I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Ro
bo
t
ics a
nd
Aut
o
m
a
t
io
n
(
I
J
RA)
Vo
l.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
,
p
p
.
2
3
4
~
2
4
6
I
SS
N:
2722
-
2
5
8
6
,
DOI
:
1
0
.
1
1
5
9
1
/i
jr
a
.
v
1
5
i
1
.
pp
2
3
4
-
246
234
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
r
a
.
ia
esco
r
e.
co
m
Edg
eRet
ina
:
H
y
b
rid multi
media
ar
chitec
ture
for dia
betic
retinopa
thy scre
e
ning
on lo
w
-
co
st
mo
biles
G
uid
o
um
Am
ina
1
,
Acho
ur
So
lt
a
na
1
,
M
a
a
m
a
r
B
o
ug
hera
r
a
1
,
Am
a
r
a
Ra
f
i
k
1
,
M
ha
m
ed
T
a
y
eb
2
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
,
H
i
g
h
e
r
N
o
r
m
a
l
S
c
h
o
o
l
o
f
K
o
u
b
a
,
A
l
g
i
e
r
s,
A
l
g
e
r
i
a
2
I
n
st
i
t
u
t
e
B
i
o
ma
t
e
r
i
a
l
s
a
n
d
Tr
a
n
s
p
o
r
t
P
h
e
n
o
m
e
n
a
La
b
o
r
a
t
o
r
y
,
M
e
d
e
a
U
n
i
v
e
r
si
t
y
,
M
e
d
e
a
,
A
l
g
e
r
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
l 2
6
,
2
0
2
5
R
ev
is
ed
Dec
4
,
2
0
2
5
Acc
ep
ted
Feb
9
,
2
0
2
6
Dia
b
e
ti
c
re
ti
n
o
p
a
th
y
(DR)
is
a
m
a
jo
r
c
a
u
se
o
f
p
re
v
e
n
tab
le
b
li
n
d
n
e
ss
,
p
a
rti
c
u
larly
i
n
a
re
a
s
with
li
m
i
ted
m
e
d
ica
l
re
so
u
rc
e
s
wh
e
re
a
c
c
e
s
s
to
o
p
h
th
a
lmo
l
o
g
ists
is
c
rit
ica
l.
E
x
ist
in
g
a
u
t
o
m
a
ted
so
l
u
ti
o
n
s
stru
g
g
le
t
o
b
a
lan
c
e
c
li
n
ica
l
p
e
rf
o
rm
a
n
c
e
,
c
o
st
-
e
ffe
c
ti
v
e
n
e
ss
,
a
n
d
r
o
b
u
st
n
e
ss
in
t
h
e
fa
c
e
o
f
fu
n
d
u
s
ima
g
e
v
a
riab
il
i
ty
—
i
n
c
lu
d
in
g
li
g
h
ti
n
g
d
iffere
n
c
e
s,
a
rti
fa
c
ts,
a
n
d
u
n
e
v
e
n
c
a
p
tu
re
q
u
a
li
t
y
.
To
a
d
d
re
ss
th
i
s
c
h
a
ll
e
n
g
e
,
we
p
ro
p
o
se
E
d
g
e
Re
ti
n
a
,
a
n
in
teg
ra
ted
s
o
lu
ti
o
n
fo
r
d
iab
e
ti
c
r
e
ti
n
o
p
a
th
y
sc
re
e
n
i
n
g
o
n
l
o
w
-
c
o
s
t
m
o
b
il
e
s
.
Ou
r
a
p
p
ro
a
c
h
c
o
m
b
in
e
s
li
g
h
t
we
ig
h
t
p
re
p
r
o
c
e
ss
in
g
(
1
2
8
×
1
2
8
re
siz
in
g
,
in
ten
sit
y
n
o
rm
a
li
z
a
ti
o
n
,
a
n
d
targ
e
ted
a
u
g
m
e
n
tati
o
n
s
sim
u
lati
n
g
r
e
a
l
-
wo
rld
c
o
n
d
i
ti
o
n
s)
wit
h
a
h
y
b
ri
d
S
q
u
e
e
z
e
Ne
t
-
M
o
b
il
e
ViT
a
rc
h
it
e
c
tu
re
(1
.
4
m
il
l
io
n
p
a
ra
m
e
ters
),
o
p
ti
m
ize
d
b
y
d
y
n
a
m
ic
th
re
sh
o
ld
c
a
li
b
ra
ti
o
n
(m
e
d
ian
:
0
.
3
)
,
m
a
x
imiz
in
g
c
li
n
ica
l
u
ti
li
t
y
.
C
li
n
i
c
a
ll
y
c
a
li
b
ra
ted
INT8
q
u
a
n
ti
z
a
ti
o
n
re
d
u
c
e
s
th
e
m
o
d
e
l
t
o
8
.
2
7
M
B
(
-
9
2
%
)
with
o
u
t
a
lt
e
rin
g
d
iag
n
o
stic
p
e
rfo
rm
a
n
c
e
(se
n
siti
v
it
y
o
f
9
0
.
7
%
f
o
r
re
fe
ra
b
l
e
d
iab
e
ti
c
re
ti
n
o
p
a
th
ies
),
w
h
il
e
p
re
se
rv
in
g
c
o
m
p
a
ti
b
il
i
ty
wit
h
flo
a
ti
n
g
p
o
i
n
t
3
2
(F
P
3
2
)
-
b
a
se
d
g
ra
d
ie
n
t
-
we
ig
h
ted
c
las
s
a
c
ti
v
a
ti
o
n
m
a
p
p
i
n
g
(G
ra
d
-
CAM)
v
is
u
a
li
z
a
ti
o
n
s.
E
v
a
lu
a
ted
o
n
t
h
e
APT
OS
2
0
1
9
d
a
tas
e
t,
t
h
is
so
lu
ti
o
n
a
c
h
iev
e
s
a
n
AU
C
o
f
0
.
9
6
with
a
late
n
c
y
(
in
fe
re
n
c
e
ti
m
e
)
o
f
1
5
.
4
3
m
s,
re
d
u
c
in
g
CP
U
c
o
n
su
m
p
ti
o
n
b
y
4
3
%
c
o
m
p
a
re
d
t
o
F
P
3
2
.
Th
e
d
y
n
a
m
ic
th
re
sh
o
l
d
/INT8
c
o
u
p
li
n
g
d
e
c
re
a
se
s
f
a
lse
p
o
siti
v
e
s
b
y
7
1
.
4
%
.
Th
is
p
ip
e
li
n
e
t
h
u
s
e
n
a
b
les
a
c
c
u
ra
te,
a
c
c
e
ss
ib
le,
a
n
d
e
a
rly
sc
r
e
e
n
in
g
o
f
d
iab
e
ti
c
re
ti
n
o
p
a
t
h
y
o
n
l
o
w
-
c
o
s
t
m
o
b
il
e
d
e
v
ice
s,
c
o
m
b
in
i
n
g
o
p
e
ra
ti
o
n
a
l
e
fficie
n
c
y
a
n
d
d
iag
n
o
stic
re
li
a
b
il
it
y
i
n
c
o
n
stra
i
n
e
d
e
n
v
ir
o
n
m
e
n
ts,
wh
ich
is
c
ru
c
ial
to
p
re
v
e
n
t
a
v
o
i
d
a
b
le
b
li
n
d
n
e
ss
.
K
ey
w
o
r
d
s
:
C
o
m
p
r
ess
io
n
Dy
n
am
ic
th
r
esh
o
l
d
E
d
g
eRetin
a
L
o
w
-
co
s
t m
o
b
iles
P
r
ep
r
o
ce
s
s
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
:
Gu
id
o
u
m
Am
in
a
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
,
Hig
h
er
N
o
r
m
al
Sch
o
o
l o
f
Ko
u
b
a
Alg
ier
s
,
Alg
er
ia
E
m
ail:
am
in
a.
g
u
id
o
u
m
@
g
.
en
s
-
k
o
u
b
a.
d
z
1.
I
NT
RO
D
UCT
I
O
N
Diab
etic
r
etin
o
p
ath
y
(
DR
)
is
a
d
ev
astatin
g
co
m
p
licatio
n
o
f
d
iab
etes,
af
f
ec
tin
g
m
o
r
e
t
h
an
1
4
0
m
illi
o
n
p
eo
p
le
wo
r
ld
wid
e
a
n
d
c
o
n
s
titu
tin
g
th
e
lead
in
g
ca
u
s
e
o
f
p
r
ev
en
tab
le
b
li
n
d
n
ess
in
w
o
r
k
i
n
g
-
ag
e
ad
u
lts
.
T
h
is
p
r
o
b
lem
is
p
ar
ticu
lar
ly
ac
u
te
in
r
eso
u
r
ce
-
lim
ited
s
ettin
g
s
,
wh
er
e
ac
ce
s
s
to
o
p
h
th
alm
o
lo
g
is
ts
is
cr
itical
(
<1
s
p
ec
ialis
t
p
er
1
m
illi
o
n
in
h
ab
it
an
ts
in
s
u
b
-
Sah
ar
an
Af
r
ica)
an
d
wh
er
e
8
0
%
o
f
d
ia
b
etes
-
r
elate
d
b
lin
d
n
ess
co
u
l
d
b
e
p
r
ev
en
te
d
th
r
o
u
g
h
ea
r
l
y
d
etec
tio
n
[
1
]
.
Als
o
co
n
s
id
er
ed
as
a
co
m
m
o
n
co
m
p
licatio
n
o
f
d
iab
etes
m
ellitu
s
,
m
an
if
esti
n
g
as
r
etin
al
d
a
m
ag
e
th
at
ca
n
lead
to
v
is
io
n
lo
s
s
.
L
ate
d
iag
n
o
s
is
o
f
t
h
is
co
n
d
itio
n
ca
n
lea
d
to
b
lin
d
n
ess
,
wh
ile
ea
r
ly
d
etec
tio
n
an
d
tr
ea
tm
e
n
t c
an
s
ig
n
if
ica
n
tly
r
ed
u
ce
th
e
r
is
k
o
f
ir
r
ev
er
s
i
b
le
v
is
io
n
d
am
a
g
e.
Ma
n
u
al
d
ia
g
n
o
s
is
o
f
f
u
n
d
u
s
im
ag
es
b
y
o
p
h
th
al
m
o
lo
g
is
t
s
is
o
f
ten
c
o
s
tly
,
tim
e
-
c
o
n
s
u
m
in
g
,
a
n
d
d
em
an
d
in
g
,
u
n
lik
e
co
m
p
u
ter
-
aid
ed
d
iag
n
o
s
tic
s
y
s
tem
s
.
I
n
r
ec
en
t
y
ea
r
s
,
m
ed
ical
im
ag
e
an
aly
s
is
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
E
d
g
eR
etin
a
:
Hyb
r
id
mu
ltime
d
ia
a
r
ch
itectu
r
e
fo
r
d
ia
b
etic
r
etin
o
p
a
th
y
s
creen
in
g
…
(
Gu
id
o
u
m
A
min
a
)
235
class
if
icatio
n
h
av
e
s
ee
n
m
ajo
r
ad
v
an
ce
s
th
an
k
s
to
d
ee
p
le
ar
n
in
g
.
Am
o
n
g
th
ese
ap
p
r
o
ac
h
es,
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
h
av
e
em
er
g
ed
as
th
e
s
tan
d
ar
d
to
o
ls
f
o
r
m
ed
ical
im
a
g
e
a
n
aly
s
is
,
p
r
o
v
id
in
g
au
to
m
ated
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
class
if
icatio
n
ca
p
ab
ilit
ies.
As
h
ig
h
lig
h
ted
in
[
2
]
,
m
u
ltime
d
ia
p
r
o
ce
s
s
in
g
n
atu
r
ally
i
n
teg
r
ates
im
ag
e
p
r
o
ce
s
s
in
g
a
s
o
n
e
o
f
its
f
u
n
d
am
e
n
tal
co
m
p
o
n
en
ts
.
T
h
eir
wo
r
k
d
e
m
o
n
s
tr
ates
th
at
d
ee
p
lear
n
in
g
tech
n
iq
u
es
ar
e
n
o
w
th
e
p
r
ef
er
r
e
d
ap
p
r
o
ac
h
f
o
r
im
ag
e
p
r
o
ce
s
s
in
g
in
th
e
m
u
ltime
d
ia
f
ield
,
en
a
b
lin
g
s
ig
n
if
ican
t
ad
v
a
n
ce
s
in
im
ag
e
r
ec
o
g
n
itio
n
,
s
em
an
tic
s
eg
m
en
tatio
n
an
d
im
ag
e
s
y
n
th
esis
.
T
h
is
co
n
ce
p
tu
al
h
ier
ar
ch
y
-
wh
e
r
e
d
ee
p
lear
n
i
n
g
is
p
ar
t
o
f
im
ag
e
p
r
o
ce
s
s
in
g
,
its
elf
a
b
r
an
ch
o
f
m
u
ltime
d
ia
p
r
o
ce
s
s
in
g
-
estab
lis
h
es
th
e
th
eo
r
etica
l
f
r
am
ewo
r
k
with
in
wh
ich
o
u
r
E
d
g
eRetin
a
ap
p
r
o
ac
h
f
o
r
d
ia
b
etic
r
etin
o
p
ath
y
s
cr
ee
n
in
g
is
e
m
b
ed
d
e
d
.
E
x
is
tin
g
AI
s
o
lu
tio
n
s
s
u
f
f
er
f
r
o
m
two
cr
itical
lim
itatio
n
s
:
o
v
er
-
p
ar
am
eter
ized
ar
c
h
itectu
r
es
(
>1
0
M
p
ar
am
eter
s
)
[
3
]
,
[
4
]
r
e
q
u
ir
in
g
co
m
p
lex
p
r
e
-
p
r
o
ce
s
s
in
g
in
co
m
p
atib
le
with
m
o
b
ile
d
ev
ices
,
an
d
in
ap
p
r
o
p
r
iate
clin
ical
tr
ad
e
-
o
f
f
s
g
en
er
atin
g
eith
er
ex
ce
s
s
iv
e
f
alse
p
o
s
itiv
es
(
o
v
er
d
iag
n
o
s
is
)
o
r
d
an
g
er
o
u
s
f
alse
n
eg
ativ
es
(
u
n
d
er
-
d
etec
tio
n
o
f
r
ef
er
a
b
le
ca
s
es).
