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.
15
,
No
.
4
,
A
u
g
u
s
t
20
26
,
p
p
.
3
7
4
5
~
3
7
6
0
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
4
.
p
p
3
7
4
5
-
3
7
6
0
3745
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Ada
ptive
bina
ry
c
a
psule networ
k f
o
r comp
le
x
ima
g
e
recog
nition
M
a
v
is
Serwa
a
Yebo
a
h
1,
2
,
P
a
t
rick
K
wa
bena
M
ens
a
h
1
,
Ad
eba
y
o
F
elix
Adeko
y
a
3
,
M
ig
h
t
y
Abra
Ay
idzo
e
1,
4
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
I
n
f
o
r
ma
t
i
c
s,
U
n
i
v
e
r
s
i
t
y
o
f
E
n
e
r
g
y
a
n
d
N
a
t
u
r
a
l
R
e
s
o
u
r
c
e
s,
S
u
n
y
a
n
i
,
G
h
a
n
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
,
S
u
n
y
a
n
i
T
e
c
h
n
i
c
a
l
U
n
i
v
e
r
si
t
y
,
S
u
n
y
a
n
i
,
G
h
a
n
a
3
F
a
c
u
l
t
y
o
f
C
o
mp
u
t
i
n
g
,
E
n
g
i
n
e
e
r
i
n
g
a
n
d
M
a
t
h
e
m
a
t
i
c
a
l
S
c
i
e
n
c
e
s
,
C
a
t
h
o
l
i
c
U
n
i
v
e
r
si
t
y
o
f
G
h
a
n
a
,
S
u
n
y
a
n
i
,
G
h
a
n
a
4
S
c
h
o
o
l
o
f
T
e
c
h
n
o
l
o
g
y
,
J
a
m
k
U
n
i
v
e
r
s
i
t
y
o
f
A
p
p
l
i
e
d
S
c
i
e
n
c
e
s,
Jy
v
ä
sk
y
l
ä
,
F
i
n
l
a
n
d
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Dec
5
,
2
0
2
4
R
ev
is
ed
J
u
n
3
,
2
0
2
6
Acc
ep
ted
J
u
l
9
,
2
0
2
6
G
lau
c
o
m
a
a
n
d
c
a
tara
c
ts
a
re
lea
d
in
g
c
a
u
se
s
o
f
b
li
n
d
n
e
ss
wo
rld
wi
d
e
,
e
m
p
h
a
siz
in
g
th
e
n
e
e
d
f
o
r
e
a
rly
d
e
tec
ti
o
n
.
C
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s
(CNN
s)
h
a
v
e
sh
o
wn
p
r
o
m
ise
in
d
e
tec
ti
n
g
o
c
u
lar
d
ise
a
se
s,
b
u
t
re
q
u
ire
e
x
ten
siv
e
trai
n
in
g
d
a
tas
e
ts.
Ho
w
e
v
e
r,
m
e
d
ica
l
d
a
tas
e
ts
a
re
sc
a
rc
e
,
li
m
it
e
d
,
a
n
d
imb
a
lan
c
e
d
,
p
ro
m
p
ti
n
g
th
e
u
se
o
f
ti
m
e
-
c
o
n
su
m
i
n
g
d
a
ta
a
u
g
m
e
n
tatio
n
tec
h
n
iq
u
e
s.
To
a
d
d
re
ss
th
e
se
li
m
it
a
ti
o
n
s,
a
c
o
m
p
u
tatio
n
a
ll
y
e
ff
icie
n
t
a
n
d
ro
b
u
st
c
a
p
su
le
n
e
tw
o
rk
(Ca
p
sN
e
t)
m
o
d
e
l
wa
s
p
r
o
p
o
se
d
.
Th
e
m
o
d
e
l
fe
a
tu
re
s
a
n
o
v
e
l
a
d
a
p
ti
v
e
c
o
n
tras
t
sp
a
ti
a
l
f
il
terin
g
(ACSF
)
a
lg
o
rit
h
m
a
n
d
i
n
c
o
rp
o
ra
tes
a
lo
c
a
l
b
i
n
a
ry
p
a
tt
e
rn
(LB
P
)
a
l
g
o
rit
h
m
to
e
n
h
a
n
c
e
r
o
b
u
stn
e
ss
.
T
h
is
st
u
d
y
a
c
h
iev
e
d
g
o
o
d
re
c
o
g
n
i
ti
o
n
a
c
c
u
ra
c
y
,
wit
h
sc
o
re
s
o
f
9
6
.
9
0
%
o
n
th
e
c
o
m
b
in
e
d
d
a
tas
e
t,
9
6
.
4
5
%
o
n
th
e
c
a
tara
c
t
-
o
n
ly
d
a
tas
e
t,
a
n
d
9
6
.
7
8
%
o
n
th
e
g
lau
c
o
m
a
-
o
n
l
y
d
a
tas
e
t.
T
h
e
m
o
d
e
l'
s
p
e
rf
o
rm
a
n
c
e
is
c
o
m
p
a
ra
b
le
to
sta
te
-
of
-
th
e
-
a
rt
m
o
d
e
ls,
d
e
m
o
n
stra
ti
n
g
it
s
p
o
ten
ti
a
l
t
o
su
p
p
o
rt
o
p
h
th
a
lmo
l
o
g
ists
i
n
d
iag
n
o
sin
g
c
a
tara
c
ts an
d
g
lau
c
o
m
a
-
re
late
d
e
y
e
issu
e
s.
K
ey
w
o
r
d
s
:
Ad
ap
tiv
e
co
n
tr
ast
C
ap
s
u
le
n
etwo
r
k
C
o
m
p
lex
im
ag
es
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
L
o
ca
l b
in
ar
y
p
atter
n
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
:
Ma
v
is
Ser
waa
Yeb
o
ah
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
I
n
f
o
r
m
atics,
Un
iv
er
s
ity
o
f
E
n
er
g
y
an
d
Natu
r
al
R
eso
u
r
ce
s
Su
n
y
an
i,
Gh
a
n
a
E
m
ail:
m
av
is
.
s
er
waa
y
eb
o
ah
@
s
tu
.
ed
u
.
g
h
1.
I
NT
RO
D
UCT
I
O
N
Glau
co
m
a
r
ep
r
esen
ts
a
cr
itical
g
lo
b
al
h
ea
lth
ch
allen
g
e
as
a
lead
in
g
ca
u
s
e
o
f
ir
r
ev
er
s
ib
le
b
lin
d
n
ess
,
ch
ar
ac
ter
ized
b
y
p
r
o
g
r
ess
iv
e
o
p
tic
n
er
v
e
d
eg
en
e
r
atio
n
th
at,
u
n
lik
e
ca
tar
ac
ts
,
ca
n
n
o
t
b
e
r
ev
er
s
ed
th
r
o
u
g
h
s
u
r
g
ical
in
ter
v
en
tio
n
.
W
ith
p
r
o
jectio
n
s
esti
m
atin
g
th
at
1
1
1
m
illi
o
n
p
eo
p
le
will
b
e
af
f
ec
ted
b
y
g
lau
co
m
a
b
y
2
0
4
0
[
1
]
,
e
ar
ly
d
etec
tio
n
h
as
b
ec
o
m
e
p
a
r
am
o
u
n
t
to
p
r
ev
e
n
tin
g
p
er
m
an
e
n
t
v
is
io
n
lo
s
s
.
T
h
is
u
r
g
en
c
y
h
as
d
r
iv
en
th
e
d
e
v
elo
p
m
e
n
t
o
f
in
tellig
en
t
s
cr
ee
n
in
g
an
d
d
ia
g
n
o
s
tic
s
y
s
tem
s
ca
p
ab
le
o
f
id
en
tify
in
g
o
cu
lar
ab
n
o
r
m
alities
b
e
f
o
r
e
ir
r
ev
e
r
s
ib
le
d
am
ag
e
o
cc
u
r
s
.
Ho
we
v
er
,
th
e
ef
f
ec
tiv
e
n
ess
o
f
t
h
ese
s
y
s
tem
s
d
ep
en
d
s
f
u
n
d
am
e
n
tally
o
n
t
h
eir
ab
ilit
y
to
ac
cu
r
ately
d
is
tin
g
u
is
h
b
et
wee
n
m
u
ltip
le
o
cu
lar
c
o
n
d
itio
n
s
wh
ile
o
p
er
atin
g
with
in
th
e
co
n
s
tr
ain
ts
o
f
r
ea
l
-
wo
r
ld
clin
ical
s
ettin
g
s
.
Dee
p
lear
n
in
g
alg
o
r
ith
m
s
,
p
a
r
ticu
lar
ly
co
n
v
o
lu
tio
n
al
n
e
u
r
a
l
n
etwo
r
k
s
(
C
NNs),
h
av
e
d
em
o
n
s
tr
ated
r
em
ar
k
ab
le
s
u
cc
ess
in
id
en
tif
y
in
g
o
cu
lar
a
b
n
o
r
m
alities
f
r
o
m
r
etin
al
im
ag
es.
R
ec
en
t
ad
v
an
ce
s
in
clu
d
e
h
y
b
r
id
C
NN
-
ad
ap
tiv
e
b
o
o
s
tin
g
(
Ad
a
B
o
o
s
t
)
m
o
d
els
ac
h
iev
in
g
9
2
.
9
6
%
ac
cu
r
ac
y
o
n
g
lau
co
m
a
d
atasets
[
2
]
,
tr
an
s
f
er
lear
n
in
g
ap
p
r
o
ac
h
es
with
p
r
e
-
tr
ain
ed
ar
ch
itectu
r
es
(
r
esid
u
al
n
etwo
r
k
5
0
(
R
esNet
-
50)
,
v
is
u
al
g
eo
m
etr
y
g
r
o
u
p
1
9
(
VGG
-
19
)
,
d
en
s
ely
co
n
n
ec
ted
co
n
v
o
lu
ti
o
n
al
n
etwo
r
k
2
0
1
(
Den
s
eNe
t
-
201
)
)
y
iel
d
in
g
ac
cu
r
ac
ies
u
p
to
9
9
.
5
7
%
[
3
]
,
an
d
f
in
e
-
tu
n
e
d
R
esNet
-
5
0
m
o
d
els
r
ea
ch
in
g
9
8
.
4
8
%
ac
cu
r
ac
y
th
r
o
u
g
h
d
ata
au
g
m
en
tatio
n
s
tr
ateg
ies
[
4
]
.
Desp
ite
th
ese
ac
h
iev
em
en
ts
,
C
NNs
f
ac
e
s
i
g
n
if
ican
t
ch
allen
g
es
r
elate
d
t
o
d
ata
s
ca
r
city
an
d
q
u
ality
.
Me
d
ical
im
ag
e
d
atasets
ar
e
ty
p
ically
lim
ited
in
s
ize
an
d
ch
ar
ac
ter
ized
b
y
class
im
b
alan
ce
,
lead
in
g
to
o
v
er
f
itti
n
g
a
n
d
r
ed
u
ce
d
g
e
n
er
aliza
tio
n
ca
p
ab
ilit
y
[
5
]
.
W
h
il
e
d
ata
au
g
m
en
tatio
n
tech
n
iq
u
es
p
ar
tially
ad
d
r
ess
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.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
7
4
5
-
3
7
6
0
3746
th
ese
lim
itatio
n
s
[
5
]
,
th
ey
d
o
n
o
t
f
u
n
d
am
e
n
tally
r
eso
lv
e
th
e
ar
ch
itectu
r
al
co
n
s
tr
ain
ts
o
f
C
NNs
in
ca
p
tu
r
in
g
s
p
atial
h
ier
ar
ch
ies an
d
p
a
r
t
-
w
h
o
le
r
elatio
n
s
h
ip
s
in
m
e
d
ical
i
m
ag
es.
C
ap
s
u
le
n
etwo
r
k
s
(
C
ap
s
Nets)
h
av
e
e
m
er
g
ed
as
a
p
r
o
m
is
in
g
alter
n
ativ
e
ar
c
h
itectu
r
e
t
h
at
r
et
ain
s
C
NN
s
tr
en
g
th
s
wh
ile
ad
d
r
ess
in
g
d
ata
-
r
elate
d
ch
allen
g
es
th
r
o
u
g
h
th
eir
u
n
iq
u
e
ab
ilit
y
to
p
r
eser
v
e
s
p
atial
r
elatio
n
s
h
ip
s
an
d
ad
ap
t
to
s
m
aller
d
atasets
.
I
n
itial
ap
p
li
ca
tio
n
s
to
o
cu
lar
d
is
ea
s
e
d
iag
n
o
s
is
h
av
e
s
h
o
wn
en
co
u
r
a
g
in
g
r
esu
lts
:
th
e
o
r
ig
i
n
al
C
ap
s
Net
with
d
y
n
am
ic
r
o
u
tin
g
ac
h
iev
ed
9
0
.
9
0
%
ac
c
u
r
ac
y
o
n
g
la
u
co
m
a
d
atasets
[
6
]
,
m
o
d
if
ie
d
3
D
C
ap
s
Nets
at
tain
ed
0
.
8
9
s
en
s
itiv
ity
o
n
o
p
tical
co
h
er
en
ce
to
m
o
g
r
a
p
h
(
OC
T
)
im
ag
es
[
7
]
,
a
n
d
en
h
a
n
ce
d
C
ap
s
Net
s
with
f
ea
tu
r
e
p
r
o
ce
s
s
i
n
g
alg
o
r
ith
m
s
d
em
o
n
s
tr
ated
p
r
ec
is
io
n
v
al
u
es
o
f
0
.
