I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
,
p
p
.
2554
~
2
5
6
6
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
3
.
p
p
2
5
5
4
-
2
5
6
6
2554
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Im
a
g
e enhancem
ent
co
mbin
ed wi
t
h Eff
ic
ient
Net
-
B7
f
o
r gra
ding
cla
ss
ificatio
n of di
a
betic
re
tinopa
th
y
Nina
Sev
a
ni
1,
2
,
E
dy
K
ristia
n
t
o
1,
3
,
Albert
Sa
lo
m
o
1
1
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
c
s
,
F
a
c
u
l
t
y
o
f
En
g
i
n
e
e
r
i
n
g
a
n
d
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
K
r
i
d
a
W
a
c
a
n
a
C
h
r
i
st
i
a
n
U
n
i
v
e
r
s
i
t
y
,
J
a
k
a
r
t
a
,
I
n
d
o
n
e
si
a
2
R
e
s
e
a
r
c
h
C
e
n
t
e
r
f
o
r
K
n
o
w
l
e
d
g
e
M
a
n
a
g
e
me
n
t
a
n
d
C
o
l
l
a
b
o
r
a
t
i
v
e
I
n
n
o
v
a
t
i
o
n
,
K
r
i
d
a
W
a
c
a
n
a
C
h
r
i
s
t
i
a
n
U
n
i
v
e
r
si
t
y
,
Ja
k
a
r
t
a
,
I
n
d
o
n
e
s
i
a
3
R
e
s
e
a
r
c
h
C
e
n
t
e
r
f
o
r
C
o
n
n
e
c
t
e
d
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
S
e
c
u
r
i
t
y
,
K
r
i
d
a
W
a
c
a
n
a
C
h
r
i
st
i
a
n
U
n
i
v
e
r
s
i
t
y
,
Ja
k
a
r
t
a
,
I
n
d
o
n
e
s
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
1
7
,
2
0
2
4
R
ev
is
ed
Feb
9
,
2
0
2
6
Acc
ep
ted
Ma
r
5
,
2
0
2
6
Dia
b
e
ti
c
r
e
ti
n
o
p
a
th
y
(DR)
is
a
c
o
m
p
li
c
a
ti
o
n
c
a
u
se
d
b
y
p
o
o
rl
y
m
a
n
a
g
e
d
d
iab
e
tes
th
a
t
a
ffe
c
ts
th
e
e
y
e
s.
Ac
c
o
rd
in
g
to
th
e
Wo
rl
d
He
a
lt
h
Org
a
n
iza
ti
o
n
(W
HO
),
4
2
2
m
il
li
o
n
p
e
o
p
le
wo
rld
wid
e
h
a
v
e
su
ffe
re
d
fro
m
DR
i
n
th
e
p
a
st
ten
y
e
a
rs.
M
a
n
u
a
l
d
e
tec
ti
o
n
u
si
n
g
re
ti
n
a
l
fu
n
d
u
s
ima
g
e
s
is
ti
m
e
-
c
o
n
su
m
i
n
g
a
n
d
re
q
u
ires
e
x
p
e
rien
c
e
d
o
p
h
t
h
a
lmo
lo
g
ists.
T
h
is
st
u
d
y
p
ro
p
o
se
s
a
d
e
e
p
lea
rn
in
g
m
e
th
o
d
u
si
n
g
th
e
p
re
-
tra
in
e
d
m
o
d
e
l
Eff
icie
n
tNe
t
-
B
7
t
o
id
e
n
ti
fy
th
is
d
ise
a
se
a
u
to
m
a
ti
c
a
ll
y
.
F
i
v
e
lev
e
ls
o
f
DR
will
b
e
c
las
sified
:
n
o
-
DR,
m
il
d
-
DR
,
m
o
d
e
ra
te
-
D
R,
se
v
e
re
-
DR,
a
n
d
p
ro
li
fe
ra
ti
v
e
-
DR
.
Th
e
m
o
d
e
l
w
a
s
train
e
d
u
sin
g
"
AP
TOS
2
0
1
9
b
l
in
d
n
e
ss
d
e
tec
ti
o
n
"
d
a
tas
e
t,
a
n
d
ima
g
e
a
u
g
m
e
n
tatio
n
wa
s
p
e
rfo
rm
e
d
.
Im
a
g
e
s
e
g
m
e
n
tatio
n
tec
h
n
iq
u
e
s
su
c
h
a
s
c
o
n
tra
st
li
m
it
e
d
a
d
a
p
ti
v
e
h
ist
o
g
ra
m
e
q
u
a
li
z
a
ti
o
n
(CLAHE)
a
n
d
re
a
l
e
n
h
a
n
c
e
d
su
p
e
r
re
so
lu
ti
o
n
g
e
n
e
ra
ti
v
e
a
d
v
e
rsa
rial
n
e
two
r
k
(Re
a
l
-
ES
RG
AN
)
we
r
e
a
p
p
li
e
d
d
u
ri
n
g
p
re
p
ro
c
e
ss
in
g
t
o
imp
r
o
v
e
th
e
m
o
d
e
l'
s
a
c
c
u
ra
c
y
sig
n
ifi
c
a
n
tl
y
.
T
h
e
imp
lem
e
n
tatio
n
o
f
CLAHE
re
s
u
lt
e
d
in
t
h
e
v
a
li
d
a
ti
o
n
a
c
c
u
ra
c
y
imp
ro
v
e
m
e
n
t
fro
m
7
6
.
6
%
to
8
3
.
4
%
c
o
m
p
a
re
d
to
n
o
se
g
m
e
n
tatio
n
,
wh
il
e
th
e
c
o
m
b
in
a
ti
o
n
o
f
Re
a
l
-
ES
RG
AN
a
n
d
CLAHE
in
c
re
a
se
d
th
e
a
c
c
u
ra
c
y
to
9
3
.
7
%
.
F
u
tu
r
e
re
se
a
rc
h
c
a
n
e
x
p
lo
re
th
e
c
o
m
b
in
a
ti
o
n
o
f
CLAHE
wit
h
o
t
h
e
r
ima
g
e
p
ro
c
e
ss
in
g
tec
h
n
i
q
u
e
s a
p
a
rt
fr
o
m
th
e
Re
a
l
-
ES
RG
AN
m
o
d
e
l.
K
ey
w
o
r
d
s
:
C
o
n
tr
ast lim
ited
ad
ap
tiv
e
h
is
to
g
r
am
eq
u
aliza
tio
n
E
f
f
icien
tNet
-
B7
Fu
n
d
u
s
I
m
ag
e
s
eg
m
en
tatio
n
R
ea
l e
n
h
an
ce
d
s
u
p
er
-
r
eso
lu
tio
n
g
en
er
ativ
e
ad
v
er
s
ar
ial
n
etwo
r
k
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
:
Nin
a
Sev
an
i
D
e
p
a
r
t
m
e
n
t
o
f
I
n
f
o
r
m
a
t
i
c
s
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
K
r
i
d
a
W
a
c
a
n
a
C
h
r
i
s
t
i
a
n
U
n
i
v
e
r
s
i
t
y
T
an
ju
n
g
D
u
r
en
R
ay
a
No
.
4
,
W
est
J
ak
ar
ta,
DKI
J
ak
ar
ta,
I
n
d
o
n
esia
E
m
ail: n
in
a.
s
ev
an
i@
u
k
r
id
a.
ac
.
id
1.
I
NT
RO
D
UCT
I
O
N
Diab
etic
r
etin
o
p
ath
y
(
DR
)
is
th
e
lead
in
g
ca
u
s
e
o
f
b
lin
d
n
ess
in
p
er
s
o
n
s
o
f
m
at
u
r
e
a
g
e.
W
o
r
k
wo
r
ld
wid
e
is
r
esp
o
n
s
ib
le
f
o
r
m
o
r
e
th
an
2
4
,
0
0
0
ca
s
es
o
f
b
lin
d
n
ess
p
er
y
ea
r
[
1
]
.
T
h
e
W
o
r
ld
Hea
lth
Or
g
an
izatio
n
(
W
HO)
r
ep
o
r
ted
th
at
4
2
2
m
illi
o
n
p
e
o
p
le
s
u
f
f
e
r
ed
f
r
o
m
DR
wo
r
ld
wid
e
in
th
e
last
ten
y
ea
r
s
.
I
n
2
0
1
7
,
th
e
r
e
wer
e
4
2
5
m
illi
o
n
DR
p
atien
ts
wo
r
ld
wid
e
,
with
an
esti
m
atio
n
th
at
6
0
0
m
illi
o
n
p
e
o
p
le
will
h
av
e
d
iab
etes in
th
e
y
ea
r
2
0
4
0
,
with
o
n
e
-
th
ir
d
o
f
ca
s
es e
x
p
er
ien
cin
g
DR
[
2
]
.
DR
is
a
co
m
p
licatio
n
o
f
t
y
p
e
1
an
d
ty
p
e
2
d
ia
b
etes,
p
r
im
a
r
ily
in
p
atien
ts
wit
h
ch
r
o
n
ic
h
y
p
er
g
ly
ce
m
ia,
p
o
o
r
g
ly
ce
m
ic
co
n
tr
o
l,
an
d
en
h
an
ce
m
en
t
o
f
b
lo
o
d
p
r
ess
u
r
e
[
2
]
.
DR
ca
n
class
if
ied
in
to
two
s
tag
es,
n
am
ely
m
ild
-
DR
,
s
t
ag
e
b
eg
in
n
i
n
g
DR
,
wh
ich
is
m
ar
k
ed
b
y
ex
is
tin
g
m
icr
o
an
eu
r
y
s
m
(
MA
)
,
p
r
o
life
r
ativ
e
-
DR
is
s
tag
e
s
ec
o
n
d
DR
an
d
ca
n
r
esu
lt
in
lo
s
s
o
f
v
is
io
n
[
3
]
.
Mo
r
eo
v
er
,
DR
c
an
in
cr
ea
s
e
th
e
r
is
k
o
f
d
am
ag
e,
d
y
s
f
u
n
ctio
n
,
an
d
f
ailu
r
e
in
o
r
g
an
s
an
d
b
o
d
y
tis
s
u
es,
s
u
ch
as
k
id
n
ey
d
is
o
r
d
er
s
,
h
ea
r
t
d
is
ea
s
e,
an
d
th
e
n
er
v
o
u
s
s
y
s
tem
[
2
]
.
