I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
11
,
No
.
4
,
A
u
g
u
s
t
2021
,
p
p
.
3
4
9
2
~
3
5
0
1
I
SS
N:
2
0
8
8
-
8708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v
1
1
i
4
.
pp
3
4
9
2
-
3
5
0
1
3492
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
A t
ra
nsf
er learni
ng
w
ith
deep
neu
ra
l net
w
o
rk
appr
o
a
ch f
o
r
dia
betic
r
etinopa
t
hy
clas
sifica
tion
M
o
ha
m
m
e
d Al
-
S
m
a
di,
M
a
hm
o
u
d
H
a
mm
a
d,
Q
a
nita
B
a
ni B
a
k
er
,
Sa
’
a
d A.
Al
-
Z
bo
o
n
Co
ll
e
g
e
o
f
Co
m
p
u
ter an
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
y
,
Jo
rd
a
n
Un
iv
e
rsity
o
f
S
c
ien
c
e
a
n
d
T
e
c
h
n
o
lo
g
y
,
Irb
id
,
Jo
r
d
a
n
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
l
20
,
2
0
2
0
R
ev
i
s
ed
Dec
9
,
2
0
2
0
A
cc
ep
ted
J
an
1
3
,
2
0
2
1
Dia
b
e
ti
c
re
ti
n
o
p
a
th
y
is
a
n
e
y
e
d
ise
a
se
c
a
u
se
d
b
y
h
ig
h
b
lo
o
d
s
u
g
a
r
a
n
d
p
re
ss
u
re
w
h
ich
d
a
m
a
g
e
s
th
e
b
lo
o
d
v
e
ss
e
ls
in
th
e
e
y
e
.
Dia
b
e
ti
c
r
e
ti
n
o
p
a
t
h
y
i
s
th
e
ro
o
t
c
a
u
se
o
f
m
o
r
e
th
a
n
1
%
o
f
th
e
b
li
n
d
n
e
ss
w
o
rld
w
id
e
.
Early
d
e
tec
ti
o
n
o
f
th
is
d
ise
a
se
is
c
ru
c
ial
a
s
it
p
re
v
e
n
ts
it
f
ro
m
p
ro
g
re
ss
in
g
to
a
m
o
re
se
v
e
r
e
lev
e
l.
H
o
w
e
v
e
r
,
th
e
c
u
rre
n
t
m
a
c
h
in
e
lea
rn
in
g
-
b
a
se
d
a
p
p
r
o
a
c
h
e
s
f
o
r
d
e
tec
ti
n
g
th
e
se
v
e
rit
y
le
v
e
l
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
a
re
e
it
h
e
r,
i)
re
l
y
o
n
m
a
n
u
a
ll
y
e
x
trac
ti
n
g
f
e
a
tu
re
s
w
h
ich
m
a
k
e
s
a
n
a
p
p
ro
a
c
h
u
n
p
ra
c
ti
c
a
l,
o
r
ii
)
t
ra
in
e
d
o
n
s
m
a
ll
d
a
tas
e
t
th
u
s c
a
n
n
o
t
b
e
g
e
n
e
r
a
li
z
e
d
.
In
th
is stu
d
y
,
w
e
p
ro
p
o
se
a
tran
s
f
e
r
lea
rn
in
g
-
b
a
se
d
a
p
p
r
o
a
c
h
f
o
r
d
e
t
e
c
ti
n
g
th
e
se
v
e
rit
y
le
v
e
l
o
f
th
e
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
w
it
h
h
ig
h
a
c
c
u
ra
c
y
.
Ou
r
m
o
d
e
l
is
a
d
e
e
p
lea
rn
in
g
m
o
d
e
l
b
a
se
d
o
n
g
lo
b
a
l
a
v
e
r
a
g
e
p
o
o
li
n
g
(G
A
P
)
tec
h
n
i
q
u
e
w
it
h
v
a
rio
u
s
p
re
-
train
e
d
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
t
-
w
o
rk
(CNN
)
m
o
d
e
ls.
T
h
e
e
x
p
e
ri
m
e
n
tal
re
su
lt
s
o
f
o
u
r
a
p
p
r
o
a
c
h
,
i
n
w
h
ich
o
u
r
b
e
st
m
o
d
e
l
a
c
h
iev
e
d
8
2
.
4
%
q
u
a
d
ra
ti
c
w
e
ig
h
ted
k
a
p
p
a
(QW
K),
c
o
rro
b
o
ra
te
th
e
a
b
il
it
y
o
f
o
u
r
m
o
d
e
l
to
d
e
tec
t
th
e
se
v
e
rit
y
lev
e
l
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
e
ff
ici
e
n
tl
y
.
K
ey
w
o
r
d
s
:
Dee
p
lear
n
in
g
Diab
etic
r
etin
o
p
ath
y
I
m
ag
e
cla
s
s
i
f
icatio
n
Me
d
ical
i
m
a
g
e
p
r
o
ce
s
s
in
g
T
r
an
s
f
er
lear
n
in
g
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
:
Mo
h
a
m
m
ed
A
l
-
S
m
ad
i
C
o
lleg
e
o
f
C
o
m
p
u
ter
an
d
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
J
o
r
d
an
Un
i
v
er
s
it
y
o
f
Scien
ce
a
n
d
T
ec
h
n
o
lo
g
y
I
r
b
id
,
J
o
r
d
an
2
2
1
1
0
E
m
ail:
m
as
m
ad
i
@
j
u
s
t.e
d
u
.
j
o
1.
I
NT
RO
D
UCT
I
O
N
Diab
etic
r
etin
o
p
ath
y
i
s
an
e
y
e
d
is
ea
s
e
ca
u
s
ed
b
y
h
ig
h
b
lo
o
d
s
u
g
ar
an
d
p
r
ess
u
r
e
w
h
ich
d
a
m
ag
e
s
t
h
e
b
lo
o
d
v
ess
els
i
n
t
h
e
b
ac
k
o
f
t
h
e
e
y
e.
B
ased
o
n
[
1
]
,
ar
o
u
n
d
4
0
.
3
%
o
f
Un
i
ted
State
s
ad
u
lts
4
0
y
ea
r
s
a
n
d
o
ld
er
s
u
f
f
er
f
r
o
m
r
etin
o
p
a
t
h
y
w
ith
8
.
2
%
h
a
v
e
v
is
io
n
-
t
h
r
ea
te
n
in
g
r
et
in
o
p
ath
y
.
D
iab
etic
r
etin
o
p
at
h
y
is
t
h
e
r
o
o
t
ca
u
s
e
o
f
m
o
r
e
t
h
an
1%
of
th
e
b
lin
d
n
es
s
w
o
r
ld
w
id
e.
P
eo
p
le
w
it
h
Diab
etic
r
eti
n
o
p
ath
y
ar
e
at
t
h
e
g
r
ea
t
r
is
k
of
d
e
v
elo
p
in
g
o
th
er
e
y
e
d
is
ea
s
e
s
s
u
c
h
a
s
g
la
u
co
m
a
a
n
d
C
atar
ac
ts
.
T
h
i
s
d
is
ea
s
e
is
p
r
o
g
r
ess
i
v
e
m
ea
n
in
g
t
h
at
it
ad
v
a
n
ce
s
f
r
o
m
o
n
e
s
ta
g
e
to
a
m
o
r
e
s
er
io
u
s
s
tag
e,
i
f
it
h
as
n
o
t
tr
ea
ted
w
ell
.
E
ar
ly
d
etec
t
io
n
w
it
h
e
f
f
ec
ti
v
e
tr
ea
t
m
en
t
o
f
t
h
e
d
iab
etic
r
etin
o
p
ath
y
ca
n
r
ed
u
c
e
v
is
io
n
lo
s
s
b
y
9
0
%
[
2
]
.
T
o
o
v
er
co
m
e
t
h
e
af
o
r
e
m
e
n
tio
n
ed
p
r
o
b
lem
an
d
d
etec
t
d
iab
etic
r
etin
o
p
ath
y
ea
r
l
y
an
d
e
f
f
icien
tl
y
,
w
e
h
av
e
d
e
v
elo
p
ed
a
d
ee
p
lear
n
i
n
g
m
o
d
el
th
at
i
s
ca
p
ab
le,
w
it
h
h
i
g
h
ac
cu
r
ac
y
,
to
d
etec
t
if
an
e
y
e
s
u
f
f
er
s
f
r
o
m
d
iab
etic
r
etin
o
p
ath
y
o
r
n
o
t.
I
f
th
e
e
y
e
s
u
f
f
er
s
f
r
o
m
d
iab
etic
r
etin
o
p
ath
y
,
o
u
r
m
o
d
el
d
etec
t
s
th
e
s
e
v
er
it
y
lev
el
o
f
th
e
d
is
ea
s
e
an
d
h
en
ce
p
r
ev
e
n
ti
n
g
th
e
d
is
ea
s
e
f
r
o
m
p
r
o
g
r
es
s
in
g
.
Ma
c
h
i
n
e
lear
n
in
g
a
n
d
m
ain
l
y
d
ee
p
lear
n
i
n
g
h
av
e
i
m
p
r
o
v
ed
d
r
asti
ca
ll
y
d
u
r
in
g
t
h
e
p
ast
d
ec
ad
e
[
3
]
.
De
ep
lear
n
in
g
alg
o
r
it
h
m
s
ad
v
a
n
ce
d
m
a
n
y
r
e
s
ea
r
ch
f
ield
s
s
u
c
h
as sp
ee
c
h
r
ec
o
g
n
it
i
o
n
,
d
ec
is
io
n
m
a
k
i
n
g
,
an
d
i
m
a
g
e
p
r
o
ce
s
s
in
g
.
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
et
w
o
r
k
(
C
NN)
is
a
d
ee
p
n
eu
r
al
n
et
w
o
r
k
m
o
d
el
t
h
at
is
w
id
el
y
u
s
ed
in
c
o
m
p
u
ter
v
is
io
n
a
n
d
i
m
ag
e
c
lass
i
f
icatio
n
s
.
C
NN
co
n
s
i
s
ts
of
t
h
r
ee
m
ai
n
co
m
p
o
n
e
n
t
s
:
i
)
s
i
n
g
le
or
m
u
l
tip
le
c
o
n
v
o
l
u
tio
n
a
l
b
lo
ck
s
w
h
ic
h
is
a
ce
n
tr
al
co
m
p
o
n
e
n
t
of
C
NN,
ii
)
s
a
m
p
l
in
g
la
y
er
s
(
p
o
o
lin
g
la
y
er
s)
s
u
c
h
as
m
ax
-
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
A
tr
a
n
s
fer lea
r
n
in
g
w
ith
d
ee
p
n
eu
r
a
l n
etw
o
r
k
a
p
p
r
o
a
c
h
fo
r
…
(
Mo
h
a
mme
d
A
l
-
S
ma
d
i
)
3493
s
a
m
p
li
n
g
an
d
m
ea
n
-
s
a
m
p
lin
g
,
an
d
iii
)
n
u
m
b
er
o
f
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
.
I
m
ag
e
clas
s
if
icat
io
n
ca
n
b
e
d
ef
i
n
ed
as
th
e
p
r
o
ce
s
s
of
lab
eli
n
g
i
m
a
g
es
w
it
h
a
ca
teg
o
r
y
f
r
o
m
a
p
r
ed
ef
in
ed
s
et
of
ca
te
g
o
r
ies.
T
h
e
p
r
o
ce
s
s
of
im
a
g
e
class
i
f
icatio
n
co
n
s
i
s
ts
o
f
m
a
n
y
p
h
a
s
es
s
tar
tin
g
f
r
o
m
co
ll
ec
tin
g
a
d
ataset
o
f
i
m
a
g
e
s
,
lab
elin
g
t
h
e
m
,
p
r
e
-
p
r
o
ce
s
s
in
g
t
h
e
i
m
a
g
es,
i
m
a
g
e
s
eg
m
e
n
tatio
n
,
f
ea
t
u
r
es
e
x
tr
ac
ti
o
n
,
an
d
f
i
n
all
y
,
o
b
j
ec
t
class
if
i
ca
tio
n
u
s
i
n
g
a
d
ee
p
lear
n
in
g
m
o
d
el
[
4
]
.
Ma
n
y
r
ese
ar
ch
er
s
b
u
i
lt
v
ar
io
u
s
d
ee
p
lear
n
in
g
ar
c
h
itect
u
r
es
b
a
s
ed
o
n
C
NN
s
u
ch
as
L
e
Net
-
5
[
5
]
,
A
lex
Net
[
6
]
,
Z
FNet
[
7
]
,
VGGN
et
[
8
]
,
Go
o
g
leNe
t
[
9
]
,
R
esNet
[
1
0
]
,
I
n
ce
p
tio
n
V2
[
1
1
]
,
I
n
ce
p
tio
n
V3
[
1
2
]
,
I
n
ce
p
tio
n
R
es
Net
(
I
n
cp
tio
n
V4
)
[
1
3
]
,
Den
s
eNe
t
[
1
4
]
,
Gap
Net
[
1
5
]
,
SNet
[
1
6
]
,
Xce
p
tio
n
[
1
7
]
,
E
f
f
icien
tNet
[
1
8
]
.
T
h
ese
d
ee
p
lear
n
in
g
ar
ch
i
tectu
r
e
s
ar
e
u
s
ed
f
o
r
b
u
ild
in
g
v
ar
io
u
s
d
ee
p
lear
n
in
g
m
o
d
el
s
.
In
th
is
r
esear
c
h
,
we
h
a
v
e
d
esig
n
ed
an
d
d
ev
elo
p
ed
6
d
if
f
er
en
t
tr
an
s
f
er
lear
n
in
g
tech
n
iq
u
es
to
d
etec
t
t
h
e
s
ev
er
it
y
lev
e
l
o
f
t
h
e
d
iab
etic
r
etin
o
p
ath
y
to
s
to
p
b
lin
d
n
e
s
s
b
ef
o
r
e
it
is
to
o
late.
