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
.
2709
~
2
7
2
2
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
7
0
9
-
2
7
2
2
2709
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Visio
nEy
eNet
:
a c
usto
mized de
ep le
a
rning
f
ra
mewo
r
k f
o
r early
dia
g
no
sis
of kerat
itis and uv
ei
tis
So
m
a
s
hek
ha
r
B
a
nn
ur
M
a
y
ig
o
wda
1
,
Ra
g
ha
v
endra
K
o
da
nd
a
ra
m
a
2
,
Su
dh
a
m
a
ni M
a
lla
ia
h
3
,
M
a
njuna
t
h Na
g
a
nn
a
1
,
J
a
m
u
na
4
,
K
ira
n K
um
a
r
B
.
S.
3
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
V
i
d
y
a
V
i
k
a
s I
n
st
i
t
u
t
e
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
M
y
s
u
r
u
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
-
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
M
a
h
a
r
a
j
a
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
M
y
su
r
u
,
I
n
d
i
a
3
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
A
p
p
l
i
c
a
t
i
o
n
s,
V
i
d
y
a
V
i
k
a
s I
n
st
i
t
u
t
e
o
f
En
g
i
n
e
e
r
i
n
g
a
n
d
T
e
c
h
n
o
l
o
g
y
,
M
y
su
r
u
,
I
n
d
i
a
4
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
V
i
d
y
a
V
i
k
a
s
I
n
st
i
t
u
t
e
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
M
y
s
u
r
u
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Sep
1
7
,
2
0
2
5
R
ev
is
ed
Ma
r
2
,
2
0
2
6
Acc
ep
ted
Ap
r
2
2
,
2
0
2
6
Ke
ra
ti
ti
s
a
n
d
u
v
e
it
is
a
re
in
c
re
a
sin
g
ly
p
re
v
a
len
t
o
c
u
lar
d
iso
rd
e
rs,
o
f
ten
li
n
k
e
d
to
d
e
lay
e
d
d
e
tec
ti
o
n
a
n
d
l
imit
e
d
sp
e
c
ialist
a
c
c
e
ss
,
p
a
rti
c
u
larly
in
ru
ra
l
h
e
a
lt
h
c
a
re
se
tt
in
g
s.
T
h
e
se
d
ise
a
se
s
c
a
n
lea
d
to
se
v
e
re
v
isu
a
l
im
p
a
irme
n
t
o
r
irrev
e
rsib
le
b
li
n
d
n
e
ss
if
n
o
t
i
d
e
n
ti
fied
a
t
a
n
e
a
rly
sta
g
e
.
Trad
it
io
n
a
l
d
iag
n
o
stic
a
p
p
ro
a
c
h
e
s
a
re
m
a
n
u
a
l,
ti
m
e
-
c
o
n
su
m
i
n
g
,
a
n
d
p
ro
n
e
to
h
u
m
a
n
e
rro
r,
m
a
k
i
n
g
t
h
e
m
c
h
a
ll
e
n
g
in
g
fo
r
lar
g
e
-
sc
a
le
sc
re
e
n
in
g
.
To
a
d
d
re
ss
th
e
se
li
m
it
a
ti
o
n
s,
th
is
st
u
d
y
p
re
se
n
ts
Visio
n
E
y
e
Ne
t,
a
fra
m
e
wo
rk
fo
r
a
u
t
o
m
a
ti
c
c
las
sifica
ti
o
n
o
f
k
e
ra
ti
ti
s
a
n
d
u
v
e
it
is.
Visi
o
n
E
y
e
Ne
t
i
n
teg
ra
tes
M
o
b
il
e
Ne
tV2
a
n
d
De
n
se
Ne
t1
2
1
with
in
a
fu
sio
n
a
rc
h
it
e
c
tu
re
,
a
lo
n
g
wit
h
ima
g
e
e
n
h
a
n
c
e
m
e
n
t
m
e
th
o
d
s
su
c
h
a
s
a
d
a
p
ti
v
e
g
a
m
m
a
c
o
rre
c
ti
o
n
a
n
d
sp
e
c
u
lar
re
flec
ti
o
n
su
p
p
re
ss
io
n
.
T
h
e
m
o
d
e
l
wa
s
train
e
d
a
n
d
e
v
a
lu
a
ted
o
n
a
c
u
ra
ted
d
a
tas
e
t
o
f
1
,
8
6
0
slit
-
lam
p
ima
g
e
s
(9
6
0
u
v
e
it
is
a
n
d
9
0
0
k
e
ra
ti
t
i
s)
u
sin
g
a
p
a
ti
e
n
t
-
wise
sp
li
t
(7
1
.
5
%
trai
n
in
g
,
8
.
4
%
v
a
li
d
a
ti
o
n
,
a
n
d
2
0
%
tes
ti
n
g
).
On
th
e
in
d
e
p
e
n
d
e
n
t
tes
t
se
t,
it
a
c
h
ie
v
e
d
9
8
.
0
%
a
c
c
u
ra
c
y
(
9
5
%
CI:
9
7
.
1
–
9
8
.
8
%
)
wi
th
b
a
lan
c
e
d
p
e
rfo
rm
a
n
c
e
a
c
ro
ss
c
las
se
s.
In
fe
re
n
c
e
a
n
a
ly
sis
sh
o
we
d
a
n
a
v
e
ra
g
e
p
ro
c
e
ss
in
g
ti
m
e
o
f
5
1
±2
m
s
p
e
r
ima
g
e
,
su
p
p
o
rti
n
g
re
a
l
-
ti
m
e
u
se
.
Th
e
se
re
su
lt
s
in
d
ica
te
th
a
t
Visi
o
n
E
y
e
Ne
t
h
a
s
stro
n
g
p
o
ten
ti
a
l
a
s
a
c
li
n
ica
ll
y
u
se
fu
l
d
e
c
isio
n
-
su
p
p
o
rt
t
o
o
l,
p
a
rti
c
u
larly
in
re
so
u
rc
e
-
li
m
it
e
d
se
tt
in
g
s
.
K
ey
w
o
r
d
s
:
C
o
m
p
u
ter
-
aid
ed
-
d
iag
n
o
s
is
k
er
atitis
Dee
p
lear
n
in
g
Go
o
d
h
ea
lth
Ocu
lar
s
u
r
f
ac
e
d
is
o
r
d
er
s
Uv
eitis
Vis
io
n
E
y
eNe
t
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
:
Kir
an
Ku
m
ar
B
.
S
.
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
A
p
p
licatio
n
s
,
Vid
y
a
Vik
as I
n
s
titu
te
o
f
E
n
g
in
ee
r
in
g
an
d
T
ec
h
n
o
l
o
g
y
My
s
u
r
u
,
I
n
d
ia
E
m
ail: k
ir
an
k
u
m
ar
k
r
n
2
0
0
0
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Ker
atitis
an
d
u
v
eitis
ar
e
two
clin
ically
s
ig
n
if
ican
t
an
ter
io
r
-
s
eg
m
en
t
d
is
o
r
d
e
r
s
th
at
m
ay
r
esu
lt
in
p
er
m
an
en
t
v
is
u
al
im
p
air
m
en
t
wh
en
ea
r
ly
th
er
a
p
eu
tic
in
ter
v
en
tio
n
is
d
elay
ed
.
Star
tin
g
o
f
f
,
ty
p
ical
d
iag
n
o
s
is
ap
p
r
o
ac
h
es tak
e
to
o
m
u
ch
tim
e,
r
ely
h
ea
v
ily
o
n
h
u
m
a
n
ef
f
o
r
t,
an
d
d
ep
en
d
s
tr
o
n
g
ly
o
n
a
d
o
cto
r
’
s
s
k
ill,
lead
in
g
f
r
eq
u
e
n
tly
to
late
tr
ea
tm
en
ts
a
n
d
u
n
ev
en
ass
ess
m
en
ts
.
B
ec
a
u
s
e
o
f
th
ese
is
s
u
es,
alo
n
g
wit
h
to
o
f
ew
s
p
ec
ialis
ts
ac
ce
s
s
ib
le
ac
r
o
s
s
v
ar
io
u
s
a
r
e
as,
th
er
e
a
r
is
es
a
clea
r
d
em
a
n
d
f
o
r
f
aster
,
s
elf
-
o
p
er
atin
g
t
o
o
ls
th
at
ass
is
t
in
d
iag
n
o
s
is
.
L
ately
,
p
r
o
g
r
ess
in
ar
tific
ial
in
tellig
en
ce
an
d
r
ec
en
t
ad
v
an
ce
d
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
h
av
e
s
ig
n
if
ican
tly
im
p
r
o
v
ed
a
u
to
m
ated
a
n
aly
s
is
o
f
s
lin
t
lam
p
an
d
c
o
r
n
ea
l
im
a
g
in
a
r
y
,
p
ar
ticu
la
r
ly
f
o
r
lesi
o
n
lo
ca
lizatio
n
an
d
p
ath
o
lo
g
y
d
is
cr
im
in
atio
n
,
m
ak
in
g
it
p
o
s
s
ib
le
to
s
o
r
t
d
is
ea
s
es
q
u
ick
ly
wh
ile
m
ain
tain
in
g
co
n
s
is
ten
cy
[
1
]
,
[
2
]
.
Ho
wev
er
,
m
o
s
t
ex
is
tin
g
m
o
d
els
f
o
cu
s
o
n
s
in
g
le
o
cu
lar
p
ath
o
lo
g
ies
o
r
r
ely
o
n
c
o
m
p
u
tatio
n
ally
h
ea
v
y
ar
ch
itectu
r
es
t
h
at
lack
g
en
er
aliza
tio
n
u
n
d
er
r
ea
l
-
w
o
r
ld
co
n
d
itio
n
s
[
3
]
.
Ov
er
d
ep
en
d
en
ce
o
n
id
ea
l
s
lit
-
lam
p
im
ag
es,
class
im
b
alan
c
e,
an
d
h
ig
h
in
f
er
en
ce
tim
e
f
u
r
th
er
lim
it
clin
ical
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
7
0
9
-
2
7
2
2
2710
ap
p
licab
ilit
y
,
leav
in
g
a
r
esear
c
h
g
ap
f
o
r
lig
h
tweig
h
t,
ac
cu
r
ate
,
an
d
r
o
b
u
s
t
d
iag
n
o
s
tic
m
o
d
els.
T
o
a
d
d
r
ess
th
ese
lim
itatio
n
s
,
we
p
r
o
p
o
s
e
Vis
io
n
E
y
eNe
t,
a
h
y
b
r
id
d
ee
p
lear
n
in
g
f
r
am
ewo
r
k
th
at
f
u
s
es
Mo
b
ileNetV2
[
4
]
an
d
Den
s
eNe
t1
2
1
[
5
]
f
o
r
d
u
al
-
c
lass
k
er
atitis
an
d
u
v
eitis
d
e
tectio
n
.
T
h
e
m
o
d
el
em
p
lo
y
s
ad
ap
tiv
e
g
am
m
a
co
r
r
ec
tio
n
a
n
d
s
p
ec
u
la
r
r
ef
lect
io
n
s
u
p
p
r
ess
io
n
[
6
]
to
en
h
a
n
ce
im
ag
e
q
u
ality
a
n
d
im
p
r
o
v
e
f
e
atu
r
e
ex
tr
ac
tio
n
.
