
e
c
hnol
.2017.8308186.
[
9]
O
.
R
onne
be
r
ge
r
,
P
.
F
i
s
c
he
r
,
a
nd
T
.
B
r
ox,
“
U
-
N
e
t
:
c
onvol
ut
i
ona
l
ne
t
w
or
ks
f
or
bi
om
e
di
c
a
l
i
m
a
ge
s
e
gm
e
nt
a
t
i
on,”
I
E
E
E
A
c
c
e
s
s
,
vol
. 9, pp. 16591
–
16603, M
a
y 2015, doi
:
10.1109/
A
C
C
E
S
S
.2021.3053408.
[
10]
D
.
M
i
l
e
a
e
t
al
.
,
“
A
r
t
i
f
i
c
i
a
l
i
nt
e
l
l
i
ge
nc
e
t
o
de
t
e
c
t
P
a
pi
l
l
e
de
m
a
f
r
om
oc
ul
a
r
f
undu
s
phot
ogr
a
phs
,”
N
e
w
E
ngl
and
J
our
nal
of
M
e
di
c
i
ne
,
vol
. 382, no. 18, pp. 1687
–
1695, 2020, doi
:
10.1056/
ne
j
m
oa
1917130.
[
11]
T
.
S
a
ba
,
S
.
A
kba
r
,
H
.
K
ol
i
va
nd,
a
nd
S
.
A
.
B
a
ha
j
,
“
A
ut
om
a
t
i
c
de
t
e
c
t
i
on
of
pa
pi
l
l
e
de
m
a
t
hr
ough
f
undus
r
e
t
i
na
l
i
m
a
ge
s
us
i
ng
de
e
p
l
e
a
r
ni
ng,”
M
i
c
r
os
c
opy
R
e
s
e
ar
c
h and T
e
c
hni
que
, vol
. 84, no. 12, pp. 3066
–
3077
, 2021, doi
:
10.1002/
j
e
m
t
.23865.
[
12]
M
.
K
okul
u
a
nd
H
.
G
öke
r
,
“
D
e
t
e
c
t
i
on
of
P
a
pi
l
l
e
de
m
a
s
e
ve
r
i
t
y
f
r
om
c
ol
or
f
undus
i
m
a
ge
s
us
i
ng
t
r
a
ns
f
e
r
l
e
a
r
ni
ng
a
ppr
oa
c
he
s
,”
A
k
s
ar
ay
U
ni
v
e
r
s
i
t
y
J
ou
r
nal
of
Sc
i
e
nc
e
and E
ngi
ne
e
r
i
ng
, vol
. 7, no. 2, pp. 53
–
61, 2023, doi
:
10.29002/
a
s
uj
s
e
.1280766.
[
13]
A
.
A
l
-
A
z
z
a
w
i
,
O
.
B
a
ya
t
,
S
.
A
l
-
j
um
a
i
l
i
,
S
.
K
u
r
na
z
,
A
.
D
.
D
u
r
u,
a
nd
O
.
N
.
U
ç
a
n,
“
P
s
e
udopa
pi
l
l
e
de
m
a
d
i
a
gnos
i
s
us
i
ng
de
e
p
t
r
a
ns
f
e
r
l
e
a
r
ni
ng
a
pp
r
oa
c
he
s
,”
2023
7t
h
I
nt
e
r
n
at
i
o
nal
Sy
m
pos
i
um
on
M
ul
t
i
di
s
c
i
pl
i
n
a
r
y
St
udi
e
s
a
nd
I
nnov
a
t
i
v
e
T
e
c
hnol
ogi
e
s
(
I
S
M
S
I
T
)
,
pp. 1
-
6,
2
023.
[
14]
A
.
A
be
da
l
l
a
,
M
.
A
bdul
l
a
h,
M
.
A
l
-
A
yyoub,
a
nd
E
.
B
e
nkhe
l
i
f
a
,
“
C
he
s
t
X
-
r
a
y
pne
um
ot
hor
a
x
s
e
gm
e
nt
a
t
i
on
us
i
ng
U
-
N
e
t
w
i
t
h
E
f
f
i
c
i
e
nt
N
e
t
a
nd R
e
s
N
e
t
a
r
c
hi
t
e
c
t
ur
e
s
,”
P
e
e
r
J
C
om
put
e
r
Sc
i
e
nc
e
, vol
. 7, pp. 1
–
36, 2021, doi
:
10.7717/
pe
e
r
j
-
c
s
.607.