Scr
ee
n
in
g
r
eg
u
lar
tr
ea
tm
en
t
f
o
r
DR
ca
s
es
i
s
n
ee
d
ed
f
o
r
ea
r
ly
d
etec
tio
n
an
d
m
an
ag
em
en
t t
o
p
r
e
v
en
t b
lin
d
n
ess
[
3
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
ma
g
e
en
h
a
n
ce
men
t c
o
mb
in
ed
w
ith
E
fficien
tN
et
-
B
7
fo
r
g
r
a
d
in
g
cla
s
s
ifica
tio
n
o
f d
ia
b
etic
…
(
N
in
a
S
ev
a
n
i)
2555
C
u
r
r
en
tly
,
o
p
h
th
alm
o
lo
g
y
d
ia
g
n
o
s
es
an
d
ass
ess
e
s
th
e
s
ev
er
i
ty
o
f
DR
th
r
o
u
g
h
v
is
u
al
ev
alu
atio
n
with
d
ir
ec
t
in
s
p
ec
tio
n
a
n
d
e
v
alu
ati
o
n
o
f
th
e
e
y
e
with
th
e
in
s
p
ec
t
io
n
f
u
n
d
o
s
co
p
y
[
4
]
.
T
h
is
in
s
p
ec
tio
n
ex
am
in
atio
n
an
d
d
eter
m
in
atio
n
o
f
t
h
e
s
ev
er
ity
o
f
DR
is
u
s
u
ally
d
o
n
e
b
y
an
aly
zin
g
lesi
o
n
f
ea
tu
r
es
f
o
u
n
d
o
n
th
e
p
atien
t'
s
f
u
n
d
u
s
im
ag
e
[
5
]
.
Ho
wev
er
,
t
h
is
m
eth
o
d
is
tim
e
-
co
n
s
u
m
in
g
,
lab
o
r
-
in
te
n
s
iv
e,
an
d
p
r
o
n
e
to
h
u
m
an
er
r
o
r
[
5
]
.
C
o
n
s
id
er
in
g
f
u
n
d
u
s
co
p
ic
ex
a
m
in
atio
n
s
lik
e
th
is
s
till
r
ely
o
n
o
p
h
th
alm
o
l
o
g
y
e
x
p
er
tis
e
f
o
r
i
n
ter
p
r
etatio
n
,
e
r
r
o
r
s
in
DR
d
etec
tio
n
r
esu
lts
ca
n
o
cc
u
r
[
2
]
.
I
t
is
cr
u
cial
to
im
p
r
o
v
e
th
e
co
n
s
is
ten
cy
o
f
p
r
e
d
ictio
n
r
esu
lts
an
d
th
e
lev
el
o
f
p
r
ec
is
io
n
an
d
e
f
f
icien
cy
in
d
iag
n
o
s
is
DR
,
u
s
in
g
th
e
h
elp
o
f
a
r
tific
ial
in
tellig
en
ce
(
AI
)
tech
n
o
l
o
g
y
.
Pre
v
io
u
s
r
esear
ch
s
h
o
ws
th
at
th
e
r
ap
id
d
ev
elo
p
m
en
t
o
f
A
I
ca
n
p
r
o
d
u
ce
a
DR
d
is
ea
s
e
d
etec
tio
n
m
o
d
el
th
at
ca
n
p
r
o
v
id
e
r
esu
lts
q
u
ick
l
y
with
co
n
tin
u
o
u
s
ly
im
p
r
o
v
in
g
le
v
els
o
f
ac
cu
r
ac
y
.
Ma
ch
in
e
lear
n
in
g
an
d
d
ee
p
lear
n
in
g
ar
e
c
o
m
m
o
n
ly
u
s
ed
f
o
r
DR
d
etec
tio
n
.
Gen
er
ally
,
ac
cu
r
ac
y
a
n
d
p
r
ec
is
io
n
a
r
e
th
e
f
ac
to
r
s
m
o
s
t
wid
ely
u
s
ed
to
ass
ess
th
e
p
er
f
o
r
m
an
ce
o
f
d
ee
p
lear
n
i
n
g
m
o
d
els
in
d
etec
tin
g
DR
.
T
h
e
u
s
e
o
f
tech
n
o
lo
g
y
lik
e
th
is
ca
n
also
h
elp
th
e
task
o
f
o
p
h
th
alm
o
lo
g
y
in
th
e
ea
r
ly
d
etec
tio
n
o
f
t
h
e
s
ev
er
ity
o
f
DR
d
is
ea
s
e
in
p
atien
ts
[
6
]
.
Ultim
ately
,
tech
n
o
lo
g
y
f
o
r
d
etec
tin
g
th
e
s
ev
e
r
ity
o
f
DR
d
is
ea
s
e
will
h
elp
ar
ea
s
with
lim
ited
n
u
m
b
er
s
o
f
o
p
h
th
alm
o
l
o
g
is
ts
p
r
o
v
id
e
s
er
v
ices to
th
e
co
m
m
u
n
ity
.
I
n
d
e
t
e
c
t
i
n
g
DR
d
i
s
e
as
e
,
s
e
v
e
r
it
y
l
e
v
e
l
c
a
n
b
e
c
l
as
s
i
f
i
e
d
i
n
t
o
s
e
v
e
r
a
l
l
e
v
e
ls
.
R
e
s
e
a
r
c
h
b
y
Q
i
an
e
t
a
l
.
[
7
]
,
wh
ich
d
etec
ts
DR
d
is
ea
s
e
in
to
f
iv
e
lev
els
u
s
in
g
a
co
m
b
i
n
atio
n
o
f
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
ar
ch
itectu
r
es,
n
am
ely
R
es2
Net
an
d
Den
s
eNe
t,
also
p
r
o
d
u
ce
s
p
r
ed
ictio
n
ac
cu
r
ac
y
a
s
h
ig
h
as
8
3
.
2
%.
Den
s
eNe
t
-
1
2
1
is
also
u
s
ed
in
DR
d
is
ea
s
e
d
etec
tio
n
r
esear
ch
,
an
d
p
r
o
d
u
cin
g
an
ac
cu
r
ac
y
o
f
9
8
.
3
6
%
[
8
]
.
Oth
e
r
s
tu
d
ies
wer
e
ca
r
r
ied
o
u
t
b
y
[
9
]
u
s
in
g
th
e
E
f
f
icien
tNet
-
B
7
ar
ch
itectu
r
e
to
g
et
a
lev
el
o
f
p
r
e
d
ictio
n
ac
cu
r
ac
y
as
h
ig
h
as
8
9
.
1
%.
Me
an
wh
ile,
o
th
er
s
tu
d
ies
also
tr
y
to
co
m
p
ar
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
s
ev
er
al
d
ee
p
lear
n
in
g
ar
ch
itectu
r
es
in
d
etec
tin
g
DR
d
is
ea
s
e,
s
u
ch
as
th
e
c
o
m
p
a
r
is
o
n
b
etwe
en
R
esNet
-
3
4
,
VG
G
-
1
6
,
I
n
ce
p
tio
n
-
V3
,
an
d
E
f
f
icien
tNetB
4
[
1
0
]
o
r
co
m
p
ar
in
g
b
etwe
en
R
esNet
-
5
0
,
R
esNet
-
1
5
2
,
an
d
Sq
u
ee
ze
Net1
[
1
1
]
.
Pre
v
io
u
s
r
esear
ch
also
s
h
o
ws
th
at
o
th
er
s
tag
es
n
ee
d
to
b
e
ca
r
r
ied
o
u
t
t
o
p
r
o
d
u
ce
a
d
ee
p
lear
n
in
g
m
o
d
el
with
g
o
o
d
ac
cu
r
ac
y
p
er
f
o
r
m
a
n
ce
.
Go
o
d
d
ee
p
lea
r
n
in
g
m
o
d
el
p
er
f
o
r
m
an
ce
is
o
b
tain
ed
n
o
t
o
n
ly
b
y
r
ely
in
g
o
n
th
e
p
r
o
p
er
ar
r
an
g
e
m
en
t
o
f
co
n
v
o
lu
tio
n
b
lo
c
k
s
an
d
h
y
p
er
p
ar
a
m
eter
s
s
ettin
g
s
.
I
m
ag
e
q
u
ality
d
ata
is
v
ital
an
d
ca
n
b
e
o
p
tim
ized
b
y
ap
p
ly
in
g
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es,
s
u
ch
as
im
ag
e
en
h
an
ce
m
en
t
an
d
s
eg
m
en
tatio
n
.
I
n
p
r
o
ce
s
s
in
g
an
im
ag
e,
th
e
p
r
e
p
r
o
ce
s
s
in
g
s
tag
e
h
elp
s
r
ed
u
ce
co
m
p
lex
ity
p
ar
am
eter
s
an
d
in
cr
ea
s
e
th
e
ac
cu
r
ac
y
o
f
a
m
o
d
el.
I
n
ar
ch
itectu
r
e
d
ee
p
lear
n
in
g
,
th
e
m
o
s
t
p
r
o
m
in
e
n
t
f
ea
tu
r
es
ar
e
s
elec
ted
t
o
test
ea
ch
class
o
n
a
d
ataset.
Sp
ec
if
ically
,
in
p
r
o
ce
s
s
in
g
im
ag
e
d
ata,
im
ag
e
en
h
a
n
ce
m
en
t,
an
d
s
eg
m
en
tatio
n
ar
e
cr
u
cial
f
o
r
s
ep
a
r
atin
g
an
d
ex
tr
ac
tin
g
p
ar
ts
o
f
t
h
e
im
ag
e.
I
n
p
r
in
cip
le,
DR
is
a
co
n
d
it
io
n
th
at
d
e
v
elo
p
s
f
r
o
m
u
n
tr
e
ated
d
iab
etes
m
ellitu
s
.
T
h
e
s
y
m
p
to
m
s
ex
p
er
ien
ce
d
b
y
DR
p
atien
ts
c
an
v
ar
y
d
ep
en
d
in
g
o
n
th
e
s
ev
er
ity
o
f
th
e
c
o
n
d
itio
n
.
Patien
ts
with
m
ild
-
DR
m
ay
n
o
t e
x
p
er
ien
ce
an
y
s
y
m
p
to
m
s
,
m
ak
in
g
it d
if
f
icu
lt to
r
ea
lize
th
e
p
r
esen
ce
o
f
th
e
d
is
ea
s
e.
As
DR
b
ec
o
m
es m
o
r
e
s
ev
er
e,
p
atien
ts
m
a
y
s
u
f
f
er
f
r
o
m
v
is
u
al
d
is
tu
r
b
an
ce
s
,
r
an
g
in
g
f
r
o
m
b
lu
r
r
ed
o
r
d
is
to
r
ted
v
is
io
n
to
co
m
p
lete
lo
s
s
o
f
v
is
io
n
.