T
h
e
m
ajo
r
co
n
t
r
ib
u
tio
n
s
o
f
t
h
is
s
tu
d
y
in
clu
d
e:
a
n
o
v
el
Mo
b
ileNet
–
Den
s
eNe
t
f
u
s
io
n
ac
h
iev
in
g
h
ig
h
ac
cu
r
ac
y
with
ef
f
icien
t
co
m
p
u
tatio
n
;
a
r
o
b
u
s
t
p
r
e
p
r
o
ce
s
s
in
g
p
ip
elin
e
en
s
u
r
i
n
g
co
n
s
is
ten
t
illu
m
in
atio
n
an
d
ar
tifa
ct
-
f
r
ee
in
p
u
ts
;
c
o
m
p
r
e
h
en
s
iv
e
p
er
f
o
r
m
a
n
ce
ev
alu
atio
n
with
co
n
f
id
en
ce
in
ter
v
als
an
d
in
f
er
en
ce
-
tim
e
an
aly
s
is
;
an
d
g
r
ad
ien
t
-
weig
h
t
ed
class
ac
tiv
atio
n
m
ap
p
i
n
g
(
Gr
ad
-
C
AM
)
–
b
ased
in
ter
p
r
eta
b
ilit
y
[
7
]
en
h
an
cin
g
clin
ical
tr
u
s
t
an
d
tr
a
n
s
p
ar
en
c
y
.
T
o
t
h
e
b
est
o
f
o
u
r
k
n
o
wle
d
g
e,
th
is
is
th
e
f
ir
s
t
h
y
b
r
i
d
f
u
s
io
n
f
r
am
ewo
r
k
f
o
r
r
ea
l
-
tim
e
k
er
atitis
an
d
u
v
eitis
class
if
icatio
n
o
p
tim
ized
f
o
r
cli
n
ical
d
ep
lo
y
m
en
t.
2.
RE
L
AT
E
D
WO
RK
Dee
p
lear
n
in
g
(
DL
)
h
as
b
ee
n
wid
ely
ap
p
lied
to
co
r
n
ea
l
d
i
s
ea
s
e
d
iag
n
o
s
is
.
T
h
e
au
th
o
r
in
[
1
]
,
[
8
]
d
em
o
n
s
tr
ated
C
NN
-
b
ased
d
et
ec
tio
n
o
f
f
u
n
g
al
a
n
d
b
ac
ter
ia
l
k
er
atitis
u
s
in
g
co
n
f
o
ca
l
an
d
s
lit
-
lam
p
im
ag
es,
r
esp
ec
tiv
ely
,
an
d
Ku
o
et
a
l.
[
2
]
co
m
p
ar
ed
m
u
ltip
le
C
NNs
o
n
ex
te
r
n
al
ey
e
p
h
o
to
g
r
ap
h
s
,
s
h
o
win
g
th
e
b
en
e
f
its
o
f
h
y
b
r
id
m
o
d
els.
Gh
o
s
h
et
a
l.
[
3
]
in
tr
o
d
u
ce
d
Dee
p
Ke
r
atitis
f
o
r
b
ac
ter
ial
–
f
u
n
g
al
d
is
cr
im
in
atio
n
,
an
d
Ass
af
et
a
l.
[
9
]
s
u
r
v
ey
e
d
co
m
p
u
ter
-
v
is
io
n
ap
p
r
o
ac
h
es
ac
r
o
s
s
in
f
ec
tio
u
s
k
er
atitis
s
tu
d
ies.
L
ar
g
e
-
s
ca
le
an
d
m
u
lticen
ter
v
alid
atio
n
s
,
s
u
ch
as
L
i
et
a
l.
[
1
0
]
s
u
p
p
o
r
t
g
e
n
er
aliza
b
ilit
y
ac
r
o
s
s
im
ag
in
g
d
ev
ices
an
d
s
ites
.
Su
b
s
eq
u
en
t
wo
r
k
s
tar
g
eted
s
p
ec
if
ic
p
ath
o
g
en
s
o
r
p
ath
o
lo
g
ies
[
1
1
]
,
[
1
2
]
o
r
s
o
u
g
h
t
p
r
o
b
ab
ilis
tic
p
ath
o
g
en
p
r
ed
ictio
n
f
r
o
m
s
lit
-
lam
p
im
a
g
er
y
[
1
3
]
.
Z
h
an
g
et
a
l.
[
1
4
]
,
L
i
u
et
a
l.
[
1
5
]
,
a
n
d
L
i
et
a
l.
[
1
6
]
e
x
p
lo
r
e
d
f
u
s
io
n
s
tr
ateg
ies,
atten
tio
n
m
e
ch
an
is
m
s
,
an
d
ar
c
h
itectu
r
al
v
ar
ia
n
ts
to
im
p
r
o
v
e
s
en
s
itiv
ity
an
d
r
o
b
u
s
tn
ess
.
C
lin
ical
an
d
ep
id
e
m
io
lo
g
ical
r
ev
iews
[
1
7
]
,
[
1
8
]
em
p
h
asize
d
iag
n
o
s
tic
d
elay
s
an
d
an
tim
icr
o
b
ial
-
r
esis
tan
ce
ch
allen
g
es
th
at
m
o
tiv
ate
au
to
m
ated
s
cr
ee
n
in
g
.
Fo
r
p
r
ac
tical
d
ep
lo
y
m
en
t,
lig
h
tweig
h
t
b
ac
k
b
o
n
es
s
u
ch
as
Mo
b
ileNetV2
[
4
]
a
n
d
Den
s
eNe
t1
2
1
[
5
]
ar
e
f
r
eq
u
e
n
tly
ad
o
p
ted
to
b
ala
n
ce
s
p
ee
d
an
d
r
ep
r
esen
tatio
n
;
co
m
m
o
n
l
y
u
s
ed
p
r
e
p
r
o
ce
s
s
in
g
in
clu
d
es
co
n
tr
ast
-
lim
ited
ad
ap
tiv
e
h
is
to
g
r
a
m
e
q
u
aliza
tio
n
(
C
L
AHE
)
,
co
n
t
r
a
s
t
en
h
an
ce
m
en
t
[
6
]
an
d
Ad
a
m
o
p
tim
izatio
n
f
o
r
s
tab
le
tr
ain
in
g
[
1
9
]
.
I
n
ter
p
r
e
tab
ilit
y
m
eth
o
d
s
(
Gr
ad
-
C
AM
)
ar
e
ap
p
lied
to
v
alid
ate
m
o
d
el
atten
tio
n
o
n
p
ath
o
lo
g
ical
r
eg
io
n
s
[
2
0
]
,
an
d
r
ec
eiv
er
o
p
er
atin
g
ch
a
r
ac
ter
is
tic
(
R
O
C
)
/
ar
ea
u
n
d
er
th
e
cu
r
v
e
(
AUC
)
r
em
ain
s
tan
d
ar
d
ev
alu
atio
n
m
etr
ics
[
7
]
.
Ou
r
cu
r
ated
Kag
g
le
co
llectio
n
[
2
1
]
alig
n
s
with
r
ec
e
n
t
b
e
n
ch
m
ar
k
in
g
an
d
m
eta
-
a
n
aly
s
is
ef
f
o
r
ts
[
2
2
]
;
s
m
ar
tp
h
o
n
e
a
n
d
p
o
r
tab
le
-
im
a
g
in
g
s
tu
d
ies
[
2
3
]
;
a
n
d
m
u
ltimo
d
al
p
ip
elin
es
[
2
4
]
d
em
o
n
s
tr
ate
f
ea
s
ib
ilit
y
f
o
r
f
ield
d
ep
lo
y
m
e
n
t.
T
h
e
m
o
s
t
r
ec
en
t
co
m
p
r
e
h
en
s
iv
e
r
ev
iew
[
9
]
h
ig
h
lig
h
ts
co
n
s
is
ten
t
g
ain
s
in
ac
cu
r
ac
y
b
u
t
also
r
ec
u
r
r
in
g
lim
itatio
n
s
:
r
elian
c
e
o
n
s
in
g
le
b
ac
k
b
o
n
es,
lim
it
ed
ex
ter
n
al
v
alid
ati
o
n
,
an
d
t
r
ad
e
-
o
f
f
s
b
etwe
en
ac
cu
r
ac
y
an
d
in
f
er
en
ce
co
s
t.
T
h
ese
g
ap
s
m
o
tiv
ate
Vis
io
n
E
y
eNe
t’
s
f
u
s
io
n
o
f
Mo
b
ileNetV2
an
d
Den
s
eNe
t1
2
1
tr
ain
ed
o
n
t
h
e
Kir
an
et
a
l.
d
at
aset
[
2
1
]
to
ac
h
iev
e
h
ig
h
d
iag
n
o
s
tic
p
r
ec
is
io
n
with
co
m
p
u
ta
tio
n
al
s
ca
lab
ilit
y
f
o
r
k
er
atitis
an
d
u
v
eitis
s
cr
ee
n
in
g
.
Mu
q
it
et
a
l.
[
2
5
]
r
ep
o
r
ted
th
at
k
er
atitis
an
d
g
r
a
n
u
lo
m
ato
u
s
an
ter
io
r
u
v
eitis
m
ay
co
ex
is
t in
in
f
lam
m
ato
r
y
o
cu
lar
d
is
o
r
d
e
r
s
ass
o
ciate
d
wit
h
im
m
u
n
e
-
co
m
p
le
x
v
ascu
liti
s
.
Desp
ite
th
e
p
r
o
g
r
ess
r
ep
o
r
te
d
in
th
e
p
r
io
r
s
tu
d
ies,
s
ev
er
al
lim
itatio
n
s
r
em
ain
.
Ma
n
y
ap
p
r
o
ac
h
es
r
ely
on
a
s
in
g
le
-
b
ac
k
b
o
n
e
ar
c
h
itectu
r
e
an
d
d
o
n
o
t
ex
p
licitly
a
d
d
r
ess
in
tr
a
-
class
s
im
i
lar
ity
b
etwe
en
ea
r
ly
-
s
tag
e
k
er
atitis
an
d
m
ild
an
ter
io
r
u
v
e
itis
.
Ad
d
itio
n
ally
,
lim
ited
em
p
h
asis
h
as
b
ee
n
p
lace
d
o
n
illu
m
in
atio
n
v
ar
ia
b
ilit
y
an
d
cr
o
s
s
-
d
ev
ice
g
e
n
er
atio
n
.
T
h
ese
g
ap
s
m
o
tiv
ate
th
e
d
ev
el
o
p
m
en
t
o
f
a
f
u
s
io
n
-
b
ased
ar
ch
itectu
r
e
tailo
r
ed
f
o
r
r
o
b
u
s
t f
ea
tu
r
e
ex
tr
ac
tio
n
in
an
t
ir
o
r
s
eg
m
en
t
p
ath
o
lo
g
y
.
3.
M
E
T
H
O
D
Vis
io
n
E
y
eNe
t
in
teg
r
ates
Mo
b
ileNetV2
[
4
]
an
d
Den
s
eNe
t1
2
1
[
5
]
t
o
e
x
tr
ac
t
c
o
m
p
lem
en
ta
r
y
f
e
atu
r
e
r
ep
r
esen
tatio
n
s
f
r
o
m
o
cu
lar
i
m
ag
es.