7
9
6
f
o
r
g
lau
c
o
m
a
a
n
d
0
.
8
1
8
f
o
r
ca
tar
ac
ts
.
Ho
wev
er
,
C
ap
s
Nets
r
em
ain
r
elativ
ely
u
n
e
x
p
lo
r
ed
f
o
r
o
c
u
lar
d
is
ea
s
e
r
ec
o
g
n
itio
n
an
d
s
u
f
f
e
r
f
r
o
m
cr
itical
lim
itatio
n
s
in
th
eir
u
p
p
e
r
en
co
d
e
r
n
etwo
r
k
[
8
]
an
d
in
s
u
f
f
icien
t
f
ea
tu
r
e
p
r
o
ce
s
s
in
g
ca
p
ab
ilit
ies
[
9
]
,
r
estrictin
g
th
eir
clin
ical
a
p
p
licab
ilit
y
d
esp
ite
th
eir
th
e
o
r
etica
l a
d
v
an
tag
es.
T
h
is
p
ap
er
ad
d
r
ess
es
th
ese
li
m
itatio
n
s
b
y
p
r
o
p
o
s
in
g
a
n
en
h
an
ce
d
C
ap
s
Net
ar
ch
itectu
r
e
s
p
ec
if
ically
d
esig
n
ed
f
o
r
m
u
lti
-
class
o
cu
l
ar
d
is
ea
s
e
d
iag
n
o
s
is
,
d
is
tin
g
u
is
h
in
g
b
etwe
en
g
lau
c
o
m
a,
ca
tar
ac
ts
,
an
d
h
ea
lth
y
co
n
d
itio
n
s
.
T
h
is
s
tu
d
y
i
n
tr
o
d
u
ce
th
e
ad
ap
tiv
e
co
n
tr
ast
s
p
a
tial
f
ilter
in
g
(
AC
SF
)
alg
o
r
ith
m
to
s
tr
en
g
th
en
th
e
u
p
p
er
e
n
co
d
e
r
n
etwo
r
k
th
r
o
u
g
h
th
r
ee
in
teg
r
ated
co
m
p
o
n
e
n
ts
:
ad
ap
tiv
e
th
r
esh
o
l
d
in
g
f
o
r
u
n
if
o
r
m
f
ea
tu
r
e
m
ap
d
is
tr
ib
u
tio
n
,
Gau
s
s
ian
k
er
n
el
g
en
er
atio
n
f
o
r
n
o
is
e
-
ad
ap
ti
v
e
s
m
o
o
th
i
n
g
,
a
n
d
s
p
atial
f
i
lter
in
g
f
o
r
c
o
n
tr
ast
en
h
an
ce
m
e
n
t.
Ad
d
itio
n
ally
,
th
is
s
tu
d
y
in
co
r
p
o
r
ate
a
lo
ca
l
b
in
ar
y
p
atter
n
(
L
B
P)
lay
er
to
c
ap
tu
r
e
f
in
e
-
g
r
ain
ed
tex
tu
r
al
d
etails,
im
p
r
o
v
in
g
c
o
m
p
u
tatio
n
al
ef
f
icie
n
cy
an
d
r
o
b
u
s
tn
ess
to
illu
m
in
atio
n
v
a
r
iatio
n
s
co
m
m
o
n
ly
en
co
u
n
ter
e
d
in
clin
ical
f
u
n
d
u
s
im
ag
in
g
.
T
h
e
p
r
im
ar
y
c
o
n
tr
ib
u
tio
n
s
o
f
th
is
r
esear
ch
in
clu
d
e:
i)
Dev
elo
p
m
en
t
o
f
a
c
o
m
p
u
tatio
n
ally
ef
f
icien
t
d
ee
p
lear
n
in
g
m
o
d
el
s
u
itab
le
f
o
r
co
m
p
lex
r
ea
l
-
wo
r
l
d
m
ed
ical
im
ag
e
r
ec
o
g
n
itio
n
with
lim
ited
d
ata
av
ailab
ilit
y
.
ii)
I
n
tr
o
d
u
ctio
n
o
f
th
e
AC
SF
alg
o
r
ith
m
th
at
en
h
an
ce
s
f
ea
tu
r
e
r
ep
r
esen
tatio
n
q
u
ality
wh
ile
r
ed
u
cin
g
n
o
is
e
in
ter
f
er
en
ce
.
iii)
I
n
teg
r
atio
n
o
f
lo
ca
l
tex
t
u
r
e
p
atter
n
e
x
tr
ac
tio
n
to
im
p
r
o
v
e
m
o
d
el
r
esil
ien
ce
a
g
ain
s
t
illu
m
in
atio
n
v
ar
iatio
n
s
.
iv
)
Pro
v
is
io
n
o
f
a
s
y
s
tem
atic
ap
p
r
o
ac
h
t
o
ad
d
r
ess
th
e
en
co
d
e
r
n
etwo
r
k
wea
k
n
ess
es
in
h
er
e
n
t
in
s
tan
d
ar
d
C
ap
s
Net
ar
ch
itectu
r
es f
o
r
o
cu
l
ar
d
is
ea
s
e
an
aly
s
is
.
T
h
e
ex
p
e
r
im
en
tal
r
esu
lts
d
e
m
o
n
s
tr
ate
p
er
f
o
r
m
an
ce
co
m
p
ar
ab
le
to
s
tate
-
of
-
th
e
-
ar
t
te
ch
n
iq
u
es,
o
f
f
er
in
g
m
ed
ical
s
p
ec
ialis
ts
a
r
eliab
le
d
iag
n
o
s
tic
to
o
l f
o
r
ac
cu
r
ate
i
d
e
n
tific
atio
n
o
f
g
lau
co
m
a
an
d
ca
tar
ac
ts
.
T
h
e
r
em
ain
d
e
r
o
f
th
is
p
ap
er
i
s
o
r
g
an
ized
as
f
o
ll
o
ws
.
Sectio
n
2
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
o
lo
g
y
.
Sectio
n
3
p
r
esen
ts
ex
p
e
r
im
en
t
al
r
esu
lts
an
d
d
is
cu
s
s
es
th
e
f
in
d
in
g
s
.
Fin
ally
,
s
ec
tio
n
4
co
n
cl
u
d
es th
e
p
ap
e
r
.
2.
M
E
T
H
O
D
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
th
e
p
r
o
p
o
s
ed
m
o
d
el,
th
e
d
ataset
d
escr
ip
tio
n
,
an
d
th
e
ex
p
er
im
en
tal
s
etu
p
.
2
.
1
.
P
r
o
po
s
ed
f
ea
t
ure
pro
ce
s
s
ing
m
et
ho
d
T
h
e
p
r
o
p
o
s
ed
f
ea
tu
r
e
p
r
o
ce
s
s
in
g
m
eth
o
d
is
ter
m
ed
AC
SF
(
s
ee
Alg
o
r
ith
m
s
1
an
d
2
)
,
wh
ich
is
im
p
lem
en
ted
as
a
lay
er
in
th
e
en
co
d
er
n
etwo
r
k
o
f
th
e
p
r
o
p
o
s
ed
C
ap
s
Net
m
o
d
el.
T
h
is
s
tu
d
y
ad
o
p
t
th
e
co
n
ce
p
t
o
f
ad
a
p
tiv
e
co
n
tr
ast
an
d
s
p
atial
f
ilter
in
g
to
d
ev
el
o
p
o
u
r
p
r
o
p
o
s
ed
m
et
h
o
d
.
T
h
e
p
r
o
p
o
s
e
d
im
ag
e
p
r
o
ce
s
s
in
g
alg
o
r
ith
m
is
d
esig
n
ed
t
o
en
h
a
n
ce
th
e
co
n
t
r
ast
o
f
in
p
u
t
im
ag
es
wh
ile
r
ed
u
cin
g
n
o
is
e.
T
h
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
co
n
s
is
ts
o
f
th
r
ee
m
ain
s
tep
s
:
ad
ap
tiv
e
th
r
esh
o
ld
in
g
,
Gau
s
s
ian
k
er
n
el
g
en
e
r
atio
n
,
an
d
s
p
a
tial
f
ilter
in
g
.
I
n
th
e
ad
ap
tiv
e
th
r
esh
o
ld
in
g
s
tep
(
1
)
to
(
3
)
,
th
e
f
ea
tu
r
e
m
ap
s
a
r
e
n
o
r
m
alize
d
to
h
av
e
ze
r
o
m
ea
n
(
1
)
an
d
u
n
it
v
ar
ian
ce
(
2
)
.
T
h
is
is
ac
h
iev
e
d
b
y
s
u
b
tr
ac
tin
g
th
e
m
ea
n
v
alu
e
f
r
o
m
ea
ch
p
ix
el
a
n
d
th
en
d
iv
id
in
g
b
y
th
e
s
tan
d
ar
d
d
ev
iatio
n
.
T
h
e
r
esu
ltin
g
f
ea
tu
r
e
m
ap
s
h
av
e
a
u
n
if
o
r
m
d
is
tr
ib
u
tio
n
,
m
ak
in
g
th
e
m
s
u
itab
le
f
o
r
s
p
atial
f
ilter
in
g
.
I
n
th
e
Gau
s
s
ian
k
er
n
el
g
en
er
atio
n
s
tep
(
4
)
to
(
9
)
,
a
Gau
s
s
ian
k
er
n
el
is
g
en
er
ated
with
a
s
p
ec
if
ied
s
ize
an
d
s
tan
d
ar
d
d
ev
iatio
n
(
5
)
to
(
7
)
.
T
h
e
Gau
s
s
ian
k
er
n
els
ar
e
s
m
o
o
th
in
g
f
ilter
s
th
at
r
ed
u
ce
n
o
is
e
in
th
e
f
ea
tu
r
e
m
ap
s
.
T
h
e
s
ize
o
f
th
e
k
er
n
els
d
eter
m
in
es
th
e
am
o
u
n
t
o
f
s
m
o
o
th
in
g
,
wh
ile
th
e
s
tan
d
ar
d
d
ev
iatio
n
co
n
tr
o
ls
th
e
s
p
r
ea
d
o
f
t
h
e
k
er
n
els.
T
h
e
Gau
s
s
ian
k
er
n
el
g
en
er
ati
o
n
allo
ws
f
o
r
ad
ju
s
tab
le
s
m
o
o
t
h
in
g
,
en
ab
lin
g
t
h
e
alg
o
r
ith
m
to
a
d
ap
t to
d
if
f
er
e
n
t
n
o
is
e
lev
els.
I
n
th
e
s
p
atial
f
ilter
in
g
s
tep
(
1
0
)
to
(
1
5
)
,
th
e
n
o
r
m
alize
d
f
ea
tu
r
e
m
ap
s
ar
e
co
n
v
o
lv
ed
with
th
e
Gau
s
s
ian
k
er
n
el
t
o
p
r
o
d
u
ce
e
n
h
an
ce
d
f
ea
tu
r
e
m
ap
s
.
T
h
is
s
tep
co
m
b
in
es
th
e
n
o
r
m
alize
d
f
e
atu
r
e
m
a
p
s
with
th
e
s
m
o
o
th
in
g
f
ilter
s
,
r
esu
ltin
g
in
f
ea
tu
r
e
m
ap
s
with
en
h
an
ce
d
co
n
tr
ast
an
d
r
ed
u
ce
d
n
o
is
e.
W
ith
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
,
th
is
s
tu
d
y
ac
h
iev
e
t
h
e
f
o
llo
win
g
:
i)
I
m
p
r
o
v
ed
co
n
tr
ast:
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
en
h
an
ce
s
th
e
c
o
n
tr
ast
o
f
th
e
f
ea
t
u
r
e
m
a
p
s
,
m
ak
in
g
it
m
o
r
e
v
is
u
ally
ap
p
ea
lin
g
.
ii)
R
ed
u
ce
d
n
o
is
e:
th
e
Gau
s
s
ian
k
er
n
el
ef
f
ec
tiv
ely
r
e
d
u
ce
s
n
o
i
s
e,
r
esu
ltin
g
in
clea
n
er
an
d
m
o
r
e
ac
cu
r
ate
f
ea
tu
r
e
m
ap
s
.
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
A
d
a
p
tive
b
in
a
r
y
ca
p
s
u
le
n
etw
o
r
k
fo
r
co
mp
lex
ima
g
e
r
ec
o
g
n
itio
n
(
Ma
vis S
erw
a
a
Yeb
o
a
h
)
3747
iii)
Ad
ap
tab
ilit
y
:
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
ca
n
ad
ap
t
to
d
if
f
e
r
en
t
n
o
is
e
lev
els
an
d
f
ea
tu
r
e
m
ap
ch
ar
ac
ter
is
tics
,
m
ak
in
g
it a
r
o
b
u
s
t a
n
d
ef
f
ec
tiv
e
s
o
lu
tio
n
.
Alg
o
r
ith
m
1
: G
au
s
s
ian
k
er
n
el
g
en
er
atio
n
1:
=
⊲
2:
=
(
−
2
⁄
+
1
,
2
⁄
+
1
)
⊲
ℎ
3
: y
=
(
−
2
⁄
+
1
,
2
⁄
+
1
)
⊲
ℎ
4
:
,
=
ℎ
(
,
)
⊲
2
5
:
=
(
−
(
2
+
2
)
2
.