T
h
e
s
ev
er
ity
o
f
DR
ca
n
b
e
id
en
tifie
d
b
y
ex
am
in
in
g
f
u
n
d
u
s
im
ag
es
o
b
tain
ed
t
h
r
o
u
g
h
f
u
n
d
o
s
co
p
ic
ex
a
m
in
atio
n
.
So
m
e
p
ar
ts
o
f
th
e
f
u
n
d
u
s
u
s
ed
t
o
d
etec
t
DR
ar
e
MA
,
h
em
o
r
r
h
ag
es
(
HE
M)
,
h
ar
d
ex
u
d
ates,
an
d
co
tto
n
wo
o
l
s
p
o
ts
(
C
W
S)
[
1
2
]
.
MA
is
an
ea
r
l
y
s
ig
n
o
f
DR
,
ty
p
ically
id
en
tifie
d
b
y
th
e
p
r
esen
ce
o
f
s
m
all,
r
o
u
n
d
,
a
n
d
d
ar
k
r
ed
d
o
ts
o
n
th
e
ey
e'
s
f
u
n
d
u
s
.
HE
M
r
ef
er
s
to
b
leed
in
g
th
at
ca
u
s
es th
e
r
ed
d
o
t to
g
r
o
w
lar
g
er
with
ir
r
eg
u
lar
ed
g
es.
HE
M
is
y
ello
wis
h
in
ap
p
ea
r
an
c
e
with
an
ir
r
eg
u
lar
an
d
s
h
in
y
cir
cu
lar
r
in
g
s
h
ap
e.
C
W
S a
r
e
g
r
ey
is
h
-
wh
ite
im
ag
e
s
o
n
r
etin
a
o
f
ey
e
with
ir
r
eg
u
la
r
ed
g
es,
r
esem
b
lin
g
a
r
o
ll o
f
f
l
u
f
f
y
c
o
tto
n
[
1
3
]
.
C
u
r
r
en
tly
,
r
esear
ch
u
s
in
g
f
u
n
d
u
s
im
ag
es
to
d
etec
t
th
e
s
ev
er
it
y
o
f
DR
d
is
ea
s
e
i
s
m
o
s
tly
em
p
lo
y
ed
th
e
d
ee
p
lear
n
in
g
tech
n
o
l
o
g
y
c
o
m
b
in
ed
with
v
ar
i
o
u
s
o
th
er
te
ch
n
iq
u
es.
Pre
v
io
u
s
r
esear
ch
a
ls
o
d
em
o
n
s
tr
ated
th
e
u
s
e
o
f
v
ar
io
u
s
d
ee
p
lear
n
in
g
a
r
ch
itectu
r
es
f
o
r
d
etec
tin
g
th
e
s
ev
er
ity
o
f
DR
d
is
ea
s
e
with
im
p
r
o
v
e
d
r
esu
lts
o
v
er
tim
e.
T
h
e
d
ee
p
lear
n
in
g
ar
ch
itectu
r
es
u
s
ed
in
clu
d
e
R
esNe
t,
Den
s
eNe
t,
E
f
f
ici
en
tNet,
VGG,
an
d
I
n
ce
p
tio
n
[
7
]
,
[
8
]
,
[
1
0
]
,
[
1
1
]
,
[
1
4
]
–
[
1
8
]
.
T
h
e
u
s
e
o
f
a
co
m
b
in
atio
n
o
f
two
C
NN
-
b
ased
ar
ch
itectu
r
es,
n
am
ely
R
esNet
an
d
Den
s
eNe
t,
h
as
als
o
b
ee
n
ca
r
r
ied
o
u
t
a
n
d
g
a
v
e
t
h
e
b
est
r
esu
lts
at
a
v
alu
e
o
f
8
3
.
2
%
[
7
]
.
R
esear
ch
u
s
in
g
R
esNet
s
h
o
ws
th
at
R
e
s
Net
-
1
5
2
p
r
o
v
id
es
ac
cu
r
ac
y
r
e
s
u
lts
o
f
9
4
.
4
0
%,
th
is
v
al
u
e
is
h
ig
h
er
t
h
an
u
s
in
g
R
esNet
-
50
[
1
1
]
.
Oth
er
r
esear
c
h
u
s
in
g
Den
s
eNe
t1
2
1
also
g
a
v
e
g
o
o
d
r
esu
lts
,
r
ea
ch
in
g
9
8
.
3
6
%
[
8
]
.
T
h
e
u
s
e
o
f
I
n
ce
p
tio
n
-
V3
also
p
r
o
v
id
es
im
p
r
o
v
e
d
ac
cu
r
ac
y
r
esu
lts
wh
en
co
m
b
in
ed
with
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es
s
u
ch
as
co
n
tr
ast
lim
ited
ad
ap
tiv
e
h
is
to
g
r
am
e
q
u
aliza
tio
n
(
C
L
AHE
)
s
h
o
win
g
an
im
p
r
o
v
em
en
t
i
n
ac
cu
r
ac
y
v
alu
es
o
f
4
.
4
%.
T
h
is
v
alu
e
is
b
etter
th
a
n
wh
en
C
L
AHE
was
u
s
ed
to
g
eth
er
with
E
f
f
icien
tNet
-
B
4
w
h
ich
o
n
ly
p
r
o
v
id
ed
an
im
p
r
o
v
em
e
n
t
in
v
alu
e
o
f
2
.
3
%
[
1
0
]
.
Sev
er
al
p
r
e
v
io
u
s
s
tu
d
ies
also
s
h
o
w
th
at
t
h
er
e
a
r
e
o
th
er
d
ee
p
lear
n
in
g
ar
ch
itectu
r
es
th
at
ar
e
u
s
ed
to
d
etec
t
s
p
ec
if
ic
DR
d
is
ea
s
es
b
ased
o
n
HE
M
a
n
d
MA
[
1
8
]
,
[
1
9
]
an
d
p
r
o
v
id
e
th
e
b
est s
en
s
itiv
ity
r
ea
ch
in
g
9
4
%
[
2
0
]
.
E
f
f
icien
tNet
[
2
1
]
is
an
ar
ch
it
ec
tu
r
e
d
ee
p
lear
n
in
g
th
at
u
tili
ze
s
co
n
ce
p
ts
o
f
s
ca
lin
g
u
p
an
d
s
ca
lin
g
d
o
wn
o
n
ea
ch
n
eu
r
al
n
etwo
r
k
(
NN)
a
n
d
p
r
o
d
u
ce
s
a
lig
h
tweig
h
t
m
o
d
el
a
n
d
g
o
o
d
p
er
f
o
r
m
an
ce
.
E
f
f
icien
tNet
g
o
o
d
p
er
f
o
r
m
a
n
ce
is
d
u
e
to
th
is
ar
ch
itectu
r
e
u
tili
zin
g
tr
a
n
s
f
er
lear
n
in
g
tech
n
iq
u
es,
w
h
ich
th
e
m
o
d
el
h
as
alr
ea
d
y
tr
ain
ed
o
n
d
atasets
I
m
ag
eNe
t.
C
u
r
r
en
tly
,
E
f
f
icien
tNet
h
as
d
ev
elo
p
ed
f
r
o
m
E
f
f
icien
tNet
-
B
0
to
E
f
f
icien
tNet
-
B
7
.
E
f
f
icien
tNet
-
B
7
m
ak
es
u
s
e
o
f
b
lo
ck
c
o
n
v
o
lu
ti
o
n
s
to
ca
p
tu
r
e
cr
u
cial
im
ag
e
f
ea
tu
r
es,
r
esu
ltin
g
in
im
p
r
o
v
ed
ac
cu
r
ac
y
o
f
8
4
.
4
% a
n
d
f
aster
p
r
o
ce
s
s
in
g
tim
e
[
2
1
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
5
5
4
-
2
5
6
6
2556
So
m
e
r
esear
ch
es
s
h
o
w
th
at
d
e
ep
lear
n
in
g
a
r
ch
itectu
r
es
ar
e
c
o
m
b
in
ed
with
p
r
ep
r
o
ce
s
s
in
g
m
eth
o
d
s
o
r
o
th
er
m
eth
o
d
s
to
p
r
o
v
i
d
e
b
etter
ac
cu
r
ac
y
.
On
e
o
f
th
em
is
th
e
u
s
e
o
f
K
-
m
ea
n
s
s
eg
m
en
tat
io
n
in
ar
ch
itectu
r
e
C
NN
in
class
if
y
in
g
ca
tar
ac
t d
is
ea
s
e
[
2
2
]
.
T
h
e
r
esear
ch
r
esu
lts
s
h
o
w
th
at
s
eg
m
e
n
tatio
n
,
wh
i
ch
is
ca
r
r
ie
d
o
u
t
as
a
p
r
e
p
r
o
ce
s
s
in
g
tec
h
n
iq
u
e
,
a
p
ar
t
f
r
o
m
p
r
o
v
id
in
g
g
o
o
d
ac
cu
r
ac
y
,
ca
n
also
s
h
o
r
te
n
th
e
d
ata
tim
e
tr
ain
in
g
m
o
d
els.
T
h
er
e
a
r
e
also
o
th
er
s
tu
d
ies
th
at
co
m
p
ar
e
s
ev
er
al
p
r
ep
r
o
ce
s
s
in
g
tech
n
i
q
u
e
s
an
d
o
b
tain
t
h
at
C
L
AHE
wo
r
k
s
well
to
im
p
r
o
v
e
im
ag
e
q
u
ality
f
u
n
d
u
s
[
2
3
]
.
T
h
e
ap
p
licatio
n
o
f
C
L
AHE
to
in
cr
ea
s
e
ac
cu
r
ac
y
in
DR
d
is
ea
s
e
d
etec
tio
n
was
al
s
o
ca
r
r
ied
o
u
t
b
y
[
1
0
]
,
[
1
1
]
,
[
2
4
]
with
th
e
h
ig
h
est
ac
cu
r
ac
y
o
b
t
ain
ed
b
ein
g
9
4
.
4
%
u
s
in
g
R
esNet
-
1
5
2
.
C
L
AHE
c
an
im
p
r
o
v
e
ac
c
u
r
ac
y
b
ec
au
s
e
it
ca
n
in
cr
ea
s
e
c
o
n
tr
ast
h
is
to
g
r
am
e
q
u
aliza
tio
n
(
HE
)
wh
ich
will
m
ak
e
th
e
i
m
ag
e
m
o
r
e
b
ea
u
tif
u
l
[
2
5
]
.