T
h
e
f
u
s
ed
f
ea
tu
r
es
ar
e
p
ass
ed
th
r
o
u
g
h
c
o
n
v
o
lu
tio
n
al
an
d
f
u
lly
co
n
n
ec
ted
lay
er
s
to
class
if
y
ea
ch
s
am
p
le
as
k
er
atitis
o
r
u
v
eitis
,
with
s
ig
m
o
id
ac
tiv
atio
n
p
r
o
d
u
cin
g
th
e
f
in
al
p
r
o
b
a
b
ilit
y
s
co
r
e.
T
h
e
m
o
d
el
is
tr
ain
ed
u
s
in
g
b
in
ar
y
cr
o
s
s
-
en
tr
o
p
y
l
o
s
s
an
d
o
p
tim
ized
f
o
r
ef
f
icien
t
m
ed
ical
im
ag
e
d
iag
n
o
s
is
.
3
.
1
.
Da
t
a
s
et
c
o
llect
i
o
n
T
h
e
d
ataset
was
o
b
tain
ed
f
r
o
m
Kag
g
le
an
d
co
n
tain
s
lo
w
-
r
eso
lu
tio
n
o
c
u
lar
im
a
g
es
o
f
k
e
r
atitis
an
d
u
v
eitis
ac
r
o
s
s
d
iv
er
s
e
p
atien
t
d
em
o
g
r
ap
h
ics
[
2
1
]
.
T
h
e
k
e
r
atitis
s
am
p
les
in
clu
d
e
b
ac
ter
i
al,
v
ir
al,
a
n
d
f
u
n
g
al
ca
s
es,
wh
ile
th
e
u
v
eitis
s
et
c
o
v
er
s
an
ter
io
r
,
in
ter
m
ed
iate,
p
o
s
ter
io
r
,
an
d
p
an
-
u
v
eitis
p
r
e
s
en
tatio
n
s
.
I
m
ag
es
wer
e
clin
ically
cr
o
s
s
-
v
er
if
ied
,
r
esized
to
2
2
4
×2
2
4
,
n
o
r
m
ali
ze
d
,
an
d
au
g
m
en
ted
th
r
o
u
g
h
f
lip
p
in
g
,
r
o
tatio
n
,
zo
o
m
,
an
d
b
r
i
g
h
tn
ess
ad
ju
s
t
m
en
t.
Min
o
r
class
im
b
ala
n
ce
was
ad
d
r
ess
ed
th
r
o
u
g
h
b
ala
n
cin
g
s
tr
ateg
ies
t
o
im
p
r
o
v
e
m
o
d
el
r
o
b
u
s
tn
ess
.
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
V
is
io
n
E
ye
N
et:
a
cu
s
to
miz
ed
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ea
r
ly
…
(
S
o
ma
s
h
ek
h
a
r
B
a
n
n
u
r
Ma
yig
o
w
d
a
)
2711
T
ab
le
1
d
ep
icts
th
e
two
d
if
f
er
en
t
class
es
o
f
o
c
u
lar
d
is
ea
s
es
co
n
s
id
er
ed
in
th
is
s
tu
d
y
,
n
am
e
ly
k
er
atitis
an
d
u
v
eitis
,
with
9
6
0
a
n
d
9
0
0
im
ag
es
,
r
esp
ec
tiv
ely
.
T
h
is
ac
cu
m
u
lated
d
ataset
was
u
s
ed
f
o
r
tr
ain
i
n
g
an
d
ev
alu
atio
n
o
f
t
h
e
p
r
o
p
o
s
ed
Vi
s
io
n
E
y
eNe
t
m
o
d
el.
T
h
ese
d
ep
ict
th
e
n
o
r
m
al
e
y
e
an
d
u
v
eitis
-
in
f
ec
ted
e
y
e
,
a
n
d
th
e
im
ag
e
d
is
p
lay
s
th
e
d
if
f
er
en
ce
b
etwe
en
th
e
ey
es
,
s
h
o
win
g
h
o
w
o
n
e
an
d
an
o
th
er
is
in
f
e
cted
b
y
th
e
v
ir
u
s
an
d
en
v
ir
o
n
m
en
tal
co
n
d
itio
n
s
.
T
h
e
f
ea
tu
r
es
o
f
t
h
ese
two
it h
as b
e
en
d
is
p
lay
ed
in
th
e
im
ag
e.
T
ab
le
1
.
T
h
e
d
if
f
e
r
en
t c
ateg
o
r
i
es o
f
ey
e
d
is
ea
s
es: ac
cu
m
u
lated
d
ataset
C
l
a
s
s
To
t
a
l
U
v
e
i
t
i
s
9
6
0
K
e
r
a
t
i
t
i
s
9
0
0
3.
2
.
Da
t
a
p
re
pro
ce
s
s
ing
3.
2
.
1
.
Ada
ptiv
e
g
a
mm
a
co
rr
ec
t
io
n
T
o
im
p
r
o
v
e
im
ag
e
v
is
ib
ilit
y
u
n
d
er
lo
w
-
lig
h
t
o
r
u
n
e
v
en
illu
m
in
atio
n
c
o
n
d
itio
n
s
,
an
ad
a
p
tiv
e
g
am
m
a
co
r
r
ec
tio
n
tech
n
iq
u
e
was
ap
p
lied
[
6
]
.
All
in
p
u
t
im
ag
es
wer
e
n
o
r
m
alize
d
t
o
th
e
r
a
n
g
e
[
0
,
1
]
p
r
io
r
to
en
h
an
ce
m
e
n
t.
T
h
e
c
o
r
r
ec
ted
p
ix
el
in
ten
s
ity
at
th
e
lo
ca
tio
n
(
,
)
is
d
ef
in
ed
as
(
1
)
a
n
d
(
2
)
.
(
,
)
=
(
,
)
(
,
)
(
1
)
(
,
)
=
1
+
∙
(
,
)
−
(
,
)
−
∈
(
2
)
I
n
tu
itiv
ely
,
d
ar
k
er
r
eg
io
n
s
(
µ
(
,
)
<
µ
)
r
ec
eiv
e
a
g
am
m
a
v
alu
e
<1
,
lead
in
g
to
b
r
ig
h
ten
i
n
g
,
wh
ile
b
r
ig
h
ter
r
eg
io
n
s
(
µ
(
,
)
>
µ
)
ar
e
en
h
an
c
ed
with
>
1
,
p
r
e
v
en
tin
g
o
v
e
r
am
p
lific
atio
n
.
T
h
is
ad
ap
tiv
ity
en
s
u
r
es
lo
ca
l
co
n
tr
ast
im
p
r
o
v
em
en
t
wh
ile
p
r
eser
v
i
n
g
g
lo
b
al
im
ag
e
b
alan
ce
.
Alter
n
ati
v
e
m
eth
o
d
:
as
a
n
alter
n
ativ
e
o
r
co
m
p
lem
en
tar
y
ap
p
r
o
ac
h
,
C
L
AHE
[
6
]
ca
n
b
e
em
p
lo
y
ed
to
im
p
r
o
v
e
lo
ca
l
co
n
tr
ast
wh
ile
co
n
tr
o
llin
g
n
o
is
e
am
p
lific
atio
n
.
I
f
s
p
ec
u
la
r
h
i
g
h
lig
h
ts
o
r
o
cc
lu
s
io
n
s
ar
e
r
em
o
v
e
d
v
ia
in
p
ain
tin
g
,
m
eth
o
d
s
s
u
ch
as Na
v
ier
–
Sto
k
es
-
b
ased
in
p
ai
n
tin
g
.
3.
2
.
2
.
Sp
ec
ula
r
re
f
lect
io
n su
pp
re
s
s
io
n (
g
la
re
re
m
o
v
a
l)
E
y
e
im
ag
es
o
f
ten
co
n
tain
b
r
ig
h
t
s
p
o
ts
d
u
e
to
s
p
ec
u
lar
r
ef
lec
tio
n
,
wh
ich
m
ay
in
ter
f
er
e
with
ac
cu
r
ate
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
o
ad
d
r
ess
th
is
,
a
b
in
ar
y
m
ask
(
,
)
is
g
en
er
ated
b
ased
o
n
an
in
te
n
s
ity
th
r
esh
o
ld
(
3
)
.
(
,
)
=
{
1
(
,
)
>
0
ℎ
(
3
)
W
h
er
e
(
,
)
is
b
in
ar
y
m
ask
i
n
d
ica
tin
g
s
p
ec
u
lar
r
eg
io
n
s
,
(
,
)
is
p
ix
el
in
ten
s
ity
at
(
,
)
,
an
d
is
in
ten
s
ity
th
r
esh
o
ld
f
o
r
g
lar
e
d
etec
tio
n
.
Pix
els
id
en
tifie
d
as
g
lar
e
(
wh
e
r
e
(
,
)
=
1
)
ar
e
r
em
o
v
ed
,
a
n
d
th
e
m
is
s
in
g
r
eg
io
n
s
a
r
e
r
ec
o
n
s
tr
u
cted
u
s
in
g
Nav
ier
–
Sto
k
es
in
p
ain
tin
g
.
T
h
e
in
p
ain
tin
g
p
r
o
c
ess
f
o
llo
ws
th
e
p
a
r
tial
d
if
f
er
en
tial
(
4
)
.
=
∇
⊥
∙
∇
(
4
)
W
h
er
e
is
im
ag
e
in
ten
s
ity
f
u
n
ctio
n
,
is
i
ter
atio
n
(
in
p
ain
tin
g
t
im
e
s
tep
)
,
is
g
r
ad
ien
t
o
f
th
e
i
m
ag
e
,
an
d
∇
⊥
is
o
r
th
o
g
o
n
al
g
r
a
d
ien
t o
f
th
e
s
tr
ea
m
f
u
n
ctio
n
.
In
(
3
)
an
d
(
4
)
,
e
n
s
u
r
e
th
at
s
p
e
cu
lar
r
eg
io
n
s
ar
e
d
etec
ted
an
d
s
ea
m
less
ly
r
ec
o
n
s
tr
u
cted
.
T
h
is
p
r
o
ce
s
s
ef
f
ec
tiv
ely
s
u
p
p
r
ess
es
g
lar
e
s
p
o
ts
,
th
er
eb
y
p
r
e
v
en
tin
g
m
is
lead
in
g
f
ea
tu
r
es
d
u
r
i
n
g
C
NN
-
b
ased
class
if
icatio
n
.
Ad
ap
tiv
e
g
a
m
m
a
c
o
r
r
ec
tio
n
(
1
)
,
(
2
)
en
h
a
n
ce
s
im
ag
e
c
o
n
tr
a
s
t
u
n
d
er
u
n
e
v
en
lig
h
tin
g
,
wh
il
e
s
p
ec
u
lar
r
ef
lectio
n
s
u
p
p
r
ess
io
n
(
3
)
,
(
4
)
r
em
o
v
es
g
lar
e
an
d
r
esto
r
es
co
r
n
ea
l
tex
t
u
r
e.
T
o
g
eth
er
,
t
h
ese
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
p
r
o
d
u
c
e
illu
m
in
atio
n
-
b
alan
ce
d
,
a
r
tifa
c
t
-
f
r
ee
im
a
g
es,
im
p
r
o
v
in
g
t
h
e
clar
ity
o
f
p
ath
o
l
o
g
ical
r
eg
io
n
s
an
d
en
a
b
lin
g
Vis
io
n
E
y
eNe
t to
ex
tr
ac
t r
o
b
u
s
t a
n
d
clin
ically
r
eliab
le
f
e
atu
r
e
s
.
3.
3
.