T
o
ad
d
r
ess
th
is
d
u
al
c
h
allen
g
e,
we
p
r
o
p
o
s
e
E
d
g
eRetin
a,
an
in
teg
r
ate
d
f
r
am
ewo
r
k
co
m
b
in
in
g
:
a
m
i
n
im
alis
t
p
r
e
-
p
r
o
ce
s
s
in
g
(
1
2
8
×1
2
8
r
esizin
g
,
in
ten
s
ity
n
o
r
m
aliza
tio
n
,
tar
g
eted
au
g
m
en
tatio
n
s
s
im
u
latin
g
f
i
eld
co
n
s
tr
ain
ts
)
,
a
h
y
b
r
id
S
q
u
ee
ze
Net
-
Mo
b
ileViT
ar
ch
it
ec
tu
r
e
(
1
.
4
m
illi
o
n
p
ar
am
eter
s
)
co
m
b
in
i
n
g
lo
ca
l c
o
m
p
r
ess
io
n
an
d
co
n
tex
tu
al
m
o
d
elin
g
,
an
d
a
d
y
n
am
ic
ca
lib
r
at
io
n
o
f
th
e
d
ec
is
io
n
th
r
esh
o
ld
(
0
.
3
)
o
p
tim
izin
g
cli
n
ical
u
tili
ty
v
ia
m
ed
ically
v
a
lid
ated
I
NT
8
(
8
b
its
in
teg
er
)
q
u
an
tizatio
n
[
2
0
]
.
E
v
alu
ated
o
n
th
e
APTO
S
2
0
1
9
b
e
n
ch
m
ar
k
,
o
u
r
s
o
lu
tio
n
d
e
m
o
n
s
tr
ates
ex
ce
p
tio
n
al
p
e
r
f
o
r
m
an
ce
(
AUC=0
.
9
6
,
laten
cy
=1
5
.
4
3
m
s
)
with
a
s
en
s
itiv
ity
o
f
9
0
.
7
%
f
o
r
r
ef
er
ab
l
e
DR
s
an
d
a
7
1
.
4
%
r
ed
u
ctio
n
in
f
alse
p
o
s
itiv
es,
wh
ile
en
ab
lin
g
lo
w
-
c
o
s
t
m
o
b
ile
d
ep
l
o
y
m
en
t
th
an
k
s
to
m
o
d
el
c
o
m
p
r
ess
io
n
to
8
.
2
7
MB
(
-
9
2
%),
a
4
3
%
r
ed
u
ctio
n
in
C
PU
co
n
s
u
m
p
tio
n
,
an
d
th
e
p
r
eser
v
atio
n
o
f
g
r
a
d
ien
t
-
weig
h
ted
class
ac
tiv
atio
n
m
ap
p
in
g
(
Gr
ad
-
C
AM
)
ex
p
lain
ab
le
v
is
u
aliza
ti
o
n
s
f
lo
atin
g
p
o
in
t 3
2
(
FP
3
2
)
.
T
h
is
ap
p
r
o
ac
h
v
alid
ates a
n
ew
s
cr
ee
n
in
g
p
ar
ad
ig
m
ac
ce
s
s
ib
le
f
o
r
co
n
s
tr
ain
ed
e
n
v
ir
o
n
m
en
ts
,
p
a
v
in
g
t
h
e
way
f
o
r
m
ass
iv
e
p
r
ev
en
tio
n
ca
m
p
aig
n
s
.
Gau
r
et
a
l
.
[
5
]
p
r
o
p
o
s
e
d
an
au
to
m
ated
m
eth
o
d
f
o
r
d
etec
tin
g
d
iab
etic
DR
an
d
class
if
y
in
g
it
s
d
if
f
er
en
t
s
tag
es
f
r
o
m
r
etin
al
im
a
g
es
u
s
in
g
a
C
NN,
m
o
r
e
p
r
ec
is
ely
th
e
Den
s
eNe
t1
6
9
ar
ch
itectu
r
e.
T
h
e
m
o
d
el
is
tr
ain
e
d
o
n
th
e
APTO
S
2
0
1
9
d
ataset
co
n
tain
in
g
a
p
p
r
o
x
im
ately
1
3
,
0
0
0
f
u
n
d
u
s
im
ag
es
an
n
o
tated
ac
co
r
d
in
g
t
o
f
iv
e
lev
els
o
f
DR
s
ev
er
ity
(
f
r
o
m
0
:
n
o
DR
to
4
:
p
r
o
life
r
ativ
e
D
R
)
.
T
h
e
im
ag
es
u
n
d
er
g
o
p
r
e
p
r
o
ce
s
s
in
g
(
r
esizin
g
,
n
o
r
m
aliza
tio
n
,
e
n
co
d
i
n
g
)
b
ef
o
r
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
an
d
class
if
icatio
n
v
ia
Den
s
eNe
t1
6
9
.
T
h
e
m
o
d
el
ac
h
iev
es
an
ac
cu
r
ac
y
o
f
8
2
%
f
o
r
class
if
icatio
n
in
to
f
iv
e
class
es,
an
d
9
8
%
f
o
r
b
in
ar
y
d
etec
tio
n
(
p
r
e
s
en
ce
o
r
ab
s
en
ce
o
f
DR
)
,
o
u
tp
er
f
o
r
m
in
g
o
th
er
m
et
h
o
d
s
s
u
ch
as c
lass
ical
C
NN
an
d
XGBo
o
s
t.
Su
s
h
ith
et
a
l
.
[
6
]
p
r
o
p
o
s
e
d
a
n
o
v
el
ap
p
r
o
ac
h
f
o
r
th
e
ea
r
ly
d
etec
tio
n
o
f
DR
,
d
esig
n
in
g
a
h
y
b
r
id
d
ee
p
lear
n
in
g
m
o
d
el,
ca
lled
tem
p
o
r
al
awa
r
e
h
y
b
r
id
d
ee
p
lear
n
i
n
g
(
T
AHDL
)
,
co
m
b
in
in
g
C
NNs
f
o
r
s
p
atial
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
r
ec
u
r
r
e
n
t
n
e
u
r
al
n
etwo
r
k
s
(
R
NNs)
—
wit
h
atten
tio
n
m
ec
h
an
is
m
s
—
to
an
aly
ze
tem
p
o
r
al
d
ep
en
d
e
n
cies
b
etwe
en
s
u
cc
e
s
s
iv
e
r
etin
al
im
ag
es.
T
h
is
m
o
d
el
lev
e
r
ag
es
th
e
tem
p
o
r
al
p
r
o
g
r
ess
io
n
o
f
th
e
d
is
ea
s
e
to
m
o
r
e
ac
cu
r
ately
d
et
ec
t
th
e
ea
r
ly
s
ig
n
s
o
f
DR
.
T
h
e
u
s
e
o
f
ad
v
an
ce
d
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es
s
u
ch
as
co
lo
r
ad
ap
tiv
e
h
is
to
g
r
am
eq
u
aliza
tio
n
(
C
L
AHE
)
,
n
o
r
m
aliza
tio
n
,
an
d
d
ata
au
g
m
e
n
tatio
n
h
elp
s
im
p
r
o
v
e
d
etec
tio
n
q
u
ality
.
T
h
e
m
o
d
el
was
ev
alu
ated
u
s
in
g
p
u
b
lic
b
en
ch
m
ar
k
d
atasets
,
in
clu
d
in
g
DR
I
VE
,
Kag
g
le
Diab
etic
R
etin
o
p
ath
y
,
an
d
E
y
ePAC
S,
an
d
d
em
o
n
s
tr
ate
d
s
u
p
er
io
r
ac
c
u
r
ac
y
c
o
m
p
ar
ed
to
co
n
v
en
tio
n
al
ap
p
r
o
ac
h
es
(
C
NN,
VGG1
9
,
I
n
ce
p
tio
n
V3
,
Mo
b
ileNetV3
,
an
d
ViT
)
,
ac
h
iev
in
g
u
p
to
9
7
.
5
%
ac
cu
r
ac
y
o
n
DR
I
VE
,
9
4
.
0
4
% o
n
Kag
g
le,
a
n
d
9
6
.
9
% o
n
E
y
ePAC
S.
Nan
d
h
in
i
et
a
l
.
[
7
]
p
r
o
p
o
s
e
d
a
m
eth
o
d
f
o
r
d
etec
tin
g
d
iab
et
ic
r
etin
o
p
ath
y
b
ased
o
n
a
m
o
d
el
ca
lled
DiaNe
t
m
o
d
el
(
DNM
)
.
Du
r
in
g
th
e
p
r
ep
r
o
ce
s
s
in
g
s
tag
e,
a
Gab
o
r
f
ilter
is
u
s
ed
to
en
h
an
ce
th
e
v
is
ib
ilit
y
o
f
b
lo
o
d
v
ess
els
in
r
etin
al
im
ag
es.
T
h
is
f
ilter
al
s
o
co
n
tr
ib
u
tes
to
tex
tu
r
e
an
aly
s
is
,
o
b
ject
r
ec
o
g
n
itio
n
,
f
ea
t
u
r
e
ex
tr
ac
tio
n
,
an
d
im
ag
e
co
m
p
r
e
s
s
io
n
.
Du
r
in
g
th
e
d
ata
au
g
m
e
n
tatio
n
s
tag
e,
th
e
in
p
u
t
d
im
en
s
io
n
s
o
f
th
e
d
ataset
ar
e
r
ed
u
ce
d
u
s
in
g
p
r
in
cip
al
co
m
p
o
n
e
n
t
an
aly
s
is
(
PC
A)
,
wh
i
ch
r
ed
u
ce
s
th
e
n
u
m
b
er
o
f
f
ea
t
u
r
es
to
b
e
p
r
o
ce
s
s
ed
with
o
u
t c
o
m
p
r
o
m
is
in
g
m
o
d
el
p
er
f
o
r
m
an
ce
.
An
av
er
a
g
e
ac
cu
r
ac
y
o
f
9
0
.
0
2
% wa
s
ac
h
iev
ed
.
Sap
r
o
o
et
a
l
.
[
8
]
p
r
o
p
o
s
e
d
a
d
ee
p
lear
n
i
n
g
b
in
ar
y
class
if
ic
atio
n
s
y
s
tem
,
b
ased
o
n
tr
an
s
f
er
lear
n
in
g
,
f
o
r
ea
r
ly
d
etec
tio
n
o
f
DR
o
n
r
etin
al
im
ag
es.
I
t
co
m
b
in
es
th
r
ee
r
o
b
u
s
t
d
atab
ases
(
E
y
ePAC
S,
I
DR
iD,
APTO
S
-
2
0
1
9
)
an
n
o
tated
b
y
o
p
h
t
h
alm
o
lo
g
is
ts
an
d
ap
p
lies
p
r
ep
r
o
ce
s
s
in
g
in
clu
d
in
g
d
en
o
is
in
g
,
n
o
r
m
aliza
tio
n
an
d
d
ata
au
g
m
en
tatio
n
to
im
p
r
o
v
e
r
o
b
u
s
tn
ess
.
Af
ter
ev
alu
atin
g
2
0
p
r
e
-
tr
ain
ed
n
etwo
r
k
s
(
Ser
ial,
DAG,
lig
h
tweig
h
t
ca
teg
o
r
ies)
v
ia
c
o
m
p
r
e
h
en
s
iv
e
m
etr
ics
(
ac
cu
r
ac
y
,
s
en
s
itiv
ity
,
an
d
AUC
-
R
OC
)
,
it
d
em
o
n
s
tr
ates
th
at
th
e
R
esNet1
0
1
(
DAG)
m
o
d
el
ac
h
i
ev
es a
g
o
o
d
r
esu
lt
.
B
ask
ar
et
a
l
.
[
9
]
s
h
o
w
ed
th
a
t
d
iab
etic
r
etin
o
p
ath
y
,
a
p
r
ev
alen
t
o
cu
lar
p
ath
o
lo
g
y
a
f
f
ec
ti
n
g
r
etin
al
v
ess
els
in
d
iab
etics,
af
f
ec
ts
a
p
p
r
o
x
im
ately
3
.
9
m
illi
o
n
p
e
o
p
le
wo
r
ld
wid
e,
h
ig
h
lig
h
tin
g
th
e
u
r
g
en
t
n
ee
d
f
o
r
ea
r
ly
d
iag
n
o
s
is
f
o
r
ef
f
ec
tiv
e
m
an
ag
em
en
t.
I
t
d
em
o
n
s
tr
ates
h
o
w
d
ee
p
lear
n
in
g
,
p
ar
ticu
lar
l
y
tr
an
s
f
er
lear
n
in
g
,
ca
n
class
if
y
th
e
f
iv
e
s
tag
es
o
f
th
e
d
is
ea
s
e
(
n
o
r
m
al,
m
ild
,
m
o
d
er
ate,
s
ev
er
e
an
d
p
r
o
life
r
ativ
e)
v
ia
Alex
Net
a
n
d
Den
s
eNe
t
-
1
6
9
ar
ch
itectu
r
es tr
ain
ed
o
n
t
h
e
APTO
S2
0
1
9
an
d
d
iab
etic
r
etin
o
p
ath
y
co
m
p
etitio
n
d
atab
ases
.
Af
ter
f
in
e
-
tu
n
in
g
o
n
2
0
,
1
6
3
im
ag
es
(
9
,
0
0
0
n
o
r
m
al,
2
,
8
0
8
m
ild
,
6
,
2
8
7
m
o
d
er
ate,
1
,
0
6
5
s
ev
er
e,
1
,
0
0
3
p
r
o
life
r
ativ
e)
an
d
v
alid
atio
n
o
n
2
,
0
1
7
im
a
g
es
(
9
0
0
n
o
r
m
al,
2
8
1
m
ild
,
6
2
9
m
o
d
er
ate,
1
0
7
s
ev
er
e,
1
0
0
p
r
o
life
r
ativ
e
)
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2
5
8
6
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
:
234
-
246
236
Den
s
eNe
t
-
1
6
9
p
r
o
v
ed
s
u
p
e
r
io
r
,
with
F1
-
s
co
r
es
o
f
0
.