T
h
e
p
r
e
-
tr
a
in
ed
m
o
d
els
ar
e:
i
)
R
esNet
[
1
0
]
,
ii
)
I
n
ce
p
tio
n
V3
[
1
2
]
,
iii
)
I
n
ce
p
tio
n
R
esNet
(
I
n
c
ep
tio
n
V4
)
[
1
3
]
,
i
v
)
Den
s
eNe
t
[
1
4
]
,
(
v
)
Xce
p
tio
n
[
1
7
]
,
an
d
vi
)
E
f
f
icie
n
tNet
[
1
8
]
.
T
h
e
p
r
o
p
o
s
ed
tech
n
iq
u
e
s
w
er
e
tr
ain
ed
an
d
ev
alu
ated
o
n
r
ea
l
-
w
o
r
ld
m
ed
ical
i
m
a
g
es
[
1
9
]
.
E
ac
h
i
m
ag
e
i
n
t
h
e
tr
ain
in
g
d
ataset
i
s
m
a
n
u
all
y
l
ab
eled
w
ith
i
ts
s
e
v
er
it
y
lev
el
b
y
a
clin
icia
n
.
T
o
s
u
m
m
ar
ize,
t
h
is
p
ap
er
m
ak
es th
e
f
o
llo
w
in
g
co
n
tr
ib
u
tio
n
s
:
I
n
n
o
v
a
tiv
e
tr
a
n
s
f
er
lear
n
i
n
g
m
o
d
el:
w
e
h
av
e
le
v
er
ag
ed
v
ar
io
u
s
s
ta
te
-
of
-
th
e
-
ar
t
C
N
N
ar
ch
it
ec
tu
r
es
to
b
u
ild
v
ar
io
u
s
tr
an
s
f
er
lear
n
i
n
g
-
b
ase
d
m
o
d
els.
T
h
e
C
NN
ar
c
h
itect
u
r
es
h
av
e
b
ee
n
u
s
ed
as
p
r
e
-
tr
ain
ed
m
o
d
els
to
o
u
r
m
o
d
els.
C
NN
-
b
ased
m
o
d
e
ls
h
a
v
e
b
ee
n
u
s
ed
s
u
cc
e
s
s
f
u
ll
y
in
i
m
a
g
e
cla
s
s
i
f
icatio
n
tas
k
s
.
T
h
eo
r
y
:
i
n
th
i
s
p
ap
er
,
w
e
s
h
o
w
t
h
at
lev
er
a
g
i
n
g
tr
an
s
f
er
l
ea
r
n
in
g
i
m
p
r
o
v
es
t
h
e
p
er
f
o
r
m
an
ce
o
f
d
ee
p
lear
n
in
g
m
o
d
el
s
an
d
in
cr
ea
s
es
its
d
etec
tio
n
ac
cu
r
ac
y
.
E
x
p
er
i
m
e
n
ts
:
w
e
h
av
e
co
n
d
u
c
ted
s
ev
er
al
e
x
p
er
i
m
e
n
ts
o
n
a
l
ar
g
e
m
ed
ical
i
m
a
g
e
d
ata
s
et.
O
u
r
ex
p
er
i
m
en
ta
l
r
esu
lt
s
s
h
o
w
t
h
e
h
i
g
h
ab
ilit
y
o
f
o
u
r
m
o
d
el
f
o
r
d
etec
tin
g
t
h
e
s
ev
er
it
y
lev
el
o
f
d
iab
etic
r
etin
o
p
ath
y
d
is
ea
s
e
.
I
n
ad
d
itio
n
,
w
e
co
m
p
ar
ed
th
e
p
er
f
o
r
m
a
n
ce
o
f
s
ix
d
i
f
f
er
en
t tr
an
s
f
er
lear
n
i
n
g
-
b
ased
m
o
d
el
s
.
T
h
e
r
em
ai
n
d
er
o
f
t
h
i
s
p
ap
er
is
o
r
g
an
ized
as
f
o
llo
w
s
.
Sect
io
n
2
p
r
o
v
id
es
a
n
o
v
er
v
ie
w
o
f
t
h
e
r
ela
ted
w
o
r
k
i
n
th
e
m
ed
ical
i
m
a
g
e
p
r
o
ce
s
s
in
g
f
iled
.
Sectio
n
3
d
escr
ib
es
o
u
r
m
et
h
o
d
to
d
esig
n
a
n
d
d
ev
elo
p
a
d
e
e
p
lear
n
in
g
m
o
d
el
to
d
etec
t
th
e
s
ev
er
it
y
of
a
d
iab
etic
r
etin
o
p
ath
y
e
y
e.
Sectio
n
4
p
r
esen
ts
o
u
r
ex
p
er
i
m
e
n
tal
r
esu
lts
an
d
f
i
n
d
in
g
s
a
n
d
w
e
d
is
c
u
s
s
t
h
e
m
i
n
s
ec
t
io
n
5
.
Fin
all
y
,
t
h
e
p
ap
er
co
n
clu
d
es
w
it
h
av
e
n
u
es
o
f
f
u
tu
r
e
w
o
r
k
o
n
s
ec
tio
n
6.
2.
RE
L
A
T
E
D
WO
RK
Me
d
ical
i
m
a
g
e
cla
s
s
i
f
icatio
n
i
s
o
n
e
o
f
t
h
e
m
o
s
t
i
m
p
o
r
tan
t
i
m
ag
e
s
p
r
o
ce
s
s
i
n
g
tas
k
s
.
I
t
s
m
ain
g
o
al
i
s
to
class
i
f
y
m
ed
ical
i
m
a
g
es
in
t
o
d
if
f
er
en
t
ca
teg
o
r
ies
to
h
elp
p
h
y
s
icia
n
s
a
n
d
clin
ician
s
to
d
iag
n
o
s
e
p
atien
t
s
f
a
s
t
e
r
[
2
0
]
.
P
h
y
s
icia
n
s
r
el
y
o
n
t
h
eir
p
r
ac
tical
ex
p
er
ien
ce
a
s
w
ell
as
m
a
n
u
a
ll
y
s
p
o
tti
n
g
o
f
v
ar
io
u
s
f
ea
t
u
r
es i
n
a
n
im
ag
e
to
d
eter
m
i
n
e
it
s
m
ed
ical
co
n
d
itio
n
.
Su
c
h
a
p
r
o
ce
s
s
is
er
r
o
r
p
r
o
n
e
an
d
ted
io
u
s
task
.
T
h
e
r
ef
o
r
e,
th
e
m
ed
ical
i
m
a
g
e
class
i
f
icat
io
n
e
m
er
g
ed
to
h
elp
p
h
y
s
icia
n
s
cla
s
s
i
f
y
m
ed
ical
i
m
a
g
es
f
aster
an
d
m
o
r
e
co
n
v
e
n
ie
n
tl
y
.
To
k
e
e
p
th
e
p
ap
er
co
n
ci
s
e
an
d
r
ea
d
ab
le,
we
w
i
ll
co
m
p
ar
e
a
n
d
co
n
tr
ast
b
et
w
ee
n
o
u
r
r
esear
ch
w
o
r
k
in
lig
h
t
of
t
h
e
r
elate
d
r
esear
ch
ef
f
o
r
t in
d
etec
ti
n
g
d
ia
b
etic
r
etin
o
p
ath
y
u
s
i
n
g
m
ac
h
i
n
e
an
d
d
ee
p
lear
n
in
g
m
eth
o
d
s
.
Diab
etic
r
etin
o
p
at
h
y
is
o
n
e
o
f
th
e
ey
e’
s
d
is
ea
s
es
th
a
t
i
s
t
h
e
r
o
o
t
ca
u
s
e
o
f
b
li
n
d
n
e
s
s
ar
o
u
n
d
th
e
w
o
r
ld
.
Dete
ctin
g
th
e
s
ev
er
it
y
le
v
el
of
d
iab
etic
r
etin
o
p
ath
y
e
y
e
ea
r
l
y
is
cr
u
cial
f
o
r
p
r
ev
en
ti
n
g
p
o
s
s
i
b
le
ad
v
an
ce
m
e
n
t
o
f
th
is
d
is
ea
s
e.
D
u
e
to
th
e
i
m
p
o
r
tan
ce
of
th
is
p
r
o
b
le
m
,
m
a
n
y
r
esear
ch
er
s
h
a
v
e
d
ev
elo
p
ed
v
ar
io
u
s
m
ac
h
i
n
e
lear
n
in
g
tec
h
n
iq
u
es
f
o
r
d
etec
tin
g
d
iab
etic
r
eti
n
o
p
ath
y
i
n
cl
u
d
i
n
g
[
2
1
-
3
1
]
.
Ou
r
r
esear
c
h
s
h
ar
es
w
it
h
th
e
p
r
ev
io
u
s
r
e
s
ea
r
ch
ef
f
o
r
t
t
h
e
id
ea
o
f
d
ete
cti
n
g
d
iab
etic
r
et
in
o
p
ath
y
b
u
t
it
is
s
i
g
n
i
f
ica
n
tl
y
d
if
f
er
en
t.
Fo
r
ex
a
m
p
le,
o
u
r
d
ataset
co
n
tai
n
s
3
,
5
6
2
o
r
ig
in
al
i
m
a
g
es
w
h
er
ea
s
m
a
n
y
p
r
e
v
i
o
u
s
w
o
r
k
tr
ain
ed
th
eir
m
o
d
el
o
n
a
s
m
all
d
ata
s
et
w
i
th
les
s
t
h
a
n
5
0
0
i
m
ag
e
s
s
u
c
h
a
s
[
2
1
-
2
7
]
.
Ot
h
er
r
esear
ch
w
o
r
k
tr
ain
ed
th
eir
m
o
d
el
s
o
n
a
b
ig
g
er
d
ataset
s
u
ch
a
s
[
2
8
]
an
d
[
2
9
]
w
it
h
1
,
2
0
0
im
a
g
es
a
n
d
[
3
0
-
3
2
]
w
i
t
h
ar
o
u
n
d
3
5
,
0
0
0
im
a
g
es.
Ne
v
er
t
h
eles
s
,
o
u
r
r
esear
ch
w
o
r
k
o
u
tp
er
f
o
r
m
ed
th
e
s
e
r
esear
ch
e
f
f
o
r
ts
in
m
a
n
y
w
a
y
s
.
F
o
r
ex
a
m
p
le,
[
2
8
]
an
d
[
2
9
]
u
s
ed
tr
ad
itio
n
al
m
ac
h
i
n
e
lear
n
i
n
g
s
u
c
h
as
SVM
a
n
d
A
d
aB
o
o
s
t.
I
n
[
3
0
]
an
d
[
3
1
]
u
s
ed
a
s
in
g
le
C
N
N
m
o
d
el.
Si
m
ilar
l
y
,
[
3
2
]
u
s
ed
m
ai
n
l
y
t
w
o
d
if
f
er
e
n
t
m
o
d
els
an
d
th
eir
b
est
o
b
tain
ed
m
o
d
el
ac
h
iev
ed
a
k
ap
p
a
s
co
r
e
of
0
.
7
2
.
H
o
w
e
v
e
r
,
in
th
i
s
r
esear
ch
,
we
h
a
v
e
u
tili
ze
d
7
d
if
f
er
e
n
t
s
tate
-
of
-
t
h
e
-
ar
t
d
ee
p
lear
n
in
g
m
o
d
el
s
.
Fin
a
ll
y
,
w
e
h
a
v
e
d
ev
e
lo
p
ed
a
t
r
an
s
f
er
d
ee
p
lear
n
in
g
m
o
d
el
a
n
d
o
u
r
r
esu
lts
o
u
tp
er
f
o
r
m
ed
th
e
p
r
ev
io
u
s
r
esear
ch
ef
f
o
r
ts
.
3.
RE
S
E
ARCH
M
E
T
H
O
D
3
.
1
.
T
he
da
t
a
s
et
T
o
tr
ain
an
d
ev
alu
a
te
o
u
r
d
ee
p
lear
n
in
g
m
o
d
el,
w
e
h
a
v
e
u
tili
ze
d
th
e
d
ataset
a
v
ailab
le
f
o
r
th
e
A
P
T
O
S
2
0
1
9
B
lin
d
n
ess
Dete
c
tio
n
Kag
g
le
co
m
p
et
itio
n
[
1
9
]
.
T
h
e
d
ata
s
et
i
s
a
r
ea
l
-
w
o
r
ld
d
ataset
o
b
ta
in
ed
f
r
o
m
m
u
l
tip
le
clin
ic
s
i
n
I
n
d
ia
u
s
i
n
g
d
i
f
f
er
e
n
t
ca
m
er
as
o
v
er
a
p
er
io
d
o
f
ti
m
e.
T
h
e
i
m
ag
e
s
ar
e
lab
eled
b
y
e
x
p
er
ts
.
H
o
w
e
v
e
r
,
th
e
y
m
i
g
h
t c
o
n
tai
n
s
o
m
e
n
o
is
e
in
b
o
th
th
e
i
m
a
g
es a
n
d
th
e
lab
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
4
,
A
u
g
u
s
t 2
0
2
1
:
3
4
9
2
-
3501
3494
T
h
e
clin
ic
s
lab
eled
t
h
e
i
m
a
g
e
s
in
t
h
e
d
ataset
w
it
h
th
e
s
e
v
er
it
y
lev
el
of
t
h
e
d
iab
etic
r
eti
n
o
p
ath
y
s
t
a
r
t
i
n
g
f
r
o
m
n
o
r
m
al
e
y
e
i
m
a
g
e
to
p
r
o
lif
er
ati
v
e
d
iab
etic
r
etin
o
p
ath
y
e
y
e
i
m
ag
e.
T
a
b
l
e
1
s
h
o
w
s
th
e
la
b
els o
f
t
h
e
im
a
g
es
in
th
e
d
ataset
alo
n
g
w
it
h
th
e
n
u
m
b
er
o
f
i
m
ag
e
s
th
at
b
elo
n
g
to
ea
ch
lab
el.