P
re
pro
ce
s
s
ing
ev
a
lua
t
io
n a
nd
v
is
ua
l c
o
m
pa
riso
n
Ad
ap
tiv
e
g
am
m
a
c
o
r
r
ec
tio
n
was
ap
p
lied
to
ad
d
r
ess
illu
m
i
n
atio
n
an
d
in
co
n
s
is
ten
cies
,
f
o
llo
wed
b
y
r
ef
r
ac
tio
n
s
u
p
p
r
ess
io
n
to
r
ed
u
ce
g
lar
e
wh
ile
p
r
eser
v
in
g
co
r
n
ea
l
d
etails.
T
h
ese
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
en
h
an
ce
d
co
n
tr
ast
d
is
tr
ib
u
tio
n
a
n
d
im
p
r
o
v
ed
class
if
icatio
n
ac
cu
r
ac
y
f
r
o
m
9
5
.
8
%
to
9
8
.
0
%.
T
h
e
v
is
u
al
d
if
f
er
e
n
ce
s
af
ter
th
e
s
tag
es a
r
e
p
r
esen
ted
in
Fig
u
r
e
1
,
w
h
er
e
Fig
u
r
e
1
(
a)
s
h
o
w
s
th
e
k
er
atitis
an
d
Fig
u
r
e
(
b
)
s
h
o
ws th
e
u
v
eitis
.
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
7
0
9
-
2
7
2
2
2712
(
a)
(
b
)
Fig
u
r
e
1
.
Vis
u
al
co
m
p
ar
is
o
n
o
f
p
r
ep
r
o
ce
s
s
in
g
s
tag
es f
o
r
o
cu
l
ar
im
ag
es
of
(
a)
k
e
r
atitis
an
d
(
b
)
u
v
eitis
3.
4
.
Da
t
a
s
et
s
pli
t
t
ing
T
h
e
d
ataset
in
clu
d
e
d
1
,
8
6
0
s
lit
-
lam
p
im
ag
es
co
n
s
is
tin
g
o
f
9
6
0
u
v
eitis
an
d
9
,
0
0
0
k
er
atitis
ca
s
es.
T
o
en
s
u
r
e
b
alan
cin
g
lea
r
n
in
g
,
s
tr
atif
ied
p
ar
titi
o
n
in
g
was
ap
p
lie
d
d
u
r
in
g
d
ataset
d
iv
is
io
n
.
Ap
p
r
o
x
im
ately
7
1
.
5
%
o
f
s
am
p
les
wer
e
c
o
llected
f
o
r
tr
ain
in
g
,
wh
ile
a
s
ep
ar
ate
v
alid
atio
n
an
d
in
d
ep
en
d
en
t
test
s
et
wer
e
r
eser
v
ed
f
o
r
p
er
f
o
r
m
an
ce
ass
ess
m
en
t.
I
m
p
o
r
tan
tly
,
p
ar
titi
o
n
in
g
was
p
er
f
o
r
m
e
d
at
th
e
p
atien
t
lev
el
to
p
r
e
v
en
t
im
ag
e
o
v
er
lap
a
n
d
in
f
o
r
m
atio
n
leak
a
g
e
.
3.
5
.
P
r
o
po
s
ed
s
y
s
t
em
T
h
e
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
d
e
p
icts
th
e
o
v
er
all
f
lo
w
o
f
th
e
wo
r
k
,
as
s
h
o
wn
in
Fig
u
r
e
2
.
Fig
u
r
e
2
ill
u
s
tr
at
es
t
h
e
Vis
i
o
n
E
y
e
Net
m
o
d
el
,
w
h
i
ch
i
n
te
g
r
at
es
f
e
atu
r
es
f
r
o
m
M
o
b
ile
Net
V2
a
n
d
De
n
s
eNe
t
1
2
1
to
im
p
r
o
v
e
th
e
class
if
icatio
n
o
f
k
er
atitis
an
d
u
v
eitis
.
I
n
p
u
t
im
ag
es
u
n
d
er
g
o
p
r
e
p
r
o
ce
s
s
in
g
a
n
d
ar
e
th
en
p
ass
ed
th
r
o
u
g
h
b
o
t
h
n
etwo
r
k
s
,
af
te
r
wh
ich
th
eir
f
ea
tu
r
es
ar
e
co
m
b
in
ed
.
T
h
e
f
u
s
ed
f
ea
tu
r
es
ar
e
p
r
o
ce
s
s
ed
f
o
r
f
i
n
al
class
if
icatio
n
,
r
esu
ltin
g
in
h
ig
h
ac
cu
r
ac
y
with
m
in
im
al
in
f
e
r
en
ce
tim
e.
illu
s
tr
ates
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
,
wh
ich
o
u
tlin
es
th
e
tr
ain
in
g
an
d
test
in
g
p
r
o
ce
s
s
o
f
th
e
m
o
d
el.
T
h
e
ap
p
r
o
ac
h
em
p
h
asize
s
th
e
f
u
s
io
n
o
f
two
ar
ch
itectu
r
es,
Mo
b
ileNetV2
[
4
]
an
d
Den
s
eNe
t1
2
1
[
5
]
to
cla
s
s
if
y
ey
e
d
is
ea
s
es.
T
h
e
in
teg
r
ated
m
o
d
el,
r
ef
e
r
r
ed
to
as
Vis
io
n
E
y
eNe
t,
is
d
esig
n
ed
to
ac
h
iev
e
s
u
p
er
io
r
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
Fin
all
y
,
th
e
class
if
icatio
n
o
u
tp
u
t
d
eter
m
in
es
wh
eth
e
r
th
e
ey
e
co
n
d
itio
n
is
k
er
atitis
o
r
u
v
eitis
,
en
s
u
r
in
g
im
p
r
o
v
ed
ac
cu
r
ac
y
a
n
d
r
e
d
u
ce
d
in
f
er
en
ce
tim
e
c
o
m
p
ar
e
d
to
in
d
iv
id
u
al
m
o
d
els.
Fig
u
r
e
2
.
Ar
c
h
itectu
r
e
d
iag
r
a
m
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
3.
6
.
T
ra
ini
ng
s
cheduli
ng
T
r
ain
in
g
was p
er
f
o
r
m
e
d
in
tw
o
p
h
ases
:
i)
Featu
r
e
ex
tr
ac
tio
n
p
h
ase:
all
co
n
v
o
l
u
tio
n
al
lay
er
s
o
f
Mo
b
il
eNe
tV2
an
d
Den
s
eNe
t1
2
1
b
a
ck
b
o
n
es
wer
e
f
r
o
ze
n
f
o
r
t
h
e
f
ir
s
t 1
0
e
p
o
ch
s
.
On
ly
th
e
f
u
lly
-
co
n
n
ec
ted
lay
er
s
wer
e
tr
ain
ed
.
ii)
Fin
e
-
tu
n
in
g
p
h
ase:
all
lay
er
s
wer
e
u
n
f
r
o
ze
n
,
a
n
d
th
e
m
o
d
e
l
was
tr
ain
ed
en
d
-
to
-
en
d
f
o
r
t
h
e
r
em
ain
in
g
ep
o
ch
s
u
s
in
g
a
r
ed
u
ce
d
lear
n
i
n
g
r
ate.
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
V
is
io
n
E
ye
N
et:
a
cu
s
to
miz
ed
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ea
r
ly
…
(
S
o
ma
s
h
ek
h
a
r
B
a
n
n
u
r
Ma
yig
o
w
d
a
)
2713
T
h
e
Ad
am
o
p
tim
izer
(
lear
n
in
g
r
ate
1
e−
4
,
weig
h
t
d
ec
ay
1
e
−5
)
was
u
s
ed
with
a
s
tep
lear
n
in
g
r
ate
s
ch
e
d
u
ler
(
f
ac
to
r
0
.
1
,
a
n
d
p
atien
ce
1
0
ep
o
ch
s
)
.
T
o
ad
d
r
ess
m
ild
cla
s
s
im
b
alan
ce
,
weig
h
ted
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
was
em
p
lo
y
ed
with
weig
h
ts
in
v
er
s
ely
p
r
o
p
o
r
tio
n
al
to
class
f
r
eq
u
en
cy
.
3.
7
.
Vis
io
nE
y
eNe
t
a
rc
hite
ct
ure
Fig
u
r
e
3
s
h
o
ws
th
e
Mo
b
ileN
etV2
an
d
Den
s
eNe
t1
2
1
.
Feat
u
r
es
f
r
o
m
b
o
th
n
etwo
r
k
s
ar
e
f
u
s
ed
,
th
en
p
ass
ed
th
r
o
u
g
h
a
f
u
lly
c
o
n
n
e
cted
lay
er
an
d
a
s
o
f
tm
ax
cla
s
s
if
ier
.
I
n
teg
r
atio
n
o
f
d
ee
p
-
lear
n
in
g
m
o
d
els
b
y
eq
u
atio
n
s
,
t
h
e
in
p
u
t im
ag
e
is
r
ep
r
esen
ted
as
(
5
)
.
224
×
224
×
3
(
5
)
W
h
er
e
in
(
5
)
2
2
4
×2
2
4
d
en
o
t
es
th
e
s
p
atial
r
eso
lu
tio
n
,
an
d
3
co
r
r
esp
o
n
d
s
to
t
h
e
R
GB
ch
an
n
els.
Mo
b
ileNet
ex
tr
ac
ts
lig
h
tweig
h
t y
et
d
is
cr
i
m
in
ativ
e
f
ea
tu
r
es f
r
o
m
th
e
in
p
u
t im
ag
e
(
6
)
to
(
9
)
.
(
)
=
(
6
)
(
)
=
121
(
7
)
(
)
=
(
)
⊕
(
)
(
8
)
=
(
(
)
)
(
9
)
Similar
ly
,
Den
s
eNe
t1
2
1
ca
p
t
u
r
es
d
en
s
e
h
ier
a
r
ch
ical
f
ea
tu
r
es
(
7
)
.
T
h
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
ar
e
th
en
f
u
s
ed
to
co
m
b
in
e
co
m
p
l
em
en
tar
y
in
f
o
r
m
atio
n
,
wh
er
e
⊕
d
en
o
tes
th
e
co
n
ca
ten
atio
n
o
r
f
u
s
io
n
o
p
er
ati
o
n
(
8
)
.
Fin
ally
,
t
h
e
f
u
s
ed
f
ea
tu
r
es
ar
e
p
ass
ed
th
r
o
u
g
h
a
s
o
f
tm
ax
lay
er
to
o
b
tain
th
e
class
if
icatio
n
o
u
tp
u
t
(
9
)
.
Fig
u
r
e
3
.
Ar
c
h
itectu
r
e
o
f
Vis
io
n
E
y
eNe
t f
o
r
k
e
r
atitis
an
d
u
v
e
itis
clas
s
if
icatio
n
3.
8
.
F
e
a
t
ure
f
us
io
n a
nd
po
o
lin
g
T
o
m
er
g
e
th
e
lea
r
n
ed
r
ep
r
ese
n
tatio
n
s
f
r
o
m
b
o
t
h
C
NN
b
ac
k
b
o
n
es,
we
a
p
p
ly
g
lo
b
al
a
v
er
a
g
e
p
o
o
lin
g
(
GAP)
to
th
e
f
ea
tu
r
e
m
ap
s
(
1
0
)
.