5
5
(
n
o
r
m
al)
,
0
.
3
8
(
m
ild
)
,
0
.
4
0
(
m
o
d
e
r
ate)
,
0
.
6
0
(
s
ev
er
e)
,
an
d
0
.
6
9
(
p
r
o
life
r
ativ
e)
.
Ku
m
ar
et
a
l
.
[
10
]
p
r
o
p
o
s
e
d
an
a
p
p
r
o
ac
h
co
m
b
i
n
in
g
two
ad
v
an
ce
d
d
ee
p
lear
n
i
n
g
a
r
ch
itectu
r
es:
I
n
ce
p
tio
n
V3
a
n
d
R
esNet5
0
.
T
h
e
I
n
ce
p
tio
n
V3
m
o
d
el
en
ab
les
m
u
lti
-
s
ca
le
f
ea
tu
r
e
ex
tr
ac
tio
n
,
m
ak
in
g
it
p
ar
ticu
lar
ly
ef
f
ec
tiv
e
at
d
ete
ctin
g
b
o
th
s
m
all
lesi
o
n
s
s
u
c
h
as
m
icr
o
an
eu
r
y
s
m
s
an
d
la
r
g
e
ab
n
o
r
m
alities
.
R
esNet5
0
,
f
o
r
its
p
ar
t,
f
ac
ilit
a
tes
th
e
tr
ain
in
g
o
f
d
ee
p
n
etwo
r
k
s
b
y
a
v
o
id
in
g
th
e
v
a
n
is
h
in
g
g
r
ad
ien
t
p
r
o
b
lem
th
r
o
u
g
h
s
k
ip
p
in
g
co
n
n
ec
tio
n
s
b
etwe
en
lay
er
s
.
T
h
e
r
esu
lts
in
d
icate
th
at
I
n
ce
p
tio
n
V3
ac
h
ie
v
es
an
ac
cu
r
ac
y
o
f
9
5
%,
wh
ile
R
esNet5
0
ac
h
iev
es
9
4
%,
th
ese
m
etr
ics
p
r
o
v
id
e
a
co
m
p
r
eh
e
n
s
iv
e
ass
ess
m
en
t
o
f
th
e
m
o
d
els'
d
iag
n
o
s
tic
ca
p
ab
ilit
ies ac
r
o
s
s
d
if
f
er
en
t
p
er
f
o
r
m
an
ce
d
im
en
s
io
n
s
.
Dasar
i
et
a
l
.
[
11
]
f
o
cu
s
e
d
o
n
t
h
e
au
to
m
atic
ass
ess
m
en
t
o
f
th
e
s
ev
er
ity
o
f
d
iab
etic
r
etin
o
p
at
h
y
u
s
in
g
a
tr
an
s
f
er
lear
n
i
n
g
a
p
p
r
o
ac
h
.
Sp
ec
if
ically
,
a
p
r
e
-
tr
ain
ed
an
d
th
en
f
in
e
-
t
u
n
ed
R
esNet5
0
m
o
d
el
was
ap
p
lied
t
o
th
e
APTO
S2
0
1
9
d
ataset,
tak
in
g
i
n
to
ac
co
u
n
t
ch
allen
g
es
r
elate
d
to
m
ed
ical
an
n
o
tatio
n
a
n
d
p
r
iv
ac
y
c
o
n
s
tr
ain
ts
.
T
is
h
a
et
a
l
.
[
1
2
]
p
r
o
p
o
s
ed
a
n
o
p
tim
ized
m
o
d
el
f
o
r
t
h
e
ac
c
u
r
ate
class
if
icatio
n
o
f
r
etin
al
d
is
o
r
d
er
s
,
in
cl
u
d
in
g
d
iab
etic
r
etin
o
p
at
h
y
-
a
c
o
m
m
o
n
co
m
p
licatio
n
o
f
d
iab
etes
m
ellitu
s
th
at
ca
u
s
es
p
o
ten
tially
b
lin
d
in
g
r
etin
al
lesi
o
n
s
in
th
e
ab
s
en
ce
o
f
ad
eq
u
ate
s
cr
ee
n
i
n
g
an
d
tr
ea
tm
en
t.
T
o
th
is
e
n
d
,
th
ey
d
e
v
el
o
p
ed
an
ap
p
r
o
ac
h
co
m
b
in
in
g
d
ee
p
lear
n
in
g
an
d
atten
tio
n
m
ec
h
an
is
m
s
.
Sev
er
al
p
r
e
-
tr
ain
e
d
ar
ch
itectu
r
es
(
R
esNet1
5
2
,
E
f
f
icien
tNetB
7
,
Mo
b
ileNetV3
L
ar
g
e)
wer
e
ev
alu
ated
b
y
th
e
au
th
o
r
s
o
n
a
d
iv
er
s
e
d
ataset
o
f
r
etin
al
im
ag
es.
T
h
eir
wo
r
k
d
em
o
n
s
tr
ates
th
at
th
e
i
m
p
r
o
v
e
d
v
er
s
io
n
o
f
R
esNet1
5
2
,
in
co
r
p
o
r
atin
g
a
s
p
atial
atten
tio
n
m
o
d
u
le,
ac
h
ie
v
es
(
ac
cu
r
ac
y
:
8
7
.
6
5
%,
p
r
ec
is
io
n
:
8
9
.
8
8
%,
r
ec
all:
9
0
.
6
9
%,
F1
-
s
co
r
e:
8
8
.
5
8
%).
T
h
e
au
th
o
r
s
h
ig
h
lig
h
t
t
h
e
co
m
p
etitiv
en
ess
o
f
th
eir
m
o
d
el
th
r
o
u
g
h
a
c
o
m
p
ar
ativ
e
an
aly
s
is
with
p
r
ev
io
u
s
wo
r
k
.
T
h
eir
r
esear
ch
co
n
tr
ib
u
tes
t
o
ad
v
a
n
ce
s
in
th
e
au
to
m
ated
d
iag
n
o
s
is
o
f
r
etin
al
p
ath
o
lo
g
ies an
d
co
u
l
d
im
p
r
o
v
e
p
atien
t c
ar
e
as we
ll a
s
d
iag
n
o
s
tic
ac
cu
r
ac
y
in
o
p
h
th
alm
o
lo
g
y
.
Ma
tth
ew
et
a
l
.
[
1
3
]
p
r
esen
ted
a
s
y
s
tem
u
s
in
g
m
ac
h
in
e
lear
n
in
g
was
d
ev
elo
p
e
d
to
class
if
y
d
iab
etic
r
etin
o
p
ath
y
u
s
in
g
tr
an
s
f
er
lear
n
in
g
,
b
ased
o
n
t
h
e
E
f
f
icien
tNet
-
B
0
m
o
d
el.
T
h
is
m
o
d
el
was
in
teg
r
ated
in
to
an
An
d
r
o
id
m
o
b
ile
ap
p
licatio
n
d
e
s
ig
n
ed
to
en
ab
le
h
ea
lth
ca
r
e
p
r
o
f
ess
io
n
als to
p
er
f
o
r
m
a
d
iag
n
o
s
is
u
s
in
g
a
s
im
p
le
s
m
ar
tp
h
o
n
e,
a
2
0
D
len
s
,
a
n
d
a
f
ew
b
asic
m
e
d
icatio
n
s
,
in
a
r
ea
s
lack
in
g
ad
e
q
u
ate
m
ed
ical
in
f
r
astru
ctu
r
e.
T
h
e
r
esu
lts
o
b
tain
ed
s
h
o
w
th
at
th
e
E
f
f
icien
tNet
-
B
0
m
o
d
el
ac
h
ie
v
es a
n
ac
cu
r
ac
y
o
f
9
1
.
8
5
% f
o
r
t
h
e
class
if
icatio
n
o
f
th
e
th
r
ee
d
is
ea
s
e
ca
teg
o
r
ies:
n
o
DR
,
n
o
n
-
p
r
o
life
r
ativ
e
DR
,
an
d
p
r
o
life
r
ativ
e
DR
.
Z
h
ao
et
a
l
.
[
1
4
]
p
r
esen
ts
a
r
ev
iew
o
f
m
eth
o
d
s
f
o
r
d
ep
l
o
y
in
g
d
ee
p
lear
n
in
g
o
n
m
o
b
il
e
d
ev
ices,
h
ig
h
lig
h
tin
g
th
eir
b
e
n
ef
its
in
ter
m
s
o
f
d
ata
p
r
iv
ac
y
an
d
o
p
er
atio
n
al
ef
f
icien
cy
.
I
t
d
escr
i
b
es
an
o
p
tim
izatio
n
p
ip
elin
e
co
m
b
in
in
g
m
o
d
el
-
o
r
i
en
ted
tech
n
iq
u
es (
s
u
ch
as q
u
a
n
tizatio
n
)
an
d
h
ar
d
war
e/so
f
twar
e
m
ec
h
an
is
m
s
.
C
r
itical
a
n
aly
s
i
s
:
No
p
r
ev
io
u
s
wo
r
k
h
as
co
m
b
in
e
d
d
y
n
a
m
ic
th
r
esh
o
ld
in
g
,
e
x
p
lain
ab
ilit
y
,
an
d
I
NT
8
q
u
an
tizatio
n
f
o
r
DR
.
ex
is
tin
g
s
o
lu
tio
n
s
ar
e
o
v
er
-
p
ar
a
m
eter
ized
,
r
eq
u
ir
e
co
m
p
lex
p
r
e
-
p
r
o
ce
s
s
in
g
,
an
d
lack
ap
p
r
o
p
r
iate
clin
ical
t
r
ad
e
-
o
f
f
s
.
W
e
n
o
w
ex
p
licitly
s
tate
t
h
a
t
E
d
g
eRetin
a
a
d
d
r
ess
es
th
ese
g
ap
s
t
h
r
o
u
g
h
th
e
in
n
o
v
ativ
e
in
teg
r
ati
o
n
o
f
lig
h
tweig
h
t
h
y
b
r
id
ar
ch
itectu
r
e,
d
y
n
am
ic
th
r
esh
o
l
d
ca
lib
r
atio
n
,
an
d
I
NT
8
q
u
an
tific
atio
n
,
s
p
ec
if
ically
d
esig
n
ed
f
o
r
lo
w
-
c
o
s
t m
o
b
ile
d
e
p
lo
y
m
en
t.
2.
M
E
T
H
O
D
I
n
th
is
s
tu
d
y
,
we
p
r
o
p
o
s
e
E
d
g
eRetin
a,
a
d
ee
p
lear
n
in
g
f
r
a
m
ewo
r
k
o
p
tim
ized
f
o
r
th
e
d
e
tectio
n
o
f
r
ef
er
ab
le
d
iab
etic
r
etin
o
p
ath
y
.
T
h
e
E
d
g
eRetin
a
f
r
am
ewo
r
k
in
teg
r
ates
lig
h
tweig
h
t
p
r
e
p
r
o
ce
s
s
in
g
,
a
h
y
b
r
id
Sq
u
ee
ze
Net
-
Mo
b
ileViT
ar
ch
it
ec
tu
r
e,
d
y
n
am
ic
th
r
esh
o
ld
ca
l
ib
r
atio
n
,
a
n
d
I
NT
8
q
u
a
n
tizatio
n
to
en
ab
le
r
ea
l
-
tim
e
d
iab
etic
r
etin
o
p
ath
y
s
cr
e
en
in
g
o
n
l
o
w
-
co
s
t
m
o
b
ile
d
e
v
ices
.
R
etin
al
im
ag
es
ar
e
ex
tr
ac
ted
f
r
o
m
th
e
p
u
b
lic
APTO
S
2
0
1
9
d
ataset
[
1
5
]
.
T
h
e
co
llected
im
ag
es
u
n
d
er
g
o
a
m
in
im
alis
t
p
r
ep
r
o
ce
s
s
in
g
in
clu
d
in
g
r
esizin
g
to
1
2
8
×1
2
8
p
ix
els,
in
te
n
s
ity
n
o
r
m
aliza
tio
n
,
a
n
d
ta
r
g
eted
au
g
m
en
tatio
n
(
r
an
d
o
m
h
o
r
izo
n
t
al
f
lip
s
an
d
±
1
0
%
co
n
tr
ast
v
ar
iatio
n
s
)
to
en
h
a
n
c
e
r
o
b
u
s
tn
ess
to
f
ield
co
n
d
itio
n
s
.
T
h
e
o
p
tim
ized
im
ag
es
ar
e
th
en
p
r
o
ce
s
s
ed
b
y
an
in
n
o
v
ativ
e
h
y
b
r
i
d
ar
ch
itectu
r
e
co
m
b
in
in
g
Sq
u
ee
ze
Net
Fire
m
o
d
u
les
(
ef
f
icien
t
l
o
ca
l
f
ea
tu
r
e
ex
tr
ac
tio
n
)
a
n
d
a
Mo
b
ileViT
b
lo
ck
(
co
n
tex
t
u
al
m
o
d
elin
g
b
y
atten
tio
n
m
ec
h
an
is
m
o
n
1
6
×1
6
p
atch
es).
Do
wn
s
tr
ea
m
,
a
d
y
n
am
ic
ca
lib
r
atio
n
o
f
th
e
d
ec
is
io
n
th
r
esh
o
ld
(
o
p
tim
ized
to
0
.
3
v
ia
F
1
-
s
co
r
e
m
a
x
im
izatio
n
)
a
n
d
a
m
ed
ically
v
alid
ated
I
NT
8
q
u
an
tizatio
n
ar
e
a
p
p
lie
d
to
en
ab
le
em
b
ed
d
ed
d
ep
l
o
y
m
en
t.