0
2
⁄
)
⊲
6
:
=
(
)
⁄
⊲
7
: o
u
tp
u
t g
Alg
o
r
ith
m
2
:
AC
SF
1
:
,
=
{
,
1
,
,
2
,
…
,
,
−
1
}
⊲
2
: Fo
r
k
in
,
3:
,
=
(
,
)
⊲
4:
,
=
(
,
−
,
)
,
⁄
∗
⊲
ℎ
ℎ
Gau
s
s
ian
k
er
n
el
g
e
n
e
r
atio
n
5:
,
=
(
,
,
)
⊲
=
,
=
,
=
ℎ
⊲
6
: r
etu
r
n
,
2
.
2
.
P
r
o
o
f
f
o
r
t
he
a
da
ptiv
e
c
o
ntr
a
s
t
s
pa
t
ia
l f
ilte
ring
a
lg
o
rit
hm
T
h
eo
r
em
: th
e
AC
SF
alg
o
r
ith
m
en
h
an
ce
s
th
e
c
o
n
tr
ast o
f
t
h
e
i
n
p
u
t im
ag
e
wh
ile
r
ed
u
ci
n
g
n
o
i
s
e.
Pro
o
f
:
L
et
(
,
)
b
e
th
e
in
p
u
t im
ag
e
,
an
d
(
,
)
b
e
th
e
o
u
tp
u
t im
ag
e.
Ad
ap
tiv
e
th
r
esh
o
l
d
in
g
:
L
et
(
,
)
b
e
th
e
ad
a
p
tiv
e
th
r
esh
o
ld
i
n
g
o
u
t
p
u
t.
(
,
)
=
(
(
,
)
−
)
/
(
∗
)
(
1
)
W
h
er
e
is
th
e
m
ea
n
,
is
th
e
s
tan
d
ar
d
d
ev
iatio
n
,
an
d
is
th
e
s
ca
lin
g
f
ac
to
r
.
L
em
m
a
1
:
th
e
ad
ap
tiv
e
th
r
es
h
o
ld
in
g
s
tep
n
o
r
m
alize
s
th
e
in
p
u
t
im
ag
e
to
h
a
v
e
ze
r
o
m
ea
n
an
d
u
n
it
v
a
r
ian
ce
.
Pro
o
f
:
[
(
,
)
]
=
[
(
(
,
)
−
)
/
(
∗
)
]
=
0
(
)
(
2
)
[
(
,
)
]
=
[
(
(
,
)
−
)
/
(
∗
)
]
=
1
(
)
(
3
)
Gau
s
s
ian
k
er
n
el
g
en
e
r
atio
n
:
L
et
(
,
)
b
e
th
e
Gau
s
s
ian
k
er
n
el
as in
(
4
)
.
(
,
)
=
(
1
/
(
2
∗
∗
2
)
)
∗
(
−
(
2
+
2
)
/
(
2
∗
2
)
)
(
4
)
L
em
m
a
2
:
th
e
Gau
s
s
ian
k
e
r
n
el
is
a
s
m
o
o
th
in
g
f
ilter
t
h
at
r
ed
u
ce
s
n
o
is
e.
Pro
o
f
:
L
et
(
,
)
b
e
th
e
n
o
is
e
co
m
p
o
n
en
t
o
f
th
e
in
p
u
t im
ag
e.
[
(
,
)
]
=
0
(
)
(
5
)
[
(
,
)
]
=
2
(
)
(
6
)
[
(
∗
)
(
,
)
]
=
0
(
)
(
7
)
=
2
(
)
∗
∑
(
,
)
(
,
)
2
(
8
)
[
(
∗
)
(
,
)
]
=
≤
2
(
)
(
9
)
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.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
7
4
5
-
3
7
6
0
3748
Sp
atial
f
ilter
in
g
:
L
et
(
,
)
b
e
th
e
o
u
tp
u
t im
a
g
e.
(
,
)
=
(
∗
)
(
,
)
(
1
0
)
L
em
m
a
3
:
th
e
s
p
atial
f
ilter
in
g
s
tep
co
m
b
in
es
th
e
n
o
r
m
alize
d
in
p
u
t
im
a
g
e
with
th
e
Gau
s
s
ian
k
er
n
el,
r
esu
ltin
g
in
an
o
u
tp
u
t im
ag
e
with
en
h
a
n
ce
d
co
n
tr
ast an
d
r
ed
u
ce
d
n
o
is
e.
Pro
o
f
:
(
,
)
=
(
∗
)
(
,
)
=
(
(
,
)
−
)
/
(
∗
)
∗
(
,
)
(
1
1
)
[
(
,
)
]
=
[
(
(
,
)
−
)
/
(
∗
)
∗
(
,
)
]
=
0
(
)
(
1
2
)
1
=
[
(
(
,
)
−
)
(
∗
)
∗
(
,
)
⁄
]
(
1
3
)
2
=
2
∗
∑
(
,
)
(
,
)
2
(
1
4
)
[
(
,
)
]
=
1
=
2
≤
2
(
)
(
1
5
)
2
.
3
.
L
o
ca
l bina
ry
pa
t
t
er
n
T
h
e
L
B
P
alg
o
r
ith
m
,
in
tr
o
d
u
ce
d
b
y
[
1
0
]
ex
tr
ac
ts
cr
u
cial
tex
tu
r
al
in
f
o
r
m
atio
n
f
r
o
m
i
m
ag
es
b
y
an
aly
zin
g
th
e
r
elatio
n
s
h
ip
s
b
etwe
en
n
eig
h
b
o
r
in
g
p
ix
els.
T
h
is
is
ac
h
iev
ed
t
h
r
o
u
g
h
a
th
r
esh
o
ld
in
g
p
r
o
ce
s
s
,
wh
er
e
th
e
b
in
ar
y
r
e
p
r
esen
tatio
n
s
o
f
n
eig
h
b
o
r
in
g
p
ix
els
ar
e
co
m
p
u
ted
an
d
co
m
p
a
r
ed
to
th
e
ce
n
tr
al
p
ix
el.
T
h
e
L
B
P
ap
p
r
o
ac
h
was
ch
o
s
en
f
o
r
th
is
r
esear
ch
d
u
e
t
o
its
ef
f
icien
cy
an
d
in
n
o
v
ativ
e
m
eth
o
d
o
f
g
ath
er
in
g
ch
r
o
m
in
a
n
ce
d
ata
f
r
o
m
im
a
g
e
s
.
T
h
e
L
B
P d
escr
ip
to
r
ca
lcu
latio
n
in
v
o
l
v
es f
o
u
r
s
tag
es:
i)
Selectin
g
F n
eig
h
b
o
r
in
g
p
ix
els at
a
r
ad
iu
s
o
f
G
f
o
r
ea
ch
p
i
x
el
(
x
,
y
)
.
ii)
C
alcu
latin
g
th
e
s
tr
en
g
th
d
if
f
er
en
ce
b
etwe
en
th
e
ce
n
tr
al
p
ix
el
an
d
its
n
eig
h
b
o
r
s
.
iii)
T
h
r
esh
o
ld
in
g
th
e
d
i
f
f
er
en
ce
s
t
o
cr
ea
te
a
b
in
a
r
y
v
ec
to
r
.
iv
)
C
o
n
v
er
tin
g
th
is
v
ec
to
r
to
a
d
e
cim
al
v
alu
e,
wh
ich
r
ep
lace
s
th
e
o
r
ig
in
al
s
tr
en
g
t
h
v
alu
e
at
(
x
,
y
)
.
Fig
u
r
e
1
s
h
o
ws
th
e
ex
am
p
le
il
lu
s
tr
atio
n
s
h
o
win
g
th
e
L
B
P
tr
an
s
f
o
r
m
atio
n
p
r
o
ce
s
s
o
n
a
f
u
n
d
u
s
im
ag
e.
T
h
e
o
r
i
g
in
al
ey
e
im
ag
e
(
lef
t)
u
n
d
e
r
g
o
es
p
ix
el
-
by
-
p
ix
el
i
n
ten
s
ity
co
m
p
ar
is
o
n
with
its
n
eig
h
b
o
r
s
.
T
h
e
p
ix
el
r
ep
r
esen
tatio
n
m
atr
ix
(
ce
n
ter
)
s
h
o
ws
in
ten
s
ity
v
alu
es
wi
th
th
e
ce
n
tr
al
th
r
esh
o
ld
v
al
u
e
o
f
1
1
4
.
E
ac
h
n
eig
h
b
o
r
in
g
p
ix
el
is
co
m
p
ar
e
d
to
th
is
th
r
esh
o
ld
to
g
en
er
ate
th
e
b
in
ar
y
r
ep
r
esen
tatio
n
(
r
ig
h
t
ce
n
ter
)
,
wh
er
e
v
alu
es
≤
1
1
4
ar
e
s
et
to
1
an
d
v
alu
es
<
1
1
4
a
r
e
s
et
to
0
.
T
h
e
b
in
a
r
y
v
alu
es
ar
e
th
en
c
o
n
v
er
ted
to
d
ec
im
al
r
ep
r
esen
tatio
n
(
r
ig
h
t)
u
s
in
g
weig
h
ted
p
o
s
itio
n
al
en
c
o
d
in
g
,
p
r
o
d
u
cin
g
th
e
f
i
n
al
L
B
P
-
tr
an
s
f
o
r
m
ed
te
x
tu
r
e
p
atter
n
s
h
o
wn
i
n
th
e
g
r
ay
s
c
ale
o
u
tp
u
t
im
a
g
e
(
b
o
tto
m
le
f
t)
.
T
h
is
tr
an
s
f
o
r
m
atio
n
ca
p
t
u
r
es
lo
ca
l
tex
tu
r
e
in
f
o
r
m
atio
n
wh
ile
m
ain
tain
in
g
in
v
ar
ian
ce
to
illu
m
in
atio
n
v
ar
iatio
n
s
.
Fig
u
r
e
1
illu
s
tr
ates
an
ex
am
p
l
e
o
f
an
L
B
P
-
p
r
o
ce
s
s
ed
im
ag
e
.
T
h
e
L
B
P
d
escr
ip
to
r
is
m
ath
em
atica
lly
r
ep
r
esen
ted
as
(
1
6
)
.
(
,
)
=
∑
2
−
1
=
0
(
−
)
(
1
6
)
W
h
er
e
r
ep
r
esen
ts
th
e
n
eig
h
b
o
r
in
g
p
ix
els,
is
th
e
r
ad
iu
s
,
an
d
d
en
o
te
th
e
in
ten
s
ities
o
f
t
h
e
cu
r
r
en
t
an
d
n
eig
h
b
o
r
in
g
p
ix
els,
r
esp
ec
tiv
e
ly
,
an
d
is
a
s
ig
n
f
u
n
ctio
n
d
e
f
in
ed
as f
o
llo
ws:
(
)
=
{
1
≥
0
0
}
2
.
4
.
Ca
ps
ule net
wo
rk
T
h
e
co
n
c
ep
t
o
f
c
a
p
s
u
l
e
s
o
r
ig
i
n
a
t
e
d
f
r
o
m
t
h
e
w
o
r
k
o
f
H
i
n
t
o
n
e
t
a
l
.
[
1
1
]
a
n
d
w
a
s
l
a
t
e
r
e
x
p
an
d
e
d
u
p
o
n
b
y
S
ab
o
u
r
e
t
a
l
.
[
1
2
]
.
T
h
e
s
e
p
i
o
n
e
er
i
n
g
s
tu
d
i
e
s
s
h
o
w
c
a
s
e
d
C
ap
s
N
e
t
s
a
s
a
p
r
o
m
is
i
n
g
a
p
p
r
o
a
c
h
f
o
r
i
m
a
g
e
c
l
a
s
s
i
f
i
c
a
t
io
n
,
g
e
n
er
a
ti
n
g
s
u
b
s
t
an
t
i
a
l
in
t
e
r
e
s
t
a
m
o
n
g
r
e
s
e
ar
c
h
er
s
a
cr
o
s
s
v
a
r
io
u
s
d
i
s
c
ip
l
i
n
e
s
a
n
d
p
r
o
m
p
t
in
g
f
u
r
th
e
r
ex
p
lo
r
a
t
io
n
o
f
t
h
i
s
m
e
th
o
d
o
lo
g
y
.
A
ca
p
s
u
l
e
i
s
a
v
e
c
t
o
r
-
b
a
s
e
d
r
ep
r
e
s
e
n
t
a
t
io
n
o
f
a
n
o
b
j
e
c
t
's
a
t
t
r
ib
u
t
e
s
,
w
i
t
h
m
u
lt
i
p
l
e
c
ap
s
u
l
e
s
f
o
r
m
i
n
g
a
c
ap
s
u
l
e
l
ay
e
r
.
A
C
ap
s
N
e
t
a
r
ch
i
t
e
c
t
u
r
e
co
m
p
r
i
s
e
s
s
t
a
c
k
e
d
c
o
n
v
o
l
u
t
io
n
a
l
l
ay
e
r
s
,
c
ap
s
u
l
e
l
ay
er
s
(
in
c
l
u
d
i
n
g
p
r
i
m
ar
y
a
n
d
c
la
s
s
c
a
p
s
u
le
la
y
e
r
s
)
,
an
d
f
u
l
l
y
c
o
n
n
e
c
t
ed
la
y
er
s
.
T
h
e
en
c
o
d
er
n
e
t
w
o
r
k
c
o
n
s
i
s
t
s
o
f
c
o
n
v
o
lu
t
i
o
n
a
l
a
n
d
ca
p
s
u
l
e
l
ay
e
r
s
,
wh
e
r
e
a
s
t
h
e
d
e
c
o
d
er
c
o
m
p
r
i
s
e
s
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
s
.