T
h
e
p
o
o
led
v
ec
to
r
s
ar
e
co
n
c
aten
ated
to
f
o
r
m
th
e
f
u
s
ed
r
ep
r
esen
tatio
n
:
wh
er
e
an
d
ar
e
th
e
f
ea
tu
r
e
d
im
en
s
io
n
s
o
f
Mo
b
ileNet
an
d
De
n
s
eNe
t1
2
1
af
ter
p
o
o
lin
g
(
1
1
)
.
T
h
is
f
u
s
io
n
en
r
ich
es
th
e
in
p
u
t
to
t
h
e
class
i
f
icatio
n
h
ea
d
with
b
o
th
lig
h
tw
eig
h
t
an
d
d
ee
p
h
ier
ar
c
h
ical
f
e
atu
r
es.
E
ac
h
d
en
s
e
lay
er
was
f
o
llo
wed
b
y
a
R
eL
U
ac
tiv
atio
n
an
d
a
Dr
o
p
o
u
t
lay
er
(
p
=0
.
5
)
to
r
ed
u
ce
o
v
er
f
itti
n
g
.
L
et
(
)
d
en
o
te
th
e
p
r
o
p
o
s
ed
Vis
io
n
E
y
eNe
t m
o
d
el
p
ar
am
ete
r
ized
b
y
weig
h
ts
θ
.
T
h
e
f
in
al
p
r
ed
ictio
n
is
g
iv
en
as (
1
2
)
.
=
(
(
)
)
,
=
(
(
)
)
(
1
0
)
=
[
,
]
∈
+
(
1
1
)
̂
=
(
(
)
)
(
1
2
)
W
h
er
e
̂
∈
2
r
ep
r
esen
ts
th
e
p
r
ed
icte
d
p
r
o
b
ab
ilit
ies f
o
r
k
e
r
atitis
an
d
u
v
eitis
.
3.
9
.
Cla
s
s
if
ica
t
io
n
h
ea
d
Fo
llo
win
g
Alg
o
r
ith
m
1
,
t
h
e
f
u
s
ed
f
ea
tu
r
es
F
ar
e
p
ass
ed
th
r
o
u
g
h
two
f
u
lly
co
n
n
ec
ted
la
y
er
s
with
R
eL
U
ac
tiv
atio
n
an
d
Dr
o
p
o
u
t
r
eg
u
lar
izatio
n
(
1
3
)
an
d
(
1
4
)
.
1
=
(
1
+
1
)
,
̀
1
(
1
,
0
.
6
)
(
1
3
)
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
7
0
9
-
2
7
2
2
2714
2
=
(
2
̀
+
2
)
,
̀
2
(
2
,
0
.
2
)
(
1
4
)
Fin
ally
,
th
e
o
u
tp
u
t la
y
er
a
p
p
li
es a
s
o
f
tm
a
x
ac
tiv
atio
n
(
1
5
)
.
̂
=
(
3
̀
2
+
3
)
(
1
5
)
W
h
er
e
̂
∈
2
g
iv
es
th
e
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
o
v
er
th
e
two
cla
s
s
es
(
k
er
atiti
s
an
d
u
v
eitis
)
.
T
h
e
m
o
d
el
is
tr
ain
ed
b
y
m
in
im
izin
g
th
e
ca
te
g
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
l
o
s
s
(
1
6
)
.
=
−
1
∑
∑
,
2
=
1
(
̂
,
)
=
1
(
1
6
)
W
h
er
e
,
is
th
e
tr
u
e
lab
el
an
d
̂
,
is
th
e
p
r
ed
icted
p
r
o
b
ab
ilit
y
f
o
r
c
lass
.
Alg
o
r
ith
m
1
.
T
r
ain
in
g
a
lg
o
r
ith
m
f
o
r
Vis
io
n
E
y
eNe
t
a
nd
m
at
h
em
atica
l f
o
r
m
u
latio
n
o
f
o
u
tp
u
t
I
n
p
u
t:
p
r
ep
r
o
ce
s
s
ed
in
p
u
t im
a
g
e
224
×
224
×
3
Ou
tp
u
t:
p
r
ed
icted
class
lab
el
(
k
er
atitis
o
r
u
v
eitis
)
1:
L
o
ad
Mo
b
ileNetV2
a
n
d
Den
s
eNe
t1
2
1
with
o
u
t t
o
p
lay
er
s
2:
Fre
ez
e
all
co
n
v
o
lu
tio
n
al
lay
er
s
in
b
o
th
n
etwo
r
k
s
3:
E
x
tr
ac
t f
ea
tu
r
es:
←
(
)
,
←
(
)
4:
Ap
p
ly
GAP:
←
(
)
,
←
(
)
5:
C
o
n
ca
ten
ate
p
o
o
led
f
ea
tu
r
es:
←
[
‖
]
6:
C
o
m
p
u
te
1
←
(
1
+
1
)
7:
Ap
p
ly
d
r
o
p
o
u
t:
̀
1
←
(
1
,
0
.
5
)
8:
C
o
m
p
u
te
2
←
(
2
̀
1
+
2
)
9:
Ap
p
ly
d
r
o
p
o
u
t:
̀
2
←
(
2
,
0
.
3
)
10:
C
o
m
p
u
te
o
u
tp
u
t:
̂
←
(
3
̀
2
+
3
)
11:
R
etu
r
n
p
r
ed
icted
class
:
(
̂
)
.
T
h
e
o
p
tim
izatio
n
o
b
jectiv
e
is
d
ef
in
ed
as
(
1
7
)
a
n
d
(
1
8
)
.
∗
=
\
(
(
)
+
|
|
|
|
2
)
(
1
7
)
∈
×
224
×
224
×
3
,
∈
{
0
,
1
}
(
1
8
)
W
h
er
e
λ
i
s
th
e
r
eg
u
lar
izatio
n
p
ar
am
eter
an
d
‖
‖
2
is
th
e
L
2
(
1
7
)
r
eg
u
lar
izatio
n
ter
m
to
p
r
e
v
en
t
o
v
er
f
itti
n
g
.
L
et
th
e
d
ataset
b
e
r
ep
r
esen
ted
as:
wh
er
e
is
th
e
s
et
o
f
n
i
n
p
u
t
im
ag
es
,
an
d
co
n
tain
s
th
e
c
o
r
r
esp
o
n
d
in
g
b
in
a
r
y
lab
els (
0
=
k
er
atitis
,
1
=
u
v
eitis
)
(
1
8
)
.
3.
10
.
L
o
s
s
f
un
ct
io
n a
nd
o
pti
m
iza
t
io
n
T
h
e
m
o
d
el
is
tr
ain
e
d
u
s
in
g
t
h
e
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
l
o
s
s
in
(
1
9
)
.
θ
=
−
1
∑
∑
,
2
=
1
l
og
(
̂
,
)
=
1
(
1
9
)
T
o
av
o
id
o
v
e
r
f
itti
n
g
,
L
2
r
eg
u
lar
izatio
n
is
ad
d
ed
(
2
0
)
.
∗
=
\
(
(
)
+
|
|
2
)
(
2
0
)
3
.
1
1
.
M
o
del f
us
io
n j
us
t
if
ica
t
io
n (
inte
g
ra
t
ed
wit
h prio
r
lite
ra
t
ure)
R
ec
en
t
s
tu
d
ies
in
o
cu
lar
im
ag
in
g
h
a
v
e
h
ig
h
lig
h
ted
th
e
s
tr
o
n
g
p
e
r
f
o
r
m
an
ce
o
f
Mo
b
ile
NetV2
an
d
Den
s
eNe
t1
2
1
f
o
r
d
etec
tin
g
a
n
ter
io
r
s
eg
m
en
t
d
is
o
r
d
er
s
s
u
ch
as
k
er
atitis
an
d
u
v
eitis
.
W
o
r
k
s
b
y
Ku
o
et
a
l.
[
2
]
,
Gh
o
s
h
et
a
l.
[
3
]
,
an
d
L
i
et
a
l.
[
1
6
]
d
em
o
n
s
tr
ated
th
at
lig
h
t
weig
h
t
C
NNs
lik
e
Mo
b
ileNet
V2
d
eliv
er
r
eliab
le
ac
cu
r
ac
y
o
n
s
lit
-
lam
p
d
atasets
wh
ile
r
em
ain
in
g
s
u
itab
le
f
o
r
r
ea
l
-
tim
e
u
s
e.
L
ik
ewise,
Den
s
eNe
t1
2
1
h
as
b
ee
n
wid
ely
ad
o
p
ted
in
m
ed
ical
im
ag
in
g
[
1
0
]
,
[
1
1
]
,
[
1
4
]
b
ec
a
u
s
e
its
d
en
s
e
c
o
n
n
ec
tio
n
s
en
h
an
c
e
f
ea
tu
r
e
r
e
u
s
e
an
d
ca
p
tu
r
e
s
u
b
tle
p
at
h
o
lo
g
ical
te
x
tu
r
es.
T
h
eo
r
etica
l
r
atio
n
ale:
Mo
b
ileNetV2
ca
p
tu
r
es
f
in
e
s
p
atial
d
etails
th
r
o
u
g
h
d
ep
th
wis
e
s
ep
ar
ab
le
c
o
n
v
o
lu
t
io
n
s
[
4
]
,
wh
er
ea
s
Den
s
eNe
t1
2
1
en
s
u
r
es
g
lo
b
al
f
ea
tu
r
e
co
n
tin
u
ity
an
d
s
tab
le
g
r
ad
ien
ts
[
5
]
.
T
h
eir
f
u
s
io
n
co
m
b
in
es
lo
ca
lized
d
etail
an
d
b
r
o
ad
er
s
tr
u
ctu
r
al
cu
es,
im
p
r
o
v
i
n
g
f
ea
tu
r
e
d
iv
e
r
s
ity
with
o
u
t
h
ea
v
y
co
m
p
u
tatio
n
al
co
s
t.
E
x
p
er
im
e
n
tal
ev
id
en
ce
:
Vis
io
n
E
y
eNe
t
ac
h
iev
ed
9
8
.
0
%
ac
cu
r
ac
y
a
n
d
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
V
is
io
n
E
ye
N
et:
a
cu
s
to
miz
ed
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ea
r
ly
…
(
S
o
ma
s
h
ek
h
a
r
B
a
n
n
u
r
Ma
yig
o
w
d
a
)
2715
0
.
9
9
AUROC
,
s
u
r
p
ass
in
g
b
o
t
h
s
in
g
le
b
ac
k
b
o
n
es
an
d
alter
n
ativ
e
p
air
in
g
s
.
T
h
ese
f
i
n
d
in
g
s
ar
e
co
n
s
is
ten
t
with
On
g
et
a
l.
[
2
2
]
a
n
d
So
leim
a
n
i
et
a
l.
[
2
3
]
,
c
o
n
f
ir
m
i
n
g
th
e
ef
f
icien
cy
o
f
f
ea
tu
r
e
-
lig
h
t
m
o
d
els.
Gr
ad
-
C
AM
r
esu
lts
[
2
0
]
f
u
r
th
e
r
v
alid
ate
d
th
at
Vis
io
n
E
y
eNe
t
f
o
cu
s
es
o
n
clin
ically
r
elev
an
t
r
eg
io
n
s
,
r
ein
f
o
r
cin
g
in
ter
p
r
etab
ilit
y
[
9
]
,
[
2
2
]
.
3.
12
.