T
h
e
co
m
p
lete
f
lo
w
o
f
th
e
E
d
g
eRetin
a
m
eth
o
d
o
l
o
g
y
is
illu
s
tr
ated
in
F
ig
u
r
e
1
.
2
.
1
.
Da
t
a
s
et
a
nd
mo
del a
rc
hite
ct
ure
T
h
e
d
ataset
u
s
ed
is
APTO
S
2
0
1
9
c
o
n
tain
s
3
,
6
6
2
r
etin
al
im
ag
es
an
n
o
tated
b
y
o
p
h
th
alm
o
lo
g
ical
ex
p
er
ts
ac
co
r
d
in
g
to
th
e
E
T
D
R
S
s
ev
er
ity
s
ca
le:
C
lass
0
:
N
o
d
iab
etic
r
etin
o
p
ath
y
(
DR
)
,
C
las
s
1
:
Mild
D
R
,
C
las
s
2
: M
o
d
er
ate
DR
,
C
lass
3
: Sev
er
e
DR
,
C
lass
4
: Pr
o
life
r
ativ
e
DR
.
T
o
ad
a
p
t
to
th
e
clin
ical
n
ee
d
f
o
r
ea
r
l
y
d
etec
tio
n
o
f
tr
ea
ta
b
le
ca
s
es,
we
p
er
f
o
r
m
a
b
in
a
r
y
g
r
o
u
p
in
g
o
f
class
es
:
C
o
n
v
er
s
io
n
o
f
t
h
e
5
s
tag
es
o
f
d
iab
etic
r
etin
o
p
ath
y
(
DR
)
in
to
a
b
in
ar
y
p
r
o
b
lem
:
cl
ass
0
:
Hea
lth
y
/m
ild
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
E
d
g
eR
etin
a
:
Hyb
r
id
mu
ltime
d
ia
a
r
ch
itectu
r
e
fo
r
d
ia
b
etic
r
etin
o
p
a
th
y
s
creen
in
g
…
(
Gu
id
o
u
m
A
min
a
)
237
s
tag
es
(
g
r
ad
es
0
-
1
)
:
with
o
u
t
R
D
2
5
9
5
im
a
g
es
an
d
class
1
:
Mo
d
er
ate/sev
er
e/p
r
o
life
r
ativ
e
s
tag
es
(
g
r
ad
e
2
-
4
)
with
R
D
1
,
0
6
7
im
ag
es.
Fig
u
r
e
1
.
E
d
g
eRetin
a
ar
ch
itect
u
r
e
an
d
s
y
s
tem
f
lo
w
d
iag
r
a
m
W
e
p
r
o
p
o
s
e
an
in
n
o
v
ativ
e
E
d
g
eRetin
a
ar
ch
itectu
r
e
(
Fig
u
r
e
1
)
f
o
r
m
ed
ical
im
ag
e
class
if
icatio
n
co
m
b
in
in
g
th
e
p
ar
am
et
r
ic
ef
f
i
cien
cy
o
f
Fire
m
o
d
u
les
with
th
e
g
lo
b
al
co
n
tex
tu
al
m
o
d
elin
g
o
f
tr
an
s
f
o
r
m
er
s
.
T
h
e
n
etwo
r
k
p
r
o
c
ess
es
1
2
8
×1
2
8
r
etin
al
im
a
g
es
th
r
o
u
g
h
f
o
u
r
Fire
m
o
d
u
les,
p
r
o
g
r
ess
iv
ely
ex
tr
ac
tin
g
h
ier
ar
ch
ical
f
ea
tu
r
es.
A
Mo
b
i
leViT
b
lo
ck
in
s
er
ted
a
f
ter
`
f
i
r
e5
_
co
n
ca
t`
p
r
o
ce
s
s
es
1
6
×1
6
p
atch
es
b
y
m
u
lti
-
h
ea
d
s
elf
-
atten
tio
n
(
4
h
ea
d
s
,
1
2
8
-
d
im
p
r
o
jectio
n
)
.
T
h
e
tr
an
s
f
o
r
m
er
o
u
tp
u
t
is
r
esam
p
led
b
y
b
ilin
ea
r
u
p
s
am
p
lin
g
an
d
f
u
s
ed
v
ia
a
r
e
s
id
u
al
co
n
n
ec
tio
n
.
A
g
lo
b
al
a
v
er
ag
e
p
o
o
lin
g
a
n
d
a
d
r
o
p
o
u
t
(
5
0
%)
p
r
ec
e
d
e
th
e
f
in
al
class
if
icatio
n
lay
er
.
−
Sq
u
ee
ze
Net
[
1
6
]
is
u
s
ed
f
o
r
its
lig
h
tweig
h
t
d
esig
n
v
ia
"Fir
e"
m
o
d
u
les
(
1
x
1
co
m
p
r
e
s
s
io
n
an
d
3
×
3
ex
p
an
s
iv
e
co
n
v
o
lu
tio
n
s
)
.
−
A
Mo
b
ileViT
[
1
7
]
b
lo
ck
is
in
teg
r
ated
d
o
wn
s
tr
ea
m
to
ca
p
t
u
r
e
g
lo
b
al
d
e
p
en
d
e
n
cies
u
s
in
g
m
u
lti
-
h
ea
d
s
elf
-
atten
tio
n
o
n
im
a
g
e
p
atch
es (
1
6
x
1
6
p
i
x
els).
−
T
h
e
o
u
tp
u
t is d
r
o
p
o
u
t
r
eg
u
lar
i
ze
d
(
5
0
%)
b
ef
o
r
e
b
in
a
r
y
class
if
icatio
n
u
s
in
g
a
d
en
s
e
s
ig
m
o
id
lay
er
.
=
(
1
∗
+
1
)
ℎ
∶
1
×
1
,
(
1
)
1
=
(
2
∗
+
2
)
ℎ
1
×
1
(
2
)
3
=
(
3
∗
+
3
)
ℎ
3
×
3
(
3
)
=
1
⊕
3
ℎ
ℎ
(
4
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2
5
8
6
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
:
234
-
246
238
T
h
e
Fire
m
o
d
u
le
(
s
q
u
ee
ze
Net)
p
r
o
ce
ed
s
in
th
r
e
e
p
h
ase
s
:
as
in
(
1
)
a
1
×1
co
n
v
o
lu
tio
n
lay
er
(
“sq
u
ee
ze
”)
r
ed
u
ce
s
th
e
d
im
e
n
s
io
n
ality
o
f
th
e
f
ea
tu
r
es,
f
o
llo
wed
b
y
an
ex
p
a
n
s
io
n
s
tep
(
“E
x
p
an
d
”
)
ap
p
ly
in
g
1
×1
an
d
3
×
3
co
n
v
o
lu
tio
n
s
in
p
ar
allel
to
ca
p
tu
r
e
m
u
lti
-
s
ca
le
f
ea
tu
r
es
as
in
(
2
)
a
n
d
(
3
)
.
T
h
e
o
u
tp
u
ts
o
f
th
e
two
b
r
an
ch
es a
r
e
m
er
g
e
d
b
y
co
n
ca
ten
atio
n
(
“c
o
n
ca
t”)
as in
(
4
)
.
=
_
ℎ
(
,
=
16
×
16
,
=
16
)
ℎ
ℎ
=
×
+
(
5
)
H
er
e
,
P is
m
atr
ix
o
f
e
x
tr
ac
ted
p
atch
es (
1
6
×
1
6
p
i
x
els)
T
h
e
Mo
b
ileViT
b
lo
ck
s
tar
ts
w
ith
a
1
6
×1
6
p
atch
ex
t
r
ac
tio
n
d
ec
o
m
p
o
s
in
g
th
e
im
ag
e
in
t
o
s
p
atial
u
n
its
as
in
(
5
)
.
T
h
ese
p
atch
es
u
n
d
e
r
g
o
a
lin
ea
r
p
r
o
jectio
n
in
to
a
l
aten
t
s
p
ac
e
Z
b
ef
o
r
e
b
ein
g
p
r
o
ce
s
s
ed
b
y
a
m
u
lti
-
h
ea
d
atten
tio
n
m
ec
h
a
n
is
m
ca
p
tu
r
in
g
g
lo
b
al
d
ep
e
n
d
en
cies.
A
s
p
atial
r
ec
o
n
s
tr
u
ctio
n
b
y
b
ilin
ea
r
u
p
s
am
p
lin
g
th
en
r
esto
r
es
th
e
o
r
ig
in
al
r
eso
lu
tio
n
,
wh
ile
a
r
esid
u
al
co
n
n
e
ctio
n
co
m
b
i
n
es
th
ese
tr
an
s
f
o
r
m
ed
f
ea
tu
r
es
with
th
e
in
p
u
t f
ea
t
u
r
es.
∗
=
1
(
,
)
ℎ
ℎ
ℎ
ℎ
,
1
:
1
(
6
)
=
1
(
)
≥
=
0
Her
e,
y
_
p
r
ed
icted
is
th
e
p
r
o
b
a
b
ilit
ies p
r
ed
icted
b
y
th
e
m
o
d
el
an
d
y
_
tr
u
e
is
th
e
t
r
u
e
lab
els (
g
r
o
u
n
d
tr
u
t
h
)
.
T
h
r
ee
o
p
tim
izatio
n
tech
n
iq
u
es we
r
e
em
p
lo
y
e
d
as sh
o
wn
in
T
ab
le
1
.
−
Data
au
g
m
en
tatio
n
,
in
clu
d
in
g
r
an
d
o
m
h
o
r
iz
o
n
tal
f
lip
p
i
n
g
an
d
±
1
0
% c
o
n
tr
ast
v
ar
iatio
n
.
−
C
las
s
weig
h
tin
g
with
a
weig
h
t
o
f
3
f
o
r
t
h
e
m
in
o
r
ity
class
(
R
D)
to
co
r
r
ec
t
f
o
r
im
b
alan
ce
.
−
Dy
n
am
ic
o
p
tim
izatio
n
o
f
th
e
class
if
icatio
n
th
r
esh
o
ld
r
ec
alib
r
ated
at
ea
ch
ep
o
c
h
d
u
r
i
n
g
t
r
ain
in
g
,
wh
er
e
a
ca
llb
ac
k
ev
alu
ates
5
0
th
r
esh
o
ld
s
(
b
etwe
en
0
.
3
an
d
0
.
7
)
o
n
th
e
v
alid
atio
n
s
et
an
d
r
e
tain
s
th
e
v
alu
e
m
ax
im
izin
g
th
e
F1
-
s
co
r
e
as
in
(
6
)
.
Alth
o
u
g
h
th
e
o
p
tim
al
t
h
r
esh
o
ld
v
a
r
ies
b
etwe
en
0
.
3
an
d
0
.
6
9
ac
r
o
s
s
ep
o
ch
s
,
th
e
m
e
d
ian
v
al
u
e
o
f
0
.
3
was r
etain
ed
f
o
r
its
clin
ical
b
alan
ce
b
etwe
en
s
en
s
itiv
ity
an
d
s
p
ec
if
icity
.
T
ab
le
1
.
T
r
ai
n
in
g
p
ar
am
eter
s
P
a
r
a
me
t
e
r
V
a
l
u
e
D
e
scri
p
t
i
o
n
i
mg
_
si
z
e
(
1
2
8
,
1
2
8
)
I
n
p
u
t
i
m
a
g
e
r
e
s
o
l
u
t
i
o
n
b
a
t
c
h
_
s
i
z
e
32
N
u
mb
e
r
o
f
sam
p
l
e
s
p
e
r
g
r
a
d
i
e
n
t
u
p
d
a
t
e
e
p
o
c
h
s
50
N
u
mb
e
r
o
f
t
r
a
i
n
i
n
g
i
t
e
r
a
t
i
o
n
s
v
a
l
_
s
p
l
i
t
0
.
1
5
F
r
a
c
t
i
o
n
o
f
d
a
t
a
r
e
s
e
r
v
e
d
f
o
r
v
a
l
i
d
a
t
i
o
n
c
l
a
ss
_
w
e
i
g
h
t
s
{0
:
1
.
0
,
1
:
3
.
0
}
W
e
i
g
h
t
i
n
g
t
o
h
a
n
d
l
e
c
l
a
ss
i
m
b
a
l
a
n
c
e
o
p
t
i
m
i
z
e
r
.
l
e
a
r
n
i
n
g
_
r
a
t
e
1e
-
3
A
d
a
m
o
p
t
i
mi
z
e
r
l
e
a
r
n
i
n
g
r
a
t
e
a
u
g
me
n
t
a
t
i
o
n
.
r
a
n
d
o
m_
f
l
i
p
Tr
u
e
En
a
b
l
e
r
a
n
d
o
m
h
o
r
i
z
o
n
t
a
l
f
l
i
p
p
i
n
g
a
u
g
me
n
t
a
t
i
o
n
.
r
a
n
d
o
m_
c
o
n
t
r
a
st
0
.
1
C
o
n
t
r
a
st
v
a
r
i
a
t
i
o
n
r
a
n
g
e
(
±
1
0
%)
T
h
e
em
b
ed
d
e
d
d
e
p
lo
y
m
e
n
t p
ip
elin
e
im
p
lem
e
n
ts
p
o
s
t
-
tr
ain
in
g
I
NT
8
q
u
an
tizatio
n
[
1
8
]
,
[
1
9
]
,
r
ed
u
cin
g
th
e
m
o
d
el
s
ize.
I
NT
8
q
u
a
n
tizatio
n
,
was
ap
p
lied
u
s
in
g
a
f
o
u
r
-
s
tep
ap
p
r
o
ac
h
.
T
h
e
Ker
as
m
o
d
el
[
1
8
]
was
f
ir
s
t
co
n
v
er
ted
to
T
en
s
o
r
Flo
w
L
ite
f
o
r
m
at
[
2
0
]
with
s
tan
d
a
r
d
o
p
tim
izatio
n
s
en
ab
led
.