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
A
d
a
p
tive
b
in
a
r
y
ca
p
s
u
le
n
etw
o
r
k
fo
r
co
mp
lex
ima
g
e
r
ec
o
g
n
itio
n
(
Ma
vis S
erw
a
a
Yeb
o
a
h
)
3749
T
h
e
m
ag
n
itu
d
e
o
f
a
ca
p
s
u
le'
s
v
ec
to
r
s
ig
n
if
ies th
e
p
r
esen
ce
o
r
ab
s
en
ce
o
f
an
en
tity
'
s
f
ea
tu
r
es,
with
th
e
v
ec
to
r
v
alu
es
g
en
er
ated
b
y
n
e
u
r
o
n
s
with
in
th
e
co
n
v
o
l
u
tio
n
a
l
lay
er
s
.
I
n
th
e
p
r
im
ar
y
ca
p
s
u
l
e
lay
er
,
ca
p
s
u
les
ar
e
tr
an
s
f
o
r
m
ed
u
s
in
g
a
tr
an
s
f
o
r
m
atio
n
m
atr
ix
(
1
7
)
,
wh
i
ch
en
c
o
d
es
p
r
o
p
er
ties
s
u
ch
as
r
o
tatio
n
a
n
d
s
ca
lin
g
.
T
h
is
tr
an
s
f
o
r
m
atio
n
y
ield
s
p
r
ed
ictio
n
v
ec
to
r
s
̂
,
wh
ich
ac
tiv
ely
s
ee
k
p
ar
en
t
ca
p
s
u
les
in
th
e
clas
s
ca
p
s
u
le
lay
er
d
u
r
in
g
ea
c
h
iter
a
tio
n
.
̂
=
∗
(
1
7
)
C
o
u
p
lin
g
co
ef
f
icien
ts
,
ar
e
ca
lcu
lated
to
estab
lis
h
lin
k
ag
es
b
etwe
en
ch
ild
̂
an
d
p
ar
en
t
ca
p
s
u
les
,
in
d
icatin
g
wh
eth
er
a
ch
ild
ca
p
s
u
le'
s
p
r
o
p
er
ties
ar
e
p
r
esen
t
in
a
p
a
r
en
t
ca
p
s
u
l
e'
s
f
ea
tu
r
e
s
p
ac
e.
T
h
e
len
g
th
o
f
th
e
m
o
d
if
ied
p
r
ed
ictio
n
v
ec
to
r
s
d
eter
m
in
es
if
two
ca
p
s
u
les
s
h
ar
e
s
im
ilar
p
r
o
p
er
ties
.
I
f
th
e
len
g
th
o
f
is
clo
s
e
to
o
n
e,
it
in
d
icate
s
s
im
ilar
ity
;
o
th
er
wis
e,
it
i
s
s
q
u
ash
ed
to
ze
r
o
.
T
h
is
s
q
u
ash
in
g
is
co
m
p
u
ted
in
(
1
8
)
to
(
2
0
)
.
=
(
)
∑
(
)
(
1
8
)
=
∑
̂
=
1
(
1
9
)
=
|
|
|
|
2
1
+
|
|
|
|
2
|
|
|
|
(
2
0
)
T
h
e
f
in
al
o
u
tp
u
ts
ar
e
tr
an
s
f
o
r
m
ed
in
to
p
r
o
b
ab
ilit
ies
b
y
u
s
in
g
th
e
s
o
f
tm
ax
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
e
en
co
d
er
n
etwo
r
k
p
e
r
f
o
r
m
s
f
ea
tu
r
e
ex
tr
ac
tio
n
,
p
r
o
ce
s
s
in
g
,
an
d
class
if
icatio
n
,
wh
ile
th
e
d
ec
o
d
er
n
et
wo
r
k
r
ec
o
n
s
tr
u
cts
th
e
in
p
u
t im
a
g
e.
Fig
u
r
e
1
.
L
B
P p
r
o
ce
s
s
in
g
d
em
o
n
s
tr
atio
n
2
.
5
.
P
r
o
po
s
ed
m
o
del
Fig
u
r
e
2
illu
s
tr
ates
th
e
p
r
o
p
o
s
ed
m
o
d
el,
wh
ich
c
o
n
s
is
ts
o
f
t
wo
AC
SF
lay
er
s
,
o
n
e
L
B
P
lay
er
with
a
3
×
3
f
ilter
,
th
r
ee
co
n
v
o
lu
tio
n
al
lay
er
s
with
th
e
f
ir
s
t
an
d
s
ec
o
n
d
lay
er
s
h
av
in
g
6
4
an
d
2
5
6
f
ilter
s
o
f
s
ize
3
×
3
r
esp
ec
tiv
ely
,
a
p
r
im
ar
y
ca
p
s
u
le
lay
er
with
ca
p
s
u
les
o
f
d
im
en
s
io
n
8
,
a
class
ca
p
s
u
le
lay
er
with
ca
p
s
u
les
o
f
d
im
en
s
io
n
2
,
an
d
th
r
ee
f
u
lly
c
o
n
n
ec
ted
lay
er
s
with
5
1
2
,
1
,
0
2
4
,
a
n
d
2
,
3
5
2
n
eu
r
o
n
s
r
esp
ec
t
iv
ely
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el'
s
ar
ch
itectu
r
e
is
d
esig
n
ed
to
e
f
f
ec
tiv
ely
ex
tr
ac
t
f
ea
t
u
r
es
an
d
class
if
y
o
cu
lar
im
a
g
e
s
,
with
ea
ch
lay
e
r
b
u
ild
in
g
o
n
th
e
p
r
e
v
io
u
s
o
n
e
to
en
ab
le
r
o
b
u
s
t
an
d
ac
cu
r
ate
o
cu
lar
im
ag
e
an
aly
s
is
.
T
h
e
n
etwo
r
k
ar
ch
itectu
r
e
p
r
o
ce
s
s
es
in
p
u
t
im
a
g
es
(
2
8
×
28
×
3
)
th
r
o
u
g
h
two
s
eq
u
en
ti
al
AC
SF
lay
er
s
.
An
L
B
P
lay
e
r
ex
tr
ac
ts
in
v
ar
ian
t
tex
tu
r
e
f
ea
t
u
r
es.
T
h
r
ee
co
n
v
o
l
u
tio
n
al
lay
er
s
p
r
o
g
r
ess
iv
ely
e
x
tr
ac
t
h
ier
ar
c
h
ical
f
ea
tu
r
es:
C
o
n
v
1
,
C
o
n
v
2
,
a
n
d
C
o
n
v
3
.
AC
SF
lay
er
s
ar
e
m
a
r
k
ed
i
n
y
ello
w,
th
e
L
B
P
lay
er
in
y
ello
w
-
g
r
ee
n
,
c
o
n
v
o
lu
tio
n
al
lay
er
s
in
g
r
ee
n
,
ca
p
s
u
le
lay
er
s
in
b
lu
e,
an
d
d
ec
o
d
er
lay
e
r
s
in
o
r
a
n
g
e.
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.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
7
4
5
-
3
7
6
0
3750
Fig
u
r
e
2
.
Ar
c
h
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
a
d
ap
tiv
e
b
in
ar
y
C
ap
s
Net
m
o
d
el
2
.
6
.
Da
t
a
s
et
des
cr
iptio
n a
nd
prepro
ce
s
s
i
ng
T
h
e
p
r
o
p
o
s
ed
m
o
d
el’
s
p
er
f
o
r
m
an
ce
was e
v
alu
ated
u
s
in
g
tw
o
d
atasets
s
o
u
r
ce
d
f
r
o
m
Ka
g
g
l
e.
T
h
e
f
ir
s
t
d
ataset
in
clu
d
ed
1
3
4
im
ag
es
o
f
ey
es
af
f
ec
te
d
b
y
g
lau
co
m
a
an
d
3
8
6
im
ag
es
o
f
ey
es
with
o
u
t
g
lau
c
o
m
a.
T
h
e
s
ec
o
n
d
d
ataset
co
n
tain
ed
3
0
6
im
ag
es
o
f
ca
tar
ac
t
-
af
f
ec
ted
ey
es
an
d
3
0
6
im
ag
es
o
f
n
o
n
-
ca
t
ar
ac
t
-
af
f
ec
ted
ey
es.
T
h
ese
d
atasets
wer
e
m
er
g
ed
t
o
tr
ain
th
e
p
r
o
p
o
s
ed
m
o
d
el.
F
ig
u
r
e
3
p
r
esen
ts
s
am
p
le
im
ag
es
f
r
o
m
th
e
d
ataset.
Fig
u
r
e
3
(
a
)
p
r
esen
ts
a
s
am
p
le
n
o
n
-
i
n
f
ec
ted
g
lau
co
m
a
e
y
e,
wh
ile
Fig
u
r
e
3
(
b
)
p
r
esen
ts
a
g
lau
co
m
ato
u
s
ey
e.
Fig
u
r
e
3
(
c
)
p
r
esen
ts
a
ca
tar
a
ct
-
in
f
ec
ted
e
y
e,
wh
ile
Fig
u
r
e
3
(
d
)
p
r
esen
ts
a
n
o
n
-
i
n
f
ec
ted
ca
tar
ac
t
ey
e
.
T
h
e
d
ataset
was
s
p
lit
u
s
in
g
a
7
0
:
2
0
:1
0
r
atio
f
o
r
tr
ain
i
n
g
,
v
alid
at
io
n
,
an
d
test
in
g
,
r
esp
ec
tiv
ely
.
T
o
test
th
e
m
o
d
el’
s
g
en
er
aliza
b
ilit
y
,
th
e
C
I
FAR
-
1
0
d
ataset
was
em
p
lo
y
ed
.
T
h
is
d
ataset
co
n
s
is
ts
o
f
1
0
ca
teg
o
r
ies:
0
:
air
p
lan
e,
1
: a
u
to
m
o
b
ile,
2
: b
ir
d
,
3
: c
at,
4
: d
ee
r
,
5
: d
o
g
,
6
:
f
r
o
g
,
7
: h
o
r
s
e,
8
: sh
ip
,
an
d
9
: tr
u
ck
.
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
3
.
R
ep
r
esen
tativ
e
s
am
p
les f
r
o
m
th
e
o
c
u
lar
d
is
ea
s
e
d
at
asets
o
f
(
a)
n
o
n
-
g
lau
c
o
m
a,
(
b
)
g
lau
co
m
a,
(
c)
ca
tar
ac
ts
,
an
d
(
d
)
n
o
n
-
ca
tar
ac
ts
Acc
o
r
d
in
g
t
o
Fig
u
r
e
3
,
s
am
p
le
f
u
n
d
u
s
an
d
ey
e
im
a
g
es
d
em
o
n
s
tr
atin
g
th
e
f
o
u
r
class
if
icatio
n
ca
teg
o
r
ies
u
s
ed
in
m
o
d
el
tr
ai
n
in
g
an
d
ev
al
u
atio
n
.
No
n
-
g
la
u
co
m
a
:
h
ea
lth
y
r
etin
al
f
u
n
d
u
s
s
h
o
win
g
n
o
r
m
al
o
p
tic
d
is
c
c
u
p
p
in
g
a
n
d
u
n
if
o
r
m
v
ascu
lar
d
is
tr
ib
u
tio
n
(
Fig
u
r
e
3
(
a)
)
.
Glau
co
m
a:
d
is
ea
s
ed
f
u
n
d
u
s
ex
h
ib
itin
g
ch
ar
ac
ter
is
tic
g
lau
co
m
ato
u
s
o
p
tic
n
er
v
e
d
am
a
g
e
with
in
cr
ea
s
ed
cu
p
-
to
-
d
is
c
r
atio
an
d
n
eu
r
o
r
etin
al
r
im
th
in
n
in
g
(
Fig
u
r
e
3
(
b
)
)
.
C
atar
ac
ts
:
an
ter
io
r
ey
e
p
h
o
to
g
r
ap
h
s
h
o
win
g
len
s
o
p
ac
ity
ch
a
r
ac
ter
is
tic
o
f
ca
tar
ac
t
f
o
r
m
atio
n
with
v
is
ib
le
clo
u
d
in
g
Fig
u
r
e
(
3
(
c
)
)
.
No
n
-
ca
tar
ac
ts
:
h
ea
lth
y
an
ter
io
r
e
y
e
with
clea
r
len
s
an
d
n
o
r
m
al
ir
is
s
tr
u
ctu
r
e
(
Fig
u
r
e
3
(
d
)
)
.
T
h
ese
im
ag
es
illu
s
tr
ate
th
e
v
is
u
al
ch
ar
ac
ter
is
tics
an
d
d
iag
n
o
s
tic
f
ea
tu
r
es
th
at
th
e
m
o
d
el
m
u
s
t d
is
tin
g
u
is
h
b
etwe
en
class
es
.
2
.
7
.
E
x
perim
ent
a
l set
up
T
h
e
ex
p
er
im
en
tal
s
y
s
tem
u
tili
ze
d
an
NVI
DI
A
GeFo
r
ce
1
,
0
6
0
g
r
ap
h
ics
ca
r
d
,
c
o
m
p
le
m
en
ted
b
y
8
g
ig
a
b
y
tes
o
f
r
a
n
d
o
m
-
ac
ce
s
s
m
em
o
r
y
,
t
o
f
ac
ilit
ate
ef
f
icien
t
p
r
o
ce
s
s
in
g
a
n
d
d
ata
h
an
d
lin
g
.