T
h
e
tab
le
s
h
o
w
s
th
at
th
e
d
ataset
h
a
s
t
w
o
p
r
o
b
le
m
s
.
First,
t
h
e
d
atas
et
is
u
n
b
ala
n
ce
d
.
T
h
e
i
m
a
g
es
th
at
b
elo
n
g
to
t
h
e
“
NO
DR
”
c
lass
ar
e
m
o
r
e
th
a
n
h
al
f
of
t
h
e
d
ataset.
T
h
er
ef
o
r
e,
if
we
t
r
ai
n
a
clas
s
if
ier
on
t
h
is
d
ataset,
th
e
cla
s
s
i
f
ier
w
i
ll
be
b
ias
to
w
ar
d
t
h
is
cla
s
s
.
Seco
n
d
,
th
e
d
ataset
is
r
elati
v
el
y
s
m
al
l
f
o
r
d
ee
p
lear
n
in
g
tas
k
s
.
T
o
s
o
lv
e
th
e
f
ir
s
t p
r
o
b
le
m
,
i
m
b
a
lan
ce
d
d
ata,
we
lev
er
ag
ed
a
d
ata
-
o
v
er
s
a
m
p
lin
g
tech
n
iq
u
e.
T
o
o
v
er
co
m
e
t
h
e
s
ec
o
n
d
p
r
o
b
lem
,
s
m
all
d
ataset,
w
e
u
s
ed
a
d
a
t
a
au
g
m
e
n
tatio
n
tech
n
iq
u
e.
Ne
x
t
s
ec
tio
n
d
escr
ib
es
t
h
ese
t
w
o
tech
n
iq
u
es
i
n
m
o
r
e
d
etail.
Fig
u
r
e
1
s
h
o
w
s
a
n
illu
s
tr
ativ
e
e
x
a
m
p
le
o
f
ea
c
h
d
iab
etic
r
etin
o
p
ath
y
s
e
v
er
it
y
lev
el.
T
h
e
s
ize
o
f
t
h
e
i
m
ag
e
s
h
as
b
ee
n
r
es
h
ap
ed
to
f
it t
h
e
p
ag
e.
T
ab
le
1
.
Data
s
et
in
f
o
r
m
a
tio
n
S
e
v
e
r
i
t
y
l
e
v
e
l
L
a
b
e
l
#
I
mag
e
s
N
o
d
i
a
b
e
t
i
c
r
e
t
i
n
o
p
a
t
h
y
N
O
D
R
1
8
0
5
M
i
l
d
d
i
a
b
e
t
i
c
r
e
t
i
n
o
p
a
t
h
y
M
i
l
d
3
7
0
M
o
d
e
r
a
t
e
d
i
a
b
e
t
i
c
r
e
t
i
n
o
p
a
t
h
y
M
o
d
e
r
a
t
e
9
9
9
S
e
v
e
r
e
d
i
a
b
e
t
i
c
r
e
t
i
n
o
p
a
t
h
y
S
e
v
e
r
e
1
9
3
P
r
o
l
i
f
e
r
a
t
i
v
e
d
i
a
b
e
t
i
c
r
e
t
i
n
o
p
a
t
hy
P
r
o
l
i
f
e
r
a
t
i
v
e
D
R
1
9
5
S
e
v
e
r
i
t
y
l
e
v
e
l
L
a
b
e
l
#
I
mag
e
s
T
o
t
a
l
i
mag
e
s
3
5
6
2
Fig
u
r
e
1
.
I
m
a
g
es
f
r
o
m
t
h
e
d
at
aset s
h
o
w
in
g
e
y
e
s
ca
n
s
w
it
h
d
if
f
er
en
t se
v
er
it
y
le
v
els o
f
d
iab
etic
r
etin
o
p
ath
y
3
.
2
.
Da
t
a
prepr
o
ce
s
s
ing
T
h
is
s
ec
tio
n
d
escr
ib
es
th
e
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
tec
h
n
iq
u
es
w
e
le
v
er
ag
ed
to
n
o
r
m
a
lize
th
e
i
m
ag
e
s
as
w
ell
a
s
to
en
lar
g
e
t
h
e
d
ataset
t
o
m
a
k
e
it r
ea
d
y
f
o
r
d
ee
p
lear
n
in
g
ta
s
k
s
.
3
.
2
.
1
.
I
m
a
g
e
no
r
m
a
liza
t
io
n
T
h
e
im
a
g
es
in
o
u
r
d
ataset
ar
e
co
lo
r
ed
im
a
g
es
w
it
h
r
ed
,
g
r
ee
n
,
an
d
b
lu
e
c
h
an
n
el
s
(
R
GB
)
an
d
th
eir
s
iz
e
v
ar
ies
f
r
o
m
o
n
e
i
m
ag
e
to
an
o
t
h
er
.
T
h
er
ef
o
r
e,
to
s
tan
d
ar
d
ize
th
e
s
ize
o
f
t
h
e
i
m
ag
e
s
,
w
e
r
es
h
ap
ed
th
e
i
m
a
g
es
to
5
1
2
x
5
1
2
p
ix
els.
W
e
ch
o
o
s
e
th
is
s
ize
s
i
n
ce
it
s
m
o
r
e
e
f
f
ici
en
t
to
r
u
n
o
n
o
u
r
co
m
p
u
ter
s
an
d
to
h
av
e
e
n
o
u
g
h
f
ea
t
u
r
es f
o
r
th
e
m
o
d
el
to
lear
n
ab
o
u
t th
e
i
m
ag
e
s
.
I
m
ag
e
n
o
r
m
al
izatio
n
is
a
cr
u
c
ial
s
tep
i
n
d
ee
p
lear
n
i
n
g
t
h
at
allo
w
s
t
h
e
g
r
ad
ie
n
t
d
ec
en
t
al
g
o
r
ith
m
to
co
n
v
er
g
e
f
a
s
ter
an
d
h
en
ce
i
m
p
r
o
v
in
g
th
e
p
er
f
o
r
m
an
ce
of
t
h
e
d
ee
p
lear
n
in
g
m
o
d
el.
T
h
er
e
ar
e
s
ev
er
al
m
e
th
o
d
s
to
n
o
r
m
alize
i
m
a
g
es,
a
f
ter
co
n
v
e
r
tin
g
t
h
e
m
to
i
n
teg
er
v
ec
to
r
s
,
s
u
c
h
as
:
i
)
d
i
v
id
in
g
ea
ch
p
i
x
el
in
an
i
m
a
g
e
b
y
t
h
e
m
ea
n
of
th
a
t
i
m
ag
e
v
ec
to
r
,
ii
)
s
u
b
tr
ac
t
t
h
e
m
ea
n
p
er
ch
a
n
n
e
l
ca
lcu
lated
o
v
er
all
i
m
a
g
es
in
t
h
e
d
ataset,
or
iii
)
i
n
p
ictu
r
e
i
m
a
g
es
d
atase
ts
,
di
v
id
i
n
g
ea
c
h
p
ix
el
by
255
is
a
s
i
m
p
le
an
d
ef
f
icie
n
t
tec
h
n
iq
u
e.
T
h
e
th
ir
d
ap
p
r
o
ac
h
h
as
b
ee
n
u
s
ed
to
n
o
r
m
al
ize
th
e
i
m
ag
es i
n
o
u
r
d
ataset.
3
.
2
.
2
.
Da
t
a
o
v
er
-
s
a
m
pli
ng
Ov
er
-
s
a
m
p
l
in
g
i
s
a
g
r
o
u
p
o
f
te
ch
n
iq
u
es
to
s
o
lv
e
t
h
e
i
m
b
a
lan
ce
d
d
ata
p
r
o
b
lem
.
S
u
ch
a
tec
h
n
iq
u
e
tr
ies
to
m
ak
e
th
e
d
ata
s
et
b
alan
ce
d
w
it
h
eq
u
al
n
u
m
b
er
o
f
i
n
s
ta
n
ce
s
in
ea
c
h
cla
s
s
.
T
h
e
o
v
er
-
s
a
m
p
lin
g
tech
n
iq
u
e
th
at
w
e
u
s
ed
in
t
h
i
s
r
esear
ch
is
b
a
s
ed
o
n
i
m
p
le
m
en
t
in
g
a
s
i
m
p
l
e
d
u
p
licate
o
f
r
an
d
o
m
r
ec
o
r
d
s
f
r
o
m
th
e
m
i
n
o
r
it
y
class
es.
Fi
g
u
r
e
2
co
m
p
ar
es
b
et
w
ee
n
t
h
e
d
ataset
b
ef
o
r
e
le
v
er
ag
in
g
t
h
e
o
v
er
-
s
a
m
p
li
n
g
tec
h
n
iq
u
e,
Fi
g
u
r
e
2
(
a
)
,
a
n
d
af
ter
lev
er
a
g
i
n
g
t
h
e
d
ata
o
v
er
-
s
a
m
p
li
n
g
tech
n
iq
u
e,
F
ig
u
r
e
2
(
b
)
.
A
s
s
h
o
w
n
i
n
t
h
e
f
i
g
u
r
e,
a
f
ter
th
e
d
ata
o
v
e
r
-
s
a
m
p
li
n
g
,
t
h
e
r
es
u
lted
d
ataset
is
b
alan
ce
d
.
T
h
e
to
tal
n
u
m
b
er
of
i
m
ag
e
s
in
t
h
e
d
ata
s
et
af
ter
t
h
e
o
v
er
-
s
a
m
p
li
n
g
i
s
7
,
9
3
5
i
m
ag
e
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
A
tr
a
n
s
fer lea
r
n
in
g
w
ith
d
ee
p
n
eu
r
a
l n
etw
o
r
k
a
p
p
r
o
a
c
h
fo
r
…
(
Mo
h
a
mme
d
A
l
-
S
ma
d
i
)
3495
(
a)
(
b
)
Fig
u
r
e
2
.
Sev
er
it
y
d
i
s
tr
ib
u
tio
n
o
f
i
m
a
g
es b
ef
o
r
e
an
d
a
f
ter
ap
p
ly
in
g
t
h
e
o
v
er
s
a
m
p
lin
g
tec
h
n
iq
u
e
,
(
a)
Or
ig
in
al
d
ataset
,
an
d
(
b
)
Data
s
et
af
ter
o
v
er
-
s
a
m
p
li
n
g
3
.
2
.
3
.
Da
t
a
a
ug
m
ent
a
t
io
n
T
r
ain
in
g
d
ee
p
lear
n
in
g
m
o
d
els
s
u
c
h
as
De
n
s
eNe
t,
R
e
s
Net,
o
r
E
f
f
icie
n
tNet,
r
eq
u
ir
e
lar
g
e
d
ataset
t
o
p
r
o
d
u
ce
s
tab
le
m
o
d
els.
S
m
all
d
atasets
p
r
o
d
u
c
e
m
o
d
els
t
h
at
o
v
er
f
it
t
h
e
tr
ai
n
i
n
g
d
ata
s
et
an
d
h
en
ce
t
h
eir
r
es
u
lt
s
ca
n
n
o
t
b
e
g
e
n
er
alize
d
.
T
o
a
v
o
id
s
u
ch
a
p
r
o
b
le
m
,
w
e
h
a
v
e
p
er
f
o
r
m
ed
a
d
ata
-
au
g
m
e
n
tatio
n
tec
h
n
iq
u
e
to
en
lar
g
e
t
h
e
d
ataset.
Data
au
g
m
en
tatio
n
i
s
a
p
r
o
ce
s
s
o
f
g
en
e
r
atin
g
(
m
an
u
f
ac
t
u
r
in
g
)
d
ata
f
r
o
m
t
h
e
ex
is
ti
n
g
d
ata
to
in
cr
ea
s
e
th
e
d
iv
er
s
i
t
y
a
n
d
t
h
e
n
u
m
b
er
o
f
th
e
i
n
s
ta
n
ce
s
in
th
e
d
ata
s
et
w
h
ile
m
a
in
ta
in
i
n
g
t
h
e
s
a
m
e
lab
el
o
f
th
e
o
r
ig
in
a
l
i
m
a
g
e.
Data
a
u
g
m
en
tatio
n
tec
h
n
iq
u
es
p
er
f
o
r
m
v
ar
io
u
s
o
p
er
atio
n
s
on
i
m
a
g
es
in
c
lu
d
i
n
g
i
m
a
g
e
s
c
a
l
in
g
,
g
eo
m
etr
ic
tr
an
s
f
o
r
m
a
tio
n
,
ad
d
in
g
n
o
i
s
e
to
i
m
ag
e
s
,
ch
an
g
i
n
g
t
h
e
lig
h
ti
n
g
co
n
d
iti
o
n
s
o
f
th
e
i
m
a
g
es,
i
m
ag
e
s
f
lip
p
in
g
.
As
d
ep
icted
in
Fi
g
u
r
e
3,
we
p
e
r
f
o
r
m
ed
v
ar
io
u
s
d
ata
au
g
m
e
n
t
atio
n
o
p
er
atio
n
s
on
th
e
d
ataset
in
cl
u
d
i
n
g
f
lip
t
h
e
i
m
a
g
e
h
o
r
izo
n
tall
y
,
f
li
p
an
i
m
ag
e
v
er
tica
ll
y
,
s
ca
le
t
h
e
s
ize
o
f
an
i
m
ag
e,
r
o
tate
t
h
e
i
m
ag
e,
s
h
ea
r
in
g
a
n
i
m
a
g
e,
an
d
elas
tic
an
d
p
er
s
p
ec
t
iv
e
tr
a
n
s
f
o
r
m
atio
n
w
h
ich
tr
ie
s
to
p
r
o
j
ec
t
an
o
b
j
ec
t
of
an
i
m
ag
e
in
a
d
if
f
er
e
n
t
p
o
i
n
t
o
f
v
i
e
w
.