Dy
n
am
ic
c
alib
r
atio
n
was th
en
p
er
f
o
r
m
ed
o
n
1
0
0
b
atch
es
o
f
v
alid
atio
n
im
a
g
es
to
f
in
e
-
t
u
n
e
th
e
q
u
an
tizatio
n
p
ar
am
e
ter
s
.
T
h
e
p
r
o
ce
s
s
-
m
ain
tain
ed
co
m
p
atib
ilit
y
with
m
o
b
ile
p
r
o
ce
s
s
o
r
s
th
an
k
s
to
s
u
p
p
o
r
t
f
o
r
s
tan
d
ar
d
an
d
o
p
ti
m
ized
T
en
s
o
r
Flo
w
o
p
er
ato
r
s
.
Fin
ally
,
au
to
m
atic
in
p
u
t
a
n
d
o
u
tp
u
t
co
n
v
er
s
io
n
m
ec
h
an
is
m
s
wer
e
im
p
lem
e
n
ted
to
m
an
a
g
e
th
e
tr
an
s
itio
n
b
etwe
en
n
u
m
er
ic
f
o
r
m
ats d
u
r
in
g
in
f
er
e
n
ce
.
T
h
e
ev
al
u
atio
n
f
r
am
ewo
r
k
m
ea
s
u
r
es:
i
)
C
lin
ical
p
er
f
o
r
m
a
n
ce
v
ia
(
AUC
[
2
1
]
,
F1
-
s
co
r
e
,
p
r
ec
is
io
n
an
d
r
ec
all)
[
2
2
]
,
[
2
3
]
at
th
e
o
p
tim
al
th
r
esh
o
ld
;
ii
)
E
x
p
lain
ab
ilit
y
v
ia
Gr
ad
-
C
AM
h
ea
tm
ap
s
[
2
4
]
lo
ca
lizin
g
p
ath
o
lo
g
ical
r
eg
io
n
s
;
iii
)
R
OC
cu
r
v
es
[
2
5
]
an
d
c
o
n
f
u
s
io
n
m
atr
ices
f
o
r
FP
3
2
(
Ker
as
)
a
n
d
I
NT
8
(
T
en
s
o
r
F
lo
w)
m
o
d
els;
iv
)
Op
er
atio
n
al
p
er
f
o
r
m
an
ce
v
ia
laten
cy
an
d
C
PU
u
tili
za
tio
n
m
ea
s
u
r
ed
o
n
an
I
n
tel
C
o
r
e
i7
-
9
7
5
0
H
(
6
co
r
es,
2
.
6
GHz
)
s
y
s
tem
with
o
u
t
h
a
r
d
war
e
ac
ce
le
r
atio
n
.
P
er
-
f
r
am
e
late
n
cy
(
d
ef
in
ed
as
th
e
in
ter
v
al
b
etwe
en
th
e
s
tar
t
o
f
in
p
u
t
d
ata
tr
an
s
f
er
an
d
th
e
co
m
p
letio
n
o
f
p
r
e
d
ictio
n
s
)
is
q
u
an
tifie
d
with
m
icr
o
s
ec
o
n
d
p
r
ec
is
io
n
v
ia
tim
e.
p
er
f
_
c
o
u
n
ter
(
)
a
n
d
t
h
en
n
o
r
m
alize
d
p
er
b
atch
.
C
PU
u
tili
za
tio
n
is
ca
lcu
lated
with
th
e
p
s
u
til
lib
r
ar
y
v
ia
s
n
ap
s
h
o
ts
b
ef
o
r
e/af
ter
ea
ch
i
n
f
er
en
ce
,
with
th
e
f
in
al
v
alu
e
r
e
p
r
esen
tin
g
th
e
weig
h
ted
a
r
ith
m
etic
m
ea
n
o
v
er
th
e
f
u
ll
v
alid
atio
n
s
et
(
5
4
9
s
am
p
l
es).
Mo
d
el
s
ize
is
m
ea
s
u
r
ed
u
s
in
g
a
s
tan
d
ar
d
ized
m
eth
o
d
o
l
o
g
y
:
.
k
er
as
f
ile
f
o
r
FP
3
2
(
tf
.
k
er
as.m
o
d
els.s
av
e_
m
o
d
el(
)
)
an
d
.
tf
lite
f
ile
f
o
r
I
NT
8
p
o
s
t
-
q
u
an
tizatio
n
.
T
h
is
ap
p
r
o
ac
h
in
teg
r
ates
all
th
e
co
m
p
o
n
e
n
ts
n
ec
ess
ar
y
f
o
r
d
ep
lo
y
m
en
t
(
ar
ch
itectu
r
e,
weig
h
ts
,
o
p
er
atio
n
al
m
etad
ata)
f
o
r
a
r
ea
lis
tic
m
ea
s
u
r
em
en
t o
f
th
e
m
em
o
r
y
f
o
o
tp
r
in
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
E
d
g
eR
etin
a
:
Hyb
r
id
mu
ltime
d
ia
a
r
ch
itectu
r
e
fo
r
d
ia
b
etic
r
etin
o
p
a
th
y
s
creen
in
g
…
(
Gu
id
o
u
m
A
min
a
)
239
Fig
u
r
e
2
im
p
lem
en
ts
th
e
Gr
a
d
-
C
AM
m
eth
o
d
ac
co
r
d
in
g
to
f
o
u
r
k
e
y
p
r
i
n
cip
les:
i
)
ex
clu
s
iv
e
u
s
e
o
f
th
e
FP
3
2
m
o
d
el
f
o
r
in
ter
p
r
eta
b
ilit
y
,
ii
)
tar
g
etin
g
th
e
f
ir
e5
_
co
n
ca
t
la
y
er
(
last
co
n
v
o
lu
tio
n
al
o
u
tp
u
t)
,
iii
)
co
m
p
u
tatio
n
o
f
g
r
ad
ien
t
-
w
eig
h
ted
ac
tiv
atio
n
m
ap
s
,
an
d
iv
)
g
en
er
atio
n
o
f
clin
ically
in
ter
p
r
etab
le
h
ea
tm
ap
s
f
o
r
DR
d
etec
tio
n
.
Sy
s
tem
atic
an
aly
s
is
co
n
f
ir
m
s
th
at
ac
tiv
ated
r
eg
io
n
s
co
r
r
esp
o
n
d
to
r
elev
an
t
an
ato
m
ical
s
tr
u
ctu
r
es
(
m
ac
u
la,
r
etin
al
v
es
s
els)
an
d
ch
ar
ac
ter
is
tic
lesi
o
n
s
o
f
d
iab
etic
r
etin
o
p
ath
y
,
with
o
u
t
f
o
r
m
al
clin
ical
v
alid
atio
n
b
y
o
p
h
th
alm
o
lo
g
is
ts
.
T
h
e
ar
ch
ite
ct
u
r
al
d
esi
g
n
o
f
t
h
e
E
d
g
eR
eti
n
a
f
r
a
m
e
wo
r
k
f
a
cil
itat
es
i
ts
d
e
p
l
o
y
m
en
t
w
it
h
i
n
r
o
b
o
tic
a
n
d
au
t
o
m
at
e
d
s
c
r
e
e
n
i
n
g
s
y
s
te
m
s
.
T
h
is
i
n
t
e
g
r
ati
o
n
r
el
ies
o
n
a
s
im
p
li
f
ie
d
,
cl
o
s
e
d
-
lo
o
p
w
o
r
k
f
l
o
w
,
t
r
a
n
s
f
o
r
m
i
n
g
a
r
o
b
o
ti
c
p
l
at
f
o
r
m
i
n
t
o
a
n
au
to
n
o
m
o
u
s
d
i
ag
n
o
s
ti
c
u
n
it
.
T
h
e
p
r
o
c
ess
b
e
g
i
n
s
wit
h
th
e
ac
q
u
is
it
i
o
n
o
f
r
eti
n
al
i
m
ag
es
v
ia
a
r
o
b
o
tic
f
u
n
d
u
s
c
am
er
a.
T
h
ese
i
m
a
g
es
a
r
e
t
h
e
n
p
r
o
c
ess
e
d
b
y
t
h
e
E
d
g
eRe
ti
n
a
p
i
p
eli
n
e
f
o
r
i
m
m
e
d
ia
te
an
al
y
s
is
.
T
h
e
c
lass
i
f
i
ca
ti
o
n
r
es
u
lts
d
ir
ec
t
ly
f
ee
d
in
to
t
h
e
s
y
s
t
em
'
s
d
ec
is
i
o
n
l
o
g
ic
,
t
r
i
g
g
e
r
i
n
g
p
r
e
d
e
f
i
n
e
d
r
o
b
o
tic
ac
ti
o
n
s
.
T
h
ese
ac
ti
o
n
s
c
a
n
r
a
n
g
e
f
r
o
m
g
e
n
e
r
a
ti
n
g
a
n
al
er
t f
o
r
a
s
p
ec
i
alis
t
a
n
d
p
r
o
v
id
in
g
f
ee
d
b
ac
k
t
o
t
h
e
p
a
tie
n
t
,
to
a
u
t
o
n
o
m
o
u
s
l
y
i
n
iti
ati
n
g
f
u
r
t
h
e
r
i
m
a
g
i
n
g
ex
am
i
n
ati
o
n
s
.
T
h
is
e
n
d
-
to
-
en
d
wo
r
k
f
l
o
w
l
e
v
e
r
a
g
es
t
h
e
m
o
d
el'
s
co
m
p
u
tati
o
n
a
l
ef
f
i
cie
n
cy
t
o
e
n
ab
l
e
ac
ti
o
n
ab
le
d
ia
g
n
o
s
es
in
r
e
al
tim
e,
a
t
th
e
p
a
tie
n
t'
s
b
ed
s
id
e
.
Fig
u
r
e
2
.
Gr
a
d
_
ca
m
wo
r
k
f
lo
w
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
e
x
p
e
r
i
m
e
n
t
a
l
r
e
s
u
l
t
s
o
f
E
d
g
e
R
e
ti
n
a
a
r
e
p
r
e
s
e
n
t
e
d
o
n
t
h
e
A
P
T
OS
2
0
1
9
v
a
l
i
d
a
t
i
o
n
s
et
(
n
=
5
4
9
)
,
w
i
t
h
a
b
i
n
a
r
y
d
i
s
t
r
i
b
u
t
i
o
n
:
5
7
%
h
ea
l
t
h
y
c
a
s
es
(
3
1
3
s
a
m
p
l
e
s
)
a
n
d
4
3
%
r
e
f
e
r
a
b
l
e
r
e
t
i
n
o
p
a
t
h
i
es
(
2
3
6
s
a
m
p
l
e
s
)
.
T
h
e
c
o
m
p
a
r
a
t
i
v
e
e
v
a
l
u
a
ti
o
n
o
f
t
h
e
o
r
i
g
i
n
a
l
FP
3
2
m
o
d
e
l
a
n
d
i
t
s
q
u
a
n
t
i
z
e
d
v
e
r
s
i
o
n
I
N
T
8
c
o
v
e
r
s
:
i
)
cl
i
n
i
c
al
m
e
t
r
ics
(
A
U
C
,
F
1
-
s
c
o
r
e
,
p
r
e
c
is
i
o
n
,
r
e
c
a
l
l
)
,
ii
)
o
p
e
r
a
t
i
o
n
a
l
p
e
r
f
o
r
m
a
n
c
e
(
s
i
z
e
,
la
t
e
n
c
y
,
C
P
U
)
,
a
n
d
iii
)
d
i
a
g
n
o
s
t
ic
v
i
s
u
a
l
iz
a
t
i
o
n
s
(
c
o
n
f
u
s
i
o
n
m
a
t
r
ic
e
s
,
R
OC
c
u
r
v
e
s
,
G
r
a
d
-
C
A
M
m
a
p
s
)
.
T
a
b
l
e
2
s
u
m
m
a
r
i
z
es
t
h
es
e
m
e
t
r
ic
s
.
Fig
u
r
e
3
titl
ed
"Ro
c
cu
r
v
e
KE
R
AS M
ODE
L
F
P3
2
"
p
r
esen
ts
th
e
R
OC
cu
r
v
e
o
f
th
e
m
o
d
el.
T
h
e
x
-
a
x
is
r
ep
r
esen
ts
th
e
f
alse
p
o
s
itiv
e
r
ate
wh
ile
th
e
y
-
ax
is
r
ep
r
esen
ts
th
e
tr
u
e
p
o
s
itiv
e
r
ate.
Fr
o
m
Fig
u
r
e
3
,
it
is
o
b
s
er
v
ed
th
at
th
e
AUC
=
0
.
9
4
s
h
o
ws
an
ex
ce
llen
t
d
is
cr
im
in
ato
r
y
ca
p
ac
ity
.
T
h
e
in
ter
p
r
etatio
n
in
d
icate
s
th
at
th
e
m
o
d
el
ef
f
ec
tiv
ely
d
is
tin
g
u
is
h
es h
ea
lth
y
ca
s
es f
r
o
m
d
is
ea
s
ed
ca
s
es (
9
4
%
o
f
ac
c
u
r
ac
y
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2
5
8
6
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
:
234
-
246
240
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
tab
le
b
et
wee
n
FP
3
2
an
d
I
NT
8
m
o
d
els
M
e
t
r
i
c
F
P
3
2
I
N
T8
I
mp
r
o
v
e
m
e
n
t
AUC
0
.
9
3
5
0
.
9
6
3
+3
%
S
c
o
r
e
0
.
8
2
6
0
.
8
9
9
+8
.
8
%
P
r
e
c
i
s
i
o
n
0
.
7
1
7
0
.
8
9
2
+
2
4
,
4
%
R
e
c
a
l
l
0
.
9
7
5
0
.
9
0
7
-
7
%
(
*
)
S
i
z
e
(
Mo
)
98
.
66
8
.
27
-
9
1
.
6
%
La
t
e
n
c
y
(
ms
)
15
.
10
15
.
43
+2
.
2
%
C
P
U
%
37
.
7
21
.
5
-
43
%
(
*
)
Th
e
h
i
g
h
F
P
3
2
r
e
c
a
l
l
(
9
7
.