Ad
d
itio
n
ally
,
th
e
b
atch
s
ize
was
s
et
to
1
0
0
,
th
e
lear
n
in
g
r
ate
was
in
itiali
ze
d
at
0
.
0
0
1
,
an
d
th
e
lear
n
in
g
d
ec
ay
r
ate
was
co
n
f
ig
u
r
ed
to
0
.
9
,
th
er
e
b
y
es
tab
lis
h
in
g
th
e
o
p
tim
al
p
ar
am
eter
s
f
o
r
th
e
tr
ai
n
in
g
p
r
o
ce
s
s
.
Fu
r
th
er
m
o
r
e,
all
co
d
in
g
was
ac
co
m
p
lis
h
ed
u
s
i
n
g
th
e
Ker
as
d
ee
p
lear
n
i
n
g
f
r
am
ewo
r
k
,
lev
e
r
ag
in
g
th
e
T
e
n
s
o
r
Flo
w
b
ac
k
en
d
f
o
r
s
ea
m
less
ex
ec
u
tio
n
.
No
tab
ly
,
th
e
m
ar
g
in
lo
s
s
f
u
n
ctio
n
,
a
s
p
r
o
p
o
s
ed
b
y
[
1
2
]
,
was
ad
o
p
ted
to
ef
f
ec
tiv
ely
m
ea
s
u
r
e
an
d
m
i
n
im
ize
th
e
er
r
o
r
b
etwe
en
p
r
ed
icted
an
d
ac
tu
al
o
u
tp
u
ts
,
wh
er
ea
s
t
h
eir
m
o
d
el
was
im
p
lem
en
ted
as
a
b
aselin
e
m
o
d
el.
T
h
e
co
d
e
f
o
r
th
is
p
a
p
er
ca
n
b
e
f
o
u
n
d
at
h
ttp
s
://g
ith
u
b
.
co
m
/m
y
te
n
u
/Ad
ap
tiv
e
-
B
in
ar
y
-
C
ap
s
u
le
-
Netwo
r
k
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
A
d
a
p
tive
b
in
a
r
y
ca
p
s
u
le
n
etw
o
r
k
fo
r
co
mp
lex
ima
g
e
r
ec
o
g
n
itio
n
(
Ma
vis S
erw
a
a
Yeb
o
a
h
)
3751
3.
RE
SU
L
T
S
AND
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
a
co
m
p
r
eh
en
s
iv
e
ev
alu
atio
n
o
f
o
u
r
p
r
o
p
o
s
ed
AC
SF
C
ap
s
Net
an
d
co
m
p
ar
es
its
p
er
f
o
r
m
an
ce
ag
ai
n
s
t
th
e
b
aselin
e
m
o
d
el
p
r
o
p
o
s
ed
b
y
[
1
2
]
ac
r
o
s
s
m
u
ltip
le
o
cu
la
r
d
i
s
ea
s
e
d
ataset
s
.
T
h
e
ev
alu
atio
n
m
et
h
o
d
o
lo
g
y
en
c
o
m
p
ass
es
f
iv
e
k
ey
an
aly
tical
d
i
m
en
s
io
n
s
:
i)
tr
ain
in
g
d
y
n
am
ic
s
an
d
co
n
v
er
g
en
ce
ch
ar
ac
ter
is
tics
th
r
o
u
g
h
ac
c
u
r
ac
y
an
d
lo
s
s
cu
r
v
es
;
ii)
class
-
s
p
ec
if
ic
p
er
f
o
r
m
a
n
ce
an
aly
s
is
u
s
in
g
co
n
f
u
s
io
n
m
atr
ices
with
d
etailed
p
r
ec
is
io
n
,
r
ec
all,
an
d
s
p
ec
if
icity
m
etr
ics
;
iii)
s
tatis
tical
r
o
b
u
s
tn
ess
ass
es
s
m
en
t
th
r
o
u
g
h
5
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
with
co
n
f
id
e
n
ce
in
ter
v
al
an
aly
s
i
s
;
iv
)
s
y
s
tem
atic
ab
latio
n
s
tu
d
ies
to
q
u
a
n
tify
in
d
iv
id
u
al
co
m
p
o
n
e
n
t
co
n
tr
i
b
u
tio
n
s
;
an
d
v
)
v
is
u
al
an
d
q
u
an
titativ
e
an
aly
s
is
o
f
f
ea
tu
r
e
m
ap
s
,
clu
s
ter
in
g
b
eh
av
io
r
,
an
d
r
ec
o
n
s
tr
u
ctio
n
q
u
ality
u
s
in
g
p
ea
k
s
ig
n
al
-
to
-
n
o
is
e
r
atio
(
PS
NR
)
an
d
s
tr
u
ctu
r
al
s
im
ilar
ity
in
d
ex
m
etr
ics
(
SS
I
M)
.
Ad
d
itio
n
ally
,
th
is
s
tu
d
y
p
r
o
v
id
e
a
f
o
cu
s
ed
d
is
cu
s
s
io
n
ex
p
lain
in
g
th
e
th
eo
r
etica
l
f
o
u
n
d
atio
n
s
an
d
m
ec
h
an
is
tic
in
s
ig
h
ts
in
to
wh
y
o
u
r
p
r
o
p
o
s
ed
AC
SF
an
d
L
B
P
lay
er
s
y
ield
s
u
p
e
r
io
r
co
n
tr
ast
en
h
an
ce
m
e
n
t
an
d
g
e
n
er
aliza
tio
n
ca
p
a
b
ilit
y
.
3
.
1
.
E
x
perim
ent
a
l
curv
es a
n
d c
o
nfusi
o
n m
a
t
rix
a
na
ly
s
is
Fig
u
r
e
4
s
h
o
ws
th
e
tr
ain
in
g
p
er
f
o
r
m
a
n
ce
,
wh
ile
th
e
ac
cu
r
ac
y
as
in
Fig
u
r
e
4
(
a)
an
d
lo
s
s
cu
r
v
es
as
in
Fig
u
r
e
4
(
b
)
f
o
r
b
o
th
th
e
p
r
o
p
o
s
ed
an
d
b
aselin
e
m
o
d
els
wh
en
tr
ain
e
d
o
n
th
e
c
o
m
b
in
e
d
d
ataset.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
ac
h
ie
v
ed
ac
c
u
r
ac
ies
o
f
9
6
.
9
0
%,
9
6
.
7
8
%,
9
6
.
4
5
%,
an
d
9
4
.
8
7
%
o
n
th
e
c
o
m
b
in
e
d
,
g
lau
c
o
m
a
-
o
n
ly
,
ca
tar
ac
t
-
o
n
ly
,
a
n
d
C
I
FAR
-
1
0
d
atasets
,
r
esp
ec
tiv
ely
.
I
n
co
m
p
ar
is
o
n
,
th
e
b
aselin
e
m
o
d
el
r
e
co
r
d
ed
ac
c
u
r
ac
ies o
f
8
4
.
2
8
%,
9
0
.
0
9
%,
8
9
.
5
0
%,
an
d
6
2
.
3
5
%
o
n
th
e
s
am
e
d
at
asets
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
d
em
o
n
s
tr
ates
g
o
o
d
co
n
v
er
g
en
ce
ch
ar
a
cter
is
tics
,
a
ch
iev
in
g
tr
ain
in
g
ac
cu
r
ac
y
o
f
ap
p
r
o
x
im
ately
9
8
%
an
d
v
alid
atio
n
ac
cu
r
ac
y
o
f
9
6
.
9
0
%
b
y
ep
o
ch
5
0
,
with
s
tab
le
p
er
f
o
r
m
a
n
ce
m
ain
tain
e
d
th
r
o
u
g
h
o
u
t
s
u
b
s
eq
u
en
t
ep
o
ch
s
.
T
h
e
ac
cu
r
ac
y
cu
r
v
es
ex
h
ib
it
m
in
im
al
f
lu
ctu
atio
n
af
ter
in
itial
co
n
v
e
r
g
en
ce
,
in
d
icatin
g
r
o
b
u
s
t
lear
n
in
g
an
d
g
o
o
d
g
en
er
aliza
tio
n
ca
p
a
b
ilit
y
.
I
n
co
n
tr
ast,
th
e
b
aselin
e
m
o
d
el
s
h
o
ws
s
lo
wer
co
n
v
er
g
en
ce
,
r
ea
ch
in
g
tr
ain
in
g
ac
cu
r
ac
y
o
f
a
p
p
r
o
x
im
ately
8
5
%
an
d
v
alid
ati
o
n
ac
c
u
r
ac
y
o
f
8
4
.
2
8
%,
with
n
o
ticea
b
ly
h
ig
h
e
r
v
ar
iab
ilit
y
in
b
o
th
tr
ain
in
g
an
d
v
alid
atio
n
cu
r
v
e
s
th
r
o
u
g
h
o
u
t
th
e
tr
ain
in
g
p
r
o
ce
s
s
.
T
h
e
l
o
s
s
c
u
r
v
e
s
f
u
r
t
h
er
c
o
r
r
o
b
o
r
a
t
e
t
h
e
s
e
o
b
s
e
r
v
a
t
i
o
n
s
,
w
h
e
r
e
t
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l'
s
t
r
a
i
n
i
n
g
a
n
d
v
a
l
i
d
a
t
i
o
n
l
o
s
s
es
c
o
n
v
e
r
g
e
t
o
a
p
p
r
o
x
i
m
a
t
e
l
y
0
.
0
5
-
0
.
1
0
b
y
ep
o
ch
2
5
an
d
r
em
ai
n
s
tab
le,
wh
ile
th
e
b
aselin
e
m
o
d
el'
s
l
o
s
s
es
s
tab
ilize
at
h
ig
h
er
v
alu
e
s
(
0
.
1
5
-
0
.
2
0
)
with
m
o
r
e
p
r
o
n
o
u
n
ce
d
o
s
cillatio
n
s
,
s
u
g
g
esti
n
g
s
u
b
o
p
tim
al
f
ea
tu
r
e
lear
n
in
g
.
Fig
u
r
e
5
s
h
o
ws
t
h
e
co
n
f
u
s
io
n
m
atr
ix
f
o
r
co
m
p
ar
in
g
tr
ain
i
n
g
p
er
f
o
r
m
an
ce
.
Fig
u
r
e
5
(
a
)
s
h
o
ws
th
e
co
n
f
u
s
io
n
m
atr
ix
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el,
wh
ile
Fig
u
r
e
5
(
b
)
s
h
o
ws
th
e
co
n
f
u
s
io
n
m
atr
ix
o
f
th
e
b
aselin
e
m
o
d
el.
T
ab
les
1
an
d
2
p
r
esen
t
d
etailed
co
n
f
u
s
io
n
m
atr
ix
an
aly
s
es
f
o
r
th
e
p
r
o
p
o
s
ed
an
d
b
aselin
e
m
o
d
els,
r
esp
e
ctiv
ely
,
r
ev
ea
lin
g
th
e
m
o
d
els'
p
er
f
o
r
m
an
ce
ac
r
o
s
s
in
d
iv
id
u
al
d
is
ea
s
e
class
e
s
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
ac
h
iev
es
ex
ce
p
tio
n
al
class
if
icatio
n
p
er
f
o
r
m
an
ce
with
class
-
s
p
ec
if
ic
ac
cu
r
ac
ies
r
an
g
in
g
f
r
o
m
0
.
9
7
8
to
0
.
9
9
1
.
Fo
r
g
lau
co
m
a
d
etec
tio
n
,
th
e
m
o
d
el
attain
s
p
er
f
ec
t
p
r
ec
is
io
n
(
1
.
0
0
0
)
with
h
ig
h
s
en
s
itiv
ity
(
0
.
9
2
6
)
an
d
s
p
ec
if
i
city
(
1
.
0
0
0
)
f
o
r
g
lau
co
m
a
-
p
o
s
itiv
e
ca
s
es,
wh
ile
ac
h
iev
in
g
p
r
ec
is
io
n
o
f
0
.
9
7
4
,
s
en
s
itiv
ity
o
f
0
.
9
8
7
,
a
n
d
s
p
ec
if
icity
o
f
0
.
9
8
7
f
o
r
g
lau
co
m
a
-
n
eg
ativ
e
ca
s
es,
d
em
o
n
s
tr
atin
g
e
x
ce
llen
t
d
is
cr
im
in
ativ
e
ab
ilit
y
with
a
m
ac
r
o
-
av
er
ag
e
ar
ea
u
n
d
e
r
th
e
cu
r
v
e
(
AUC)
o
f
0
.
9
7
6
.
T
h
e
m
o
d
el
s
im
ilar
ly
ex
ce
ls
in
ca
tar
ac
t
d
etec
tio
n
,
ac
h
iev
in
g
a
p
r
ec
is
io
n
o
f
0
.
9
5
2
a
n
d
a
s
en
s
itiv
ity
o
f
0
.
9
6
7
f
o
r
ca
tar
ac
t
-
p
o
s
itiv
e
ca
s
es,
an
d
a
p
r
ec
is
io
n
o
f
0
.
9
6
7
with
a
s
en
s
i
tiv
ity
o
f
0
.
9
6
7
f
o
r
ca
tar
ac
t
-
n
eg
ativ
e
ca
s
es.
Fig
u
r
e
4
s
h
o
ws
th
e
tr
ain
in
g
p
er
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
v
er
2
0
0
e
p
o
ch
s
.