U
n
f
o
r
tu
n
ately
,
in
c
r
ea
s
in
g
im
ag
e
co
n
tr
ast
ca
n
also
in
cr
ea
s
e
th
e
co
n
tr
ast
o
f
o
th
er
p
ar
ts
th
at
ar
e
n
o
t
n
ec
ess
ar
y
f
o
r
DR
d
etec
tio
n
.
T
h
er
ef
o
r
e,
o
th
er
s
tu
d
ies
u
s
e
a
co
m
b
in
atio
n
o
f
C
L
AHE
with
en
h
an
ce
d
s
u
p
er
-
r
eso
lu
tio
n
g
e
n
er
ativ
e
ad
v
e
r
s
ar
ial
n
etwo
r
k
(
E
SR
GAN)
f
o
r
DR
d
is
ea
s
e
d
etec
tio
n
an
d
p
r
o
v
id
e
s
th
e
b
est
ac
cu
r
ac
y
r
esu
lts
o
f
9
8
.
3
6
%
[
8
]
.
E
SR
GAN
wo
r
k
s
to
im
p
r
o
v
e
im
ag
e
tex
tu
r
e
to
b
e
m
o
r
e
r
ea
lis
tic
an
d
n
atu
r
al
in
an
im
ag
e
[
2
6
]
.
Pre
v
io
u
s
r
esear
ch
s
h
o
ws
th
at
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es
an
d
d
ee
p
lea
r
n
in
g
ar
ch
itectu
r
e
ca
n
p
r
o
v
id
e
g
o
o
d
ac
c
u
r
ac
y
f
o
r
DR
d
is
ea
s
e
d
etec
tio
n
.
Mo
r
eo
v
er
,
th
e
r
e
is
s
till
a
ch
an
ce
f
o
r
in
cr
e
ased
ac
cu
r
ac
y
with
f
aster
p
r
ep
r
o
ce
s
s
in
g
tim
e.
T
h
er
ef
o
r
e,
th
is
r
esear
ch
will
u
s
e
th
e
r
ea
l
en
h
an
ce
d
s
u
p
er
r
e
s
o
lu
tio
n
g
en
er
ativ
e
ad
v
er
s
ar
ial
n
etwo
r
k
(
R
ea
l
-
E
SR
GAN)
tech
n
iq
u
e,
d
e
v
elo
p
m
en
t
o
f
E
SR
GAN,
to
b
e
c
o
m
b
in
ed
with
th
e
E
f
f
icien
tNet
-
B
7
ar
ch
itectu
r
e.
2.
M
E
T
H
O
D
T
h
e
m
e
t
h
o
d
u
s
e
d
i
n
t
h
i
s
s
t
u
d
y
b
e
g
i
n
s
w
i
t
h
s
p
l
it
t
i
n
g
t
h
e
d
a
tas
e
t
i
n
t
o
d
at
a
s
et
s
f
o
r
t
r
ai
n
i
n
g
,
te
s
t
i
n
g
,
a
n
d
v
a
l
i
d
a
t
i
o
n
.
F
o
l
l
o
we
d
b
y
p
r
e
p
r
o
c
e
s
s
i
n
g
,
n
a
m
e
l
y
t
h
e
a
p
p
l
ic
a
t
i
o
n
o
f
C
L
AH
E
,
R
e
a
l
-
E
SR
G
AN
te
c
h
n
i
q
u
e
s
,
a
n
d
d
a
ta
a
u
g
m
e
n
t
a
t
i
o
n
,
b
e
f
o
r
e
b
ei
n
g
p
r
o
c
e
s
s
e
d
b
y
E
f
f
i
ci
e
n
t
N
et
-
B
7
m
o
d
e
l
.
T
h
e
r
e
a
r
e
f
i
v
e
c
l
as
s
es
t
h
at
d
e
s
c
r
i
b
e
s
e
v
e
r
i
t
y
o
f
D
R
t
h
at
w
il
l
b
e
d
e
te
c
t
e
d
b
y
t
h
e
m
o
d
e
l
,
a
s
s
e
e
n
i
n
Fi
g
u
r
e
1
.
T
h
e
r
e
s
e
a
r
c
h
w
o
r
k
f
l
o
w
c
a
n
b
e
s
e
en
i
n
F
i
g
u
r
e
2
.
Fig
u
r
e
1
.
I
m
ag
e
b
ased
o
n
clas
s
lab
el
DR
Fig
u
r
e
2
.
T
h
e
f
lo
w
o
f
r
esear
ch
wo
r
k
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
ma
g
e
en
h
a
n
ce
men
t c
o
mb
in
ed
w
ith
E
fficien
tN
et
-
B
7
fo
r
g
r
a
d
in
g
cla
s
s
ifica
tio
n
o
f d
ia
b
etic
…
(
N
in
a
S
ev
a
n
i)
2557
2
.
1
.
Da
t
a
s
et
s
T
h
is
s
tu
d
y
u
s
ed
th
e
p
u
b
lic
d
a
taset
f
r
o
m
th
e
Kag
g
le
[
2
7
]
,
A
PTOS
2
0
1
9
d
ataset,
wh
ich
is
co
n
s
is
t
o
f
im
ag
es
o
f
DR
d
is
ea
s
e
ca
teg
o
r
i
ze
d
in
to
5
class
es,
to
talin
g
3
,
6
6
2
im
ag
es.
T
h
e
d
ataset
in
clu
d
es
PNG
f
iles
an
d
a
C
SV
f
ile
th
at
as
th
e
lab
el
f
o
r
e
ac
h
im
ag
e
ac
c
o
r
d
in
g
to
its
cla
s
s
.
T
h
ese
class
e
s
ar
e
n
u
m
b
e
r
e
d
f
r
o
m
0
t
o
4
b
ased
o
n
th
e
s
ev
er
ity
o
f
th
e
DR
d
is
ea
s
e.
T
h
e
im
ag
es in
th
e
d
ataset
ca
n
b
e
v
iewe
d
in
Fig
u
r
e
1
,
an
d
th
e
d
is
tr
ib
u
tio
n
o
f
im
ag
es f
o
r
ea
ch
class
lab
el
ca
n
b
e
s
ee
n
in
T
ab
le
1
.
T
h
e
d
ataset
will
b
e
d
iv
id
ed
in
to
th
r
ee
p
air
s
:
8
0
%
f
o
r
tr
ain
in
g
d
ata,
1
0
%
f
o
r
test
in
g
,
an
d
1
0
%
f
o
r
v
alid
atio
n
.
E
ac
h
d
ata
s
eg
m
e
n
t
will
b
e
s
ep
ar
ated
u
s
in
g
th
e
tr
ain
_
test
_
s
p
lit
m
eth
o
d
f
r
o
m
t
h
e
Sk
lear
n
lib
r
ar
y
.
W
e
will
u
s
e
au
g
m
en
tatio
n
te
ch
n
iq
u
es
to
in
cr
ea
s
e
th
e
n
u
m
b
er
o
f
im
a
g
es
b
y
f
lip
p
in
g
th
em
v
er
tically
an
d
h
o
r
izo
n
tally
,
as we
ll a
s
s
ca
lin
g
th
em
u
p
to
1
1
0
%.
E
x
am
p
les o
f
th
e
au
g
m
e
n
ted
im
ag
es c
an
b
e
s
ee
n
in
Fig
u
r
e
3
.
T
h
e
au
g
m
e
n
tatio
n
will
en
s
u
r
e
th
at
th
e
n
u
m
b
er
o
f
im
ag
es
f
o
r
ea
ch
class
r
ea
ch
es
1
,
8
0
5
,
w
h
ich
is
th
e
h
ig
h
est
n
u
m
b
er
o
f
im
a
g
es in
th
e
n
o
-
D
R
class
.
a
s
th
e
r
esu
lts
o
f
th
e
d
a
ta
au
g
m
en
tatio
n
ca
n
b
e
s
ee
n
i
n
Fig
u
r
e
4
.
T
ab
le
1
.
Nu
m
b
er
o
f
im
ag
es b
a
s
ed
o
n
DR
class
lab
el
C
l
a
s
s
A
mo
u
n
t
i
m
a
g
e
No
-
D
R
(
N
o
r
m
a
l
)
1
,
8
0
5
M
i
l
d
-
DR
3
7
0
M
o
d
e
r
a
t
e
-
DR
9
9
9
S
e
v
e
r
e
-
DR
1
9
3
P
r
o
l
i
f
e
r
a
t
i
v
e
-
DR
2
9
5
Fig
u
r
e
3
.
Au
g
m
en
ted
im
ag
es i
n
d
atasets
Fig
u
r
e
4
.
C
o
m
p
a
r
is
o
n
o
f
th
e
a
m
o
u
n
t
o
f
au
g
m
en
ted
d
ata
2
.
2
.
P
re
-
pro
ce
s
s
ing
Pre
p
r
o
ce
s
s
in
g
is
d
o
n
e
to
s
im
p
lify
th
e
s
tu
d
y
im
ag
es
a
n
d
ac
h
i
ev
e
o
p
tim
al
r
esu
lts
f
o
r
APTO
S
2
0
1
9
,
as
it
h
as
d
if
f
e
r
en
t
im
a
g
e
s
izes.
A
ll
im
ag
es
will
b
e
r
esized
to
2
2
4
×2
2
4
to
m
atch
t
h
e
in
p
u
t
s
ize
o
f
E
f
f
icien
tNet
-
B
7
m
o
d
el
an
d
c
o
n
v
e
r
ted
to
g
r
ay
s
ca
le.
Ad
d
itio
n
ally
,
p
r
ep
r
o
ce
s
s
in
g
will
in
v
o
lv
e
in
cr
ea
s
in
g
t
h
e
im
ag
e
co
n
tr
ast
u
s
in
g
C
L
AHE
.
T
h
is
will
en
h
an
ce
th
e
v
is
ib
ilit
y
o
f
d
etails
in
th
e
f
u
n
d
u
s
im
ag
e
f
o
r
DR
d
ete
ctio
n
b
y
in
cr
ea
s
in
g
co
n
tr
ast an
d
r
e
d
u
cin
g
im
ag
e
n
o
is
e.
2
.
2
.
1
.
Co
ntr
a
s
t
lim
it
ed
a
da
ptiv
e
his
t
o
g
ra
m
equa
liza
t
io
n
C
L
AHE
is
u
s
ed
to
e
n
h
an
ce
d
etail
to
a
f
in
er
,
m
o
r
e
d
etailed
tex
tu
r
ed
le
v
el,
a
n
d
t
o
in
c
r
ea
s
e
lo
w
co
n
tr
ast
f
r
o
m
im
ag
e
DR
u
s
in
g
a
d
is
tr
ib
u
tio
n
to
m
ar
k
b
r
ig
h
tn
ess
o
n
th
e
in
p
u
t
im
ag
e.