5
%)
r
e
f
l
e
c
t
s
a
b
i
a
s
t
o
w
a
r
d
s
f
a
l
se
p
o
si
t
i
v
e
s.
I
N
T8
q
u
a
n
t
i
z
a
t
i
o
n
d
r
a
s
t
i
c
a
l
l
y
r
e
d
u
c
e
s
mo
d
e
l
si
z
e
(
-
9
1
.
6
%)
w
h
i
l
e
i
m
p
r
o
v
i
n
g
A
U
C
a
n
d
p
r
e
c
i
si
o
n
.
T
h
e
sl
i
g
h
t
d
e
c
r
e
a
s
e
i
n
r
e
c
a
l
l
(
-
7
%)
i
s
o
f
f
se
t
b
y
a
si
g
n
i
f
i
c
a
n
t
i
mp
r
o
v
e
me
n
t
i
n
p
r
e
c
i
s
i
o
n
(
+
2
4
.
4
%),
r
e
su
l
t
i
n
g
i
n
a
h
i
g
h
e
r
F
1
-
sc
o
r
e
(
+
8
.
8
%).
Th
e
4
3
%
r
e
d
u
c
t
i
o
n
i
n
C
P
U
u
s
a
g
e
i
s
c
r
u
c
i
a
l
f
o
r
e
m
b
e
d
d
e
d
d
e
v
i
c
e
d
e
p
l
o
y
me
n
t
s.
C
o
m
p
r
e
ss
i
o
n
sh
o
w
s
a
9
1
.
6
%
r
e
d
u
c
t
i
o
n
i
n
mo
d
e
l
si
z
e
(
f
r
o
m
9
8
.
6
6
M
B
t
o
8
.
2
7
M
B
)
,
i
d
e
a
l
f
o
r
m
o
b
i
l
e
d
e
v
i
c
e
s.
C
P
U
u
s
a
g
e
i
s
r
e
d
u
c
e
d
b
y
4
3
%
(
f
r
o
m
3
7
.
7
%
t
o
2
1
.
5
%),
e
sse
n
t
i
a
l
f
o
r
e
n
e
r
g
y
a
u
t
o
n
o
m
y
.
La
t
e
n
c
y
s
h
o
w
s
n
e
g
l
i
g
i
b
l
e
v
a
r
i
a
t
i
o
n
(
+
0
.
3
3
ms)
,
d
e
mo
n
s
t
r
a
t
i
n
g
t
h
a
t
q
u
a
n
t
i
z
a
t
i
o
n
d
o
e
s
n
o
t
a
f
f
e
c
t
r
e
s
p
o
n
s
i
v
e
n
e
ss.
T
h
e
x
-
a
x
i
s
r
e
p
r
e
s
e
n
ts
t
h
e
f
a
l
s
e
p
o
s
i
t
i
v
e
r
a
t
e
,
w
h
i
l
e
t
h
e
y
-
a
x
i
s
r
e
p
r
e
s
e
n
t
s
t
h
e
t
r
u
e
p
o
s
it
i
v
e
r
a
t
e
.
F
r
o
m
F
i
g
u
r
e
4
,
we
s
e
e
a
n
A
UC
o
f
0
.
9
6
a
n
d
a
p
o
s
t
-
q
u
a
n
ti
z
a
ti
o
n
i
m
p
r
o
v
e
m
e
n
t
o
f
+
2
.
1
%
.
T
h
e
c
u
r
v
e
r
i
s
e
s
r
a
p
i
d
l
y
t
o
w
a
r
d
s
t
h
e
u
p
p
e
r
l
e
f
t
c
o
r
n
e
r
,
in
d
i
c
a
t
i
n
g
e
x
c
el
l
e
n
t
e
a
r
l
y
p
e
r
f
o
r
m
a
n
c
e
i
n
d
e
te
c
t
i
n
g
t
r
u
e
p
o
s
it
iv
e
s
.
At
t
h
e
o
p
t
i
m
al
p
o
i
n
t
,
f
o
r
a
f
a
l
s
e
p
o
s
i
t
i
v
e
r
a
t
e
o
f
0
.
2
,
t
h
e
t
r
u
e
p
o
s
i
t
i
v
e
r
a
t
e
a
lr
e
a
d
y
r
e
a
c
h
e
s
>
0
.
9
,
m
e
a
n
i
n
g
t
h
a
t
w
i
t
h
o
n
l
y
2
0
%
m
i
s
d
i
a
g
n
o
s
i
s
,
t
h
e
m
o
d
e
l
d
et
e
c
ts
m
o
r
e
t
h
a
n
9
0
%
o
f
t
r
u
e
d
i
a
b
eti
c
r
e
t
i
n
o
p
a
t
h
y
c
as
es
.
I
N
T
8
o
p
tim
i
z
a
t
i
o
n
i
m
p
r
o
v
es
p
e
r
f
o
r
m
a
n
c
e
d
e
s
p
i
t
e
c
o
m
p
r
e
s
s
i
o
n
.
I
n
c
o
m
p
a
r
is
o
n
,
t
h
e
I
N
T
8
c
u
r
v
e
d
o
m
i
n
a
t
es
FP
3
2
o
v
er
t
h
e
e
n
ti
r
e
r
a
n
g
e
,
s
h
o
w
i
n
g
t
h
e
b
e
s
t
s
e
n
s
i
ti
v
i
t
y
/s
p
e
c
i
f
i
c
it
y
t
r
a
d
e
o
f
f
.
T
h
e
c
o
n
v
e
x
s
h
a
p
e
o
f
t
h
e
c
u
r
v
e
s
i
n
d
i
c
a
t
e
s
t
h
a
t
t
h
e
m
o
d
e
l
s
m
a
i
n
t
a
i
n
a
g
o
o
d
b
a
l
a
n
c
e
b
e
tw
ee
n
s
e
n
s
i
t
i
v
it
y
(
t
r
u
e
p
o
s
it
i
v
e
s
)
an
d
s
p
e
c
i
f
i
c
it
y
(
f
a
ls
e
p
o
s
it
i
v
e
s
)
.
Fig
u
r
e
3
.
R
o
c
cu
r
v
e
KE
R
AS m
o
d
el
FP
3
2
Fig
u
r
e
4
.
R
o
c
cu
r
v
e
I
NT
8
q
u
a
n
tized
m
o
d
el)
Acc
o
r
d
in
g
t
o
Fig
u
r
e
5
,
th
e
r
e
a
r
e
9
1
f
alse
p
o
s
itiv
es
(
h
ea
lth
y
p
atien
ts
wr
o
n
g
ly
d
iag
n
o
s
ed
w
ith
d
iab
etic
r
etin
o
p
ath
y
)
.
A
s
tr
o
n
g
p
o
in
t
i
s
th
e
o
n
ly
6
f
alse
n
eg
ativ
es
(
6
ca
s
es
o
f
d
iab
etic
r
etin
o
p
at
h
y
n
o
t
d
etec
ted
)
.
T
h
is
in
d
icate
s
th
at
th
e
FP
3
2
m
o
d
el
p
r
io
r
itizes
s
en
s
itiv
ity
o
v
er
s
p
ec
if
icity
,
en
s
u
r
in
g
h
ig
h
d
etec
tio
n
o
f
t
r
u
e
p
ath
o
lo
g
ical
ca
s
es.
A
b
ias
is
o
b
s
er
v
ed
with
a
ten
d
en
cy
t
o
o
v
er
d
ia
g
n
o
s
es
d
iab
etic
r
etin
o
p
ath
y
(
ex
tr
em
e
s
en
s
itiv
ity
at
th
e
ex
p
en
s
e
o
f
s
p
ec
if
icity
)
.
Acc
o
r
d
in
g
to
Fig
u
r
e
6
,
f
o
r
ca
s
es
with
o
u
t
d
iab
etic
r
etin
o
p
ath
y
,
th
er
e
ar
e
2
8
7
tr
u
e
n
eg
ativ
es
an
d
2
6
f
alse
p
o
s
itiv
es,
g
iv
in
g
a
s
p
ec
if
icity
o
f
9
1
.
7
%.
Fo
r
ca
s
es
with
d
iab
etic
r
etin
o
p
ath
y
,
th
er
e
ar
e
2
1
4
tr
u
e
p
o
s
itiv
es
an
d
2
2
f
alse
n
eg
ativ
es,
co
r
r
esp
o
n
d
in
g
to
a
s
en
s
itiv
ity
o
f
9
0
.
7
%.
T
h
ese
r
esu
lts
d
em
o
n
s
tr
a
te
a
b
etter
b
alan
ce
b
etwe
en
s
en
s
itiv
ity
an
d
s
p
ec
if
icity
co
m
p
ar
e
d
to
t
h
e
FP
3
2
m
o
d
el.
T
h
e
ac
cu
r
ate
d
etec
tio
n
o
f
h
ea
lth
y
ca
s
es
is
2
8
7
h
ea
lth
y
p
atien
ts
co
r
r
ec
tly
id
en
tifie
d
o
u
t
o
f
3
0
9
.
Fals
e
p
o
s
itiv
es
ar
e
r
ed
u
ce
d
f
r
o
m
9
1
to
2
6
(
-
7
1
%)
co
m
p
ar
ed
to
t
h
e
FP
3
2
m
o
d
e
l.
Acc
u
r
ac
y
in
cr
ea
s
es
f
r
o
m
7
1
.
7
%
to
8
9
.
2
%
co
m
p
ar
ed
to
th
e
FP
3
2
m
o
d
el.
A
m
o
d
er
ate
in
cr
ea
s
e
i
n
f
alse
n
eg
ativ
es
(
f
r
o
m
6
to
2
2
)
is
o
b
s
er
v
ed
,
wh
ic
h
is
co
n
s
id
er
ed
a
cc
ep
tab
le
b
ec
a
u
s
e
it
is
co
m
p
en
s
ated
b
y
a
m
o
r
e
r
eli
ab
le
d
etec
tio
n
o
f
p
o
s
itiv
e
ca
s
e
s
.
As
co
n
clu
s
io
n
E
d
g
eRetin
a
m
ain
tain
s
a
h
ig
h
s
en
s
itiv
ity
(
9
0
.
7
%)
f
o
r
s
cr
ee
n
i
n
g
r
e
f
er
ab
l
e
d
iab
etic
r
etin
o
p
ath
y
wh
ile
r
ed
u
cin
g
f
a
ls
e
p
o
s
itiv
es
b
y
7
1
%
(
f
r
o
m
9
1
to
2
6
)
co
m
p
ar
ed
t
o
th
e
FP
3
2
r
ef
e
r
en
ce
m
o
d
e
l
th
an
k
s
to
th
e
u
s
e
o
f
a
d
y
n
am
ic
th
r
esh
o
ld
(
0
.
3
)
.
Alth
o
u
g
h
th
is
th
r
esh
o
ld
r
esu
lts
in
a
co
n
tr
o
lled
in
cr
ea
s
e
in
f
alse
n
eg
ativ
es
(
f
r
o
m
6
to
2
2
)
th
is
s
tr
ateg
y
s
ig
n
if
ican
tly
im
p
r
o
v
e
s
d
iag
n
o
s
tic
p
r
ec
is
io
n
(
+2
4
.
5
%).
T
h
is
ad
j
u
s
tab
le
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
E
d
g
eR
etin
a
:
Hyb
r
id
mu
ltime
d
ia
a
r
ch
itectu
r
e
fo
r
d
ia
b
etic
r
etin
o
p
a
th
y
s
creen
in
g
…
(
Gu
id
o
u
m
A
min
a
)
241
clin
ical
co
m
p
r
o
m
is
e
th
u
s
o
p
tim
izes
s
cr
ee
n
in
g
ef
f
icien
cy
b
y
lim
itin
g
o
v
er
d
iag
n
o
s
is
wh
ile
m
ain
tain
in
g
a
9
0
.
7
% p
ath
o
lo
g
ical
ca
s
e
d
etec
tio
n
r
ate.
T
h
e
an
aly
s
is
o
f
th
e
o
p
tim
al
cl
ass
if
icatio
n
r
atio
is
b
ased
o
n
3
2
s
am
p
les.
T
h
is
r
ep
o
r
t
c
o
r
r
es
p
o
n
d
s
t
o
th
e
o
p
tim
al
th
r
esh
o
ld
(
0
.
3
0
)
i
d
en
t
if
ied
d
u
r
in
g
tr
ain
in
g
o
n
a
v
ali
d
atio
n
s
u
b
s
et
o
f
3
2
s
am
p
les.
T
h
e
d
is
tr
ib
u
tio
n
o
f
th
e
s
u
b
s
et
is
:
6
9
%
with
o
u
t d
iab
etic
r
etin
o
p
ath
y
(
2
2
s
am
p
les)
an
d
3
1
%
with
d
iab
etic
r
etin
o
p
ath
y
(
1
0
s
am
p
les).
T
h
e
d
is
tr
ib
u
tio
n
o
f
t
h
e
s
u
b
s
e
t
is
:
6
9
%
with
o
u
t
d
iab
etic
r
e
tin
o
p
ath
y
(
2
2
s
am
p
les)
an
d
3
1
%
with
d
iab
etic
r
etin
o
p
ath
y
(
1
0
s
am
p
les).
Acc
o
r
d
in
g
t
o
T
ab
le
3
an
d
Fig
u
r
e
7
,
th
r
esh
o
l
d
o
p
tim
izatio
n
d
u
r
i
n
g
tr
ain
i
n
g
r
e
v
ea
led
a
m
ax
im
u
m
p
o
ten
tial
o
f
F1
=0
.
9
5
o
n
a
v
alid
atio
n
s
et
(
n
=
3
2
)
,
with
p
r
ec
is
io
n
=0
.
9
1
(
m
in
im
izi
n
g
f
a
ls
e
p
o
s
itiv
es)
an
d
r
ec
all=1
.
0
0
(
n
o
f
alse n
eg
ativ
e
s
)
,
f
o
r
an
o
v
e
r
all
ac
cu
r
ac
y
o
f
9
7
%.