T
o
au
g
m
e
n
t
o
u
r
i
m
a
g
es,
w
e
h
a
v
e
u
t
ilized
th
e
I
m
g
Au
g
[
3
3
]
,
a
P
y
t
h
o
n
lib
r
ar
y
f
o
r
i
m
ag
e
a
u
g
m
e
n
tatio
n
.
Fo
r
each
in
p
u
t
i
m
a
g
e,
we
g
e
n
e
r
ate
64
d
if
f
er
en
t
i
m
a
g
es
o
th
er
t
h
an
t
h
e
o
r
ig
i
n
al
o
n
e.
Af
ter
th
e
d
ata
au
g
m
e
n
tatio
n
s
tep
,
w
e
e
n
d
ed
u
p
w
it
h
5
1
5
,
7
7
5
d
if
f
er
en
t la
b
eled
i
m
a
g
e
s
.
Fig
u
r
e
3
.
I
m
a
g
e
au
g
m
e
n
tatio
n
3
.
2
.
4
.
P
er
f
o
rm
i
ng
ps
eu
do
-
la
bel
T
h
e
w
o
r
k
o
f
[
3
4
]
,
th
e
p
s
eu
d
o
-
lab
el
w
a
s
i
m
p
le
m
e
n
ted
i
n
th
is
r
esear
ch
to
en
h
a
n
ce
th
e
m
o
d
el
p
er
f
o
r
m
a
n
ce
.
p
s
eu
d
o
-
lab
el
is
a
s
im
p
le
a
n
d
ef
f
ic
ien
t
s
e
m
i
-
s
u
p
er
v
i
s
ed
lear
n
in
g
tech
n
iq
u
e
to
i
m
p
r
o
v
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
d
ee
p
n
eu
r
al
n
et
w
o
r
k
m
o
d
els.
T
h
e
m
o
d
el
th
at
u
s
es
p
s
e
u
d
o
-
lab
el
is
tr
ai
n
e
d
u
s
in
g
s
u
p
er
v
is
e
d
lear
n
in
g
m
ec
h
a
n
is
m
w
it
h
lab
e
led
an
d
u
n
lab
eled
(
test
)
d
ata
at
th
e
s
a
m
e
ti
m
e.
Fo
r
u
n
lab
eled
d
ata,
th
e
m
o
d
el
i
s
tr
ain
ed
u
s
in
g
t
h
e
lab
eled
d
ata.
T
h
en
,
th
e
tr
ain
ed
m
o
d
el
is
u
s
ed
to
p
r
ed
ict
th
e
test
(
u
n
lab
ele
d
)
d
ata.
Fin
all
y
,
w
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
4
,
A
u
g
u
s
t 2
0
2
1
:
3
4
9
2
-
3501
3496
re
-
tr
ain
t
h
e
s
a
m
e
m
o
d
el
i
n
a
s
u
p
er
v
i
s
ed
m
ec
h
a
n
is
m
u
s
i
n
g
th
e
lab
eled
an
d
p
r
ed
icted
d
ata
an
d
m
a
k
e
a
n
e
w
p
r
ed
ictio
n
o
f
th
e
te
s
t
d
ata.
S
u
ch
a
s
i
m
p
le
ap
p
r
o
ac
h
i
m
p
r
o
v
ed
th
e
p
er
f
o
r
m
a
n
ce
o
f
a
s
ta
te
-
of
-
t
h
e
-
ar
t
n
e
u
r
al
n
et
w
o
r
k
m
o
d
el
[
3
4
]
.
3
.
3
.
Tr
a
ns
f
er
lea
rning
t
e
c
h
n
i
q
u
e
I
n
s
tead
o
f
tr
ain
i
n
g
a
d
ee
p
lear
n
in
g
m
o
d
el,
m
a
in
l
y
C
NNs,
f
r
o
m
s
cr
atch
,
m
a
n
y
r
e
s
ea
r
ch
er
s
,
esp
ec
iall
y
in
t
h
e
m
ed
ical
i
m
ag
e
s
p
r
o
ce
s
s
i
n
g
,
le
v
er
a
g
e
tr
an
s
f
er
lear
n
i
n
g
t
ec
h
n
iq
u
e
to
g
en
er
ate
ef
f
icie
n
t
m
o
d
el
s
[
3
5
]
.
In
th
is
r
esear
ch
,
w
e
d
e
v
elo
p
ed
a
tr
an
s
f
er
-
lear
n
i
n
g
-
b
ased
m
o
d
el
a
f
ter
f
i
n
e
-
t
u
n
i
n
g
a
p
r
e
-
tr
ain
ed
C
N
N
m
o
d
el,
tr
ai
n
ed
o
n
d
if
f
er
e
n
t
i
m
a
g
es,
a
n
d
i
n
clu
d
ed
th
e
p
r
e
-
tr
ain
ed
m
o
d
el
as
an
i
n
p
u
t
to
o
u
r
m
o
d
el.
T
h
e
lev
er
a
g
ed
C
NN
m
o
d
el
s
ar
e
p
r
e
-
tr
ain
ed
o
n
t
h
e
I
m
ag
e
N
et
d
ataset.
T
h
e
I
m
ag
eNe
t
d
ataset
co
n
tain
s
m
as
s
i
v
e
n
u
m
b
er
o
f
i
m
ag
e
s
a
n
d
t
h
e
C
NN
m
o
d
el
s
ar
e
av
a
ilab
le
to
p
u
b
lic.
Su
c
h
a
n
id
ea
g
r
ea
tl
y
i
m
p
r
o
v
e
d
th
e
p
er
f
o
r
m
a
n
ce
o
f
o
u
r
m
o
d
el.
Nev
er
t
h
eles
s
,
an
en
s
e
m
b
le
m
o
d
el
o
u
t o
f
t
h
e
b
es
t p
er
f
o
r
m
in
g
m
o
d
els i
s
i
m
p
le
m
en
ted
.
3.
4
.
P
re
-
t
ra
ined
m
o
d
el
s
I
n
o
r
d
er
to
d
etec
t
th
e
s
e
v
er
it
y
lev
el
o
f
d
iab
etic
r
etin
o
p
ath
y
i
m
a
g
es
a
n
d
to
co
m
p
ar
e
b
et
w
e
en
v
a
r
i
o
u
s
p
r
e
-
tr
ain
ed
m
o
d
els,
w
e
h
a
v
e
lev
er
ag
ed
6
s
tate
-
of
-
th
e
-
ar
t
C
NN
m
o
d
els.
T
h
e
p
r
e
-
tr
ain
ed
m
o
d
el
s
ar
e:
i
)
R
e
s
N
e
t
[
1
0
]
,
ii
)
I
n
ce
p
tio
n
V3
[
1
2
]
,
iii
)
I
n
ce
p
tio
n
R
esNe
t
(
I
n
ce
p
tio
n
V4
)
[
1
3
]
,
iv
)
Den
s
eNe
t
[
1
4
]
,
v
)
Xce
p
tio
n
[
1
7
]
,
an
d
vi
)
E
f
f
icie
n
tNet
[
1
8
]
.
3
.
5
.
G
lo
ba
l
a
v
e
r
a
g
e
po
o
lin
g
(
G
AP
)
-
ba
s
ed
cla
s
s
if
ier
Fig
u
r
e
4
o
v
er
v
ie
w
s
o
u
r
c
lass
if
ier
f
o
r
d
etec
ti
n
g
t
h
e
s
e
v
er
it
y
le
v
el
o
f
d
iab
etic
r
eti
n
o
p
ath
y
.
As
s
h
o
w
n
i
n
th
e
f
i
g
u
r
e,
we
d
ev
elo
p
ed
v
ar
io
u
s
d
ee
p
lear
n
i
n
g
m
o
d
el
s
t
h
at
lev
er
ag
e
t
h
e
tr
an
s
f
er
-
lear
n
in
g
te
ch
n
iq
u
e
to
in
cr
ea
s
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
o
u
r
m
o
d
el
.
E
ac
h
m
o
d
el
u
s
es
o
n
e
o
f
t
h
e
C
NN
m
o
d
el
s
,
d
is
cu
s
s
ed
i
n
s
e
cti
o
n
3
.
4
.
,
as
a
p
r
e
-
tr
ain
ed
m
o
d
el.
Fig
u
r
e
4
.
Ou
r
p
r
o
p
o
s
ed
class
if
ier
,
a
GA
P
-
b
ased
d
ee
p
lear
n
in
g
n
e
u
r
al
n
et
w
o
r
k
th
at
le
v
er
a
g
e
s
tr
an
s
f
er
lear
n
i
n
g
Ou
r
d
ee
p
lear
n
in
g
m
o
d
el
lev
er
ag
es
t
h
e
G
A
P
tech
n
iq
u
e.
GA
P
r
ed
u
ce
s
th
e
co
m
p
u
tatio
n
al
p
o
w
er
r
eq
u
ir
ed
f
o
r
th
e
n
eu
r
al
n
et
wo
r
k
to
w
o
r
k
an
d
h
e
n
ce
in
cr
ea
s
in
g
its
p
er
f
o
r
m
a
n
ce
.
G
A
P
also
is
u
s
ed
as
a
r
eg
u
lar
izatio
n
la
y
er
th
at
r
ed
u
c
es th
e
o
v
er
f
i
ttin
g
p
r
o
b
le
m
a
n
d
in
cr
ea
s
es
th
e
g
e
n
er
aliza
b
ilit
y
o
f
th
e
r
e
s
u
l
ts
o
f
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
A
tr
a
n
s
fer lea
r
n
in
g
w
ith
d
ee
p
n
eu
r
a
l n
etw
o
r
k
a
p
p
r
o
a
c
h
fo
r
…
(
Mo
h
a
mme
d
A
l
-
S
ma
d
i
)
3497
class
i
f
ier
[
3
6
]
.
Nev
er
th
ele
s
s
,
GA
P
h
a
s
s
h
o
w
n
an
ab
ili
t
y
to
ac
t
as
an
a
tten
t
io
n
la
y
er
b
y
d
i
s
cr
i
m
i
n
ati
n
g
r
eg
io
n
s
o
f
in
ter
est
o
f
t
h
e
i
m
a
g
e
an
d
r
etain
th
e
m
to
th
e
f
i
n
al
la
y
er
o
f
th
e
m
o
d
el
[
3
7
]
.
T
h
er
ef
o
r
e,
th
e
GA
P
la
y
er
is
p
lace
d
af
ter
th
e
p
r
e
-
tr
ai
n
ed
o
u
tp
u
t
la
y
er
to
tr
an
s
f
er
atten
ti
v
e
k
n
o
w
led
g
e
to
th
e
s
ec
o
n
d
p
ar
t
o
f
th
e
m
o
d
el.
Ba
tch
n
o
r
m
a
l
i
z
a
t
i
o
n
[
1
1
]
an
d
D
r
o
p
o
u
t
[
3
8
]
r
eg
u
latio
n
t
ec
h
n
iq
u
es
ar
e
th
e
n
u
s
ed
to
r
ed
u
ce
th
e
o
v
er
f
itti
n
g
p
r
o
b
lem
an
d
i
n
cr
ea
s
e
th
e
lear
n
in
g
ca
p
ab
ilit
ies
o
f
th
e
cla
s
s
i
f
ier
.
T
h
e
n
ex
t
la
y
er
in
o
u
r
d
ee
p
n
eu
r
al
n
et
w
o
r
k
is
t
h
e
Den
s
e
la
y
er
,
a
f
u
l
l
y
co
n
n
ec
ted
lay
er
w
it
h
1
,
0
2
4
n
eu
r
o
n
s
.
Nex
t,
th
e
o
u
tp
u
t
o
f
th
e
D
en
s
e
la
y
er
f
ed
to
a
r
ec
tif
ied
li
n
ea
r
u
n
it
(
R
e
L
U)
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
o
s
e
last
f
i
v
e
s
tep
s
ar
e
r
ep
ea
ted
3
tim
es
in
o
u
r
c
lass
if
ier
,
d
en
o
ted
as
X
3
in
Fig
u
r
e
4
,
b
ef
o
r
e
th
eir
o
u
tp
u
t
g
o
es
to
t
h
e
n
ex
t
le
v
el.
T
h
e
d
ee
p
er
th
e
n
et
w
o
r
k
t
h
e
m
o
r
e
v
an
i
s
h
in
g
t
h
e
g
r
ad
ien
t
s
w
il
l
b
e.
T
h
er
ef
o
r
e,
a
R
e
L
U
la
y
er
is
ad
d
ed
to
th
e
en
d
of
each
of
th
e
th
r
ee
b
lo
ck
s
to
m
i
n
i
m
ize
t
h
e
e
f
f
ec
t
o
f
v
a
n
i
s
h
i
n
g
g
r
ad
ien
t
p
r
o
b
lem
,
w
h
er
e
t
h
e
g
r
ad
ien
t
s
la
y
er
af
ter
la
y
er
ar
e
g
etti
n
g
m
o
r
e
a
n
d
m
o
r
e
s
m
aller
a
n
d
ar
e
n
o
t
b
ac
k
-
p
r
o
p
ag
ated
to
t
h
e
n
et
w
o
r
k
l
a
y
er
s
,
p
r
ev
e
n
ti
n
g
t
h
e
n
et
w
o
r
k
f
r
o
m
lear
n
i
n
g
lo
w
lev
el
d
etails
o
f
t
h
e
i
m
a
g
e
s
[
1
0
,
1
4
,
3
9
]
.
A
f
t
e
r
r
e
pe
at
i
ng
t
he
pr
e
vi
o
us
s
t
e
ps
t
hr
e
e
ti
m
e
s
,
ou
r
cl
a
s
s
i
f
i
e
r
pe
r
f
or
m
s
b
a
t
c
h
n
o
r
m
a
liz
a
tio
n
a
nd
d
r
o
p
o
u
t
r
e
g
ul
at
i
o
n
t
e
c
hn
i
q
ue
s
,
t
he
n
a
no
t
he
r
D
e
ns
e
la
ye
r
w
it
h
a
R
e
L
U
a
ct
i
va
ti
o
n
f
un
c
t
i
on
.