Acc
u
r
ac
y
c
u
r
v
es:
th
e
p
r
o
p
o
s
ed
m
o
d
el
ac
h
ie
v
es
9
8
%
tr
ain
in
g
a
n
d
9
6
.
9
0
%
v
alid
atio
n
ac
cu
r
ac
y
with
s
tab
le
co
n
v
er
g
e
n
ce
,
wh
ile
th
e
b
aselin
e
r
ea
ch
es
o
n
ly
8
5
%
a
n
d
8
4
.
2
8
%
with
h
ig
h
o
s
cillatio
n
s
(
Fig
u
r
e
4
(
a)
)
.
L
o
s
s
cu
r
v
es:
p
r
o
p
o
s
ed
m
o
d
el
co
n
v
er
g
es to
0
.
0
5
-
0
.
1
0
b
y
ep
o
ch
2
5
,
wh
ile
b
aselin
e
s
tab
ilizes at
h
ig
h
er
v
alu
es (
0
.
1
5
-
0
.
2
0
)
with
f
lu
ctu
atio
n
s
in
d
icatin
g
s
u
b
o
p
tim
al
lear
n
in
g
(
Fig
u
r
e
4
(
b
)
)
.
(
a)
(
b
)
Fig
u
r
e
4
.
T
r
ain
in
g
p
er
f
o
r
m
an
c
e
co
m
p
ar
is
o
n
o
f
(
a)
ac
cu
r
ac
y
cu
r
v
es a
n
d
(
b
)
l
o
s
s
cu
r
v
es
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.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
7
4
5
-
3
7
6
0
3752
(
a)
(
b
)
Fig
u
r
e
5
.
C
o
n
f
u
s
io
n
m
atr
ices f
o
r
(
a)
p
r
o
p
o
s
ed
m
o
d
el
an
d
(
b
)
b
aselin
e
m
o
d
el
T
ab
le
1
.
An
aly
zi
n
g
th
e
c
o
n
f
u
s
io
n
m
atr
ix
v
alu
es o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
C
l
a
s
s
TP
FP
FN
TN
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
S
p
e
c
i
f
i
c
i
t
y
A
c
c
u
r
a
c
y
p
e
r
c
l
a
s
s
F1
-
sc
o
r
e
AUC
C
l
a
u
c
o
m
a
-
p
o
si
t
i
v
e
25
0
2
1
9
9
1
.
0
0
0
0
.
9
2
6
1
.
0
0
0
0
.
9
9
1
0
.
9
6
2
0
.
9
6
3
G
l
a
u
c
o
ma
-
n
e
g
a
t
i
v
e
76
2
1
1
4
7
0
.
9
7
4
0
.
9
8
7
0
.
9
8
7
0
.
9
8
7
0
.
9
8
1
0
.
9
8
7
C
a
t
a
r
a
c
t
s
-
p
o
s
i
t
i
v
e
59
3
2
1
6
2
0
.
9
5
2
0
.
9
6
7
0
.
9
8
2
0
.
9
7
8
0
.
9
5
9
0
.
9
7
5
C
a
t
a
r
a
c
t
s
-
n
e
g
a
t
i
v
e
59
2
2
1
6
3
0
.
9
6
7
0
.
9
6
7
0
.
9
8
8
0
.
9
8
2
0
.
9
6
7
0
.
9
7
8
M
a
c
r
o
-
a
v
e
r
a
g
e
A
U
C
0
.
9
7
6
T
ab
le
2
.
An
aly
zi
n
g
th
e
c
o
n
f
u
s
io
n
m
atr
ix
v
alu
es o
f
th
e
b
aselin
e
m
o
d
el
C
l
a
s
s
TP
FP
FN
TN
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
S
p
e
c
i
f
i
c
i
t
y
A
c
c
u
r
a
c
y
p
e
r
c
l
a
s
s
F1
-
sc
o
r
e
AUC
C
l
a
u
c
o
m
a
-
p
o
si
t
i
v
e
25
15
2
1
8
4
0
.
6
2
5
0
.
9
2
6
0
.
9
2
5
0
.
9
2
5
0
.
7
4
6
0
.
9
2
6
G
l
a
u
c
o
ma
-
n
e
g
a
t
i
v
e
56
9
21
1
4
0
0
.
8
6
2
0
.
7
2
7
0
.
9
4
0
0
.
8
6
7
0
.
7
8
9
0
.
9
4
0
C
a
t
a
r
a
c
t
s
-
p
o
s
i
t
i
v
e
59
3
2
1
6
2
0
.
9
5
2
0
.
9
6
7
0
.
9
8
2
0
.
9
7
8
0
.
9
6
0
0
.
9
7
5
C
a
t
a
r
a
c
t
s
-
n
e
g
a
t
i
v
e
53
6
8
1
5
9
0
.
8
3
4
0
.
8
7
2
0
.
9
5
2
0
.
9
2
7
0
.
8
4
4
0
.
9
6
4
M
a
c
r
o
-
a
v
e
r
a
g
e
A
U
C
0
.
9
1
3
T
h
e
co
n
s
is
ten
tly
h
ig
h
F1
-
s
c
o
r
es
(
0
.
9
5
9
-
0
.
9
8
1
)
an
d
A
UC
v
alu
es
(
0
.
9
6
3
-
0
.
9
8
7
)
ac
r
o
s
s
all
class
es
in
d
icate
b
alan
ce
d
p
e
r
f
o
r
m
a
n
ce
with
o
u
t
s
ig
n
if
ican
t
b
ias
to
war
d
an
y
p
ar
ticu
lar
class
,
d
esp
ite
th
e
d
ataset's
in
h
er
en
t
class
im
b
al
an
ce
.
I
n
co
n
tr
ast,
th
e
b
aselin
e
m
o
d
el
ex
h
ib
its
s
u
b
s
tan
tially
lo
wer
p
er
f
o
r
m
an
c
e
ac
r
o
s
s
all
m
etr
ics,
p
ar
ticu
lar
ly
s
tr
u
g
g
lin
g
with
g
lau
c
o
m
a
d
et
ec
tio
n
,
wh
er
e
it
ac
h
iev
es
o
n
ly
0
.
6
2
5
p
r
ec
is
io
n
f
o
r
g
lau
co
m
a
-
p
o
s
itiv
e
ca
s
es,
in
d
icatin
g
a
h
ig
h
f
alse
p
o
s
itiv
e
r
ate
(
1
5
m
is
class
if
icati
o
n
s
)
.
T
h
e
b
aselin
e
m
o
d
el'
s
s
en
s
itiv
ity
f
o
r
g
la
u
co
m
a
-
n
eg
ativ
e
ca
s
es
is
n
o
tab
ly
p
o
o
r
at
0
.
7
2
7
,
with
2
1
f
alse
n
e
g
ativ
es
r
ep
r
esen
tin
g
a
co
n
ce
r
n
i
n
g
m
is
s
r
ate
f
o
r
th
is
cr
itical
d
iag
n
o
s
tic
task
.
W
h
ile
th
e
b
aselin
e
m
o
d
el
p
er
f
o
r
m
s
r
elativ
ely
b
etter
o
n
ca
tar
ac
t
d
etec
tio
n
(
p
r
ec
is
io
n
:
0
.
9
5
2
an
d
s
en
s
itiv
ity
:
0
.
9
6
7
f
o
r
ca
tar
a
ct
-
p
o
s
itiv
e
ca
s
es),
its
o
v
er
all
m
ac
r
o
-
av
e
r
ag
e
AUC
o
f
0
.
9
1
3
f
alls
s
ig
n
if
ica
n
tly
s
h
o
r
t
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el'
s
0
.
9
7
6
,
r
ep
r
esen
tin
g
a
6
.
9
%
p
er
f
o
r
m
an
ce
g
ap
.
T
h
e
b
aselin
e's
lo
wer
F1
-
s
co
r
es
(
0
.
7
4
6
-
0
.
9
6
0
)
r
ef
lect
an
i
m
b
alan
ce
d
tr
ad
e
-
o
f
f
b
etwe
en
p
r
ec
is
io
n
an
d
r
ec
all,
s
u
g
g
esti
n
g
th
at
t
h
e
s
tan
d
a
r
d
C
ap
s
Net
ar
ch
itectu
r
e
wit
h
o
u
t
t
h
e
p
r
o
p
o
s
ed
en
h
an
ce
m
e
n
ts
s
tr
u
g
g
les
to
ca
p
tu
r
e
th
e
s
u
b
tle
d
is
cr
im
in
ativ
e
f
ea
tu
r
es
n
ec
ess
ar
y
f
o
r
ac
cu
r
ate
m
u
lti
-
class
o
cu
lar
d
is
ea
s
e
class
if
icatio
n
.
T
h
e
co
n
f
u
s
io
n
m
atr
ices
r
ev
ea
l
s
p
ec
if
ic
m
is
class
if
icatio
n
p
att
er
n
s
th
at
h
ig
h
lig
h
t
th
e
p
r
o
p
o
s
ed
m
o
d
el'
s
ad
v
an
tag
es.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
co
m
m
its
o
n
ly
2
f
alse
n
eg
ativ
es
f
o
r
g
lau
co
m
a
-
p
o
s
itiv
e
ca
s
es
an
d
1
f
alse
n
eg
ativ
e
f
o
r
g
lau
co
m
a
-
n
eg
a
tiv
e
ca
s
es,
d
em
o
n
s
tr
atin
g
h
ig
h
s
en
s
itiv
ity
cr
u
cial
f
o
r
clin
ical
s
cr
ee
n
in
g
ap
p
licatio
n
s
wh
er
e
m
is
s
in
g
p
o
s
itiv
e
ca
s
es
ca
r
r
y
s
ev
er
e
co
n
s
eq
u
en
ce
s
.
Fo
r
ca
tar
ac
t
class
i
f
icatio
n
,
th
e
m
o
d
el
ex
h
ib
its
s
y
m
m
etr
ic
p
er
f
o
r
m
a
n
ce
with
o
n
ly
2
-
3
m
is
cla
s
s
if
icatio
n
s
p
er
cla
s
s
,
in
d
icat
in
g
r
o
b
u
s
t
f
ea
tu
r
e
d
is
cr
im
in
atio
n
.
T
h
e
b
aselin
e
m
o
d
el'
s
co
n
f
u
s
io
n
m
atr
ix
s
h
o
ws
a
m
o
r
e
p
r
o
b
le
m
atic
p
att
er
n
,
with
1
5
f
alse
p
o
s
itiv
es
an
d
9
f
alse
n
eg
ativ
es
f
o
r
g
lau
co
m
a
class
es,
s
u
g
g
esti
n
g
d
if
f
icu
lty
in
lear
n
in
g
d
is
cr
im
in
ativ
e
b
o
u
n
d
ar
ies
b
etwe
en
d
is
ea
s
e
a
n
d
h
ea
lth
y
s
tates.
T
h
is
p
er
f
o
r
m
an
ce
d
is
p
ar
ity
d
ir
ec
tly
v
alid
ates
th
e
ef
f
ec
tiv
e
n
ess
o
f
o
u
r
p
r
o
p
o
s
ed
AC
SF
alg
o
r
ith
m
an
d
L
B
P
lay
er
in
en
h
an
cin
g
f
ea
tu
r
e
r
ep
r
esen
tatio
n
q
u
ality
,
en
ab
lin
g
th
e
m
o
d
el
to
ca
p
tu
r
e
f
in
e
-
g
r
ain
ed
tex
tu
r
al
an
d
s
tr
u
ctu
r
al
p
atter
n
s
th
at
d
is
tin
g
u
is
h
b
etwe
en
g
lau
co
m
a,
ca
tar
ac
ts
,
an
d
h
ea
lth
y
o
cu
lar
co
n
d
itio
n
s
.
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
A
d
a
p
tive
b
in
a
r
y
ca
p
s
u
le
n
etw
o
r
k
fo
r
co
mp
lex
ima
g
e
r
ec
o
g
n
itio
n
(
Ma
vis S
erw
a
a
Yeb
o
a
h
)
3753
Fig
u
r
e
5
s
h
o
ws
th
e
c
o
n
f
u
s
io
n
m
atr
ices
f
o
r
m
u
lti
-
class
o
cu
la
r
d
is
ea
s
e
cla
s
s
if
icatio
n
o
n
th
e
co
m
b
in
ed
test
d
ataset.
Pro
p
o
s
ed
m
o
d
el
co
n
f
u
s
io
n
m
atr
ix
s
h
o
win
g
g
o
o
d
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
ac
r
o
s
s
all
f
o
u
r
class
es
(
g
lau
co
m
a
-
p
o
s
itiv
e,
g
lau
co
m
a
-
n
eg
ativ
e,
ca
ta
r
ac
ts
-
p
o
s
itiv
e,
an
d
ca
tar
ac
ts
-
n
eg
a
tiv
e
)
as
s
h
o
wn
in
Fig
u
r
e
5
(
a)
.
T
h
e
m
o
d
el
ac
h
iev
es
h
ig
h
d
iag
o
n
al
v
alu
es
with
m
in
im
al
o
f
f
-
d
ia
g
o
n
al
m
is
class
if
icatio
n
s
.
B
aselin
e
m
o
d
el
co
n
f
u
s
io
n
m
atr
ix
r
e
v
ea
ls
s
u
b
s
tan
tially
m
o
r
e
class
if
icat
io
n
er
r
o
r
s
,
as sh
o
wn
in
Fig
u
r
e
5
(
b
)
.