T
h
e
m
o
d
el
ef
f
ec
tiv
ely
tar
g
ets at
-
r
is
k
p
atien
ts
with
o
u
t
u
n
n
ec
ess
ar
y
aler
ts
.
I
t
ac
h
iev
es
p
er
f
ec
t
p
r
ec
is
io
n
(
1
.
0
0
)
f
o
r
t
h
e
"No
DR
"
ca
teg
o
r
y
:
n
o
f
alse
p
o
s
itiv
es.
T
h
e
F1
-
s
co
r
e
ex
ce
ed
s
0
.
9
5
f
o
r
b
o
th
class
es,
in
d
icatin
g
a
well
-
b
alan
ce
d
m
o
d
el.
T
h
is
p
er
f
o
r
m
a
n
ce
d
em
o
n
s
tr
ates
th
e
m
o
d
el'
s
ab
ilit
y
to
ac
cu
r
ately
d
is
tin
g
u
is
h
h
ea
lth
y
an
d
p
ath
o
lo
g
ical
ca
s
es
wh
ile
m
in
im
izin
g
class
if
icatio
n
er
r
o
r
s
.
Fig
u
r
e
5
.
C
o
n
f
u
s
io
n
m
atr
i
x
FP
3
2
(
o
r
ig
in
al)
Fig
u
r
e
6
.
C
o
n
f
u
s
io
n
m
atr
i
x
I
N
T
8
(
q
u
an
tized
m
o
d
el)
T
ab
le
3
.
C
lass
if
icatio
n
r
ep
o
r
t
o
f
clin
ical
v
alid
atio
n
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
W
i
t
h
o
u
t
R
D
1
.
0
0
0
.
9
5
0
.
9
8
22
W
i
t
h
R
D
0
.
9
1
1
.
0
0
0
.
9
5
10
A
c
c
u
r
a
c
y
0
.
9
7
32
M
a
c
r
o
a
v
g
0
.
9
5
0
.
9
8
0
.
9
6
32
W
e
i
g
h
t
e
d
a
v
g
0
.
9
7
0
.
9
7
0
.
9
7
32
Fig
u
r
e
7
.
C
o
n
f
u
s
io
n
m
atr
i
x
o
f
clin
ical
v
alid
atio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
2
5
8
6
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
,
Vo
l
.
1
5
,
No
.
1
,
Ma
r
ch
20
2
6
:
234
-
246
242
An
aly
s
is
o
f
tr
ain
in
g
m
etr
ics
s
h
o
ws
th
at
th
e
v
alid
atio
n
AU
C
r
ea
ch
es
0
.
9
8
6
as
ea
r
ly
as
e
p
o
ch
5
an
d
s
tab
ilizes
at
>0
.
9
5
.
T
h
e
F1
-
s
co
r
e
co
n
v
er
g
es
to
wa
r
d
s
0
.
9
5
+
af
ter
2
0
ep
o
c
h
s
.
A
r
ap
i
d
c
o
n
v
er
g
en
ce
p
h
ase
is
o
b
s
er
v
ed
d
u
r
i
n
g
th
e
f
ir
s
t
1
5
ep
o
ch
s
,
f
o
llo
wed
b
y
a
s
tab
ilizatio
n
o
f
th
e
m
et
r
ics
af
ter
ep
o
ch
3
0
.
T
h
e
m
o
d
el
d
em
o
n
s
tr
ates
r
o
b
u
s
tn
ess
to
th
e
ch
o
ice
o
f
class
if
icatio
n
th
r
esh
o
ld
(
o
p
tim
al
b
etwe
en
0
.
3
a
n
d
0
.
7
)
.
I
t
is
m
o
r
e
ac
cu
r
ate
o
n
s
m
all
h
o
m
o
g
en
e
o
u
s
b
atch
es,
wh
ile
p
e
r
f
o
r
m
a
n
ce
n
o
r
m
alize
s
o
v
er
t
h
e
f
u
l
l
s
et
with
g
r
ea
te
r
v
ar
iab
ilit
y
.
T
h
e
d
is
cr
ep
an
c
y
b
etwe
en
th
e
o
p
tim
al
class
if
icatio
n
r
atio
(
F1
=
0
.
9
5
o
n
3
2
s
am
p
les)
an
d
t
h
e
o
v
er
all
r
esu
lts
(
F1
=
0
.
8
9
9
)
r
e
q
u
ir
e
s
lar
g
e
-
s
ca
le
v
alid
atio
n
.
C
lin
ical
r
o
b
u
s
tn
ess
is
co
n
f
ir
m
ed
wh
en
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
is
s
im
ilar
to
its
h
o
m
o
g
en
eo
u
s
co
m
p
o
s
itio
n
s
by
r
ep
licatio
n
o
f
r
esu
lts
o
n
5
4
9
ca
s
es,
an
aly
s
is
o
f
m
ea
s
u
r
em
en
t d
ata
(
AUC
+3
%
,
ac
cu
r
ac
y
+2
4
.
5
%),
an
d
r
e
d
u
c
tio
n
o
f
f
alse
p
o
s
itiv
es
—
th
e
m
o
s
t
d
an
g
er
o
u
s
er
r
o
r
in
th
is
s
tep
.
T
h
e
I
NT
8
q
u
an
tizatio
n
ap
p
r
o
ac
h
n
ar
r
o
wed
th
is
p
er
f
o
r
m
an
ce
-
p
o
ten
tial
g
ap
:
wh
ile
m
ain
tain
in
g
a
clin
ically
ac
ce
p
tab
le
r
ec
all
o
f
0
.
9
0
7
,
it
in
cr
ea
s
ed
p
r
ec
is
io
n
t
o
0
.
8
9
2
(
+2
4
.
5
%)
an
d
F1
-
s
co
r
e
to
0
.
8
9
9
(
+8
.
8
%)
o
n
th
e
f
u
ll
s
et.
T
h
is
d
e
m
o
n
s
tr
ates
th
at
o
p
tim
izatio
n
b
y
q
u
an
tizatio
n
n
o
t
o
n
ly
im
p
r
o
v
es
o
p
er
atio
n
al
ef
f
icien
c
y
(
9
2
% r
ed
u
ce
d
s
ize,
-
4
3
% C
PU
)
,
b
u
t a
ls
o
d
ia
g
n
o
s
tic
r
o
b
u
s
tn
e
s
s
ac
r
o
s
s
d
iv
er
s
e
p
o
p
u
latio
n
s
.
Gr
ad
-
C
AM
is
a
tech
n
i
q
u
e
u
s
ed
to
v
is
u
alize
th
e
ar
ea
s
o
f
a
n
i
m
ag
e
th
at
m
o
s
t
in
f
lu
en
ce
d
th
e
d
ec
is
io
n
o
f
a
C
NN.
I
t is o
f
ten
u
s
ed
in
m
ed
icin
e
to
u
n
d
er
s
tan
d
wh
y
a
d
ee
p
lear
n
in
g
m
o
d
el
d
etec
ted
p
at
h
o
lo
g
y
in
an
im
ag
e
.
R
ed
/y
ello
w
ar
ea
s
in
d
icate
h
ig
h
ac
tiv
atio
n
an
d
th
er
ef
o
r
e
m
a
k
e
a
s
tr
o
n
g
co
n
tr
ib
u
tio
n
to
th
e
m
o
d
el'
s
p
r
ed
ictio
n
,
wh
ile
b
lu
e
ar
ea
s
in
d
icate
lo
w
ac
tiv
atio
n
.
Fig
u
r
e
8
s
h
o
ws a
f
u
n
d
u
s
im
ag
e
with
m
ar
k
e
d
ac
tiv
atio
n
s
in
th
e
ce
n
ter
an
d
ar
o
u
n
d
th
e
o
p
tic
d
is
c.
B
r
i
g
h
t
s
p
o
ts
in
th
e
f
u
n
d
u
s
c
o
u
ld
co
r
r
esp
o
n
d
to
h
a
r
d
e
x
u
d
ates
o
r
m
icr
o
a
n
eu
r
y
s
m
s
,
s
ig
n
s
o
f
d
iab
etic
r
etin
o
p
ath
y
.
T
h
e
m
o
d
el
f
o
cu
s
es
its
atten
tio
n
o
n
s
ev
er
al
p
ath
o
lo
g
ical
r
eg
io
n
s
,
wh
ich
is
ex
p
ec
ted
in
an
au
t
o
m
ated
d
iag
n
o
s
is
.
Fig
u
r
e
9
also
s
h
o
ws
a
f
u
n
d
u
s
with
f
o
c
u
s
ar
ea
s
ar
o
u
n
d
th
e
o
p
tic
n
e
r
v
e
(
b
o
tto
m
lef
t)
an
d
p
er
ip
h
e
r
al
a
r
ea
s
,
wh
ich
co
u
ld
in
d
icate
th
at
th
e
m
o
d
el
is
lo
o
k
in
g
f
o
r
a
b
n
o
r
m
alities
in
th
e
d
is
tr
ib
u
tio
n
o
f
r
etin
al
v
ess
els
o
r
m
o
r
e
d
if
f
u
s
e
lesi
o
n
s
.
Fig
u
r
e
1
0
s
h
o
ws
th
at
th
e
m
o
d
el
f
o
c
u
s
es
m
ain
ly
o
n
th
e
o
p
tic
n
e
r
v
e
(
lef
t)
,
with
th
e
r
est
o
f
th
e
r
etin
a
b
ei
n
g
v
er
y
wea
k
ly
ac
tiv
ated
,
wh
ich
co
u
ld
m
ea
n
eith
er
th
e
ab
s
en
ce
o
f
o
b
v
i
o
u
s
p
ath
o
lo
g
ical
s
ig
n
s
o
r
th
at
th
e
cla
s
s
if
i
ca
tio
n
r
elies
o
n
th
e
o
p
tic
n
er
v
e
(
e.
g
.,
to
d
etec
t
g
lau
co
m
a)
.
T
h
e
s
o
lu
tio
n
co
m
b
in
es
clin
ical
ex
ce
llen
ce
(
AUC
0
.
9
6
)
an
d
o
p
e
r
atio
n
al
ef
f
icien
cy
(
8
.
2
7
MB,
-
4
3
% C
PU)
,
p
av
in
g
t
h
e
way
f
o
r
r
ea
l
-
tim
e
s
cr
ee
n
in
g
o
f
d
iab
e
tic
r
etin
o
p
ath
y
o
n
s
m
ar
tp
h
o
n
e
s
.
Fig
u
r
e
8
.
g
r
ad
_
c
am
0
Fig
u
r
e
9
.
Gr
a
d
_
ca
m
1
Fig
u
r
e
1
0
.
g
r
ad
_
ca
m
2
:
v
is
u
aliza
tio
n
s
b
ased
o
n
t
h
e
FP
3
2
m
o
d
el
3
.
1
.
Dis
cus
s
io
n
E
d
g
eRetin
a
f
u
n
d
a
m
en
tally
d
is
tin
g
u
is
h
es
its
elf
f
r
o
m
th
e
s
tate
o
f
th
e
ar
t
as
s
h
o
wn
in
T
ab
le
4
b
y
th
r
e
e
k
ey
in
n
o
v
atio
n
s
:
a.
T
h
e
in
tr
o
d
u
ctio
n
o
f
t
h
e
f
ir
s
t
d
y
n
am
ic
th
r
esh
o
ld
ca
lib
r
atio
n
s
y
s
tem
(
0
.
3
-
0
.
7
)
,
co
m
b
in
ed
with
clin
ically
ca
lib
r
ated
I
NT
8
q
u
an
tific
atio
n
,
r
ed
u
cin
g
f
alse p
o
s
itiv
es b
y
7
1
% wh
ile
m
ain
tain
in
g
a
s
en
s
itiv
ity
>9
0
%.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
AE
S
I
n
t
J
R
o
b
&
A
u
to
m
I
SS
N:
2722
-
2
5
8
6
E
d
g
eR
etin
a
:
Hyb
r
id
mu
ltime
d
ia
a
r
ch
itectu
r
e
fo
r
d
ia
b
etic
r
etin
o
p
a
th
y
s
creen
in
g
…
(
Gu
id
o
u
m
A
min
a
)
243
b.
A
n
o
v
el
ar
ch
itectu
r
al
h
y
b
r
id
izatio
n
(
Sq
u
ee
ze
Net
-
Mo
b
ile
Vit
)
p
r
o
d
u
cin
g
a
m
o
d
el
1
2
×
m
o
r
e
co
m
p
ac
t
(
8
.
2
7
MB)
th
an
ex
is
tin
g
m
o
b
il
e
s
o
lu
tio
n
s
.
c.
T
h
e
f
ir
s
t
c
o
m
p
lete
o
p
er
atio
n
al
v
alid
atio
n
in
clu
d
in
g
laten
cy
(
1
5
.
4
3
m
s
)
,
C
PU
(
2
1
.
5
%),
an
d
ex
p
lain
a
b
ilit
y
(
Gr
ad
-
C
AM
)
-
m
etr
ics ab
s
en
t
f
r
o
m
p
r
ev
io
u
s
s
tu
d
ies.
T
ab
le
4
.
C
o
m
p
a
r
is
o
n
ch
ar
t
of
E
d
g
eRetin
a
v
s
.
s
tate
o
f
th
e
ar
t
C
r
i
t
e
r
i
o
n
O
u
r
so
l
u
t
i
o
n
E
d
g
e
R
e
t
i
n
a
Ex
i
s
t
i
n
g
w
o
r
k
r
e
p
r
e
se
n
t
a
t
i
v
e
A
d
v
a
n
t
a
g
e
/
I
n
n
o
v
a
t
i
o
n
o
f
Ed
g
e
R
e
t
i
n
a
A
r
c
h
i
t
e
c
t
u
r
e
S
q
u
e
e
z
e
N
e
t
+
M
o
b
i
l
e
V
i
T
R
e
sN
e
t
/
I
n
c
e
p
t
i
o
n
V
3
(
~
4
0
M
p
a
r
a
m)
[
1
0
]
10
-
1
6
x
p
a
r
a
me
t
e
r
r
e
d
u
c
t
i
o
n
(
1
.