F
i
na
ll
y,
t
h
e
r
e
s
ul
t
s
of
t
he
pr
e
vi
o
us
la
ye
r
a
r
e
f
e
d
to
a
S
o
f
t
Ma
x
f
u
nc
ti
o
n
f
or
f
i
na
l
cl
a
s
s
i
f
i
ca
ti
o
n.
4.
E
XP
E
R
I
M
E
NT
A
T
I
O
N
AN
D
E
VA
L
UA
T
I
O
N
T
h
is
s
ec
tio
n
d
is
c
u
s
s
es
t
h
e
ex
p
er
im
e
n
tat
io
n
s
et
u
p
an
d
th
e
ev
alu
a
tio
n
p
r
o
ce
d
u
r
e
f
o
r
th
e
p
r
o
p
o
s
ed
m
o
d
e
l
s
.
T
h
e
r
est
o
f
t
h
i
s
s
ec
ti
o
n
is
o
r
g
a
n
ized
as
f
o
llo
w
s
:
t
h
e
ex
p
er
i
m
en
ta
tio
n
s
e
tu
p
a
n
d
th
e
p
ar
a
m
eter
s
t
h
at
w
er
e
u
s
ed
to
tr
ain
th
e
p
r
o
p
o
s
ed
tech
n
iq
u
e
s
ar
e
d
is
c
u
s
s
ed
in
s
u
b
s
ec
tio
n
4
.
1
.
,
th
e
ev
al
u
ati
o
n
m
ea
s
u
r
e
u
s
ed
to
e
v
a
l
u
a
t
e
t
h
e
p
r
o
p
o
s
ed
tech
n
iq
u
es
is
d
i
s
cu
s
s
ed
i
n
s
u
b
s
ec
tio
n
4
.
2
.
Fi
n
all
y
,
t
h
e
m
o
d
els
’
e
v
alu
a
tio
n
r
esu
lts
ar
e
p
r
esen
ted
in
s
u
b
s
ec
tio
n
4
.
3
.
4
.
1
.
E
x
peri
m
ent
a
t
io
n
s
et
up
T
h
e
p
r
o
p
o
s
ed
tech
n
iq
u
e
s
in
t
h
is
r
esear
ch
w
er
e
tr
ain
ed
u
s
in
g
th
e
p
r
o
v
id
ed
d
ataset
(
s
ee
s
ec
tio
n
3
.
1
).
A
ll
th
e
tr
an
s
f
er
lear
n
in
g
-
b
as
ed
m
o
d
els
w
er
e
f
ir
s
t
tr
ain
ed
w
i
th
o
u
t
p
s
eu
d
o
-
lab
el
lear
n
i
n
g
.
Seco
n
d
,
tr
ain
ed
m
o
d
el
s
w
er
e
u
s
ed
to
p
r
ed
ict
t
h
e
lab
el
o
f
t
h
e
test
i
n
g
i
m
a
g
es
f
o
r
p
s
e
u
d
o
-
lab
el
lear
n
i
n
g
.
P
r
ed
icted
lab
els
a
n
d
th
eir
ass
o
ciate
d
i
m
a
g
es a
r
e
th
e
n
ad
d
ed
to
tr
ain
in
g
d
ataset
a
n
d
u
s
ed
to
tr
ain
th
e
p
r
o
p
o
s
ed
m
o
d
els.
T
a
b
l
e
2
p
r
esen
ts
t
h
e
h
y
p
er
-
p
a
r
a
m
eter
s
u
s
ed
in
tr
ai
n
i
n
g
th
e
p
r
o
p
o
s
ed
m
o
d
els,
w
h
er
e,
B
S
s
tan
d
s
f
o
r
b
atch
s
ize
an
d
LR
s
tan
d
s
f
o
r
lear
n
in
g
r
ate.
A
ll
t
h
e
m
o
d
el
s
w
er
e
tr
ain
ed
w
it
h
a
lear
n
i
n
g
r
ate
o
f
1
e
-
4
f
o
r
a
m
ax
i
m
u
m
n
u
m
b
er
o
f
7
5
ep
o
c
h
s
.
A
ll
t
h
e
m
o
d
els
w
er
e
tr
ain
ed
w
it
h
i
m
a
g
e
s
ize
o
f
5
1
2
x
5
1
2
p
ix
els
ex
ce
p
t
f
o
r
th
e
E
f
f
icie
n
tNet
-
B
4
w
as
tr
ain
ed
u
s
i
n
g
3
8
0
x
3
8
0
p
ix
els.
All
t
h
e
e
x
p
er
i
m
e
n
ts
w
er
e
co
n
d
u
cte
d
u
s
i
n
g
t
h
e
Ka
g
g
le
k
er
n
el
o
f
t
h
e
ch
al
len
g
e.
S
m
all
b
atch
s
ize
s
w
er
e
u
s
ed
to
tr
ain
th
e
m
o
d
el
s
d
u
e
to
t
h
e
li
m
ited
r
eso
u
r
ce
s
p
r
o
v
id
ed
b
y
t
h
e
k
er
n
el
a
n
d
th
e
h
ig
h
c
o
m
p
lex
i
t
y
o
f
th
e
u
s
ed
m
o
d
els.
T
ab
le
2
.
Mo
d
els H
y
p
er
-
p
ar
a
m
eter
s
M
o
d
e
l
#
E
p
o
c
h
BS
LR
I
n
p
u
t
I
mag
e
S
i
z
e
#
C
h
a
n
n
e
l
s
R
e
sN
e
t
-
50
75
6
1e
-
4
5
1
2
*
5
1
2
3
I
n
c
e
p
t
i
o
n
-
V3
75
8
1e
-
4
5
1
2
*
5
1
2
3
I
n
c
e
p
t
i
o
n
R
e
sN
e
t
-
V2
75
6
1e
-
4
5
1
2
*
5
1
2
3
D
e
n
se
N
e
t
-
169
75
8
1e
-
4
5
1
2
*
5
1
2
3
X
c
e
p
t
i
o
n
75
8
1e
-
4
5
1
2
*
5
1
2
3
Ef
f
i
c
i
e
n
t
N
e
t
-
B4
75
6
1e
-
4
3
8
0
*
3
8
0
3
4
.
2
.
E
v
a
lua
t
i
o
n
m
ea
s
ure
T
h
e
q
u
ad
r
atic
w
ei
g
h
ted
k
ap
p
a
(
QW
K)
[
4
0
]
is
u
s
ed
to
ev
a
lu
ate
th
e
p
er
f
o
r
m
an
ce
o
f
t
h
e
m
o
d
els.
T
h
e
QW
K
ev
alu
ate
s
t
h
e
lev
el
o
f
ag
r
ee
m
e
n
t
b
et
w
ee
n
th
e
i
m
a
g
e
tar
g
et
lab
el
an
d
th
e
p
r
ed
icted
s
ev
er
it
y
lev
e
l.
T
h
e
Qu
ad
r
atic
w
eig
h
ted
k
ap
p
a
is
c
o
m
p
u
ted
u
s
in
g
(
1
)
.
(
1
)
w
he
r
e
,
i
r
e
pr
e
s
e
nt
s
t
he
ta
r
ge
t
la
be
l
,
j
r
e
pr
e
s
e
nt
s
t
he
pr
e
di
c
te
d
l
a
be
l
,
O
i
,
j
is
an
N*N
m
a
t
r
i
x
r
e
pr
e
s
e
nt
s
t
h
e
r
e
c
ei
ve
d
ta
r
ge
t
a
n
d
pr
e
di
c
te
d
l
a
be
l
s
,
E
i
,
j
is
an
N*N
m
a
t
r
i
x
r
e
pr
e
s
e
nt
s
t
he
e
xp
e
c
t
e
d
t
a
r
ge
t
a
nd
pr
e
di
ct
e
d
la
be
l
,
a
nd
w
i
,
j
is
an
N
*
N
m
a
t
r
i
x
r
e
pr
e
s
e
nt
s
a
w
e
i
gh
t
c
al
c
ul
at
e
d
ba
s
e
d
on
t
he
di
f
f
e
r
e
nc
e
be
t
w
e
e
n
t
he
ta
r
ge
t
a
n
d
pr
e
di
c
t
e
d
la
be
l
.
w
i
,
j
i
s
c
o
m
p
ut
e
d
us
i
n
g
(
2
)
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
4
,
A
u
g
u
s
t 2
0
2
1
:
3
4
9
2
-
3501
3498
(
2
)
w
he
r
e,
t
he
N
r
e
p
r
e
s
e
nt
s
t
he
nu
m
be
r
of
t
e
s
ti
ng
s
a
m
pl
e
s
,
i
r
e
pr
e
s
e
nt
s
t
he
t
a
r
ge
t
l
a
be
l,
a
nd
j
r
e
pr
e
s
e
nt
s
t
he
pr
e
di
c
t
e
d
l
a
be
l.
4
.
3
.
Resul
t
s
T
a
b
l
e
3
s
h
o
w
s
t
h
e
r
esu
l
ts
ac
h
i
ev
ed
b
y
o
u
r
tr
an
s
f
er
lear
n
in
g
m
o
d
el
s
.
T
h
e
b
est r
esu
lt
w
as sc
o
r
ed
b
y
th
e
I
n
ce
p
tio
n
-
V3
+
G
A
P
-
b
ased
cla
s
s
i
f
ier
w
i
th
QW
K
=
8
2
.
0
%.
T
h
e
n
e
x
t
b
es
t
s
co
r
e
w
a
s
ac
h
ie
v
e
d
b
y
t
h
e
De
n
s
e
Net
-
169
+
GA
P
-
b
ased
class
i
f
ier
w
it
h
QW
K
=
8
1
.
8
%,
w
h
er
ea
s
t
h
e
th
ir
d
lev
el
w
as
s
co
r
ed
by
t
h
e
Xce
p
tio
n
+
G
A
P
-
b
ased
class
i
f
ier
w
it
h
QW
K
=
8
0
.
9
%.
On
th
e
o
t
h
er
h
a
n
d
,
th
e
R
e
s
Net
-
50
+
GA
P
-
b
ased
cla
s
s
i
f
ie
r
s
co
r
ed
th
e
w
o
r
s
t
r
esu
lt
s
a
m
o
n
g
t
h
e
tr
ain
ed
m
o
d
els
w
it
h
a
QW
K
=
7
7
.
6
%.
As
th
is
r
esear
ch
is
b
ased
o
n
a
Kag
g
le
ch
alle
n
g
e,
an
d
i
n
o
r
d
er
to
ac
h
iev
e
h
i
g
h
r
esu
lts
a
n
d
r
an
k
,
a
n
en
s
e
m
b
le
b
ased
o
n
s
i
m
p
le
av
e
r
ag
e
o
f
th
e
p
r
ed
ictio
n
s
o
f
th
e
t
o
p
th
r
ee
p
e
r
f
o
r
m
in
g
m
o
d
els
(
i
.
e.
,
Den
s
eNe
t
-
1
6
9
,
I
n
ce
p
tio
n
-
V3
,
an
d
Xce
p
tio
n
)
w
a
s
co
m
p
u
ted
.
T
h
e
en
s
e
m
b
le
m
o
d
el
o
u
tp
er
f
o
r
m
s
t
h
e
b
est
p
er
f
o
r
m
i
n
g
m
o
d
el
(
i.e
.
,
I
n
ce
p
tio
n
-
V3
)
w
it
h
0
.
4
%
w
it
h
a
QW
K
o
f
8
2
.
4
%.
A
lt
h
o
u
g
h
w
e
f
i
n
i
s
h
ed
t
h
e
c
h
alle
n
g
e
w
it
h
a
r
a
n
k
o
f
7
1
(
tea
m
n
a
m
e
:
Data
_
Sc
ien
ce
@
J
UST
)
th
e
f
ir
s
t
tea
m
(
“
[
o
d
s
.
ai]
to
p
co
d
er
s
”)
ac
h
iev
ed
a
s
co
r
e
of
QW
K
=
8
5
.
6
%
w
it
h
o
n
l
y
3
.
2
%
o
f
ad
v
an
ce
m
e
n
t
o
v
e
r
o
u
r
ac
h
iev
ed
r
esu
lt
s
in
h
ttp
s
://
www
.
k
a
g
g
le.
co
m
/c/a
p
to
s
2
0
1
9
-
b
lin
d
n
ess
-
d
etec
tio
n
/lead
er
b
o
ar
d
.
I
t is
w
o
r
th
n
o
ti
n
g
t
h
at
2
,
9
3
1
team
s
h
a
v
e
p
ar
ticip
ated
in
th
i
s
Kag
g
le
ch
alle
n
g
e.
T
ab
le
3
.
R
esu
lts
ac
h
ie
v
ed
b
y
o
u
r
p
r
o
p
o
s
ed
m
o
d
els
M
o
d
e
l
R
e
su
l
t
s (Q
W
K
)
R
e
sN
e
t
-
50
+
GAP
-
b
a
se
d
c
l
a
ssi
f
i
e
r
7
7
.
6
%
I
n
c
e
p
t
i
o
n
R
e
sN
e
t
-
V2
+
G
A
P
-
b
a
se
d
c
l
a
s
si
f
i
e
r
7
9
.
6
%
Ef
f
i
c
i
e
n
t
N
e
t
-
B4
+
GAP
-
b
a
se
d
c
l
a
ss
i
f
i
e
r
8
0
.
0
%
X
c
e
p
t
i
o
n
+
GAP
-
b
a
se
d
c
l
a
ss
i
f
i
e
r
8
0
.
9
%
D
e
n
se
N
e
t
-
169
+
GAP
-
b
a
se
d
c
l
a
ssi
f
i
e
r
8
1
.
8
%
I
n
c
e
p
t
i
o
n
-
V3
+
GAP
-
b
a
se
d
c
l
a
ss
i
f
i
e
r
8
2
.
0
%
En
se
mb
l
e
(
D
e
n
se
N
e
t
-
1
6
9
,
I
n
c
e
p
t
i
o
n
-
V
3
,
X
c
e
p
t
i
o
n
)
8
2
.