[
15]
H
.
L
u,
Y
.
S
he
,
J
.
T
i
e
,
a
nd
S
.
X
u,
“
H
a
l
f
-
U
N
e
t
:
a
s
i
m
pl
i
f
i
e
d
U
-
N
e
t
a
r
c
hi
t
e
c
t
ur
e
f
or
m
e
di
c
a
l
i
m
a
ge
s
e
gm
e
nt
a
t
i
on,”
F
r
ont
i
e
r
s
i
n
N
e
ur
oi
nf
or
m
at
i
c
s
, vol
. 16, 2022, doi
:
10.3389/
f
ni
nf
.2022.911679.
[
16]
K
.
S
i
m
onya
n
a
nd
A
.
Z
i
s
s
e
r
m
a
n,
“
V
e
r
y
de
e
p
c
onvol
ut
i
ona
l
ne
t
w
or
ks
f
or
l
a
r
ge
-
s
c
a
l
e
i
m
a
ge
r
e
c
ogni
t
i
on,”
3r
d
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on L
e
ar
ni
ng R
e
pr
e
s
e
nt
at
i
ons
, I
C
L
R
2015
-
C
onf
e
r
e
nc
e
T
r
ac
k
P
r
oc
e
e
di
ngs
, 2015.
[
17]
K
.
H
e
,
X
.
Z
ha
ng,
S
.
R
e
n,
a
nd
J
.
S
un,
“
D
e
e
p
r
e
s
i
du
a
l
l
e
a
r
ni
ng
f
or
i
m
a
ge
r
e
c
og
ni
t
i
on,”
2016
I
E
E
E
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on
and P
at
t
e
r
n R
e
c
ogni
t
i
on (
C
V
P
R
)
, pp. 770
–
778, 2016, doi
:
10.1109/
C
V
P
R
.2016
.90.
[
18]
C
.
S
z
e
ge
dy
e
t
al
.
,
“
G
oi
ng
de
e
pe
r
w
i
t
h
c
onvol
ut
i
ons
,”
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
So
c
i
e
t
y
C
onf
e
r
e
nc
e
on
C
o
m
put
e
r
V
i
s
i
o
n
and P
at
t
e
r
n R
e
c
ogni
t
i
on
, pp. 1
–
9, 2015, doi
:
10.1109/
C
V
P
R
.2015.7298594.
[
19]
N
. V
. C
ha
w
l
a
, K
. W
.
B
ow
ye
r
,
L
. O
.
H
a
l
l
, a
nd
W
. P
. K
e
ge
l
m
e
y
e
r
, “
S
M
O
T
E
:
S
ynt
he
t
i
c
m
i
nor
i
t
y ove
r
-
s
a
m
pl
i
ng t
e
c
hni
que
,”
J
ou
r
nal
of
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
R
e
s
e
ar
c
h
, vol
. 16, pp. 321
–
357, 2002, doi
:
10.1613/
j
a
i
r
.953.
[
20]
J
.
H
u,
L
.
S
he
n,
a
nd
G
.
S
un,
“
S
que
e
z
e
-
a
nd
-
e
xc
i
t
a
t
i
on
ne
t
w
or
ks
,”
2018
I
E
E
E
/
C
V
F
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on
and
P
at
t
e
r
n
R
e
c
ogni
t
i
on
, pp. 7132
–
7141, 2018, doi
:
10.1109/
C
V
P
R
.2018.00745.
[
21]
S
.
X
i
e
,
R
.
G
i
r
s
hi
c
k,
P
.
D
ol
l
á
r
,
Z
.
T
u,
a
nd
K
.
H
e
,
“
A
ggr
e
ga
t
e
d
r
e
s
i
dua
l
t
r
a
ns
f
or
m
a
t
i
ons
f
or
de
e
p
ne
ur
a
l
ne
t
w
or
ks
,”
2017
I
E
E
E
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on and P
at
t
e
r
n R
e
c
ogni
t
i
on (
C
V
P
R
)
, pp. 5987
–
59
95, 2017, doi
:
10.1109/
C
V
P
R
.2017.634.
[
22]
L
.
R
undo
e
t
al
.