T
h
e
p
r
o
ce
s
s
o
f
h
o
w
C
L
AHE
wo
r
k
s
to
im
p
r
o
v
e
th
e
q
u
ality
o
f
f
u
n
d
u
s
im
ag
es
ca
n
b
e
s
ee
n
in
Fig
u
r
e
5
.
T
h
e
f
ir
s
t
s
tag
e
i
s
th
e
t
ile
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
5
5
4
-
2
5
6
6
2558
g
en
er
atio
n
s
tag
e,
wh
e
r
e
th
e
in
p
u
t
im
ag
e
is
d
iv
i
d
ed
i
n
to
f
o
u
r
eq
u
ally
s
ized
p
a
r
ts
.
E
ac
h
p
ar
t
o
f
th
e
im
a
g
e
g
o
es
th
r
o
u
g
h
a
HE
p
r
o
ce
s
s
es
,
wh
ich
in
v
o
lv
es
s
ix
s
tep
s
:
p
r
o
ce
s
s
in
g
h
is
to
g
r
am
s
,
ca
lcu
latin
g
ex
ce
s
s
,
d
is
tr
ib
u
tin
g
ex
ce
s
s
,
r
ed
is
tr
ib
u
tin
g
ex
ce
s
s
,
cu
m
u
lativ
e
d
is
tr
ib
u
tio
n
f
u
n
cti
o
n
(
C
DF)
,
an
d
s
ca
lin
g
an
d
m
a
p
p
in
g
p
r
o
ce
s
s
es.
I
n
ea
ch
p
ar
t
o
f
th
e
im
ag
e,
th
e
h
i
s
to
g
r
am
v
alu
es
ar
e
clip
p
ed
an
d
p
r
o
p
a
g
ated
to
th
e
s
ec
tio
n
h
is
to
g
r
am
v
alu
es
o
f
o
th
er
p
ar
ts
o
f
th
e
im
ag
e.
T
h
en
,
th
e
h
is
to
g
r
am
is
ca
lcu
lated
u
s
in
g
th
e
C
DF
p
er
p
ix
el
s
ca
le
o
f
th
e
in
p
u
t
im
ag
e,
an
d
m
ap
p
in
g
is
ex
ec
u
ted
b
ased
o
n
th
e
C
DF
v
alu
es.
T
o
i
n
cr
ea
s
e
co
n
tr
ast,
b
ilin
ea
r
i
n
ter
p
o
latio
n
is
u
s
ed
to
co
m
b
in
e
ea
c
h
p
ar
t
o
f
th
e
im
ag
e
in
to
o
n
e
wh
o
le.
T
h
is
tec
h
n
iq
u
e
n
atu
r
ally
in
cr
ea
s
es
th
e
lo
ca
l
co
n
tr
ast
an
d
m
ak
es th
e
b
o
r
d
er
lin
es a
n
d
cu
r
v
es in
th
e
im
ag
e
clea
r
e
r
.
Fig
u
r
e
5
.
W
o
r
k
f
lo
w
C
L
AHE
2
.
2
.
2
.
Rea
l e
nh
a
nced
s
up
er
r
eso
lutio
n g
ener
a
t
iv
e
a
dv
er
s
a
ria
l net
wo
rk
Af
ter
ap
p
ly
i
n
g
C
L
AHE
,
th
e
n
ex
t
s
tep
is
to
u
s
e
R
ea
l
-
E
SR
GAN,
wh
ich
h
elp
s
r
em
o
v
e
n
o
is
e
an
d
en
h
a
n
ce
th
e
q
u
ality
o
f
th
e
f
u
n
d
u
s
im
ag
e,
p
ar
ticu
lar
ly
at
th
e
ed
g
es
an
d
cu
r
v
es
o
f
th
e
lin
es.
R
ea
l
-
E
S
R
GA
N
is
b
ased
o
n
th
e
ar
ch
itectu
r
e
o
f
SR
R
e
s
Net
an
d
is
an
i
m
p
r
o
v
e
m
en
t
o
f
E
SR
GAN,
in
teg
r
atin
g
m
u
ltip
le
r
esid
u
al
-
in
-
r
esid
u
al
d
e
n
s
e
b
lo
ck
(
R
R
DB
)
elem
en
ts
.
T
h
e
u
s
e
o
f
R
ea
l
-
E
SR
GA
N
r
esu
lts
in
a
m
o
r
e
n
at
u
r
al
tex
tu
r
e
f
o
r
th
e
f
u
n
d
u
s
im
ag
e
a
n
d
en
h
a
n
ce
s
th
e
v
is
ib
ilit
y
o
f
d
etails,
wh
ich
is
cr
u
cial
f
o
r
m
ed
ical
im
ag
es
[
2
8
]
.
Fo
r
m
o
r
e
in
f
o
r
m
atio
n
o
n
h
o
w
R
ea
l
-
E
SR
GAN
wo
r
k
s
,
r
ef
er
to
Fig
u
r
e
6
.
Fig
u
r
e
6
.
W
o
r
k
f
lo
w
R
ea
l
-
E
SR
GAN
2
.
3
.
E
x
perim
ent
s
et
t
ing
s
Fo
r
ass
ess
in
g
th
e
im
p
ac
t
o
f
th
e
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
e
,
th
r
ee
ex
p
er
im
en
tal
s
ce
n
a
r
io
s
will
b
e
co
n
d
u
cte
d
as f
o
llo
ws:
‒
T
h
e
f
ir
s
t scen
a
r
io
will in
v
o
l
v
e
n
o
p
r
ep
r
o
c
ess
in
g
an
d
will u
s
e
E
f
f
icien
tNet
-
B
7
.
‒
T
h
e
s
ec
o
n
d
s
ce
n
a
r
io
will in
co
r
p
o
r
ate
C
L
AHE
with
E
f
f
icien
t
Net
-
B
7
.
‒
T
h
e
th
ir
d
s
ce
n
a
r
io
will f
ea
tu
r
e
th
e
co
m
b
i
n
ed
u
s
e
o
f
R
ea
l
-
E
SR
GAN
an
d
C
L
AHE
with
E
f
f
icien
tNet
-
B
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
ma
g
e
en
h
a
n
ce
men
t c
o
mb
in
ed
w
ith
E
fficien
tN
et
-
B
7
fo
r
g
r
a
d
in
g
cla
s
s
ifica
tio
n
o
f d
ia
b
etic
…
(
N
in
a
S
ev
a
n
i)
2559
All
ex
p
er
im
en
ts
will
u
s
e
a
C
o
lab
Pro
C
PU
with
NVI
DI
A
A
1
0
0
T
e
n
s
o
r
GPU,
wh
ich
h
as
a
R
AM
o
f
8
3
.
5
GB
f
o
r
th
e
C
PU a
n
d
4
0
GB
f
o
r
t
h
e
GPU.
2
.
4
.
E
v
a
lua
t
i
o
n per
f
o
rma
nces
T
o
m
e
a
s
u
r
e
t
h
e
r
es
u
l
ts
o
f
m
o
d
e
l
e
v
a
l
u
a
t
i
o
n
,
t
h
i
s
r
es
e
a
r
c
h
w
i
l
l
u
s
e
f
o
u
r
m
e
t
r
i
cs
,
n
a
m
e
l
y
a
c
c
u
r
a
c
y
,
p
r
e
c
i
s
i
o
n
,
r
e
c
al
l
o
r
s
e
n
s
it
i
v
i
t
y
,
a
n
d
F
1
-
s
c
o
r
e
.
T
h
e
f
o
r
m
u
l
a
s
f
o
r
t
h
e
s
e
f
o
u
r
m
e
t
r
i
cs
a
r
e
w
r
i
t
te
n
a
s
i
n
(
1
)
t
o
(
4
)
[
2
9
]
.
=
+
+
+
+
(
1
)
=
+
(
2
)
=
+
(
3
)
1
−
=
2
×
×
+
(
4
)
W
h
er
e
T
P is
tr
u
e
p
o
s
itiv
e
,
T
N
is
tr
u
e
n
eg
ativ
e
,
FP
is
f
alse p
o
s
itiv
e
,
an
d
FN is
f
alse n
eg
ativ
e
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
Co
ntr
a
s
t
lim
it
ed
a
da
pti
v
e
his
t
o
g
ra
m e
qu
a
liza
t
io
n
s
eg
m
ent
a
t
i
o
n
T
h
e
C
L
AHE
im
p
lem
en
tatio
n
f
o
llo
wed
th
e
r
esizin
g
o
f
th
e
f
u
n
d
u
s
im
a
g
e
to
2
2
4
×2
2
4
an
d
co
n
v
er
s
io
n
to
a
s
in
g
le
-
ch
an
n
el
g
r
ay
s
ca
le
f
o
r
m
at.
Gr
ay
s
ca
lin
g
was
em
p
l
o
y
ed
to
r
ed
u
ce
c
o
m
p
lex
ity
a
n
d
m
in
im
ize
m
o
d
el
tr
ain
in
g
tim
e.
Fig
u
r
e
7
d
is
p
lay
s
a
s
am
p
le
o
f
th
e
g
r
ay
s
ca
led
A
PTOS
2
0
1
9
d
ataset
im
ag
e
r
es
u
lts
.
Fig
u
r
e
7
.
C
o
lo
r
ch
an
g
e
to
g
r
ay
s
ca
lin
g
T
h
e
im
p
lem
en
tatio
n
o
f
C
L
AHE
o
n
th
e
im
a
g
e
test
ed
th
e
clip
lim
it
v
alu
e.
T
h
is
clip
lim
it
co
n
tr
o
ls
h
o
w
th
e
im
ag
e
p
ix
el
in
ten
s
ity
ch
a
n
g
es
as
th
e
co
n
tr
ast
ch
an
g
es.
Af
ter
test
in
g
clip
lim
its
r
an
g
i
n
g
f
r
o
m
5
to
1
0
,
it
was
f
o
u
n
d
th
at
t
h
e
o
p
tim
al
cl
ip
lim
it
v
alu
e
is
8
,
as
i
n
d
icat
ed
in
b
o
ld
in
T
a
b
le
2
.