4
M
p
a
r
a
me
t
e
r
s)
−
D
e
n
seN
e
t
1
6
9
(
1
3
M
p
a
r
a
m)
[
5
]
I
n
n
o
v
a
t
i
v
e
V
i
T
m
o
b
i
l
e
i
n
t
e
g
r
a
t
i
o
n
−
TA
H
D
L
(
C
N
N
+
R
N
N
)
[
6
]
,
Ef
f
i
c
i
e
n
t
N
e
t
B
3
,
Ef
f
i
c
i
e
n
t
N
e
t
V
2
B
1
,
R
e
g
N
e
t
X
0
0
8
,
R
e
g
N
e
t
X
0
8
0
,
R
e
g
N
e
t
Y
0
0
6
,
R
e
g
N
e
t
Y
0
0
8
[
4
]
P
r
e
t
r
e
a
t
me
n
t
M
i
n
i
m
a
l
i
st
(
1
2
8
×
1
2
8
+
n
o
r
m
a
l
i
z
a
t
i
o
n
)
−
C
LA
H
E
[
4
]
+
R
o
I
seg
me
n
t
a
t
i
o
n
[
6
]
El
i
m
i
n
a
t
e
s
c
o
m
p
l
e
x
st
e
p
s
−
G
a
b
o
r
f
i
l
t
e
r
s
[
7
]
,
P
C
A
[
7
]
,
G
a
u
ss
i
a
n
f
i
l
t
e
r
i
n
g
[
6
]
S
u
i
t
a
b
l
e
f
o
r
l
o
w
-
c
o
s
t
m
o
b
i
l
e
p
h
o
n
e
s
O
p
t
i
mi
z
a
t
i
o
n
D
y
n
a
mi
c
t
h
r
e
s
h
o
l
d
i
n
g
(
0
.
3
)
+
c
a
l
i
b
r
a
t
e
d
I
N
T8
N
o
t
m
e
n
t
i
o
n
e
d
[
4
]
–
[
6
]
,
[
1
0
]
-
7
1
%
f
a
l
se
p
o
si
t
i
v
e
s
S
i
z
e
mo
d
e
l
8
.
2
7
M
B
(
I
N
T8
q
u
a
n
t
i
z
e
d
)
N
o
t
a
d
d
r
e
ss
e
d
[
4
]
,
[
6
]
,
[
1
0
]
12
-
6
4
×
C
o
mp
r
e
ss
i
o
n
,
R
e
a
l
i
s
t
i
c
mo
b
i
l
e
d
e
p
l
o
y
me
n
t
C
l
i
n
i
c
a
l
p
e
r
f
o
r
m
a
n
c
e
S
e
n
s
i
t
i
v
i
t
y
:
9
0
,
7
%
i
n
t
8
A
U
C
:
0
.
9
6
(
I
N
T8
)
F
1
:
0
.
8
9
9
A
U
C
:
0
.
9
8
(
R
e
sN
e
t
1
5
2
)
[
1
2
]
P
r
e
c
i
s
:
Ef
f
i
c
a
c
i
t
é
N
e
t
B
3
8
5
,
1
%
[
4
]
F
1
:
0
.
8
8
6
(
R
e
sN
e
t
1
5
2
)
[
1
2
]
A
c
c
:
9
4
%
,
9
5
%
[
1
0
]
,
M
a
c
r
o
F
1
-
sc
o
r
e
:
0
.
6
5
,
a
c
c
:
0
.
8
2
[
5
]
P
r
e
c
i
s
:
9
4
%
K
a
g
g
l
e
D
R
d
a
t
a
set
[
6
]
F
o
c
u
s re
c
a
l
l
(
s
e
v
e
r
e
c
a
se
d
e
t
e
c
t
i
o
n
)
C
o
n
st
r
a
i
n
e
d
b
a
l
a
n
c
e
d
me
t
r
i
c
s
C
P
U
c
o
n
su
mp
t
i
o
n
2
1
.
5
%
(
I
N
T8
)
N
o
t
a
d
d
r
e
ss
e
d
[
4
]
–
[
6
]
,
[
1
0
]
,
[
1
3
]
-
4
3
%
c
o
m
p
a
r
e
d
t
o
o
u
r
F
P
3
2
m
o
d
e
D
e
p
l
o
y
me
n
t
t
a
r
g
e
t
Lo
w
-
c
o
st
smar
t
p
h
o
n
e
s
S
e
r
v
e
r
s
/
C
l
o
u
d
[
8
]
,
n
o
t
c
i
t
e
d
[
4
]
,
[
6
]
,
[
1
0
]
S
p
e
c
i
a
l
i
z
e
d
H
a
r
d
w
a
r
e
[
1
2
]
S
o
l
u
t
i
o
n
f
o
r
v
u
l
n
e
r
a
b
l
e
a
r
e
a
s
R
e
a
l
-
t
i
me
d
e
t
e
c
t
i
o
n
(
1
5
ms)
Ex
p
l
a
i
n
a
b
i
l
i
t
y
G
r
a
d
-
C
A
M
c
o
m
p
a
t
i
b
l
e
I
N
T8
/
F
P
3
2
N
o
t
a
d
d
r
e
ss
e
d
[
4
]
–
[
8
]
,
[
1
0
]
,
[
1
3
]
C
l
i
n
i
c
a
l
l
y
a
u
d
i
t
a
b
l
e
v
i
s
u
a
l
i
z
a
t
i
o
n
s
T
h
is
s
tu
d
y
ad
d
r
ess
es
th
e
f
u
n
d
am
en
tal
s
cien
tific
q
u
esti
o
n
o
f
h
o
w
to
d
ep
lo
y
ac
cu
r
ate,
e
f
f
i
cien
t,
an
d
clin
ically
v
iab
le
AI
-
b
ased
d
i
ab
etic
r
etin
o
p
ath
y
s
cr
ee
n
in
g
o
n
lo
w
-
co
s
t
m
o
b
ile
d
ev
ices
in
r
eso
u
r
ce
-
lim
ite
d
s
ettin
g
s
.
Ou
r
m
ain
f
in
d
in
g
d
em
o
n
s
tr
ates
th
at
a
h
o
lis
t
ic
ap
p
r
o
ac
h
—
co
m
b
in
i
n
g
a
lig
h
tweig
h
t
h
y
b
r
id
ar
ch
itectu
r
e
(
Sq
u
ee
ze
Net
-
Mo
b
ileViT
)
,
d
y
n
am
ic
d
ec
is
io
n
th
r
esh
o
ld
ca
lib
r
atio
n
,
an
d
p
o
s
t
-
tr
ain
in
g
I
NT
8
q
u
an
tific
atio
n
—
ac
h
iev
es
s
im
u
ltan
eo
u
s
ly
h
ig
h
clin
ical
p
e
r
f
o
r
m
a
n
ce
(
AUC=0
.
9
6
,
Sen
s
itiv
ity
=9
0
.
7
%)
a
n
d
r
em
ar
k
ab
le
o
p
er
atio
n
al
e
f
f
icie
n
cy
(
8
.
2
7
MB
m
o
d
el
s
ize,
4
3
%
r
ed
u
ctio
n
in
C
PU
u
tili
za
tio
n
)
.
T
h
ese
f
in
d
i
n
g
s
ar
e
s
u
p
p
o
r
ted
b
y
q
u
an
titativ
e
ev
id
en
ce
f
r
o
m
th
e
APTO
S
2
0
1
9
d
ataset,
in
clu
d
in
g
s
tan
d
ar
d
clin
ical
an
d
o
p
er
atio
n
al
m
etr
ics,
as
well
as
q
u
alitativ
e
v
is
u
aliza
tio
n
s
(
Gr
ad
-
C
AM
)
co
m
p
ar
in
g
o
u
r
FP
3
2
an
d
I
NT
8
m
o
d
els.
T
h
e
s
ig
n
if
ican
ce
o
f
t
h
ese
f
in
d
i
n
g
s
lies
in
estab
lis
h
in
g
a
p
r
ac
t
ical
b
lu
ep
r
in
t f
o
r
th
e
d
ev
elo
p
m
en
t
o
f
ed
g
e
-
n
ativ
e
m
ed
ical
AI
s
o
lu
tio
n
s
,
d
e
m
o
n
s
tr
atin
g
th
at
o
p
tim
izatio
n
g
u
i
d
ed
b
y
clin
ical
n
ee
d
s
ca
n
im
p
r
o
v
e
p
er
f
o
r
m
an
ce
wh
ile
r
ed
u
cin
g
th
e
co
m
p
u
tati
o
n
al
f
o
o
tp
r
in
t.
I
m
p
o
r
tan
tly
,
t
h
e
ch
o
ice
o
f
th
e
d
y
n
a
m
ic
th
r
esh
o
ld
(
m
ed
ia
n
0
.
3
)
r
ep
r
esen
ts
a
d
elib
e
r
ate
clin
ical
tr
ad
e
-
o
f
f
,
o
p
tim
ized
f
o
r
m
ass
s
cr
ee
n
in
g
s
ettin
g
s
wh
er
e
r
ed
u
cin
g
f
alse
p
o
s
itiv
es
is
cr
u
cial
to
av
o
id
o
v
er
lo
a
d
in
g
h
ea
lth
ca
r
e
s
y
s
tem
s
.
R
eg
ar
d
in
g
r
ea
l
-
wo
r
ld
im
p
lem
en
tatio
n
ch
allen
g
es,
o
u
r
E
d
g
e
R
etin
a
s
o
lu
tio
n
is
p
r
im
ar
ily
d
esig
n
ed
f
o
r
m
o
b
ile
-
b
ased
o
p
er
atio
n
to
o
v
e
r
co
m
e
co
n
n
ec
tiv
ity
is
s
u
es in
r
eso
u
r
ce
-
lim
ited
ar
ea
s
.
Ho
wev
er
,
it is
im
p
o
r
tan
t to
n
o
te
th
at
th
e
p
ip
elin
e
c
o
u
ld
a
ls
o
b
e
d
ep
lo
y
e
d
as
a
h
y
b
r
i
d
o
r
clo
u
d
-
b
ased
s
y
s
tem
f
o
r
ce
n
tr
alize
d
an
aly
s
is
if
in
ter
n
et
in
f
r
astru
ctu
r
e
is
av
ailab
le.
T
h
e
m
ain
c
h
allen
g
e
ad
d
r
ess
ed
co
n
ce
r
n
s
th
e
tr
a
d
e
-
o
f
f
in
h
er
e
n
t
in
ca
lib
r
atin
g
th
e
d
ec
is
io
n
th
r
esh
o
ld
; o
u
r
ch
o
ice
o
f
a
th
r
esh
o
l
d
o
f
0
.
3
p
r
i
o
r
itizes r
ed
u
cin
g
f
alse p
o
s
itiv
es to
av
o
id
h
ea
lth
ca
r
e
s
y
s
tem
o
v
er
lo
ad
,
w
h
ile
ac
ce
p
tin
g
a
co
n
tr
o
lled
in
c
r
ea
s
e
in
f
alse
n
eg
ativ
es
—
a
cr
u
cial
co
n
s
id
er
atio
n
f
o
r
r
ea
l
-
w
o
r
ld
s
cr
ee
n
i
n
g
p
r
o
g
r
am
s
.
T
h
ese
ch
ar
ac
ter
is
tics
ar
e
n
o
t
o
n
l
y
ad
v
an
tag
eo
u
s
f
o
r
m
o
b
ile
p
h
o
n
es
b
u
t
ar
e
also
ess
en
tial
p
r
er
eq
u
is
ites
f
o
r
th
eir
in
teg
r
atio
n
in
t
o
r
ea
l
-
ti
m
e
r
o
b
o
tic
s
y
s
tem
s
.
T
h
e
lo
w
in
f
er
e
n
ce
laten
c
y
(
1
5
.
4
3
m
s
)
en
s
u
r
es
th
at
im
ag
e
an
aly
s
is
d
o
es
n
o
t
b
ec
o
m
e
a
b
o
ttlen
ec
k
in
a
d
y
n
am
ic
r
o
b
o
tic
co
n
tr
o
l
p
r
o
ce
s
s
,
th
u
s
en
ab
lin
g
im
m
e
d
iate
f
ee
d
b
ac
k
an
d
ac
tio
n
.
T
h
e
s
ig
n
i
f
ican
tly
r
ed
u
ce
d
C
PU
co
n
s
u
m
p
tio
n
allo
ws
th
e
d
iag
n
o
s
tic
m
o
d
el
to
o
p
er
ate
e
f
f
icien
tly
alo
n
g
s
id
e
o
t
h
er
cr
iti
ca
l
r
o
b
o
tic
p
r
o
ce
s
s
es,
s
u
ch
as
s
en
s
o
r
d
ata
f
u
s
io
n
an
d
m
o
t
o
r
co
n
tr
o
l,
with
o
u
t o
v
er
lo
ad
in
g
t
h
e
s
y
s
tem
'
s
lim
ited
co
m
p
u
tin
g
r
eso
u
r
ce
s
.
T
h
is
h
o
lis
tic
ap
p
r
o
ac
h
th
u
s
f
il
ls
a
cr
itical
g
ap
f
o
r
s
cr
ee
n
in
g
in
p
r
ec
ar
io
u
s
en
v
ir
o
n
m
en
ts
.
I
t
ad
d
r
ess
es
th
e
d
u
al
ch
allen
g
e
o
f
clin
ical
p
er
f
o
r
m
a
n
ce
an
d
o
p
er
atio
n
al
co
n
s
tr
ain
ts
th
at
lim
ited
p
r
ev
io
u
s
s
o
lu
tio
n
s
.
T
h
e
in
teg
r
atio
n
o
f
d
y
n
am
ic
th
r
es
h
o
ld
in
g
with
q
u
an
tific
atio
n
r
ep
r
esen
ts
a
n
ew
p
ar
ad
ig
m
f
o
r
m
o
b
ile
h
ea
lth
ap
p
licatio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.