4
%
5.
DIS
CU
SS
I
O
N
On
e
o
f
th
e
r
i
s
k
s
f
o
r
tr
ain
i
n
g
tr
an
s
f
er
lear
n
i
n
g
w
it
h
d
ee
p
n
e
u
r
al
n
et
w
o
r
k
s
i
s
o
v
er
f
itti
n
g
.
T
h
e
r
ef
o
r
e,
th
e
ca
llb
ac
k
f
u
n
ctio
n
o
f
“
E
ar
l
y
St
o
p
p
in
g
”
in
h
ttp
s
://
k
er
as.io
/cal
lb
ac
k
s
/
f
r
o
m
Ker
as
ca
llb
ac
k
s
is
u
s
ed
to
s
to
p
th
e
m
o
d
el
tr
ain
i
n
g
w
h
en
th
e
co
m
p
u
ted
v
al
id
atio
n
lo
s
s
v
a
lu
e
is
n
o
m
o
r
e
i
m
p
r
o
v
i
n
g
.
As
d
ep
icte
d
in
Fi
g
u
r
e
5
,
th
er
e
is
n
o
g
ap
b
et
w
ee
n
th
e
co
m
p
u
t
ed
tr
ain
lo
s
s
a
n
d
t
h
e
v
alid
atio
n
lo
s
s
o
v
er
m
o
d
el
tr
ain
i
n
g
ep
o
ch
s
.
T
h
is
in
d
icate
s
th
at
th
e
m
o
d
el
w
a
s
tr
ai
n
ed
w
it
h
o
u
t
o
v
er
f
itti
n
g
.
A
l
t
h
o
u
g
h
th
e
m
a
x
i
m
u
m
n
u
m
b
er
o
f
ep
o
ch
s
en
ab
led
f
o
r
tr
ain
in
g
w
a
s
s
et
to
7
5
,
th
e
m
o
d
el
s
to
p
p
ed
tr
ain
in
g
af
ter
ep
o
ch
4
0
to
p
r
ev
en
t o
v
er
f
itt
in
g
as s
h
o
w
n
in
Fig
u
r
e
5
.
B
o
th
lo
s
s
v
alu
e
s
d
ec
li
n
ed
to
g
eth
er
d
u
r
in
g
m
o
d
el
tr
ai
n
i
n
g
d
u
e
to
th
e
m
o
d
el
la
y
er
s
r
e
s
p
o
n
s
ib
le
f
o
r
r
eg
u
lar
izatio
n
(
i.e
.
,
b
atch
n
o
r
m
a
lizatio
n
a
n
d
d
r
o
p
o
u
t o
f
Fi
g
u
r
e
4
)
.
A
s
d
is
c
u
s
s
ed
e
ar
lier
,
b
atch
n
o
r
m
aliza
tio
n
[
1
1
]
an
d
d
r
o
p
o
u
t
[
3
8
]
ar
e
r
eg
u
latio
n
tec
h
n
iq
u
es
u
s
e
d
to
r
ed
u
ce
th
e
o
v
er
f
itti
n
g
p
r
o
b
le
m
an
d
i
n
cr
ea
s
e
t
h
e
lear
n
i
n
g
ca
p
ab
ilit
ies
o
f
t
h
e
class
i
f
ier
.
I
n
o
r
d
er
to
s
h
o
w
th
e
s
tr
e
n
g
th
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
ar
ch
ite
ctu
r
e
a
n
d
t
h
e
i
m
p
o
r
tan
ce
o
f
t
h
e
p
r
o
p
o
s
ed
i
m
a
g
e
p
r
ep
r
o
ce
s
s
in
g
tec
h
n
iq
u
es
(
s
ee
s
ec
tio
n
3
)
,
an
ab
latio
n
an
al
y
s
i
s
w
a
s
co
n
d
u
cted
o
n
t
h
e
b
est
p
er
f
o
r
m
i
n
g
m
o
d
el
(
I
n
ce
p
tio
n
-
V3
+
G
A
P
-
b
ased
class
i
f
ier
)
.
As
p
r
esen
ted
in
T
a
b
l
e
4
,
r
el
y
i
n
g
o
n
tr
ai
n
i
n
g
th
e
I
n
ce
p
tio
n
-
V3
m
o
d
el
a
lo
n
e
w
it
h
o
u
t
tr
an
s
f
er
lear
n
in
g
,
t
h
e
m
o
d
el
s
co
r
ed
th
e
lo
w
est
r
e
s
u
l
ts
o
f
QW
K
=
6
3
.
8
%
w
it
h
-
1
3
.
7
%
lo
w
er
th
a
n
t
h
e
ac
h
ie
v
ed
r
esu
lt
by
th
e
tr
a
n
s
f
er
lear
n
i
n
g
w
it
h
t
h
e
I
n
ce
p
tio
n
-
V3
m
o
d
el
(
i.e
.
,
I
n
ce
p
tio
n
-
V3
+
G
AP
-
b
ased
class
i
f
ier
w
i
th
QW
K
=
8
2
.
0
%).
T
h
is
f
i
n
d
in
g
s
h
o
w
s
t
h
e
s
i
g
n
i
f
ica
n
ce
o
f
t
h
e
p
r
o
p
o
s
ed
tr
an
s
f
er
lear
n
in
g
ar
ch
itect
u
r
e.
Mo
r
eo
v
er
,
th
i
s
f
in
d
in
g
g
o
es
in
li
n
e
w
it
h
f
i
n
d
i
n
g
s
r
ep
o
r
ted
in
l
iter
atu
r
e
f
o
r
th
e
v
al
u
e
o
f
u
s
i
n
g
tr
an
s
f
er
lear
n
in
g
w
i
th
d
ee
p
n
eu
r
al
n
et
w
o
r
k
s
in
g
e
n
er
al
[
4
1
]
an
d
f
o
r
t
h
e
m
ed
ical
i
m
a
g
e
cla
s
s
i
f
icatio
n
i
n
p
ar
ticu
lar
[
4
2
]
.
T
h
e
s
ec
o
n
d
h
i
g
h
est
i
m
p
ac
t
w
a
s
t
h
e
ab
latio
n
o
f
p
s
eu
d
o
-
l
ab
el
tech
n
iq
u
e,
t
h
e
m
o
d
el
s
c
o
r
ed
QW
K
7
0
.
3
%
w
it
h
o
u
t
p
s
e
u
d
o
-
lab
el
with
a
s
co
r
e
d
ec
lin
e
by
-
1
1
.
7
%.
T
h
is
f
in
d
i
n
g
s
h
o
w
s
h
o
w
e
f
f
ici
en
t
w
a
s
th
e
p
s
e
u
d
o
-
lab
el
tech
n
iq
u
e
in
i
m
p
r
o
v
in
g
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
T
h
e
s
a
m
e
f
in
d
i
n
g
w
a
s
r
ep
o
r
ted
in
t
h
e
tech
n
iq
u
e
p
u
b
lis
h
ed
in
[
3
4
]
.
T
h
e
ab
latio
n
of
t
h
e
i
m
ag
e
a
u
g
m
en
tatio
n
s
co
r
ed
a
d
ec
li
n
e
in
t
h
e
re
s
u
lt
s
w
it
h
-
3
.
4
%
.
P
r
ev
io
u
s
r
esear
ch
h
as
d
e
m
o
n
s
tr
ated
th
e
ef
f
ec
ti
v
en
e
s
s
o
f
u
s
i
n
g
i
m
a
g
e
d
ata
au
g
m
e
n
tat
io
n
i
n
en
h
an
c
in
g
m
o
d
e
l
s
class
i
f
icatio
n
p
er
f
o
r
m
a
n
ce
[
4
3
]
.
A
lt
h
o
u
g
h
it
w
a
s
ex
p
ec
ted
th
at
i
m
a
g
e
d
ata
au
g
m
e
n
tatio
n
wo
u
ld
h
av
e
a
h
ig
h
er
i
m
p
ac
t
o
n
e
n
h
a
n
ci
n
g
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
r
es
u
lts
,
t
h
e
d
ata
a
u
g
m
e
n
tatio
n
tec
h
n
iq
u
es
w
er
e
m
ai
n
l
y
tr
ad
itio
n
al
a
n
d
s
i
m
p
le
(
i.e
.
r
o
tatin
g
,
f
lip
p
i
n
g
,
a
n
d
cr
o
p
p
in
g
o
f
in
p
u
t
i
m
a
g
es).
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
A
tr
a
n
s
fer lea
r
n
in
g
w
ith
d
ee
p
n
eu
r
a
l n
etw
o
r
k
a
p
p
r
o
a
c
h
fo
r
…
(
Mo
h
a
mme
d
A
l
-
S
ma
d
i
)
3499
Fig
u
r
e
5
.
T
h
e
lo
s
s
v
alu
e
s
f
o
r
t
h
e
tr
ain
i
n
g
an
d
v
a
lid
atio
n
d
ata
s
ets d
u
r
i
n
g
ea
c
h
ep
o
ch
o
f
tr
ai
n
in
g
t
h
e
m
o
d
el
Fin
all
y
,
t
h
e
ab
latio
n
of
th
e
o
v
e
r
s
a
m
p
li
n
g
tec
h
n
iq
u
e
h
ad
t
h
e
l
o
w
est
i
m
p
ac
t
w
it
h
a
d
ec
lin
e
in
th
e
m
o
d
el
s
co
r
e
o
f
-
0
.
7
%.
As
d
is
c
u
s
s
ed
ea
r
lier
,
a
r
an
d
o
m
o
v
er
s
a
m
p
lin
g
tec
h
n
iq
u
e
w
a
s
i
m
p
le
m
en
ted
to
s
o
lv
e
t
h
e
i
m
b
al
an
ce
d
p
r
o
b
le
m
o
f
th
e
o
r
ig
in
a
l
d
ataset.
H
o
w
e
v
e
r
,
th
e
ab
latio
n
an
al
y
s
is
s
h
o
w
s
t
h
at
ab
latin
g
th
i
s
tech
n
iq
u
e
h
as
a
v
er
y
s
m
all
a
f
f
ec
t
on
th
e
m
o
d
el
ac
h
ie
v
ed
r
esu
l
ts
w
it
h
o
n
l
y
-
0
.
7
%.
A
lt
h
o
u
g
h
t
h
e
r
an
d
o
m
o
v
er
s
a
m
p
li
n
g
tech
n
iq
u
e
ca
n
be
u
s
e
f
u
l
to
av
o
id
th
e
n
eg
ati
v
e
a
f
f
ec
t
of
i
m
b
alan
ce
d
d
ata
on
ac
h
iev
ed
m
o
d
el
r
esu
lt
s
,
it
ca
u
s
es
th
e
m
o
d
el
to
o
v
er
f
it
d
u
r
i
n
g
tr
ain
i
n
g
[
4
4
]
.
H
o
w
e
v
e
r
,
as
w
e
u
s
ed
th
e
ca
llb
ac
k
f
u
n
ctio
n
o
f
“
E
ar
l
y
Sto
p
p
in
g
”
f
r
o
m
Ker
as
ca
ll
b
ac
k
s
to
p
r
ev
e
n
t
m
o
d
el
o
v
er
f
itti
n
g
,
th
e
e
f
f
ec
t
of
th
e
r
a
n
d
o
m
o
v
er
s
a
m
p
l
in
g
was
v
er
y
lo
w
on
t
h
e
m
o
d
el
a
c
h
i
e
v
e
d
class
i
f
icatio
n
r
esu
lt
s
.
T
ab
le
4
.
A
b
latio
n
an
al
y
s
i
s
r
es
u
lts
f
o
r
th
e
I
n
ce
p
tio
n
-
V3
+
GA
P
-
b
ased
m
o
d
el
A
b
l
a
t
e
d
F
e
a
t
u
r
e
s
R
e
s
ul
t
D
i
f
f
e
r
e
n
c
e
W
i
t
h
o
u
t
O
v
e
r
-
S
a
mp
l
i
n
g
8
1
.
3
%
-
0
.
7
%
W
i
t
h
o
u
t
A
u
g
me
n
t
a
t
i
o
n
7
8
.
6
%
-
3
.
4
%
W
i
t
h
o
u
t
P
se
u
d
o
-
L
a
b
e
l
7
0
.
3
%
-
1
1
.
7
%
W
i
t
h
o
u
t
T
r
a
n
sf
e
r
L
e
a
r
n
i
n
g
6
8
.
3
%
-
1
3
.
7
%
I
n
c
e
p
t
i
o
n
-
V
3
+
G
A
P
-
b
a
se
d
c
l
a
ssi
f
i
e
r
-
F
u
l
l
8
2
.
0
%
—
6.
CO
NCLU
SI
O
N
Diab
etic
r
etin
o
p
ath
y
i
s
a
p
r
o
g
r
ess
i
v
e
e
y
e
d
is
ea
s
e
ca
u
s
ed
b
y
a
h
ig
h
b
lo
o
d
s
u
g
ar
o
r
p
r
e
s
s
u
r
e.
T
h
is
d
is
ea
s
e,
i
f
n
o
t
d
etec
ted
an
d
tr
e
ated
w
el
l
e
a
r
l
y
,
ca
n
ca
u
s
e
v
is
i
o
n
lo
s
s
.
T
o
th
at
en
d
,
w
e
h
av
e
p
r
o
p
o
s
ed
a
tr
an
s
f
er
lear
n
in
g
ap
p
r
o
ac
h
f
o
r
ac
cu
r
atel
y
d
etec
tin
g
th
e
s
e
v
er
it
y
lev
el
o
f
d
iab
etic
r
etin
o
p
ath
y
.
O
u
r
m
o
d
el
is
a
d
ee
p
lear
n
in
g
m
o
d
el
b
ased
o
n
g
lo
b
al
a
v
e
r
a
g
e
p
o
o
lin
g
(
G
A
P
)
tech
n
iq
u
e
w
it
h
v
ar
io
u
s
p
r
e
-
tr
ain
ed
C
NN
m
o
d
els.