T
h
is
v
alu
e
is
cr
u
cial
f
o
r
p
r
o
v
id
i
n
g
th
e
m
o
s
t
s
u
itab
le
p
ix
el
in
ten
s
ity
ad
ju
s
tm
en
t
b
ased
o
n
th
e
im
ag
e
c
h
ar
ac
ter
is
tic
in
APTO
S2
0
1
9
.
T
o
v
iew
th
e
r
esu
lts
,
y
o
u
ca
n
r
ef
e
r
to
Fig
u
r
e
8
,
wh
ich
s
h
o
ws
th
e
C
L
AHE
im
p
lem
en
tatio
n
with
a
clip
lim
it
o
f
8
an
d
a
tile g
r
id
s
ize
with
d
im
en
s
io
n
s
o
f
8
×8
.
T
ab
le
2
.
C
lip
lim
it c
o
m
p
a
r
is
o
n
Ti
l
e
g
r
i
d
si
z
e
C
l
i
p
l
i
m
i
t
A
c
c
u
r
a
c
y
1
st
2
n
d
3
r
d
M
e
a
n
8
×
8
5
0
.
7
8
1
0
.
7
7
0
0
.
7
5
3
0
.
7
6
8
6
0
.
7
9
8
0
.
7
8
3
0
.
7
8
5
0
.
7
8
9
7
0
.
8
0
4
0
.
8
2
7
0
.
8
1
1
0
.
8
1
4
8
0
.
8
3
1
0
.
8
2
8
0
.
8
3
4
0
.
8
3
1
9
0
.
8
2
2
0
.
8
1
8
0
.
8
0
9
0
.
8
1
6
10
0
.
8
1
9
0
.
8
1
3
0
.
8
0
7
0
.
8
1
3
I
n
Fig
u
r
e
8
,
it
is
ev
id
en
t
th
at
t
h
e
u
s
e
o
f
C
L
AHE
en
h
a
n
ce
s
th
e
v
is
ib
ilit
y
o
f
HE
M
d
etails
in
t
h
e
f
u
n
d
u
s
im
ag
e.
Similar
ly
,
HE
d
etails
also
ap
p
ea
r
clea
r
er
d
u
e
to
in
cr
ea
s
ed
co
n
tr
ast
in
th
e
im
ag
e
.
Ad
d
itio
n
ally
,
it
is
ap
p
ar
en
t
f
r
o
m
Fig
u
r
e
8
th
at
C
L
AHE
ca
n
a
u
g
m
e
n
t
th
e
co
n
t
r
ast
in
f
u
n
d
u
s
d
etails
with
lo
w
-
in
ten
s
ity
v
ar
iatio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
5
5
4
-
2
5
6
6
2560
C
L
AHE
h
as
also
b
ee
n
em
p
lo
y
ed
in
p
r
ev
io
u
s
s
tu
d
ies
to
d
ia
g
n
o
s
e
DR
d
is
ea
s
e
u
s
in
g
f
u
n
d
u
s
im
ag
es
as
p
ar
t
o
f
p
r
ep
r
o
ce
s
s
in
g
s
in
ce
it
h
as
its
ab
ilit
ies
to
en
h
an
ce
im
ag
e
q
u
ality
[
1
1
]
,
[
2
3
]
,
[
2
4
]
.
T
h
ese
f
i
n
d
in
g
s
d
em
o
n
s
tr
ate
th
at
f
o
r
APTO
S
2
0
1
9
,
th
e
im
p
lem
en
tatio
n
o
f
C
L
AHE
s
ig
n
if
ican
tly
im
p
r
o
v
es
v
is
u
aliza
tio
n
an
d
aid
s
in
th
e
d
etec
tio
n
o
f
cr
u
cial
f
ea
tu
r
es
f
o
r
id
en
tify
in
g
DR
d
is
ea
s
e.
T
h
e
u
s
e
o
f
C
L
AHE
,
wh
ich
in
cr
ea
s
es
ac
cu
r
ac
y
,
is
co
n
s
is
ten
t w
ith
ea
r
lier
r
esear
ch
[
1
0
]
,
[
2
3
]
.
Fig
u
r
e
8
.
R
esu
lts
o
f
ap
p
l
y
in
g
C
L
AHE
3
.
2
.
Rea
l e
nh
a
nced
s
up
er
re
s
o
lutio
n g
ener
a
t
iv
e
a
dv
er
s
a
r
ia
l net
wo
rk
s
eg
m
ent
a
t
io
n
R
ea
l
-
E
SR
GAN
i
s
em
p
lo
y
ed
f
o
llo
win
g
th
e
ap
p
licatio
n
o
f
C
L
AHE
to
en
h
an
ce
th
e
p
ix
el
q
u
ality
o
f
th
e
s
lig
h
tly
d
is
to
r
ted
DR
im
ag
e,
r
esu
ltin
g
in
s
m
o
o
th
er
tex
tu
r
e
a
n
d
r
ed
u
ce
d
n
o
is
e.
T
h
e
p
r
o
ce
s
s
in
v
o
lv
es
u
p
s
ca
lin
g
th
e
2
2
4
×
2
2
4
DR
im
ag
e
to
7
8
4
×
7
8
4
(
3
.
5
×
lar
g
er
)
an
d
th
en
r
e
s
izin
g
it
b
ac
k
to
2
2
4
×
2
2
4
f
o
r
t
h
e
tr
ain
in
g
m
o
d
el.
I
m
p
lem
en
tin
g
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
in
ea
ch
DR
clas
s
is
ac
h
iev
ab
le,
as d
em
o
n
s
tr
ated
in
Fig
u
r
e
9
.
I
n
Fig
u
r
e
9
,
it
is
ap
p
ar
e
n
t
th
at
th
e
ap
p
licatio
n
o
f
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
ca
n
im
p
r
o
v
e
th
e
tex
tu
r
e
o
f
th
e
f
u
n
d
u
s
im
ag
e,
m
ak
in
g
lo
w
-
in
ten
s
ity
d
etails
s
u
ch
as
HE
M
an
d
HE
m
o
r
e
v
is
ib
le.
R
ea
l
-
E
SR
GAN
im
p
r
o
v
es
im
a
g
e
tex
tu
r
e,
allo
win
g
lo
w
-
in
ten
s
ity
f
ea
tu
r
es
t
o
b
e
m
o
r
e
v
is
ib
le.
T
h
is
is
w
h
er
e
r
ea
l
-
E
SR
GAN
is
em
p
lo
y
ed
o
n
a
wid
e
r
an
g
e
o
f
im
ag
es,
n
o
t
o
n
ly
m
e
d
ical
o
n
es
[
2
6
]
,
[
2
8
]
.
T
h
e
o
v
er
all
im
ag
e
q
u
ality
is
also
well
p
r
eser
v
ed
.
Fig
u
r
e
9
.
R
esu
lts
o
f
ap
p
l
y
in
g
C
L
AHE
an
d
r
ea
l
-
E
SR
GAN
3
.
3
.
P
er
f
o
r
m
a
nce
ev
a
lua
t
io
n
T
h
e
tr
ain
in
g
m
o
d
el
was
co
n
d
u
cted
u
s
in
g
th
r
ee
d
if
f
e
r
en
t
lear
n
in
g
r
ate
v
alu
es:
0
.
0
0
0
1
,
0
.
0
0
1
,
an
d
0
.
0
1
,
with
5
0
ep
o
ch
s
f
o
r
ea
c
h
v
alu
e.
T
h
is
v
alu
e
is
ca
lcu
lated
b
y
tak
in
g
in
to
ac
c
o
u
n
t
t
h
e
s
am
e
v
alu
es
in
p
r
ev
io
u
s
r
esear
ch
[
7
]
,
[
9
]
,
[
1
0
]
,
s
o
th
at
a
f
air
er
an
aly
s
is
co
m
p
ar
is
o
n
ca
n
b
e
m
a
d
e.
T
h
e
m
o
d
el
was
tr
ain
ed
with
th
e
Ad
am
o
p
tim
izer
at
a
lear
n
in
g
r
ate
o
f
1
×
1
0
-
4
an
d
b
atch
s
ize
o
f
3
2
.
T
o
r
ed
u
ce
o
v
er
f
it
tin
g
,
th
e
o
p
tim
izer
u
s
ed
weig
h
t
d
ec
a
y
r
e
g
u
lar
izat
io
n
with
a
d
ec
ay
v
alu
e
o
f
2
×
1
0
-
6
,
ca
lcu
lated
as
th
e
r
atio
o
f
t
h
e
lear
n
in
g
r
ate
t
o
th
e
to
tal
n
u
m
b
er
o
f
tr
ain
in
g
e
p
o
ch
s
(
5
0
)
.
T
o
av
o
i
d
o
v
er
tr
ai
n
in
g
,
a
n
ea
r
ly
h
altin
g
tech
n
iq
u
e
was
u
s
ed
,
wh
ich
in
v
o
lv
ed
m
o
n
ito
r
in
g
th
e
v
alid
atio
n
lo
s
s
o
v
e
r
a
1
0
-
ep
o
c
h
p
e
r
io
d
.
W
h
en
n
o
im
p
r
o
v
em
e
n
t
in
v
alid
atio
n
lo
s
s
was
s
ee
n
with
in
th
is
tim
e
f
r
am
e,
th
e
tr
ain
in
g
p
r
o
ce
s
s
was
s
to
p
p
ed
,
an
d
th
e
m
o
d
el
weig
h
ts
co
r
r
esp
o
n
d
i
n
g
to
th
e
lo
west v
alid
atio
n
lo
s
s
wer
e
r
esto
r
ed
.
T
h
e
test
r
esu
lts
f
o
r
th
e
th
r
ee
s
ce
n
ar
io
s
ar
e
p
r
esen
ted
in
T
ab
l
e
s
3
to
5
.
T
h
e
b
est
ac
cu
r
ac
y
,
h
ig
h
lig
h
ted
in
b
o
l
d
,
was
ac
h
iev
ed
with
a
lear
n
i
n
g
r
ate
o
f
0
.
0
0
1
.
Ad
d
itio
n
ally
,
it
was
o
b
s
er
v
e
d
th
at
ac
cu
r
ac
y
in
cr
ea
s
ed
with
th
e
im
p
lem
e
n
tatio
n
o
f
p
r
ep
r
o
ce
s
s
in
g
tec
h
n
iq
u
es.
Sp
ec
if
ically
,
ac
cu
r
ac
y
im
p
r
o
v
e
d
f
r
o
m
7
6
.
6
%
with
o
u
t
p
r
ep
r
o
ce
s
s
in
g
to
8
3
.
4
%
with
C
L
AHE
,
an
d
f
u
r
th
er
t
o
9
3
.
7
%
with
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN.