3
.
2
.
Cro
s
s
-
v
a
lid
a
t
io
n a
nd
s
t
a
t
is
t
ica
l r
o
bu
s
t
nes
s
a
na
ly
s
i
s
T
o
ass
ess
t
h
e
s
t
atis
t
ica
l
r
o
b
u
s
tn
ess
an
d
g
e
n
e
r
a
liz
ati
o
n
ca
p
a
b
il
it
y
o
f
t
h
e
p
r
o
p
o
s
e
d
m
o
d
el
,
t
h
is
s
t
u
d
y
co
n
d
u
ct
ed
5
-
f
o
ld
c
r
o
s
s
-
v
al
id
at
io
n
o
n
al
l
d
at
asets
.
I
n
t
h
is
ap
p
r
o
ac
h
,
e
ac
h
d
a
tase
t
was
r
a
n
d
o
m
l
y
p
a
r
tit
io
n
ed
i
n
t
o
f
i
v
e
eq
u
a
ll
y
-
s
ize
d
s
u
b
s
ets
,
wi
t
h
e
ac
h
s
u
b
s
et
s
e
r
v
i
n
g
as
t
h
e
te
s
t
s
et
o
n
ce
wh
ile
t
h
e
r
em
ai
n
i
n
g
f
o
u
r
s
u
b
s
e
ts
we
r
e
u
s
e
d
f
o
r
t
r
ai
n
i
n
g
.
T
h
is
p
r
o
c
es
s
en
s
u
r
es
t
h
at
e
v
er
y
s
a
m
p
le
i
n
th
e
d
atas
et
is
u
s
e
d
f
o
r
b
o
th
tr
ain
in
g
an
d
test
in
g
,
p
r
o
v
i
d
i
n
g
a
c
o
m
p
r
e
h
e
n
s
i
v
e
ev
al
u
at
io
n
o
f
m
o
d
el
p
e
r
f
o
r
m
a
n
c
e
a
cr
o
s
s
d
i
f
f
e
r
e
n
t
d
ata
s
p
l
its
.
T
h
e
m
ea
n
p
e
r
f
o
r
m
a
n
c
e
m
et
r
ics
an
d
t
h
ei
r
c
o
r
r
esp
o
n
d
i
n
g
s
ta
n
d
a
r
d
d
ev
iati
o
n
s
a
cr
o
s
s
a
ll
f
i
v
e
f
o
l
d
s
wer
e
c
alc
u
l
ate
d
to
q
u
a
n
ti
f
y
t
h
e
m
o
d
e
l'
s
co
n
s
is
te
n
c
y
an
d
r
eli
ab
ilit
y
.
T
a
b
les
3
an
d
4
p
r
es
en
t
t
h
e
d
et
ail
ed
r
e
s
u
lts
o
f
th
e
5
-
f
o
l
d
cr
o
s
s
-
v
ali
d
a
ti
o
n
f
o
r
b
o
th
t
h
e
p
r
o
p
o
s
e
d
m
o
d
el
a
n
d
t
h
e
b
ase
lin
e
m
o
d
el
a
cr
o
s
s
a
ll
p
er
f
o
r
m
a
n
c
e
m
e
tr
ics
.
T
h
e
p
r
o
p
o
s
e
d
m
o
d
el
a
ch
ie
v
e
d
a
m
ea
n
ac
c
u
r
a
cy
o
f
9
5
.
7
2
%
w
it
h
a
s
ta
n
d
a
r
d
d
e
v
i
ati
o
n
o
f
0
.
7
8
%
o
n
t
h
e
c
o
m
b
i
n
ed
d
at
ase
t,
c
o
m
p
ar
e
d
to
t
h
e
b
as
el
in
e
m
o
d
el'
s
m
ea
n
a
cc
u
r
ac
y
o
f
8
4
.
8
7
%
wi
th
a
s
t
a
n
d
ar
d
d
e
v
i
ati
o
n
o
f
1
.
1
5
%
.
F
o
r
th
e
g
l
au
co
m
a
-
o
n
l
y
d
a
tas
et,
t
h
e
p
r
o
p
o
s
e
d
m
o
d
el
d
e
m
o
n
s
t
r
at
ed
s
u
p
er
io
r
p
e
r
f
o
r
m
an
ce
wi
th
a
m
e
an
a
cc
u
r
ac
y
o
f
9
6
.
4
3
% (
±
0
.
8
9
%
)
,
w
h
i
le
t
h
e
b
aseli
n
e
ac
h
i
ev
ed
8
9
.
8
5
%
(
±
1
.
2
3
%
)
.
Sim
ila
r
l
y
,
o
n
t
h
e
c
ata
r
a
cts
-
o
n
l
y
d
atas
et
,
th
e
p
r
o
p
o
s
e
d
m
o
d
el
att
ai
n
e
d
9
6
.
7
1
%
(
±
0
.
8
2
%)
a
cc
u
r
ac
y
c
o
m
p
a
r
e
d
to
t
h
e
b
ase
li
n
e'
s
8
9
.
3
4
%
(
±
1
.
1
8
%
)
.
T
h
es
e
r
es
u
lts
i
n
d
ic
at
e
th
at
t
h
e
p
r
o
p
o
s
ed
m
o
d
e
l
n
o
t
o
n
ly
a
c
h
ie
v
es
h
i
g
h
e
r
ac
cu
r
a
c
y
b
u
t
als
o
e
x
h
i
b
its
g
r
ea
t
e
r
s
ta
b
ili
ty
ac
r
o
s
s
d
i
f
f
e
r
e
n
t
d
at
a
p
ar
titi
o
n
s
,
as
e
v
i
d
e
n
c
e
d
b
y
t
h
e
l
o
w
e
r
s
t
an
d
a
r
d
d
e
v
ia
ti
o
n
s
.
T
h
is
s
t
u
d
y
c
o
m
p
u
te
d
9
5
%
co
n
f
i
d
e
n
c
e
i
n
te
r
v
als
f
o
r
th
e
m
ea
n
ac
c
u
r
ac
y
t
o
p
r
o
v
i
d
e
a
s
tat
i
s
tica
ll
y
r
ig
o
r
o
u
s
m
e
asu
r
e
o
f
m
o
d
el
r
el
ia
b
il
it
y
.
T
h
e
co
n
f
i
d
e
n
c
e
i
n
t
er
v
a
ls
w
er
e
c
al
cu
l
at
ed
u
s
in
g
:
=
±
1
.
96
×
(
/
√
)
,
wh
er
e
r
ep
r
esen
ts
th
e
n
u
m
b
e
r
o
f
f
o
ld
s
.
Fo
r
th
e
p
r
o
p
o
s
ed
m
o
d
el,
th
e
9
5
%
co
n
f
id
en
ce
in
ter
v
als
o
n
th
e
co
m
b
in
ed
d
ataset
wer
e
[
9
5
.
0
4
%,
9
6
.
4
0
%],
in
d
icatin
g
th
at
we
ca
n
b
e
9
5
%
co
n
f
i
d
en
t
th
at
t
h
e
tr
u
e
p
o
p
u
latio
n
m
ea
n
ac
cu
r
ac
y
lies
with
in
th
is
r
an
g
e.
T
h
e
c
o
r
r
esp
o
n
d
in
g
i
n
ter
v
als
f
o
r
th
e
g
lau
co
m
a
-
o
n
ly
an
d
ca
tar
ac
ts
-
o
n
ly
d
atasets
wer
e
[
9
5
.
5
4
%,
9
7
.
3
2
%]
an
d
[
9
5
.
8
9
%,
9
7
.
5
3
%],
r
esp
ec
tiv
ely
.
I
n
co
n
tr
ast,
t
h
e
b
aselin
e
m
o
d
el
ex
h
ib
ited
wid
e
r
c
o
n
f
id
e
n
ce
in
ter
v
als
o
f
[
8
3
.
8
6
%,
8
5
.
8
8
%]
f
o
r
t
h
e
co
m
b
in
ed
d
ataset,
[
8
8
.
6
2
%,
9
1
.
0
8
%]
f
o
r
g
lau
co
m
a,
a
n
d
[
8
8
.
1
6
%,
9
0
.
5
2
%]
f
o
r
ca
tar
ac
ts
.
T
h
e
n
a
r
r
o
wer
co
n
f
id
en
c
e
in
ter
v
als
o
f
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
d
em
o
n
s
tr
ate
n
o
t
o
n
ly
s
u
p
er
i
o
r
p
er
f
o
r
m
a
n
ce
b
u
t
also
g
r
ea
ter
p
r
ec
is
io
n
in
t
h
e
p
er
f
o
r
m
an
ce
esti
m
ates,
s
u
g
g
esti
n
g
th
at
th
e
im
p
r
o
v
e
m
en
ts
ar
e
s
tatis
tical
ly
r
o
b
u
s
t a
n
d
r
ep
r
o
d
u
ci
b
le.
T
o
d
eter
m
in
e
w
h
eth
er
th
e
p
er
f
o
r
m
an
ce
d
if
f
er
e
n
ce
b
etwe
en
th
e
p
r
o
p
o
s
ed
an
d
b
aselin
e
m
o
d
els
was
s
tatis
t
ically
s
ig
n
if
ican
t,
we
c
o
n
d
u
cte
d
a
p
air
ed
t
-
test
o
n
th
e
ac
cu
r
ac
y
s
co
r
es
o
b
tain
e
d
f
r
o
m
ea
ch
f
o
ld
.
T
h
e
p
air
ed
t
-
test
is
ap
p
r
o
p
r
iate
h
e
r
e
b
ec
au
s
e
ea
ch
f
o
ld
r
e
p
r
esen
ts
a
m
atch
ed
p
air
o
f
o
b
s
er
v
atio
n
s
f
r
o
m
th
e
s
am
e
d
ata
s
u
b
s
et.
T
h
e
r
esu
lts
s
h
o
wed
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
e
l
s
ig
n
if
ican
tly
o
u
tp
er
f
o
r
m
ed
th
e
b
aselin
e
m
o
d
el
ac
r
o
s
s
all
d
atasets
(
p
<
0
.
0
0
1
)
,
with
m
ea
n
ac
cu
r
ac
y
im
p
r
o
v
e
m
en
ts
o
f
1
0
.
8
5
%
o
n
th
e
co
m
b
in
ed
d
ataset,
6
.
5
8
%
o
n
th
e
g
la
u
co
m
a
d
ataset,
an
d
7
.
3
7
%
o
n
th
e
ca
tar
ac
ts
d
atas
et.
T
h
ese
s
tatis
tically
s
ig
n
if
ican
t
im
p
r
o
v
em
e
n
ts
,
co
m
b
in
ed
with
th
e
co
n
s
is
ten
t
p
er
f
o
r
m
an
ce
ac
r
o
s
s
all
f
o
ld
s
an
d
th
e
n
ar
r
o
w
co
n
f
id
e
n
ce
in
ter
v
als,
p
r
o
v
id
e
s
tr
o
n
g
ev
id
en
c
e
th
at
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
an
d
L
B
P
in
teg
r
atio
n
co
n
tr
ib
u
te
s
u
b
s
tan
tially
to
en
h
an
ce
d
d
iag
n
o
s
tic
ac
cu
r
ac
y
an
d
m
o
d
el
r
eliab
ilit
y
.
T
h
e
co
n
s
is
ten
t
s
u
p
er
io
r
p
e
r
f
o
r
m
an
ce
ac
r
o
s
s
m
u
ltip
le
in
d
ep
en
d
e
n
t
d
ata
s
p
lits
d
em
o
n
s
tr
ates
th
at
o
u
r
m
o
d
el'
s
ad
v
an
tag
es
ar
e
n
o
t
ar
tifa
cts
o
f
a
p
ar
ticu
la
r
tr
ain
-
test
s
p
lit
b
u
t
r
ep
r
esen
t g
en
u
in
e
im
p
r
o
v
e
m
e
n
ts
in
th
e
ab
ilit
y
to
lear
n
r
o
b
u
s
t f
ea
tu
r
es f
r
o
m
lim
ited
m
ed
ica
l im
ag
in
g
d
ata.
T
ab
le
3
.
5
-
f
o
ld
v
alid
atio
n
r
esu
lts
f
o
r
th
e
b
aselin
e
m
o
d
el
u
s
in
g
th
e
co
m
b
in
ed
d
ataset
F
o
l
d
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
AUC
1
0
.
8
4
9
6
0
.
8
2
0
5
0
.
8
3
8
1
0
.
8
2
8
4
0
.
9
0
8
7
2
0
.
8
3
1
9
0
.
8
1
4
2
0
.
8
2
7
3
0
.
8
2
0
2
0
.
9
0
1
5
3
0
.
8
5
8
4
0
.
8
3
1
2
0
.
8
4
5
5
0
.
8
3
7
8
0
.
9
2
0
1
4
0
.
8
4
0
7
0
.
8
1
8
9
0
.
8
2
9
8
0
.
8
2
3
9
0
.
9
0
9
4
5
0
.
8
6
2
8
0
.
8
3
5
6
0
.
8
5
0
1
0
.
8
4
2
3
0
.
9
2
5
3
M
e
a
n
0
.
8
4
8
7
0
.
8
2
4
1
0
.
8
3
8
2
0
.
8
3
0
5
0
.
9
1
3
0
S
t
d
.
d
e
v
0
.
0
1
1
5
0
.
0
0
8
5
0
.
0
0
9
3
0
.