W
e
h
av
e
u
ti
lized
6
s
tate
-
of
-
t
h
e
-
ar
t
C
NN
m
o
d
els
a
s
p
r
e
-
tr
ai
n
ed
m
o
d
els
to
o
u
r
G
A
P
-
b
ase
m
o
d
el
an
d
co
m
p
ar
ed
b
et
w
ee
n
t
h
e
m
.
O
u
r
b
est
m
o
d
el,
th
e
I
n
ce
p
tio
n
-
V3
+
G
A
P
-
b
ased
class
i
f
ier
,
ac
h
iev
ed
8
2
.
0
%
QW
K.
I
m
p
r
o
v
in
g
th
e
p
er
f
o
r
m
a
n
ce
of
o
u
r
m
o
d
els
as
w
ell
as
ap
p
l
y
i
n
g
o
u
r
tr
an
s
f
er
lear
n
i
n
g
m
o
d
el
s
to
o
t
h
er
m
ed
ical
p
r
o
b
lem
s
ar
e
in
ter
esti
n
g
av
e
n
u
e
o
f
f
u
t
u
r
e
d
ir
ec
tio
n
s
.
ACK
NO
WL
E
D
G
E
M
E
NT
S
T
h
is
r
esear
ch
is
p
ar
tiall
y
f
u
n
d
ed
b
y
J
o
r
d
an
Un
i
v
er
s
i
t
y
o
f
S
cien
ce
a
n
d
T
ec
h
n
o
lo
g
y
,
R
ese
ar
ch
Gr
an
t
Nu
m
b
er
: 2
0
1
7
0
1
0
7
an
d
2
0
1
9
0
30
6
.
RE
F
E
R
E
NC
E
S
[1
]
J.
H.
Ke
m
p
e
n
,
B.
J.
O’Co
l
m
a
i
n
,
M
.
C.
L
e
sk
e
,
S
.
M
.
Ha
ff
n
e
r
,
R.
Kle
in
,
e
t
a
l
.
,
“
T
h
e
p
re
v
a
le
n
c
e
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
a
m
o
n
g
a
d
u
lt
s i
n
th
e
u
n
it
e
d
sta
tes
.
,
”
Arc
h
ive
s o
f
o
p
h
t
h
a
l
-
mo
lo
g
y
.,
v
o
l
.
1
2
2
,
n
o
.
4
,
p
p
.
5
5
2
-
5
6
3
,
2
0
1
4
.
[2
]
C.
W
il
k
in
so
n
e
t
a
l.
,
“
P
ro
p
o
se
d
i
n
tern
a
ti
o
n
a
l
c
li
n
ica
l
d
iab
e
ti
c
re
ti
n
o
p
a
t
h
y
a
n
d
d
ia
-
b
e
ti
c
m
a
c
u
lar
e
d
e
m
a
d
ise
a
s
e
se
v
e
rit
y
s
c
a
les
,
”
Op
h
th
a
lmo
lo
g
y
,
v
o
l.
1
1
0
,
n
o
.
9
,
p
p
.
1
6
7
7
–
1
6
8
2
,
2
0
0
3
.
[3
]
Y.
L
e
Cu
n
,
Y.
Be
n
g
io
,
G
.
Hin
to
n
,
“
De
e
p
lea
rn
in
g
,
”
n
a
t
u
re
,
v
o
l.
5
2
1
,
n
o
.
7
5
5
3
,
p
p
.
4
3
6
-
4
4
4
,
2
0
1
5
.
[4
]
P
.
Ka
m
a
v
isd
a
r,
S
.
S
a
lu
ja,
S
.
Ag
ra
w
a
l,
“
A
su
rv
e
y
o
n
i
m
a
g
e
c
las
sif
i
c
a
ti
o
n
a
p
p
ro
a
c
h
e
s
a
n
d
tec
h
n
i
q
u
e
s,
”
In
te
rn
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
A
d
v
a
n
c
e
d
Res
e
a
rc
h
in
Co
mp
u
ter
a
n
d
Co
mm
u
n
ica
ti
o
n
E
n
g
i
n
e
e
rin
g
,
v
o
l.
2
,
n
o
.
1
,
p
p
.
1
0
0
5
-
1
0
0
9
,
2
0
1
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
4
,
A
u
g
u
s
t 2
0
2
1
:
3
4
9
2
-
3501
3500
[5
]
L
e
Cu
n
,
Ya
n
n
e
t
a
l
.
,
“
G
ra
d
ien
t
-
b
a
se
d
lea
rn
in
g
a
p
p
li
e
d
to
d
o
c
u
m
e
n
t
re
c
o
g
n
it
io
n
,
”
Pro
c
e
e
d
in
g
s
o
f
th
e
IEE
E
,
v
o
l.
86
,
n
o
.
11
,
p
p
.
2
2
7
8
–
2
3
2
4
,
1
9
9
8
.
[6
]
A
.
Kriz
h
e
v
sk
y
,
I.
S
u
tsk
e
v
e
r,
G
.
E.
Hin
t
o
n
,
“
Im
a
g
e
n
e
t
c
las
si
f
ica
ti
o
n
w
it
h
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,
”
Ad
v
a
n
c
e
s i
n
n
e
u
r
a
l
i
n
f
o
rm
a
ti
o
n
p
ro
c
e
ss
in
g
sy
ste
ms
,
v
o
l.
2
5
,
p
p
.
1
0
9
7
–
1
1
0
5
,
2
0
1
2
.
[7
]
M
.
D.
Zeiler,
R.
F
e
rg
u
s,
“
V
isu
a
li
z
in
g
a
n
d
u
n
d
e
rsta
n
d
i
n
g
c
o
n
v
o
lu
ti
o
n
a
l
n
e
tw
o
rk
s
,
”
Eu
ro
p
e
a
n
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
,
p
p
.
8
1
8
-
8
3
3
,
2
0
1
4
.
[8
]
K.
S
im
o
n
y
a
n
,
A
.
Zi
ss
e
r
m
a
n
,
“
V
e
ry
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
tw
o
rk
s
fo
r
larg
e
-
sc
a
le
i
m
a
g
e
re
c
o
g
n
it
io
n
,
”
a
rXiv
p
re
p
rin
t
a
rXiv:1
4
0
9
.
1
5
5
6
,
2
0
1
4
.
[9
]
C.
S
z
e
g
e
d
y
,
W
.
L
iu
,
Y.
Jia
,
P
.
S
e
r
m
a
n
e
t
,
S
.
Re
e
d
,
D.
A
n
g
u
e
lo
v
,
D.
Erh
a
n
,
V
.
V
a
n
h
o
u
c
k
e
,
A
.
Ra
b
in
o
v
ich
,
“
G
o
in
g
d
e
e
p
e
r
w
it
h
c
o
n
v
o
lu
ti
o
n
s,
”
Pro
c
e
e
d
in
g
s
o
f
th
e
IEE
E
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
a
n
d
p
a
tt
e
rn
re
c
o
g
n
it
io
n
,
2
0
1
5
,
p
p
.
1
–
9.
[1
0
]
K.
He
,
X
.
Z
h
a
n
g
,
S
.
Re
n
,
J
.
S
u
n
,
“
De
e
p
re
sid
u
a
l
lea
rn
i
n
g
f
o
r
im
a
g
e
r
e
c
o
g
n
it
io
n
,
”
Pr
o
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
a
n
d
p
a
tt
e
rn
re
c
o
g
n
it
io
n
,
2
0
1
6
,
p
p
.
7
7
0
-
7
7
8
.
[1
1
]
S
.
Io
f
f
e
,
C.
S
z
e
g
e
d
y
,
“
B
a
tch
n
o
rm
a
li
z
a
ti
o
n
:
A
c
c
e
ler
a
ti
n
g
d
e
e
p
n
e
tw
o
rk
train
in
g
b
y
re
d
u
c
in
g
in
tern
a
l
c
o
v
a
riate
sh
if
t,
”
a
rXiv p
re
p
rin
t
a
rX
iv:1
5
0
2
.
0
3
1
6
7
,
p
p
.
4
4
8
-
4
5
6
,
2
0
1
5
.
[1
2
]
C.
S
z
e
g
e
d
y
,
V
.
V
a
n
h
o
u
c
k
e
,
S
.
I
o
f
fe
,
J.
S
h
len
s,
Z.
W
o
j
n
a
,
“
Re
th
in
k
in
g
th
e
in
c
e
p
t
io
n
a
rc
h
it
e
c
tu
re
f
o
r
c
o
m
-
p
u
ter
v
isio
n
,
”
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
IEE
E
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
a
n
d
p
a
tt
e
rn
re
c
o
g
n
it
io
n
,
2
0
1
6
,
p
p
.
2
8
1
8
-
2
8
2
6
.
[1
3
]
C.
S
z
e
g
e
d
y
,
S
.
Io
ff
e
,
V
.
V
a
n
h
o
u
c
k
e
,
A
.
A
.
A
l
e
m
i,
“
In
c
e
p
ti
o
n
-
v
4
,
in
c
e
p
ti
o
n
-
re
sn
e
t
a
n
d
th
e
im
p
a
c
t
o
f
re
sid
u
a
l
c
o
n
n
e
c
ti
o
n
s o
n
lea
rn
i
n
g
,
”
Pro
c
e
e
d
in
g
s o
f
th
e
A
AA
I
C
o
n
fer
e
n
c
e
o
n
Arti
fi
c
ia
l
I
n
telli
g
e
n
c
e
,
v
o
l.
3
1
,
n
o
.
1
,
2
0
1
7
.
[1
4
]
G
.
Hu
a
n
g
,
Z.
L
iu
,
L
.
V
a
n
De
r
M
a
a
ten
,
K.
Q.
W
e
in
b
e
rg
e
r,
“
De
n
se
l
y
c
o
n
n
e
c
ted
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
tw
o
rk
s,
”
Pro
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
a
n
d
p
a
t
ter
n
re
c
o
g
n
i
ti
o
n
,
2
0
1
7
,
p
p
.
4
7
0
0
–
4
7
0
8
.
[1
5
]
S
.
S
a
b
o
u
r,
N.
F
r
o
ss
t,
G
.
E.
Hin
t
o
n
,
“
Dy
n
a
m
i
c
ro
u
ti
n
g
b
e
tw
e
e
n
c
a
p
su
les
,
”
a
rXiv
p
re
p
rin
t
a
rXiv:
1
7
1
0
.
0
9
8
2
9
,
p
p
.
3
8
5
6
-
3
8
6
6
,
2
0
1
7
.
[1
6
]
J.
Hu
,
L
.
S
h
e
n
,
G
.
S
u
n
,
“
S
q
u
e
e
z
e
-
a
n
d
-
e
x
c
it
a
ti
o
n
n
e
tw
o
rk
s
,
”
Pro
c
e
e
d
in
g
s
o
f
th
e
IEE
E
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
a
n
d
p
a
tt
e
rn
re
c
o
g
n
it
io
n
,
2
0
1
8
,
p
p
.
7
1
3
2
–
7
1
4
1
.
[1
7
]
F
.
Ch
o
ll
e
t,
“
X
c
e
p
ti
o
n
:
De
e
p
lea
rn
in
g
w
it
h
d
e
p
th
w
ise
se
p
a
ra
b
le
c
o
n
v
o
lu
ti
o
n
s,
”
Pro
c
e
e
d
in
g
s
o
f
t
h
e
IE
EE
c
o
n
fer
e
n
c
e
o
n
c
o
m
p
u
ter
v
isi
o
n
a
n
d
p
a
tt
e
rn
re
c
o
g
n
i
ti
o
n
,
2
0
1
7
,
p
p
.
1
2
5
1
-
1
2
5
8
.
[1
8
]
M
.
T
a
n
,
Q.
V
.
L
e
,
“
E
ff
icie
n
tn
e
t:
Re
th
in
k
in
g
m
o
d
e
l
sc
a
li
n
g
f
o
r
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,
”
a
rXiv
p
re
p
rin
t
a
rXiv:1
9
0
5
.
1
1
9
4
6
,
2
0
2
0
.
[1
9
]
Ka
g
g
l
e
,
A
p
to
s
2
0
1
9
b
li
n
d
n
e
ss
d
e
tec
ti
o
n
,
[
O
n
li
n
e
]
2
0
1
9
,
A
v
a
ib
le:
h
tt
p
s:/
/www
.
k
a
g
g
l
e
.
c
o
m
/
c
/ap
to
s2
0
1
9
-
b
li
n
d
n
e
ss
-
d
e
tec
ti
o
n
.
[2
0
]
Z.
L
a
i,
H.
De
n
g
,
“
M
e
d
ica
l
i
m
a
g
e
c
las
si
f
ica
ti
o
n
b
a
se
d
o
n
d
e
e
p
f
e
a
tu
re
s
e
x
tra
c
ted
b
y
d
e
e
p
m
o
d
e
l
a
n
d
sta
ti
stic
f
e
a
tu
re
f
u
sio
n
w
it
h
m
u
lt
il
a
y
e
r
p
e
rc
e
p
tro
n
,
”
Co
mp
u
ta
ti
o
n
a
l
in
telli
g
e
n
c
e
a
n
d
n
e
u
ro
sc
ien
c
e
,
2
0
1
8
.
[2
1
]
G
.
Ga
rd
n
e
r,
D.
Ke
a
ti
n
g
,
T
.
H.
W
il
li
a
m
so
n
,
A
.
T
.
El
l
io
tt
,
“
A
u
to
m
a
ti
c
d
e
tec
ti
o
n
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
u
sin
g
a
n
a
rti
f
icia
l
n
e
u
ra
l
n
e
tw
o
rk
:
a
sc
re
e
n
in
g
to
o
l.
,
”
Brit
ish
j
o
u
r
n
a
l
o
f
O
p
h
t
h
a
lmo
lo
g
y
,
v
o
l.