T
ra
ini
ng
pro
t
o
c
o
l a
nd
hy
perpa
ra
m
et
er
s
T
h
e
m
o
d
el
tr
ai
n
in
g
was
p
er
f
o
r
m
ed
u
s
in
g
th
e
Ad
am
o
p
t
im
izer
with
an
in
itial
lear
n
in
g
r
ate
o
f
1
×
10
−
4
an
d
a
weig
h
t
d
ec
ay
o
f
1
×
10
−
5
[
2
0
]
.
A
s
tep
lear
n
in
g
-
r
ate
s
ch
ed
u
ler
was
ap
p
lied
,
r
ed
u
cin
g
th
e
r
ate
b
y
a
f
ac
to
r
o
f
0
.
1
af
ter
1
0
co
n
s
ec
u
tiv
e
ep
o
c
h
s
with
o
u
t
im
p
r
o
v
em
e
n
t.
T
h
e
b
atch
s
ize
was
s
e
t
to
3
2
,
an
d
th
e
m
o
d
el
was
tr
ain
e
d
f
o
r
a
m
ax
im
u
m
o
f
1
0
0
e
p
o
ch
s
wit
h
ea
r
ly
s
to
p
p
in
g
(
p
atien
ce
=
1
5
ep
o
c
h
s
)
b
ased
o
n
v
alid
atio
n
lo
s
s
.
4.
E
XP
E
R
I
M
E
N
T
AND
E
VA
L
UATI
O
N
T
o
s
tab
ilize
co
n
v
er
g
en
ce
,
t
h
e
Mo
b
ileNetV2
an
d
Den
s
eNe
t
1
2
1
b
a
ck
b
o
n
es
wer
e
in
itially
f
r
o
ze
n
f
o
r
th
e
f
ir
s
t
1
0
ep
o
c
h
s
,
allo
win
g
o
n
ly
t
h
e
f
u
lly
co
n
n
ec
ted
la
y
er
s
to
b
e
tr
ain
ed
.
Af
ter
war
d
s
,
all
lay
er
s
wer
e
u
n
f
r
o
z
en
an
d
f
in
e
-
tu
n
ed
en
d
-
to
-
en
d
.
All
ex
p
er
im
en
ts
wer
e
co
n
d
u
cte
d
o
n
a
n
NVI
DI
A
T
esla
V1
0
0
GPU
with
3
2
GB
m
em
o
r
y
,
r
u
n
n
in
g
Py
T
o
r
ch
2
.
0
.
T
r
ai
n
in
g
ea
ch
f
o
ld
r
eq
u
ir
ed
ap
p
r
o
x
im
atel
y
2
.
5
h
o
u
r
s
.
Fo
r
r
ep
r
o
d
u
cib
ilit
y
,
r
a
n
d
o
m
s
ee
d
s
wer
e
f
ix
ed
at
4
2
ac
r
o
s
s
Nu
m
Py
an
d
Py
T
o
r
ch
li
b
r
ar
ies.
4
.
1
.
P
er
f
o
r
m
a
nce
m
e
t
rics
T
o
co
m
p
r
e
h
en
s
iv
ely
ev
alu
ate
th
e
p
r
o
p
o
s
ed
Vis
io
n
E
y
eNe
t
f
r
am
ewo
r
k
,
we
r
ep
o
r
t
b
o
th
o
v
er
all
an
d
p
er
-
class
m
etr
ics.
I
n
ad
d
itio
n
t
o
ac
cu
r
ac
y
,
th
e
f
o
llo
win
g
e
v
alu
atio
n
m
ea
s
u
r
es we
r
e
u
s
ed
in
(
2
1
)
a
n
d
(
2
2
)
.
=
/
+
,
(
)
=
/
+
(
2
1)
=
+
,
1
−
=
2
×
×
/
+
(
2
2
)
Her
e,
T
P,
T
N,
FP
,
an
d
FN
r
ep
r
esen
t
t
r
u
e
p
o
s
itiv
es,
tr
u
e
n
eg
ativ
es,
f
alse
p
o
s
itiv
es,
an
d
f
alse
n
eg
ativ
es,
r
esp
ec
tiv
ely
.
R
OC
cu
r
v
es
an
d
th
e
c
o
r
r
esp
o
n
d
in
g
AUC
wer
e
g
en
er
ated
f
o
r
ea
ch
class
.
Sin
ce
th
e
d
ataset
co
n
tain
s
class
im
b
alan
ce
ac
r
o
s
s
k
e
r
atitis
s
u
b
ty
p
es,
p
r
ec
is
io
n
–
r
ec
all
(
PR
)
c
u
r
v
es
w
er
e
also
in
cl
u
d
ed
to
b
etter
ass
ess
m
o
d
el
b
eh
av
io
r
in
m
in
o
r
ity
class
es.
Fu
r
th
er
m
o
r
e,
9
5
%
co
n
f
id
e
n
ce
in
ter
v
als
(
9
5
%
C
I
)
f
o
r
ac
cu
r
ac
y
,
F1
-
s
co
r
e
,
an
d
AUC
wer
e
esti
m
ated
u
s
in
g
1
,
0
0
0
b
o
o
ts
tr
ap
r
esam
p
les
o
f
th
e
test
s
et.
T
h
is
p
r
o
ce
d
u
r
e
p
r
o
v
id
es a
n
esti
m
ate
o
f
s
tatis
ti
ca
l u
n
ce
r
tain
ty
a
n
d
r
e
d
u
ce
s
th
e
r
is
k
o
f
o
v
er
f
itti
n
g
claim
s
.
4.
2
.
Cro
s
s
-
v
a
lid
a
t
io
n str
a
t
eg
y
A
to
tal
o
f
f
i
v
e
v
alid
atio
n
cy
c
les
wer
e
ap
p
lie
d
,
allo
ca
tin
g
8
0
%
o
f
s
am
p
les
to
m
o
d
el
d
e
v
elo
p
m
en
t
wh
ile
2
0
%
s
u
p
p
o
r
te
d
ass
ess
m
en
t
at
e
v
er
y
s
tag
e;
o
u
tco
m
es
ap
p
ea
r
as
av
er
a
g
e
al
o
n
g
s
id
e
v
ar
iatio
n
m
ea
s
u
r
es.
Sep
ar
ately
,
o
n
e
-
f
if
th
o
f
o
b
s
er
v
atio
n
s
s
tay
ed
asid
e
th
r
o
u
g
h
o
u
t,
later
u
s
ed
f
o
r
im
p
a
r
tial
o
u
tco
m
e
v
er
if
icatio
n
.
Su
ch
a
s
tr
u
ctu
r
e
lim
its
ex
ce
s
s
iv
e
ad
ap
tatio
n
to
tr
ain
in
g
p
att
er
n
s
,
p
r
o
m
o
tin
g
co
n
s
is
ten
cy
b
ey
o
n
d
in
itial
ca
s
es.
Ah
ea
d
,
s
cr
u
tin
y
will
ex
ten
d
t
o
war
d
f
u
r
th
er
o
p
en
e
y
e
-
r
elat
ed
co
llectio
n
s
,
ex
am
in
i
n
g
s
ta
b
ilit
y
ac
r
o
s
s
v
ar
ied
g
r
o
u
p
s
th
r
o
u
g
h
Vis
io
n
E
y
eNe
t
tr
ials
.
4.
3
.
Abla
t
io
n
s
t
ud
y
T
o
ev
alu
ate
th
e
co
n
tr
ib
u
tio
n
o
f
ea
ch
b
ac
k
b
o
n
e
n
etwo
r
k
,
we
co
n
d
u
cte
d
an
ab
latio
n
s
tu
d
y
c
o
m
p
ar
in
g
th
r
ee
co
n
f
ig
u
r
atio
n
s
:
Mo
b
il
eNe
tV2
o
n
ly
,
a
b
aselin
e
li
g
h
tweig
h
t
m
o
d
el.
Den
s
eNe
t1
2
1
o
n
l
y
:
d
ee
p
er
ar
ch
itectu
r
e
with
h
ig
h
er
r
ep
r
esen
tatio
n
al
ca
p
ac
ity
.
Pro
p
o
s
ed
Vis
io
n
E
y
eNe
t
(
Mo
b
ileNetV2
+
Den
s
eNe
t1
2
1
f
u
s
io
n
)
:
f
ea
tu
r
e
-
lev
el
in
teg
r
ati
o
n
o
f
b
o
th
b
ac
k
b
o
n
es
,
e
ac
h
m
o
d
el
was
tr
ain
ed
u
n
d
er
id
en
tical
p
r
ep
r
o
ce
s
s
in
g
,
au
g
m
en
tatio
n
,
an
d
h
y
p
er
p
ar
a
m
eter
s
ettin
g
s
.
Per
f
o
r
m
an
ce
was
co
m
p
ar
ed
in
ter
m
s
o
f
ac
c
u
r
ac
y
,
ar
ea
u
n
d
e
r
th
e
R
OC
cu
r
v
e
(
AUC),
F1
-
s
co
r
e,
an
d
in
f
er
e
n
ce
tim
e.
I
n
f
er
en
c
e
tim
e
was
m
ea
s
u
r
ed
as
th
e
av
er
ag
e
p
r
ed
ictio
n
laten
cy
p
er
im
ag
e
(
b
atch
s
ize
=1
)
o
n
th
e
s
am
e
h
ar
d
wa
r
e.
T
o
ass
ess
s
ta
tis
t
ical
s
ig
n
if
ican
ce
,
a
p
air
e
d
Mc
Nem
ar
’
s
test
was
ap
p
lied
to
co
m
p
ar
e
m
is
class
if
icatio
n
d
is
tr
ib
u
tio
n
s
b
etwe
en
Vis
io
n
E
y
eNe
t
an
d
t
h
e
in
d
iv
id
u
al
b
aselin
e
m
o
d
els.
Sig
n
if
ican
ce
was r
ep
o
r
ted
at
p
<
0
.
0
5
.
4.
4
.
I
nfe
re
nce
t
im
e
I
n
f
er
en
ce
tim
e
was
ass
ess
ed
b
y
c
o
m
p
u
tin
g
th
e
av
e
r
ag
e
lat
en
cy
p
er
im
a
g
e
(
b
atch
s
ize
=1
)
u
s
in
g
an
NVI
DI
A
T
esla
V1
0
0
GPU.
Acr
o
s
s
r
ep
ea
ted
r
u
n
s
,
th
e
p
r
o
p
o
s
ed
Vis
io
n
E
y
eNe
t
r
eq
u
ir
ed
ap
p
r
o
x
im
ately
5
1
±
2
m
s
p
er
im
a
g
e
,
in
d
icati
n
g
s
tab
le
co
m
p
u
tatio
n
al
b
e
h
a
v
io
r
,
s
p
ee
d
s
u
g
g
ests
th
at
th
e
f
r
am
ewo
r
k
ca
n
b
e
in
teg
r
ated
in
to
clin
ical
s
ettin
g
s
wh
er
e
tim
ely
d
ec
is
io
n
s
u
p
p
o
r
t
is
im
p
o
r
tan
t,
wh
ile
m
ain
tain
in
g
s
tr
o
n
g
d
iag
n
o
s
tic
p
er
f
o
r
m
a
n
ce
.
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
7
0
9
-
2
7
2
2
2716
4.
5
.
E
x
pla
ina
bil
it
y
a
nd
m
o
d
el
inte
rpre
t
a
bil
it
y
Gr
ad
-
C
AM
was
em
p
lo
y
ed
to
v
is
u
alize
th
e
r
eg
io
n
s
co
n
tr
ib
u
tin
g
to
Vis
io
n
E
y
eNe
t’
s
p
r
ed
ictio
n
.