,
“
U
S
E
-
N
e
t
:
i
nc
or
por
a
t
i
ng
s
que
e
z
e
-
a
nd
-
e
xc
i
t
a
t
i
on
bl
oc
ks
i
nt
o
U
-
N
e
t
f
or
pr
os
t
a
t
e
z
ona
l
s
e
gm
e
nt
a
t
i
on
of
m
ul
t
i
-
i
ns
t
i
t
ut
i
ona
l
M
R
I
da
t
a
s
e
t
s
,”
N
e
u
r
oc
om
put
i
ng
, vol
. 365, pp. 31
–
43, 2019, doi
:
1
0.1016/
j
.ne
uc
om
.2019.07.006.
[
23]
K
.
H
a
n,
Y
.
W
a
ng,
Q
.
T
i
a
n,
J
.
G
uo,
C
.
X
u,
a
nd
C
.
X
u,
“
G
hos
t
N
e
t
:
M
or
e
f
e
a
t
ur
e
s
f
r
om
c
he
a
p
ope
r
a
t
i
ons
,”
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
n
c
e
on
C
om
put
e
r
V
i
s
i
on
and
P
at
t
e
r
n
R
e
c
ogni
t
i
on
,
pp.
1577
–
1586,
2020,
doi
:
10.1109/
C
V
P
R
42600.2020.00165.
[
24]
D
.
P
.
K
i
ngm
a
a
nd
J
.
L
.
B
a
,
“
A
da
m
:
A
m
e
t
hod
f
or
s
t
oc
h
a
s
t
i
c
opt
i
m
i
z
a
t
i
on,”
3r
d
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
L
e
a
r
ni
ng
R
e
pr
e
s
e
nt
at
i
ons
, I
C
L
R
2015
-
C
onf
e
r
e
nc
e
T
r
ac
k
P
r
o
c
e
e
di
ngs
, 2015.
[
25]
A
.
P
.
B
r
a
dl
e
y,
“
T
he
us
e
of
t
he
a
r
e
a
unde
r
t
he
R
O
C
c
ur
ve
i
n
t
he
e
va
l
ua
t
i
on
of
m
a
c
hi
ne
l
e
a
r
ni
ng
a
l
gor
i
t
hm
s
,”
P
at
t
e
r
n
R
e
c
ogni
t
i
on
,
vol
. 30, no. 7, pp. 1145
–
1159, 1997, doi
:
10.1016/
S
0031
-
3203(
96
)
00142
-
2.
[
26]
A
.
T
ha
r
w
a
t
,
“
C
l
a
s
s
i
f
i
c
a
t
i
on
a
s
s
e
s
s
m
e
nt
m
e
t
hod
s
,”
A
ppl
i
e
d
C
o
m
put
i
ng
and
I
nf
or
m
at
i
c
s
,
vol
.
17,
no.
1,
pp.
168
–
192,
2018,
doi
:
10.1016/
j
.a
c
i
.2018.08.003.
B
I
O
G
R
A
P
H
I
E
S
O
F
A
U
T
H
O
R
S
Wiharto
received
obtained
a
bachelor’s
degree
in
electrical
engineering
fro
m
Universitas
Telkom,
Indonesia,
in
1999.
He
obtained
a
master’s
deg
ree
in
computer
science
from
Universitas
Gadjah
Mada,
Indonesia,
in
2004
and
a
d
octoral
degree
from
the
same
University,
in
2017.
Currently
he
works
as
a
lecturer
in
the
Department
of
Informatics,
Faculty
of
Infor
mation
Techn
ology
and
Data
Sci
ence
,
Univer
sitas
S
ebela
s
Maret
,
Suraka
rta,
Indonesia.
His
experience
and
areas
of
interest
focus
on
artificial
int
elligence,
c
omputational
intelligence
,
medical
imaging,
and
expert
systems.
H
e
can
b
e
co
ntacted
at
email
:
wiharto@
staff.uns.ac.id.
Angga
Exca
Pradipta
Syaifuddin
holds
Bachelor
of
Computer
S
cience
degree
with
3.79
GPA.
He
received
some
recognition
in
the
startup
field
suc
h
as
fundings
while
still
being
an
undergraduate.
He
holds
two
intellectual
properties
as
the
i
nventor
of
his
works
i
n
the startup
he founde
d
.
He can be contacted at email:
excaangga@
student.uns.ac.i
d.
Evaluation Warning : The document was created with Spire.PDF for Python.