80
,
n
o
.
11
,
p
p
.
9
4
0
–
9
4
4
,
1
9
9
6
.
[2
2
]
J.
Da
v
id
,
R.
Krish
n
a
n
,
S
.
Ku
m
a
r
,
“
Ne
u
ra
l
n
e
t
w
o
rk
b
a
se
d
re
ti
n
a
l
im
a
g
e
a
n
a
l
y
sis,
”
2
0
0
8
Co
n
g
re
ss
o
n
Ima
g
e
a
n
d
S
ig
n
a
l
Pro
c
e
ss
in
g
,
v
o
l.
2
,
2
0
0
8
,
p
p
.
4
9
-
53.
[2
3
]
R.
P
riy
a
,
P
.
A
ru
n
a
,
“
S
v
m
a
n
d
n
e
u
ra
l
n
e
tw
o
rk
b
a
s
e
d
d
iag
n
o
sis
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Co
mp
u
ter
A
p
p
li
c
a
ti
o
n
s
,
v
o
l.
41
,
n
o
.
1
,
2
0
1
2
.
[2
4
]
R.
P
riy
a
,
P
.
A
ru
n
a
,
“
Dia
g
n
o
sis o
f
d
iab
e
ti
c
re
ti
n
o
p
a
t
h
y
u
sin
g
m
a
c
h
in
e
lea
rn
in
g
tec
h
n
iq
u
e
s,
”
ICT
ACT
J
o
u
rn
a
l
o
n
so
ft
c
o
mp
u
ti
n
g
,
v
o
l.
3
,
n
o
.
4
,
p
p
.
5
6
3
–
575
,
2
0
1
3
.
[2
5
]
S
.
Ro
y
c
h
o
w
d
h
u
ry
,
D.
D.
Ko
o
z
e
k
a
n
a
n
i,
K.
K.
P
a
rh
i,
“
Dre
a
m
:
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
a
n
a
l
y
sis u
sin
g
m
a
c
h
in
e
lea
rn
i
n
g
,
”
IEE
E
j
o
u
rn
a
l
o
f
b
io
me
d
ic
a
l
a
n
d
h
e
a
lt
h
i
n
f
o
rm
a
ti
c
s
,
v
o
l.
18
,
n
o
.
5
,
p
p
.
1
7
1
7
-
1
7
2
8
,
2
0
1
3
.
[2
6
]
S
.
S
o
m
a
su
n
d
a
ra
m
,
P
.
A
ll
i,
“
A
m
a
c
h
in
e
lea
rn
in
g
e
n
se
m
b
le
c
l
a
ss
i
f
i
e
r
f
o
r
e
a
rl
y
p
re
d
ictio
n
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
,
”
J
o
u
rn
a
l
o
f
M
e
d
ica
l
S
y
ste
ms
,
v
o
l.
41
,
n
o
.
12
,
p
p
.
1
-
1
2
,
2
0
1
7
.
[2
7
]
Q.
A
b
b
a
s,
I.
F
o
n
d
o
n
,
A
.
S
a
r
m
i
e
n
to
,
S
.
Jim
e
´
n
e
z
,
P
.
A
le
m
a
n
y
,
“
A
u
to
m
a
ti
c
r
e
c
o
g
n
it
io
n
o
f
se
v
e
rit
y
lev
e
l
f
o
r
d
iag
n
o
sis
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
u
sin
g
d
e
e
p
v
isu
a
l
f
e
a
tu
re
s,
”
M
e
d
ica
l
&
b
io
lo
g
ica
l
e
n
g
in
e
e
rin
g
&
c
o
mp
u
ti
n
g
,
v
o
l.
55
,
n
o
.
11
,
p
p
.
1
9
5
9
-
1
9
7
4
,
2
0
1
7
.
[2
8
]
R.
Ka
¨lv
ia¨in
e
n
,
H.
Uu
si
talo
,
“
D
iare
td
b
1
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
d
a
tab
a
se
a
n
d
e
v
a
lu
a
ti
o
n
p
ro
t
o
c
o
l
,
”
M
e
d
ica
l
Ima
g
e
Un
d
e
rs
ta
n
d
in
g
a
n
d
A
n
a
lys
is,
v
o
l
.
2
0
0
7
,
p
.
6
1
,
2
0
0
7
.
[2
9
]
K.
Bh
a
ti
a
,
S
.
A
ro
ra
,
R.
T
o
m
a
r,
“
Dia
g
n
o
sis
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
u
sin
g
m
a
c
h
in
e
lea
rn
in
g
c
las
si
f
ica
t
io
n
a
lg
o
rit
h
m
,
”
2
0
1
6
2
n
d
In
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
Ne
x
t
Ge
n
e
ra
ti
o
n
C
o
mp
u
ti
n
g
T
e
c
h
n
o
l
o
g
ies
(
NGCT),
2
0
1
6
,
p
p
.
3
4
7
-
3
5
1
.
[3
0
]
D.
Do
sh
i,
A
.
S
h
e
n
o
y
,
D.
S
id
h
p
u
ra
,
P
.
G
h
a
rp
u
re
,
“
Dia
b
e
ti
c
re
ti
n
o
p
a
th
y
d
e
tec
ti
o
n
u
sin
g
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,
”
in
:
2
0
1
6
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Co
mp
u
ti
n
g
,
An
a
lytics
a
n
d
S
e
c
u
rity
T
re
n
d
s
(
CAS
T
),
2
0
1
6
,
p
p
.
2
6
1
-
2
6
6
.
[3
1
]
S
.
Du
tt
a
,
B.
C
.
M
a
n
i
d
e
e
p
,
S
.
M
.
Ba
sh
a
,
R.
D.
Ca
y
ti
les
,
N.
I
y
e
n
g
a
r,
“
Clas
sif
ic
a
ti
o
n
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
ima
g
e
s
b
y
u
sin
g
d
e
e
p
lea
rn
in
g
m
o
d
e
ls,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
Gr
id
a
n
d
Distrib
u
ted
Co
m
p
u
ti
n
g
,
v
o
l.
11
,
n
o
.
1
,
p
p
.
89
–
106
,
2
0
1
8
.
[3
2
]
Krish
n
a
n
,
A
rv
in
d
S
a
i
e
t
a
l
.
,
“
A
tran
s
f
e
r
lea
rn
in
g
a
p
p
ro
a
c
h
f
o
r
d
iab
e
ti
c
re
ti
n
o
p
a
th
y
c
las
si
f
ica
t
io
n
u
sin
g
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,
”
in
:
2
0
1
8
1
5
th
IE
EE
In
d
ia
C
o
u
n
c
il
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
(
IN
DICO
N),
2
0
1
8
,
p
p
.
1
–
6.
[3
3
]
A
.
B.
Ju
n
g
e
t
a
l.
,
[
On
l
in
e
]
,
A
v
a
ib
le:
h
tt
p
s:
//
g
it
h
u
b
.
c
o
m
/ale
ju
/i
m
g
a
u
g
,
2
0
1
9
.
[3
4
]
D.
-
H.
L
e
e
,
“
P
se
u
d
o
-
la
b
e
l:
T
h
e
si
m
p
le
a
n
d
e
ff
ici
e
n
t
se
m
i
-
su
p
e
rv
ise
d
l
e
a
rn
in
g
m
e
th
o
d
f
o
r
d
e
e
p
n
e
u
ra
l
n
e
tw
o
rk
s,
”
in
:
W
o
rk
sh
o
p
o
n
C
h
a
l
len
g
e
s i
n
Rep
r
e
se
n
ta
ti
o
n
L
e
a
r
n
in
g
,
ICM
L
,
v
o
l.
3
,
n
o
.
2
,
2
0
1
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2
0
8
8
-
8708
A
tr
a
n
s
fer lea
r
n
in
g
w
ith
d
ee
p
n
eu
r
a
l n
etw
o
r
k
a
p
p
r
o
a
c
h
fo
r
…
(
Mo
h
a
mme
d
A
l
-
S
ma
d
i
)
3501
[3
5
]
H.
-
C.
S
h
in
,
H.
R.
Ro
th
,
M
.
G
a
o
,
L
.
L
u
,
Z.
X
u
,
I.
No
g
u
e
s,
J.
Ya
o
,
D.
M
o
ll
u
ra
,
R.
M
.
S
u
m
m
e
rs,
“
De
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
f
o
r
c
o
m
p
u
ter
-
a
id
e
d
d
e
tec
ti
o
n
:
Cn
n
a
rc
h
i
tec
tu
re
s,
d
a
tas
e
t
c
h
a
ra
c
teristics
a
n
d
tran
sf
e
r
lea
rn
in
g
,
”
IEE
E
tra
n
sa
c
ti
o
n
s
o
n
me
d
ica
l
im
a
g
i
n
g
,
v
o
l.
35
,
n
o
.
5
,
p
p
.
1
2
8
5
-
1
2
9
8
,
2
0
1
6
.
[3
6
]
M
.
L
in
,
Q.
C
h
e
n
,
S
.
Ya
n
,
“
Ne
tw
o
rk
in
n
e
tw
o
rk
,”
a
rXiv p
re
p
ri
n
t
a
rXiv:1
3
1
2
.
4
4
0
0
,
2
0
1
3
.
[3
7
]
B.
Zh
o
u
,
A
.
Kh
o
sla
,
A
.
L
a
p
e
d
riz
a
,
A
.
Oliv
a
,
A
.
T
o
rra
lb
a
,
“
Lea
rn
in
g
d
e
e
p
f
e
a
tu
re
s
f
o
r
d
isc
rim
in
a
ti
v
e
lo
c
a
li
z
a
ti
o
n
,
”
in
:
Pr
o
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
a
n
d
p
a
t
ter
n
re
c
o
g
n
it
io
n
,
2
0
1
6
,
p
p
.
2
9
2
1
-
2
9
2
9
.
[3
8
]
N.
S
riv
a
sta
v
a
,
G
.
Hin
to
n
,
A
.
Kriz
h
e
v
s
k
y
,
I.
S
u
tsk
e
v
e
r,
R.
S
a
lak
h
u
t
d
in
o
v
,
“
Dro
p
o
u
t
:
a
sim
p
le
w
a
y
to
p
re
v
e
n
t
n
e
u
ra
l
n
e
tw
o
rk
s
f
ro
m
o
v
e
r
f
it
ti
n
g
,
”
T
h
e
j
o
u
rn
a
l
o
f
ma
c
h
in
e
le
a
rn
i
n
g
re
se
a
rc
h
,
v
o
l.
15
,
n
o
.
1
,
p
p
.
1
9
2
9
-
1
9
5
8
,
2
0
1
4
.
[3
9
]
K.
He
,
X
.
Z
h
a
n
g
,
S
.
Re
n
,
J.
S
u
n
,
“
Id
e
n
ti
ty
m
a
p
p
in
g
s
in
d
e
e
p
r
e
sid
u
a
l
n
e
tw
o
rk
s,
”
in
:
Eu
r
o
p
e
a
n
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
,
S
p
ri
n
g
e
r,
2
0
1
6
,
p
p
.
6
3
0
-
6
4
5
.
[4
0
]
J.
Co
h
e
n
,
“
W
e
ig
h
ted
k
a
p
p
a
:
N
o
m
in
a
l
sc
a
l
e
a
g
re
e
m
e
n
t
p
ro
v
isio
n
f
o
r
sc
a
led
d
isa
g
re
e
m
e
n
t
o
r
p
a
rti
a
l
c
re
d
it
,”
Psy
c
h
o
lo
g
ica
l
b
u
l
letin
,
v
o
l.
70
,
n
o
.
4
,
1
9
6
8
.
[4
1
]
C.
T
a
n
,
F
.
S
u
n
,
T
.
Ko
n
g
,
W
.
Zh
a
n
g
,
C.
Ya
n
g
,
C.
L
iu
,
“
A
su
rv
e
y
o
n
d
e
e
p
tran
sf
e
r
lea
rn
in
g
,
”
in
:
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Arti
fi
c
i
a
l
Ne
u
r
a
l
Ne
two
rk
s,
S
p
ri
n
g
e
r,
2
0
1
8
,
p
p
.
2
7
0
-
2
7
9
.
[4
2
]
G
.
L
it
jen
s,
T
.
Ko
o
i,
B.
E.
Be
jn
o
rd
i,
A
.
A
.
A
.
S
e
ti
o
,
F
.
Cio
m
p
i
,
M
.
G
h
a
f
o
o
rian
,
J.
A
.
V
a
n
De
r
L
a
a
k
,
B.
V
a
n
G
in
n
e
k
e
n
,
C.
I.
S
a
´
n
c
h
e
z
,
“
A
su
r
v
e
y
o
n
d
e
e
p
lea
rn
in
g
in
m
e
d
ica
l
ima
g
e
a
n
a
l
y
sis,
”
M
e
d
ica
l
ima
g
e
a
n
a
lys
is
,
v
o
l.
42
,
p
p
.
60
-
88
,
2
0
1
7
.
[4
3
]
L
.
P
e
re
z
,
J.
W
a
n
g
,
“
T
h
e
e
ffe
c
ti
v
e
n
e
ss
o
f
d
a
ta
a
u
g
m
e
n
tatio
n
i
n
i
m
a
g
e
c
l
a
ss
i
f
ica
ti
o
n
u
sin
g
d
e
e
p
l
e
a
rn
in
g
,
”
a
rXiv
p
re
p
rin
t
a
rXiv:1
7
1
2
.
0
4
6
2
1
,
2
0
1
7
.
[4
4
]
N.
Ja
p
k
o
w
icz
,
S
.
S
tep
h
e
n
,
“
T
h
e
c
las
s
i
m
b
a
lan
c
e
p
ro
b
lem
:
A
s
y
st
e
m
a
ti
c
stu
d
y
,
”
In
telli
g
e
n
t
d
a
ta
a
n
a
lys
is
,
v
o
l.
6
,
n
o
.
5
,
p
p
.
4
2
9
-
4
4
9
,
2
0
0
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.