Activ
atio
n
m
ap
s
.
E
m
p
h
asized
co
r
n
ea
l
o
p
ac
ities
an
d
in
f
lam
m
ato
r
y
f
ea
t
u
r
es
r
ath
er
th
a
n
b
ac
k
g
r
o
u
n
d
a
r
tifa
cts
,
in
teg
r
ated
g
r
ad
ien
t
p
r
o
d
u
ce
d
s
im
ilar
attr
ib
u
tio
n
p
atter
n
s
ac
r
o
s
s
s
am
p
les,
in
d
icatin
g
ta
b
le
an
d
clin
ically
m
ea
n
in
g
f
u
l
m
o
d
el
b
eh
av
i
o
r
.
4.
6
.
Da
t
a
inte
g
rit
y
a
nd
lea
ka
g
e
prev
ent
io
n
Data
a
u
g
m
en
tatio
n
,
in
cl
u
d
in
g
f
lip
s
,
r
o
tatio
n
s
,
zo
o
m
in
g
,
an
d
b
r
ig
h
tn
ess
ad
ju
s
tm
en
t
,
was
ap
p
lied
o
n
ly
to
th
e
tr
ain
in
g
s
et.
T
h
e
v
alid
a
tio
n
an
d
test
s
ets
wer
e
k
ep
t
u
n
ch
an
g
e
d
to
av
o
id
in
f
o
r
m
atio
n
leak
ag
e.
Data
s
et
s
p
litt
in
g
was
p
er
f
o
r
m
e
d
at
th
e
p
atien
t
lev
el
to
en
s
u
r
e
th
at
im
ag
es
f
r
o
m
th
e
s
am
e
in
d
iv
id
u
al
d
id
n
o
t
ap
p
ea
r
in
m
u
ltip
le
s
u
b
s
ets,
m
ain
tain
in
g
f
air
an
d
u
n
b
iased
e
v
alu
atio
n
.
4.
7
.
H
y
perpa
ra
m
e
t
er
s
ea
rc
h a
nd
re
pro
du
cibi
lity
Key
h
y
p
er
p
ar
am
eter
s
,
in
cl
u
d
in
g
lear
n
i
n
g
r
ate
,
b
atch
s
ize
,
an
d
d
r
o
p
o
u
t
,
wer
e
tu
n
ed
u
s
in
g
v
alid
atio
n
-
b
ased
g
r
id
s
ea
r
ch
.
R
an
d
o
m
s
ee
d
s
(
4
2
)
wer
e
f
ix
e
d
in
Py
th
o
n
,
Nu
m
p
y
,
an
d
Py
T
o
r
ch
to
m
ain
tain
r
ep
r
o
d
u
cib
ilit
y
in
Py
T
o
r
c
h
2
.
0
with
C
UDA
1
1
.
8
.
T
h
e
im
p
le
m
en
tatio
n
an
d
p
r
ep
r
o
ce
s
s
in
g
s
cr
ip
ts
will
b
e
m
ad
e
p
u
b
licly
av
ailab
le
o
n
GitHu
b
t
o
s
u
p
p
o
r
t
tr
an
s
p
ar
e
n
cy
.
4.
8
.
St
a
t
is
t
ica
l sig
nifica
nce
t
esting
Mc
Nem
ar
’
s
test
was
u
s
ed
to
co
m
p
ar
e
Vis
io
n
E
y
eNe
t
with
th
e
s
tr
o
n
g
est
b
aselin
e
(
E
f
f
icie
n
tNetB
0
to
R
esNet5
0
)
.
T
h
e
p
-
v
alu
e(
0
.
0
3
2
)
in
d
icate
s
a
s
tatis
t
ically
m
e
an
in
g
f
u
l
d
i
f
f
er
en
ce
in
p
er
f
o
r
m
an
ce
.
C
o
n
f
id
en
ce
in
ter
v
als
(
9
5
%)
wer
e
also
esti
m
ated
with
ac
cu
r
ac
y
at
9
8
.
0
±
0
.
6
%
an
d
AUROC
at
0
.
9
9
±
0
.
0
0
4
.
Mo
s
t
er
r
o
r
s
o
cc
u
r
r
e
d
b
etwe
en
b
ac
ter
ial
an
d
f
u
n
g
al
k
er
atitis
ca
s
es,
lik
ely
d
u
e
to
s
im
ilar
lesi
o
n
b
o
u
n
d
ar
y
ch
ar
ac
ter
is
tics
.
4.
9
.
E
t
hica
l c
o
ns
idera
t
io
ns
a
nd
da
t
a
licens
ing
T
h
e
d
ataset
u
s
ed
in
th
is
s
tu
d
y
was
s
o
u
r
ce
d
f
r
o
m
th
e
Ka
g
g
le
r
ep
o
s
ito
r
y
[
2
1
]
,
an
d
all
im
ag
es
ar
e
de
-
id
en
tifie
d
an
d
p
u
b
licly
av
a
ilab
le
u
n
d
er
Kag
g
le
’
s
ter
m
s
o
f
u
s
e.
No
p
er
s
o
n
ally
id
en
tifia
b
le
in
f
o
r
m
atio
n
was
ac
ce
s
s
ed
.
T
h
e
s
tu
d
y
ad
h
e
r
es
to
eth
ical
r
esear
ch
p
r
ac
tices,
en
s
u
r
in
g
co
m
p
lian
ce
with
d
ata
s
h
ar
in
g
an
d
u
s
ag
e
g
u
id
elin
es.
5.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
T
h
is
s
ec
tio
n
r
ep
o
r
ts
th
e
ex
p
er
i
m
en
tal
r
esu
lts
o
f
Vis
io
n
E
y
eNe
t
an
d
b
en
ch
m
ar
k
s
th
em
ag
ain
s
t
ex
is
tin
g
d
ee
p
lear
n
i
n
g
m
o
d
els
u
s
in
g
a
cc
u
r
ac
y
,
AUROC
,
in
f
er
en
ce
tim
e,
an
d
class
-
wis
e
m
etr
ics.
Vis
u
al
an
aly
s
es
s
u
ch
as lea
r
n
in
g
cu
r
v
es a
n
d
c
o
n
f
u
s
io
n
m
atr
ices a
r
e
in
clu
d
ed
t
o
s
u
p
p
o
r
t th
e
ev
alu
atio
n
.
5
.
1
.
Co
m
pa
ra
t
iv
e
m
o
del per
f
o
rm
a
nce
T
ab
le
2
s
h
o
ws
th
e
d
ataset
d
i
s
tr
ib
u
tio
n
o
f
k
er
atitis
an
d
u
v
eitis
im
ag
es
in
to
tr
ain
in
g
an
d
v
alid
atio
n
s
u
b
s
ets.
A
b
alan
ce
d
r
ep
r
esen
t
atio
n
o
f
b
o
th
class
es
is
g
u
ar
a
n
teed
b
y
th
e
s
tr
atif
ied
d
i
v
id
e.
T
ab
le
3
s
h
o
ws
h
o
w
Vis
io
n
E
y
eNe
t
p
er
f
o
r
m
s
in
co
m
p
ar
is
o
n
t
o
a
n
u
m
b
er
o
f
cu
r
r
en
t
C
NN
ar
ch
itectu
r
es.
Vis
i
o
n
E
y
eNe
t
p
er
f
o
r
m
s
b
etter
o
n
a
m
u
lti
-
class
o
cu
lar
d
ataset
th
an
m
o
d
els
lik
e
E
f
f
ici
en
tNetV2
M,
VGG1
9
,
an
d
De
n
s
eNe
t,
wh
ich
o
n
ly
m
o
d
er
ately
p
er
f
o
r
m
o
n
s
lit
-
lam
p
o
r
d
o
m
ai
n
-
s
p
ec
if
ic
d
ata
s
ets.
T
ab
le
2
s
u
m
m
ar
izes
th
e
class
-
wi
s
e
s
p
li
t
(1
,
4
8
8
tr
ain
,
3
7
2
v
alid
atio
n
)
en
s
u
r
in
g
b
alan
ce
d
d
ata.
T
a
b
le
3
co
m
p
ar
es
m
o
d
el
ac
cu
r
ac
y
an
d
AUROC
,
s
h
o
win
g
Vis
io
n
E
y
eNe
t o
u
t
p
er
f
o
r
m
in
g
all
b
aselin
es with
9
8
% a
cc
u
r
ac
y
.
T
ab
le
2
.
C
lass
-
wis
e
tr
ain
–
v
alid
atio
n
s
p
lit o
f
th
e
d
ataset
C
l
a
s
s
To
t
a
l
i
m
a
g
e
s
Tr
a
i
n
(
8
0
%)
V
a
l
i
d
a
t
i
o
n
(
2
0
%)
U
v
e
i
t
i
s
9
6
0
7
6
8
1
9
2
K
e
r
a
t
i
t
i
s
9
0
0
7
2
0
1
8
0
To
t
a
l
1
,
8
6
0
1
,
4
8
8
3
7
2
T
ab
le
3
.
C
o
m
p
a
r
is
o
n
o
f
ex
is
tin
g
m
o
d
els an
d
Vis
io
n
E
y
eNe
t
M
o
d
e
l
D
a
t
a
s
e
t
A
c
c
u
r
a
c
y
(
%)
A
U
R
O
C
Ef
f
i
c
i
e
n
t
N
e
t
V
2
M
S
l
i
t
-
l
a
m
p
74
0
.
8
5
V
G
G
1
9
D
e
e
p
k
e
r
a
t
i
t
i
s
91
0
.
9
0
Ef
f
i
c
i
e
n
t
N
e
t
B
3
S
l
i
t
-
l
a
m
p
68
0
.
8
5
D
e
n
seN
e
t
H
S
V
s
t
r
o
ma
l
96
0
.
7
3
V
i
si
o
n
E
y
e
N
e
t
(
p
r
o
p
o
se
d
)
M
u
l
t
i
-
c
l
a
ss
o
c
u
l
a
r
98
0
.
9
9
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
V
is
io
n
E
ye
N
et:
a
cu
s
to
miz
ed
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ea
r
ly
…
(
S
o
ma
s
h
ek
h
a
r
B
a
n
n
u
r
Ma
yig
o
w
d
a
)
2717
5
.
2
.
Acc
ura
cy
a
nd
infe
re
nce
t
im
e
Fig
u
r
e
4
co
m
p
ar
es
th
e
ac
cu
r
a
cy
o
f
Vis
io
n
E
y
eNe
t
with
p
r
e
-
ex
is
tin
g
m
o
d
els,
s
h
o
win
g
Vis
io
n
E
y
eNe
t
o
u
tp
er
f
o
r
m
in
g
all
b
aselin
es.
Fig
u
r
e
5
p
r
esen
ts
in
f
er
en
ce
tim
e
an
aly
s
is
,
h
ig
h
lig
h
tin
g
th
at
Vis
io
n
E
y
eNe
t
ac
h
iev
es
co
m
p
etitiv
e
s
p
ee
d
(
5
1
m
s
/im
ag
e)
wh
ile
m
ain
ta
in
in
g
th
e
h
ig
h
est
ac
cu
r
ac
y
.
Fig
u
r
e
4
p
r
esen
ts
ac
cu
r
ac
y
ac
r
o
s
s
f
iv
e
ar
c
h
itectu
r
es
—
Mo
b
ileNetV2
,
Den
s
eNe
t1
2
1
,
VGG1
9
,
E
f
f
icien
t
NetV2
M,
an
d
th
e
p
r
o
p
o
s
ed
Vis
io
n
E
y
eNe
t.