0
0
8
8
0
.
0
0
9
8
T
h
e
9
5
% c
o
n
f
id
en
ce
in
ter
v
al
f
o
r
th
e
b
ase
m
o
d
el
u
s
in
g
th
e
co
m
b
in
ed
d
ataset:
i)
Acc
u
r
ac
y
: 8
4
.
8
7
% ±
1
.
0
1
%
≥
(
8
3
.
8
6
%
-
8
5
.
8
8
%)
.
ii)
Pre
cisi
o
n
: 8
2
.
4
1
% ±
0
.
7
5
%
≥
(
8
1
.
6
6
%
-
8
3
.
1
6
%)
.
iii)
R
ec
all: 8
3
.
8
2
% ±
0
.
8
2
%
≥
(
8
3
.
0
0
%
-
8
4
.
6
4
%)
.
iv
)
F1
-
s
co
r
e
: 8
3
.
0
5
% ±
0
.
7
7
%
≥
(
8
2
.
2
8
%
-
8
3
.
8
2
%)
.
v)
AUC:
9
1
.
3
0
% ±
0
.
8
6
%
≥
(
9
0
.
4
4
%
-
9
2
.
1
6
%)
.
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.
15
,
No
.
4
,
Au
g
u
s
t
20
26
:
3
7
4
5
-
3
7
6
0
3754
T
ab
le
4
.
5
-
f
o
ld
v
alid
atio
n
r
esu
lts
f
o
r
th
e
p
r
o
p
o
s
ed
m
o
d
el
u
s
in
g
th
e
c
o
m
b
in
e
d
d
ataset
F
o
l
d
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
AUC
1
0
.
9
5
5
8
0
.
9
5
2
3
0
.
9
5
3
1
0
.
9
5
2
7
0
.
9
7
4
5
2
0
.
9
4
6
9
0
.
9
4
4
5
0
.
9
4
5
3
0
.
9
4
4
9
0
.
9
6
8
2
3
0
.
9
6
4
6
0
.
9
6
1
8
0
.
9
6
2
5
0
.
9
6
2
1
0
.
9
8
1
2
4
0
.
9
5
1
3
0
.
9
5
8
9
0
.
9
4
9
7
0
.
9
4
9
3
0
.
9
7
2
1
5
0
.
9
6
7
3
0
.
9
6
4
1
0
.
9
6
4
9
0
.
9
6
4
5
0
.
9
8
2
8
M
e
a
n
0
.
9
5
7
2
0
.
9
5
4
3
0
.
9
5
5
1
0
.
9
5
4
7
0
.
9
7
5
8
S
t
d
.
d
e
v
0
.
0
0
7
8
0
.
0
0
7
6
0
.
0
0
7
7
0
.
0
0
7
6
0
.
0
0
5
9
T
h
e
9
5
%
co
n
f
id
en
ce
in
ter
v
als
f
o
r
th
e
p
r
o
p
o
s
ed
m
o
d
el
u
s
in
g
t
h
e
co
m
b
i
n
ed
d
ataset:
i)
Acc
u
r
ac
y
: 9
5
.
7
2
% ±
0
.
6
8
%
≥
(
9
5
.
0
4
%
-
9
6
.
4
0
%)
.
ii)
Pre
cisi
o
n
: 9
5
.
4
3
% ±
0
.
6
%
≥
(
9
4
.
7
6
%
-
9
6
.
1
0
%)
.
iii)
R
ec
all: 9
5
.
5
1
% ±
0
.
6
8
%
≥
(
9
4
.
8
3
%
-
9
6
.
1
9
%)
.
iv
)
F1
-
s
co
r
e
: 9
5
.
4
7
% ±
0
.
6
7
%
≥
(
9
4
.
8
0
%
-
9
6
.
1
4
%)
.
v)
AUC:
9
7
.
5
8
% ±
0
.
5
2
%
≥
(
9
7
.
0
6
%
-
9
8
.
1
0
%)
.
3
.
3
.
Abla
t
io
n study
T
h
is
s
ec
tio
n
p
r
esen
ts
a
s
y
s
tem
atic
ab
latio
n
an
aly
s
is
o
f
o
u
r
p
r
o
p
o
s
ed
C
ap
s
Net
alg
o
r
ith
m
,
w
h
ich
aim
s
to
d
is
s
ec
t
th
e
co
n
tr
ib
u
tio
n
s
o
f
ea
ch
co
m
p
o
n
e
n
t
to
th
e
m
o
d
e
l's
p
er
f
o
r
m
an
ce
.
B
y
s
y
s
tem
atica
lly
r
em
o
v
in
g
o
r
m
o
d
if
y
in
g
in
d
iv
i
d
u
al
co
m
p
o
n
en
ts
,
we
in
v
esti
g
ate
th
e
cr
itica
l e
lem
en
ts
th
at
d
r
iv
e
th
e
ac
cu
r
ac
y
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
Pre
s
en
ted
in
T
ab
le
5
,
th
is
s
tu
d
y
o
b
s
er
v
ed
th
at
r
em
o
v
in
g
th
e
AC
SF
lay
er
1
r
esu
lted
in
a
d
ec
r
ea
s
e
in
ac
cu
r
ac
y
o
f
3
.
4
8
%
o
n
th
e
co
m
b
in
ed
d
ataset,
3
.
7
9
%
o
n
th
e
g
lau
co
m
a
d
ataset,
an
d
3
.
3
3
%
o
n
th
e
ca
tar
ac
ts
d
ataset.
R
em
o
v
in
g
AC
SF
lay
er
2
led
to
a
d
ec
r
ea
s
e
in
ac
c
u
r
a
cy
o
f
5
.
9
5
%
o
n
t
h
e
co
m
b
in
ed
d
ataset,
5
.
0
3
%
o
n
th
e
g
lau
co
m
a
d
ataset,
an
d
4
.
4
1
%
o
n
th
e
ca
tar
ac
t
d
ataset.
R
em
o
v
in
g
b
o
t
h
AC
SF
lay
er
s
1
an
d
2
r
esu
lted
in
a
s
ig
n
if
ican
t
d
ec
r
ea
s
e
in
ac
c
u
r
a
cy
o
f
1
0
.
2
8
%
o
n
th
e
co
m
b
in
e
d
d
ataset,
1
0
.
2
1
%
o
n
th
e
g
lau
co
m
a
d
ataset,
an
d
1
0
.
5
6
%
o
n
t
h
e
ca
tar
ac
ts
d
atas
et.
R
em
o
v
in
g
c
o
n
v
o
lu
tio
n
al
lay
er
s
2
an
d
3
r
esu
lted
i
n
a
d
ec
r
ea
s
e
in
ac
cu
r
ac
y
o
f
2
.
8
7
%
o
n
th
e
co
m
b
i
n
ed
d
at
aset,
2
.
0
7
%
o
n
th
e
g
lau
co
m
a
d
ataset,
an
d
1
.
6
9
%
o
n
th
e
ca
tar
ac
ts
d
ataset.
R
em
o
v
in
g
L
B
P
lay
er
1
,
in
c
o
m
b
in
atio
n
with
r
em
o
v
in
g
A
C
SF
lay
er
s
1
an
d
2
,
r
esu
lted
in
a
d
ec
r
ea
s
e
in
ac
cu
r
ac
y
o
f
5
.
5
1
%
o
n
th
e
co
m
b
in
ed
d
ataset,
5
.
5
5
%
o
n
th
e
g
lau
co
m
a
d
ataset,
an
d
5
.
1
7
%
o
n
th
e
ca
tar
ac
ts
d
ataset.
R
em
o
v
in
g
AC
SF
lay
er
1
,
co
n
v
o
lu
tio
n
al
lay
er
1
,
a
n
d
L
B
P
lay
er
1
r
esu
lted
in
a
d
ec
r
ea
s
e
in
ac
cu
r
ac
y
o
f
6
.
1
6
%
o
n
t
h
e
co
m
b
in
ed
d
at
aset,
5
.
8
1
%
o
n
t
h
e
g
lau
c
o
m
a
d
ataset,
an
d
5
.
5
9
%
o
n
t
h
e
ca
ta
r
ac
ts
d
ataset.
T
h
ese
r
esu
lts
s
u
g
g
est
th
at
th
e
AC
SF
lay
er
s
,
p
ar
ticu
lar
ly
la
y
er
1
,
a
r
e
cr
u
cial
f
o
r
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
,
w
h
ile
th
e
co
n
v
o
l
u
tio
n
al
lay
er
s
2
an
d
3
h
av
e
a
s
m
aller
im
p
ac
t o
n
ac
cu
r
ac
y
.
Ad
d
itio
n
ally
,
th
e
L
B
P
lay
er
1
h
as
a
s
ig
n
if
ican
t
im
p
ac
t
wh
en
co
m
b
in
ed
with
r
em
o
v
in
g
AC
SF
lay
er
s
.
I
n
s
im
p
le
ter
m
s
,
th
e
a
b
latio
n
s
tu
d
y
s
h
o
wed
th
at
as f
o
llo
ws
:
i)
T
h
e
AC
SF
lay
er
s
ar
e
v
er
y
im
p
o
r
tan
t f
o
r
th
e
p
r
o
p
o
s
ed
m
o
d
el's ac
cu
r
ac
y
,
esp
ec
ially
th
e
f
ir
s
t A
C
SF
lay
er
.
ii)
R
em
o
v
in
g
th
e
AC
SF
lay
er
s
s
ig
n
if
ican
tly
r
e
d
u
ce
s
th
e
p
r
o
p
o
s
ed
m
o
d
el'
s
p
er
f
o
r
m
an
ce
.
iii)
T
h
e
co
n
v
o
lu
tio
n
al
la
y
er
s
(
2
a
n
d
3
)
h
av
e
a
s
m
aller
im
p
ac
t
o
n
ac
cu
r
ac
y
.
iv
)
T
h
e
L
B
P lay
er
1
also
p
lay
s
a
s
ig
n
if
ican
t r
o
le,
esp
ec
ially
wh
e
n
co
m
b
i
n
ed
with
th
e
AC
SF
lay
er
s
.
T
ab
le
5
.
Ab
latio
n
s
tu
d
ies
La
y
e
r
s
C
o
m
b
i
n
e
d
d
a
t
a
set
(
%)
G
l
a
u
c
o
ma
d
a
t
a
s
e
t
(
%)
C
a
t
a
r
a
c
t
d
a
t
a
s
e
t
(
%)
*
A
C
S
F
1
9
2
.
1
5
9
2
.
9
9
9
3
.
1
2
*
A
C
S
F
2
8
9
.
6
8
9
1
.
7
5
9
2
.
3
4
*
A
C
S
F
1
a
n
d
2
8
5
.
3
5
8
6
.
5
7
8
5
.
8
9
*
A
C
S
F
1
a
n
d
C
o
n
v
1
8
9
.
4
7
9
1
.
6
7
9
2
.
8
6
*
A
C
S
F
1
,
2
,
a
n
d
L
B
P
l
a
y
e
r
1
8
5
.
1
2
8
6
.
2
3
8
7
.
2
8
*
C
o
n
v
2
a
n
d
3
9
2
.
7
6
9
4
.
7
1
9
4
.
7
6
*
LB
P
l
a
y
e
r
1
,
A
C
S
F
2
,
a
n
d
C
o
n
v
1
9
0
.
3
9
9
1
.
9
7
9
2
.
3
4
*
mea
n
s
r
e
m
o
v
a
l
o
f
a
l
a
y
e
r
3
.
4
.
Vis
ua
l a
na
ly
s
is
I
n
th
is
s
ec
tio
n
,
th
is
s
tu
d
y
v
is
u
ally
an
aly
ze
th
e
f
ea
tu
r
e
m
ap
s
,
clu
s
ter
s
,
an
d
r
ec
o
n
s
tr
u
cted
im
ag
es
p
r
o
d
u
ce
d
b
y
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
an
d
th
e
b
aselin
e
m
o
d
el
to
g
ain
a
d
ee
p
er
u
n
d
er
s
tan
d
in
g
o
f
th
eir
p
er
f
o
r
m
an
ce
.
Fig
u
r
e
6
s
h
o
ws
th
e
co
m
p
ar
is
o
n
o
f
f
ir
s
t
co
n
v
o
lu
tio
n
al
lay
er
f
ea
tu
r
e
m
ap
s
.
T
h
e
f
ea
tu
r
e
m
ap
s
g
en
er
ated
b
y
th
e
p
r
o
p
o
s
ed
m
o
d
el
(
Fig
u
r
e
6
(
a)
)
ar
e
s
ig
n
if
ica
n
tly
s
m
o
o
th
e
r
an
d
m
o
r
e
v
is
u
ally
ap
p
ea
lin
g
co
m
p
ar
ed
to
th
o
s
e
p
r
o
d
u
ce
d
b
y
t
h
e
b
aselin
e
m
o
d
el
(
Fig
u
r
e
6
(
b
)
)
.
T
h
e
f
ea
tu
r
e
s
in
o
u
r
p
r
o
p
o
s
ed
m
o
d
el'
s
f
ea
tu
r
e
m
ap
s
a
r
e
m
o
r
e
d
ef
in
ed
a
n
d
ea
s
ily
id
en
tifia
b
le
,
wh
er
ea
s
th
e
b
aselin
e
m
o
d
el'
s
f
ea
tu
r
e
m
a
p
s
h
av
e
a
lo
t o
f
b
la
ck
ed
-
o
u
t
ar
ea
s
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.