Su
b
s
eq
u
en
tly
,
th
e
ch
o
s
en
lea
r
n
in
g
r
ate
was
u
s
ed
in
f
u
r
th
er
ex
p
er
im
en
ts
to
ev
alu
ate
m
etr
ic
v
al
u
es.
T
h
e
p
r
ec
is
io
n
,
r
ec
all
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
ma
g
e
en
h
a
n
ce
men
t c
o
mb
in
ed
w
ith
E
fficien
tN
et
-
B
7
fo
r
g
r
a
d
in
g
cla
s
s
ifica
tio
n
o
f d
ia
b
etic
…
(
N
in
a
S
ev
a
n
i)
2561
(
s
en
s
itiv
ity
)
,
an
d
F1
-
s
co
r
e
r
esu
lts
,
ac
h
iev
ed
with
a
lear
n
i
n
g
r
ate
o
f
0
.
0
0
1
f
o
r
ea
c
h
ex
p
er
im
en
tal
s
ce
n
ar
io
u
s
in
g
E
f
f
icien
tNet
-
B
7
,
ar
e
d
is
p
lay
ed
in
Fig
u
r
es 1
0
t
o
1
2
.
T
ab
le
3
.
Acc
u
r
ac
y
r
esu
lts
with
o
u
t CLAH
E
an
d
R
ea
l
-
E
SR
GA
N
o
n
E
f
f
icien
tNet
-
B7
Le
a
r
n
i
n
g
r
a
t
e
A
c
c
u
r
a
c
y
M
e
a
n
Ti
me
e
x
e
c
u
t
i
o
n
(
s)
0
.
0
0
0
0
1
0
.
6
6
5
0
.
7
0
6
4
,
4
5
3
.
5
6
0
.
0
0
0
1
0
.
6
8
7
4
,
4
8
8
.
1
6
0
.
0
0
1
0
.
7
6
6
4
,
4
7
3
.
5
6
T
ab
le
4
.
Acc
u
r
ac
y
r
esu
lts
with
C
L
AHE
o
n
E
f
f
icien
tNet
-
B7
Le
a
r
n
i
n
g
r
a
t
e
A
c
c
u
r
a
c
y
M
e
a
n
Ti
me
e
x
e
c
u
t
i
o
n
(
s)
0
.
0
0
0
0
1
0
.
8
1
3
0
.
8
1
8
4
,
4
5
1
.
8
9
0
.
0
0
0
1
0
.
8
0
8
4
,
4
6
8
.
2
3
0
.
0
0
1
0
.
8
3
4
4
,
4
6
9
.
0
1
T
ab
le
5
.
Acc
u
r
ac
y
r
esu
lts
with
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
o
n
E
f
f
icien
tNet
-
B7
Le
a
r
n
i
n
g
r
a
t
e
A
c
c
u
r
a
c
y
M
e
a
n
Ti
me
e
x
e
c
u
t
i
o
n
(
s)
0
.
0
0
0
0
1
0
.
9
2
6
0
.
9
2
8
4
,
0
8
2
.
6
5
0
.
0
0
0
1
0
.
9
2
1
4
,
0
7
9
.
7
4
0
.
0
0
1
0
.
9
3
7
4
,
0
8
6
.
5
5
Fig
u
r
e
1
0
.
C
o
m
p
ar
is
o
n
o
f
p
r
e
cisi
o
n
o
b
tain
ed
b
y
E
f
f
icien
tNet
-
B
7
u
n
d
er
t
h
r
ee
ex
p
er
im
en
tal
s
ettin
g
s
Fig
u
r
e
1
1
.
C
o
m
p
ar
is
o
n
o
f
r
ec
a
ll o
b
tain
ed
b
y
E
f
f
icie
n
tNet
-
B
7
u
n
d
er
th
r
ee
e
x
p
er
im
en
tal
s
ettin
g
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
5
5
4
-
2
5
6
6
2562
Fig
u
r
e
1
2
.
C
o
m
p
ar
is
o
n
o
f
F1
-
s
co
r
es o
b
tain
ed
b
y
E
f
f
icien
tNet
-
B
7
u
n
d
er
t
h
r
ee
ex
p
er
im
en
ta
l settin
g
s
B
ased
o
n
th
e
ex
p
er
im
en
tal
r
esu
lts
u
s
in
g
th
e
E
f
f
icien
tNet
-
B
7
m
o
d
el,
it'
s
ev
id
en
t
th
at
ap
p
ly
in
g
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
p
r
o
d
u
ce
s
th
e
b
est
r
esu
lts
.
T
h
e
s
e
r
esu
lts
ar
e
in
lin
e
with
t
h
e
r
esu
lts
o
f
p
r
e
v
io
u
s
r
esear
ch
[
8
]
.
Alth
o
u
g
h
th
e
p
r
ev
io
u
s
r
esear
ch
s
till
u
s
ed
E
f
f
icien
tNet
-
B
4
an
d
o
n
ly
ca
lcu
l
ated
ac
cu
r
ac
y
.
T
h
e
ac
cu
r
ac
y
v
alu
e
s
ig
n
if
ica
n
tly
in
cr
ea
s
ed
f
r
o
m
0
.
7
0
6
to
0
.
9
2
8
af
ter
ap
p
l
y
in
g
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN.
Similar
ly
,
in
th
ir
d
ex
p
er
im
en
t,
wh
ich
ap
p
lied
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN,
th
e
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
v
alu
es
wer
e
0
.
9
8
3
,
0
.
9
8
9
,
an
d
0
.
9
6
7
,
r
esp
ec
tiv
ely
,
d
em
o
n
s
tr
atin
g
a
n
o
tab
le
im
p
r
o
v
em
e
n
t
co
m
p
ar
ed
to
th
e
f
ir
s
t
ex
p
er
im
en
t
th
at
d
id
n
o
t
u
tili
ze
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
e
s
an
d
ac
h
iev
ed
v
alu
es
o
f
0
.
6
6
2
,
0
.
4
9
2
,
an
d
0
.
5
6
5
f
o
r
p
r
ec
is
io
n
,
r
ec
all,
a
n
d
F1
-
s
co
r
e.
T
h
is
in
d
icate
s
a
5
0
%
in
cr
ea
s
e
in
th
e
r
ec
all
v
alu
e
af
te
r
ap
p
ly
in
g
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN.
T
h
e
s
ig
n
if
ican
t
im
p
r
o
v
em
e
n
t
in
v
alu
es
d
em
o
n
s
tr
ates
th
e
p
o
s
itiv
e
im
p
ac
t
o
f
im
p
lem
en
tin
g
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
o
n
f
u
n
d
u
s
im
ag
es
f
r
o
m
th
e
APTO
S
2
0
1
9
d
at
aset,
r
ef
lectin
g
th
e
p
o
ten
tial to
cr
ea
te
a
m
o
d
el
wit
h
a
g
o
o
d
lev
el
o
f
s
en
s
itiv
ity
.
Fu
r
th
er
m
o
r
e
,
th
e
ap
p
licatio
n
o
f
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
y
ield
ed
th
e
b
est
r
esu
lts
in
th
e
s
ev
er
e
-
DR
class
,
wi
th
a
r
ec
all
v
alu
e
o
f
0
.
9
8
9
,
wh
ile
th
e
lo
we
s
t
r
ec
all
v
alu
e
was
o
b
s
er
v
ed
in
th
e
m
o
d
e
r
ate
-
DR
class
,
at
0
.
8
7
5
.
I
n
th
e
s
ev
e
r
e
-
DR
class
,
th
e
em
p
lo
y
m
e
n
t
o
f
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
s
u
cc
ess
f
u
lly
s
m
o
o
th
ed
th
e
tex
tu
r
e
o
f
th
e
f
u
n
d
u
s
im
ag
e,
r
ev
ea
lin
g
s
y
m
p
to
m
s
o
f
d
is
r
u
p
tio
n
m
o
r
e
clea
r
ly
[
2
6
]
.
Ho
wev
e
r
,
in
th
e
m
o
d
er
ate
-
DR
class
,
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
alo
n
e
wer
e
in
s
u
f
f
icien
t
t
o
id
en
tify
p
atter
n
s
o
f
d
is
tu
r
b
an
ce
clea
r
ly
.
Fig
u
r
es
8
an
d
9
s
h
o
w
th
at
th
e
m
o
d
er
ate
-
DR
class
h
ad
ab
o
u
t
th
e
s
am
e
n
u
m
b
er
o
f
in
d
icato
r
s
o
f
d
is
tu
r
b
an
ce
in
ter
m
s
o
f
p
lace
,
s
h
ap
e,
an
d
co
lo
r
as th
e
m
ild
-
DR
cla
s
s
h
ad
.
T
h
is
ca
u
s
ed
th
e
s
y
s
tem
to
f
r
eq
u
e
n
tly
m
ak
e
wr
o
n
g
p
r
ed
ictio
n
s
,
n
o
t
p
r
e
d
ictin
g
th
e
m
o
d
er
ate
-
DR
class
.
T
h
is
d
em
o
n
s
tr
ates
wh
y
th
e
m
o
d
er
ate
-
DR
class
h
as
a
lo
w
er
r
ec
all
v
alu
e
th
an
th
e
o
th
er
class
es.
T
h
e
m
o
d
er
ate
-
DR
class
also
s
h
o
wed
th
e
lo
west
r
ec
all
v
alu
e
o
f
0
.
4
9
2
wh
en
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
wer
e
n
o
t
ap
p
lied
.
T
h
is
s
u
g
g
ests
th
at
th
e
m
o
d
er
ate
-
DR
class
is
th
e
lea
s
t
in
f
lu
en
ce
d
b
y
th
e
im
p
lem
en
tatio
n
o
f
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN.
V
is
u
al
in
s
p
ec
tio
n
o
f
f
u
n
d
u
s
im
ag
es
in
th
e
m
o
d
er
ate
-
DR
class
in
d
icate
d
n
o
o
b
v
io
u
s
s
ig
n
s
o
f
DR
,
an
d
en
h
an
cin
g
th
e
im
ag
e
th
r
o
u
g
h
co
n
tr
ast
ad
ju
s
tm
en
t
an
d
n
o
is
e
r
ed
u
cti
o
n
d
id
n
o
t
s
ig
n
if
ican
tly
alter
th
e
in
itial
im
ag
e.
Deta
ils
r
elate
d
to
DR
,
s
u
ch
as
HE
M
an
d
HE
,
d
id
n
o
t
b
ec
o
m
e
clea
r
er
.