M
o
b
ileNetV2
an
d
Den
s
eNe
t1
2
1
ac
h
iev
ed
9
6
.
5
%
an
d
9
6
.
9
%
ac
cu
r
ac
y
,
o
u
tp
er
f
o
r
m
in
g
VGG1
9
(
9
1
.
0
%)
an
d
E
f
f
icien
tNetV2
M
(
7
3
.
5
%).
T
h
e
h
y
b
r
id
Vis
io
n
E
y
eNe
t
r
ea
ch
ed
th
e
h
ig
h
est
ac
cu
r
ac
y
(
9
8
%),
v
alid
atin
g
its
co
m
p
lem
en
tar
y
f
u
s
io
n
d
esig
n
.
Fig
u
r
e
5
s
h
o
ws
in
f
er
en
ce
s
p
ee
d
co
m
p
ar
is
o
n
,
wh
er
e
Vis
io
n
E
y
eNe
t
ac
h
iev
ed
5
1
±
2
m
s
p
er
im
ag
e,
clo
s
e
t
o
Mo
b
ileNetV2
(
4
9
m
s
)
a
n
d
f
aster
th
an
Den
s
eNe
t1
2
1
(
5
5
m
s
)
,
o
f
f
er
in
g
a
n
ef
f
ec
tiv
e
b
alan
ce
b
et
wee
n
p
r
ec
is
io
n
an
d
ef
f
icien
c
y
.
Fig
u
r
e
4
.
Mo
d
el
ac
cu
r
ac
y
Fig
u
r
e
5
.
I
n
f
er
e
n
ce
tim
e
co
m
p
ar
is
o
n
5
.
3
.
Co
nfusi
o
n m
a
t
rix
a
nd
p
er
-
cla
s
s
m
et
rics
T
h
e
co
n
f
u
s
io
n
m
atr
ix
Fig
u
r
e
6
,
illu
s
tr
ates
th
e
class
if
icatio
n
r
esu
lts
f
o
r
k
er
atitis
an
d
u
v
eitis
.
T
ab
le
4
p
r
o
v
id
es
p
r
ec
is
io
n
,
r
ec
all,
a
n
d
F1
-
s
co
r
e
f
o
r
ea
c
h
class
,
s
h
o
win
g
b
alan
ce
d
an
d
r
eli
ab
le
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
e
co
n
f
u
s
io
n
m
atr
ix
in
Fig
u
r
e
5
d
em
o
n
s
tr
ates
th
e
clas
s
if
icatio
n
p
er
f
o
r
m
a
n
ce
o
f
th
e
p
r
o
p
o
s
ed
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
7
0
9
-
2
7
2
2
2718
m
o
d
el
o
n
th
e
two
class
es
:
k
er
atitis
an
d
u
v
eitis
.
C
o
r
r
ec
t
a
n
s
wer
s
lin
e
u
p
alo
n
g
th
e
d
iag
o
n
al,
f
o
r
m
i
n
g
a
clea
r
tr
ail
th
r
o
u
g
h
th
e
c
h
ar
t.
Fro
m
th
e
r
esu
lts
,
it
is
ev
id
en
t
th
at
th
e
m
ajo
r
ity
o
f
th
e
s
am
p
les
ar
e
class
if
ied
co
r
r
ec
tly
,
ac
h
iev
in
g
an
o
v
er
all
ac
cu
r
ac
y
o
f
9
8
%.
T
h
is
in
d
icate
s
th
at
th
e
m
o
d
el
is
h
ig
h
ly
ef
f
ec
ti
v
e
in
d
is
tin
g
u
is
h
in
g
b
etwe
en
k
er
atitis
an
d
u
v
eitis
with
m
in
im
al
m
is
class
if
icatio
n
.
Fig
u
r
e
6
.
C
o
n
f
u
s
io
n
m
atr
i
x
o
f
p
r
o
p
o
s
ed
m
o
d
el
o
n
th
e
test
s
et
T
ab
le
4
.
C
lass
-
wis
e
p
er
f
o
r
m
an
ce
m
etr
ics o
n
test
s
et
C
l
a
s
s
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
S
u
p
p
o
r
t
K
e
r
a
t
i
t
i
s
0
.
9
8
0
.
9
7
0
.
9
7
5
1
8
0
U
v
e
i
t
i
s
0
.
9
8
0
.
9
9
0
.
9
8
5
1
9
2
O
v
e
r
a
l
l
0
.
9
8
0
.
9
8
0
.
9
8
3
7
2
5
.
4
.
L
ea
rning
b
eha
v
io
r
Fig
u
r
es 7
an
d
8
illu
s
tr
ate
th
e
l
o
s
s
an
d
ac
cu
r
ac
y
p
r
o
g
r
ess
io
n
d
u
r
in
g
tr
ain
i
n
g
an
d
v
alid
atio
n
.
T
h
e
s
m
all
g
ap
b
etwe
en
tr
ain
i
n
g
a
n
d
v
alid
atio
n
cu
r
v
es
in
d
icate
s
m
in
im
al
o
v
e
r
f
itti
n
g
,
co
n
f
i
r
m
in
g
Vis
io
n
E
y
eNe
t’
s
r
o
b
u
s
tn
ess
.
Fig
u
r
es
7
a
n
d
8
s
h
o
w
th
at
th
e
m
o
d
el
co
n
v
er
g
ed
s
tead
ily
o
v
er
th
e
ep
o
ch
s
,
with
ac
cu
r
ac
y
in
cr
ea
s
in
g
an
d
lo
s
s
s
tead
ily
d
ec
r
ea
s
in
g
.
Valid
atio
n
lo
s
s
clo
s
ely
f
o
llo
wed
th
e
tr
ain
in
g
l
o
s
s
,
d
r
o
p
p
in
g
f
r
o
m
ar
o
u
n
d
1
.
2
t
o
u
n
d
er
0
.
2
,
i
n
d
ic
atin
g
co
n
s
is
ten
t
lear
n
in
g
with
m
in
im
al
o
v
er
f
itti
n
g
.
Acc
u
r
ac
y
f
o
r
b
o
th
tr
ain
in
g
an
d
ev
alu
atio
n
im
p
r
o
v
e
d
f
r
o
m
ab
o
u
t
6
0
%
to
o
v
er
9
0
%,
d
em
o
n
s
tr
atin
g
s
tr
o
n
g
g
en
er
ali
za
tio
n
an
d
ef
f
ec
tiv
e
f
ea
tu
r
e
ex
tr
ac
tio
n
s
u
itab
le
f
o
r
clin
ical
u
s
e.
5.
5
.
E
f
f
iciency
a
na
l
y
s
is
(
infe
re
nce
t
im
e)
T
h
e
av
er
a
g
e
p
r
ed
ictio
n
tim
e
p
er
im
ag
e
(
b
atc
h
s
ize
=1
)
o
n
an
NVI
DI
A
T
esla
V1
0
0
GP
U
(
3
2
GB
R
AM
)
was
m
ea
s
u
r
ed
to
d
eter
m
in
e
in
f
er
e
n
ce
tim
e.
E
ac
h
r
ep
o
r
ted
v
alu
e
r
e
p
r
esen
ts
th
e
m
ea
n
±
s
tan
d
ar
d
d
ev
iatio
n
ca
lcu
lated
o
v
e
r
1
0
0
in
d
ep
en
d
en
t
r
u
n
s
.
T
a
b
le
5
s
h
o
ws
in
f
er
en
ce
-
tim
e
co
m
p
ar
i
s
o
n
:
Vis
io
n
E
y
eNe
t
(
5
1
±
2
m
s
)
b
alan
ce
s
ac
cu
r
ac
y
an
d
ef
f
icien
cy
b
etwe
en
Mo
b
ileNetV2
(
4
9
m
s
)
an
d
Den
s
eNe
t1
2
1
(
5
5
m
s
)
,
s
u
p
p
o
r
tin
g
r
ea
l
-
tim
e
d
e
p
lo
y
m
en
t.
5
.
6
.
Da
t
a
s
et
lim
it
a
t
io
ns
a
nd
g
ener
a
liza
bil
it
y
Alth
o
u
g
h
Vis
io
n
E
y
eNe
t
ac
h
i
ev
ed
h
i
g
h
ac
c
u
r
ac
y
o
n
th
e
c
u
r
ated
Kag
g
le
d
ataset
o
f
1
,
8
6
0
s
lit
-
lam
p
im
ag
es
(
k
er
atitis
=9
0
0
,
u
v
eitis
=9
6
0
)
,
its
lim
ited
s
ize
an
d
s
in
g
le
-
s
o
u
r
ce
o
r
ig
in
m
a
y
co
n
s
tr
ai
n
g
en
er
aliza
b
ilit
y
.
T
h
e
im
ag
es
we
r
e
ca
p
t
u
r
ed
u
n
d
er
s
im
ilar
illu
m
in
ati
o
n
,
d
ev
i
ce
,
an
d
d
e
m
o
g
r
a
p
h
ic
co
n
d
itio
n
s
,
wh
ich
m
ay
n
o
t
r
ep
r
esen
t
th
e
v
a
r
iab
ilit
y
o
f
r
ea
l
-
wo
r
ld
clin
ical
d
ata.
As
a
r
es
u
lt,
m
o
d
el
p
e
r
f
o
r
m
an
ce
ac
r
o
s
s
d
if
f
e
r
en
t
h
o
s
p
itals
,
im
ag
in
g
s
y
s
tem
s
,
o
r
p
o
p
u
la
tio
n
s
r
em
ain
s
to
b
e
v
er
if
ied
.
T
o
r
ed
u
ce
o
v
er
f
itti
n
g
an
d
im
p
r
o
v
e
in
te
r
n
al
g
en
er
aliza
tio
n
,
we
ap
p
lied
e
x
ten
s
iv
e
d
ata
au
g
m
en
tatio
n
an
d
f
iv
e
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
,
co
n
s
is
ten
t
with
m
eth
o
d
s
u
s
ed
in
r
ec
e
n
t
o
p
h
th
a
lm
ic
-
AI
s
tu
d
ies
[
1
]
,
[
3
]
,
[
9
]
,
[
1
0
]
,
[
2
2
]
.
Fu
tu
r
e
wo
r
k
will
f
o
cu
s
o
n
co
n
s
tr
u
ctin
g
a
m
u
lti
-
ce
n
ter
,
d
e
v
ice
-
d
iv
er
s
e
d
ataset
in
co
r
p
o
r
atin
g
s
am
p
les
f
r
o
m
v
ar
ied
clin
ical
en
v
ir
o
n
m
en
ts
.
E
x
te
r
n
al
v
alid
atio
n
will
also
b
e
p
er
f
o
r
m
ed
u
s
in
g
in
d
e
p
en
d
e
n
t
co
h
o
r
ts
,
s
u
ch
as
th
o
s
e
r
ep
o
r
ted
b
y
On
g
et
a
l.
[
2
2
]
an
d
So
leim
an
i
et
a
l.
[
2
3
]
,
to
ass
ess
r
o
b
u
s
tn
ess
ac
r
o
s
s
h
eter
o
g
e
n
e
o
u
s
im
ag
in
g
c
o
n
d
itio
n
s
an
d
p
a
tien
t p
o
p
u
latio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.