Mo
r
eo
v
e
r
,
with
o
u
t
ap
p
ly
in
g
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN,
th
e
No
-
D
R
clas
s
ex
h
ib
ited
th
e
h
ig
h
est r
ec
all
r
esu
lts
,
with
a
v
alu
e
o
f
0
.
9
3
9
,
s
lig
h
tly
b
etter
th
an
wh
en
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
wer
e
ap
p
lied
,
wh
ich
ac
h
iev
ed
a
r
ec
all
v
alu
e
o
f
0
.
9
1
4
.
I
n
a
d
d
itio
n
to
t
h
e
v
is
u
al
d
if
f
e
r
en
ce
s
o
b
s
er
v
ed
in
t
h
e
no
-
D
R
class
,
th
e
ap
p
licatio
n
o
f
a
u
g
m
en
tatio
n
tech
n
iq
u
es
also
in
f
lu
en
ce
d
th
e
r
ec
all
r
esu
lts
,
alth
o
u
g
h
it
h
ad
m
in
im
al
im
p
ac
t
o
n
ac
cu
r
a
cy
,
p
r
ec
is
io
n
,
an
d
F1
-
s
co
r
e
m
etr
ics.
T
h
ese
r
esu
lts
in
d
icate
th
at
th
e
m
o
d
el
r
eq
u
ir
es
a
s
ig
n
if
ican
t
n
u
m
b
er
o
f
im
ag
es
to
p
r
o
d
u
ce
a
s
en
s
itiv
e
m
o
d
el.
N
o
tab
ly
,
th
e
no
-
DR
class
,
wh
ich
d
id
n
o
t
u
n
d
er
g
o
a
u
g
m
e
n
tatio
n
f
r
o
m
t
h
e
b
e
g
in
n
in
g
,
s
till
y
ield
ed
s
lig
h
tly
b
etter
r
ec
all
r
esu
lts
th
an
th
e
class
th
at
u
n
d
er
wen
t
au
g
m
en
tatio
n
,
b
o
th
o
f
wh
ich
ac
h
iev
e
d
v
alu
es
ab
o
v
e
0
.
9
.
Fu
r
th
er
m
o
r
e,
th
e
im
p
o
r
tan
ce
o
f
au
g
m
en
t
atio
n
is
ev
id
en
t
th
r
o
u
g
h
th
e
h
ig
h
p
r
ec
is
io
n
an
d
F1
-
s
co
r
e
r
esu
lts
in
th
e
m
o
d
er
ate
-
DR
an
d
s
ev
er
e
-
DR
class
es,
a
s
th
ese
c
lass
e
s
ex
p
er
ien
ce
d
th
e
m
o
s
t
au
g
m
en
tatio
n
to
in
co
r
p
o
r
ate
im
ag
es u
s
ed
in
th
e
m
o
d
el
tr
ain
i
n
g
p
r
o
ce
s
s
.
T
h
e
ex
p
er
im
en
ts
d
em
o
n
s
tr
ate
th
at
im
p
lem
en
tin
g
C
L
AHE
a
n
d
R
ea
l
-
E
SR
GAN
ca
n
en
h
a
n
ce
p
r
ec
is
io
n
b
y
m
o
r
e
th
a
n
4
%
an
d
in
cr
e
ase
r
ec
all
b
y
m
o
r
e
th
an
3
%.
B
alan
cin
g
r
ec
all
p
r
ec
is
io
n
ca
n
o
p
tim
ize
th
e
p
er
f
o
r
m
an
ce
o
f
m
ed
ical
d
iag
n
o
s
tic
test
s
.
E
n
h
an
ce
d
p
r
ec
is
io
n
an
d
r
ec
all
in
a
m
ed
ical
co
n
tex
t
in
d
icate
h
o
w
ef
f
ec
tiv
ely
a
m
o
d
el
id
en
tifie
s
tr
u
e
p
o
s
itiv
e
ca
s
es
am
o
n
g
all
th
e
p
o
s
itiv
e
ca
s
es
it
id
e
n
tifie
s
.
Hig
h
p
r
ec
is
io
n
is
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
I
ma
g
e
en
h
a
n
ce
men
t c
o
mb
in
ed
w
ith
E
fficien
tN
et
-
B
7
fo
r
g
r
a
d
in
g
cla
s
s
ifica
tio
n
o
f d
ia
b
etic
…
(
N
in
a
S
ev
a
n
i)
2563
cr
u
cial
to
m
i
n
im
ize
u
n
n
ec
ess
ar
y
tr
ea
tm
en
ts
,
r
ed
u
ce
p
atien
t
a
n
x
iety
,
u
tili
ze
m
ed
ical
r
eso
u
r
c
es
ef
f
icien
tly
,
an
d
en
ab
le
m
o
r
e
in
f
o
r
m
ed
clin
ical
d
ec
is
io
n
-
m
ak
in
g
.
3.
4
.
B
a
s
eline
co
m
pa
riso
n
I
n
ce
p
tio
n
-
V3
will
b
e
c
h
o
s
en
as
th
e
co
m
p
ar
is
o
n
m
o
d
e
l
in
th
is
r
esear
ch
.
E
x
p
er
im
en
ts
with
I
n
ce
p
tio
n
-
V3
will
b
e
ca
r
r
ied
o
u
t
f
o
llo
win
g
th
e
h
y
p
er
p
ar
am
eter
s
ettin
g
s
th
at
p
r
o
v
id
e
th
e
m
o
s
t
o
p
tim
al
r
esu
lts
wh
en
u
s
in
g
E
f
f
icien
tNet
-
B
7
,
n
am
ely
lear
n
in
g
r
ates
0
.
0
0
1
a
n
d
ep
o
c
h
s
5
0
.
Fo
r
C
L
AHE
im
p
lem
en
tatio
n
,
u
s
e
clip
lim
it
8
an
d
tile
g
r
id
s
ize
with
s
ize
d
im
en
s
io
n
s
8
×8
.
T
h
er
e
ar
e
th
r
ee
ad
d
itio
n
al
ex
p
er
im
en
tal
s
ce
n
ar
io
s
ca
r
r
ied
o
u
t
f
o
llo
win
g
th
e
t
h
r
e
e
m
ain
s
ce
n
ar
io
s
.
T
h
e
o
n
ly
d
i
f
f
er
en
ce
is
in
th
e
d
ee
p
lear
n
in
g
ar
ch
itectu
r
e
u
s
ed
.
T
h
ese
ad
d
itio
n
al
ex
p
er
im
en
ts
ar
e:
‒
T
h
e
f
o
u
r
th
s
ce
n
ar
io
ex
p
e
r
im
e
n
t is with
o
u
t p
r
ep
r
o
ce
s
s
in
g
with
I
n
ce
p
tio
n
-
V3
,
‒
T
h
e
s
ec
o
n
d
s
ce
n
a
r
io
ex
p
er
im
e
n
t u
s
es C
L
AHE
with
I
n
ce
p
tio
n
-
V3
,
‒
T
h
e
th
ir
d
s
ce
n
a
r
io
ex
p
er
im
en
t
u
s
es R
ea
l
-
E
S
R
GAN
an
d
C
L
A
HE
to
g
eth
er
with
I
n
ce
p
tio
n
-
V
3
.
A
co
m
p
a
r
is
o
n
o
f
c
lass
i
f
i
ca
t
io
n
r
e
p
o
r
t r
esu
lts
o
f
E
f
f
i
cie
n
t
Net
-
B
7
a
n
d
I
n
c
ep
ti
o
n
-
V
3
ca
n
b
e
s
e
en
i
n
T
a
b
l
es
6
to
8
.
Me
an
wh
ile,
Fig
u
r
e
1
3
d
is
p
lay
s
a
co
m
p
ar
is
o
n
b
etwe
en
th
e
a
cc
u
r
ac
y
an
d
tim
e
p
r
o
d
u
ce
d
b
y
th
e
s
ix
s
ce
n
ar
io
s
ca
r
r
ied
o
u
t in
th
is
r
esear
ch
.
T
ab
le
6
.
C
o
m
p
a
r
is
o
n
p
er
f
o
r
m
an
ce
o
f
E
f
f
icien
tNet
-
B
7
an
d
I
n
ce
p
tio
n
-
V3
with
o
u
t CLAH
E
an
d
R
ea
l
-
E
SR
GAN
Ef
f
i
c
i
e
n
t
N
e
t
-
B7
I
n
c
e
p
t
i
o
n
-
V3
P
r
e
c
i
s
i
o
n
0
.
7
6
3
0
.
7
1
4
R
e
c
a
l
l
0
.
7
6
6
0
.
7
1
6
F1
-
sc
o
r
e
0
.
7
6
0
0
.
7
1
3
A
c
c
u
r
a
c
y
0
.
7
6
6
0
.
7
1
4
T
ab
le
7
.
C
o
m
p
a
r
is
o
n
p
er
f
o
r
m
an
ce
o
f
E
f
f
icien
tNet
-
B
7
an
d
I
n
ce
p
tio
n
-
V3
u
s
in
g
C
L
AHE
Ef
f
i
c
i
e
n
t
N
e
t
-
B7
I
n
c
e
p
t
i
o
n
-
V3
P
r
e
c
i
s
i
o
n
0
.
8
3
1
0
.
7
8
6
R
e
c
a
l
l
0
.
8
3
4
0
.
7
8
7
F1
-
sc
o
r
e
0
.
8
3
2
0
.
7
8
6
A
c
c
u
r
a
c
y
0
.
8
3
4
0
.
7
8
7
T
ab
le
8
.
C
o
m
p
a
r
is
o
n
p
er
f
o
r
m
an
ce
o
f
E
f
f
icien
tNet
-
B
7
an
d
I
n
ce
p
tio
n
-
V3
u
s
in
g
C
L
AHE
an
d
R
ea
l
-
E
SR
GAN
Ef
f
i
c
i
e
n
t
N
e
t
-
B7
I
n
c
e
p
t
i
o
n
-
V3
P
r
e
c
i
s
i
o
n
0
.
9
3
7
0
.
8
6
1
R
e
c
a
l
l
0
.
9
3
7
0
.
8
5
2
F1
-
sc
o
r
e
0
.
9
3
7
0
.
8
5
2
A
c
c
u
r
a
c
y
0
.
9
3
7
0
.
8
5
2
Fig
u
r
e
1
3
.
C
o
m
p
ar
is
o
n
o
f
tim
e
an
d
ac
cu
r
ac
y
o
f
E
f
f
icie
n
tNet
-
B
7
an
d
I
n
ce
p
tio
n
-
V3
Evaluation Warning : The document was created with Spire.PDF for Python.