I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
15
,
No
.
1
,
M
ar
ch
20
26
,
p
p
.
414
~
427
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v15
i
1
.
p
p
4
1
4
-
4
2
7
414
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
Lig
htweight d
eep
lea
rning
appro
a
ch
for re
tinal
O
CT
ima
g
e
cla
ss
ificatio
n:
A
CNN
with
hy
brid
poo
ling
and o
pti
mized
lea
rning
P
a
rt
h R.
Da
v
e
1
.2
,
Nik
un
j
H
.
Do
m
a
diy
a
2
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
En
g
i
n
e
e
r
i
n
g
,
G
u
j
a
r
a
t
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
si
t
y
-
A
h
me
d
a
b
a
d
,
G
u
j
a
r
a
t
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
En
g
i
n
e
e
r
i
n
g
,
L.
D
.
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
A
h
m
e
d
a
b
a
d
,
G
u
j
a
r
a
t
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ap
r
1
6
,
2
0
2
5
R
ev
is
ed
Oct
1
6
,
2
0
2
5
Acc
ep
ted
No
v
5
,
2
0
2
5
Op
ti
c
a
l
c
o
h
e
re
n
c
e
to
m
o
g
ra
p
h
y
(OCT)
is
a
n
o
n
-
in
v
a
siv
e
tec
h
n
i
q
u
e
th
ro
u
g
h
wh
ich
a
re
ti
n
a
sp
e
c
ialist
c
a
n
se
e
th
e
stru
c
tu
re
b
e
h
i
n
d
t
h
e
e
y
e
.
T
h
is
tec
h
n
o
l
o
g
y
o
ffe
rs
a
k
e
y
r
o
le
to
id
e
n
ti
fy
v
a
rio
u
s
a
b
n
o
rm
a
li
ti
e
s
in
th
e
re
ti
n
a
:
Dru
se
n
,
d
iab
e
ti
c
m
a
c
u
lar
e
d
e
m
a
(DME
)
a
n
d
c
h
o
ro
i
d
a
l
n
e
o
v
a
sc
u
lariz
a
ti
o
n
(CNV
).
Ho
we
v
e
r,
m
a
n
u
a
l
a
n
a
ly
sis
o
f
OC
T
sc
a
n
s
c
a
n
b
e
ti
m
e
-
c
o
n
su
m
in
g
a
n
d
p
ro
n
e
to
v
a
riab
il
it
y
a
m
o
n
g
c
li
n
icia
n
s.
To
a
d
d
re
ss
th
is
c
h
a
ll
e
n
g
e
,
we
p
re
se
n
t
a
li
g
h
twe
i
g
h
t
a
n
d
e
x
p
lain
a
b
le
d
e
e
p
lea
rn
in
g
-
b
a
se
d
a
p
p
ro
a
c
h
f
o
r
a
u
to
m
a
ti
c
cl
a
ss
ifi
c
a
ti
o
n
o
f
re
ti
n
a
l
OCT
ima
g
e
s.
Th
e
p
rima
ry
g
o
a
l
o
f
th
is
re
s
e
a
rc
h
is
a
m
o
d
e
l
t
h
a
t
d
e
li
v
e
rs
h
i
g
h
d
iag
n
o
s
ti
c
a
c
c
u
ra
c
y
.
A
c
o
m
p
u
ter
-
a
id
e
d
s
u
g
g
e
sti
v
e
m
e
th
o
d
c
a
n
h
e
l
p
re
ti
n
a
l
d
o
c
t
o
rs
a
u
to
m
a
ti
c
a
ll
y
c
las
sify
th
e
a
n
o
m
a
li
e
s
with
m
o
re
c
o
n
fi
d
e
n
c
e
a
n
d
p
re
c
isio
n
.
I
n
t
h
is
p
a
p
e
r,
we
p
r
o
p
o
se
d
a
n
o
v
e
l
a
p
p
r
o
a
c
h
b
a
se
d
o
n
d
e
e
p
lea
rn
in
g
:
a
si
x
-
l
a
y
e
r
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
(CNN
)
i
n
t
e
g
r
a
t
e
d
w
i
t
h
h
y
b
r
i
d
p
o
o
l
i
n
g
f
o
r
e
f
fe
c
t
i
v
e
f
e
a
t
u
r
e
e
x
t
r
a
c
t
io
n
.
D
a
t
a
a
u
g
m
e
n
t
a
t
i
o
n
a
n
d
e
x
p
o
n
e
n
t
i
a
l
l
e
a
r
n
i
n
g
r
a
t
e
i
s
i
m
p
l
e
m
e
n
t
e
d
t
o
h
a
n
d
l
e
d
a
t
a
i
m
b
a
l
a
n
c
e
b
e
t
w
e
e
n
c
l
a
s
s
e
s
a
n
d
f
o
r
s
t
a
b
i
l
i
z
e
d
l
e
a
r
n
i
n
g
c
o
n
s
e
c
u
t
i
v
e
l
y
.
O
u
r
p
r
o
p
o
s
e
d
a
p
p
r
o
a
c
h
a
c
h
i
e
v
e
d
9
8
.
7
5
%
o
f
a
c
c
u
r
a
c
y
w
h
i
l
e
t
e
s
t
i
n
g
o
n
t
h
e
d
a
t
a
se
t
.
T
o
f
u
r
t
h
e
r
e
n
h
a
n
c
e
t
h
e
i
n
t
e
r
p
r
e
t
a
b
i
l
i
t
y
o
f
t
h
e
m
o
d
e
l
,
w
e
a
l
s
o
i
n
t
e
g
r
a
t
e
e
x
p
l
a
i
n
a
b
l
e
A
I
(
X
A
I)
u
s
i
n
g
c
l
a
s
s
a
c
t
i
v
a
t
i
o
n
m
a
p
p
i
n
g
(
C
A
M
)
t
o
v
i
s
u
a
l
i
z
e
t
h
e
c
r
i
t
i
c
a
l
r
e
g
i
o
n
s
i
n
t
h
e
r
e
t
i
n
a
t
h
a
t
c
o
n
t
r
i
b
u
t
e
t
o
t
h
e
c
l
a
s
s
i
f
i
c
a
t
i
o
n
d
e
c
i
s
io
n
s
.
K
ey
w
o
r
d
s
:
Ar
tific
ial
in
tellig
en
ce
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
Ma
ch
in
e
lear
n
in
g
Op
tical
co
h
er
en
ce
to
m
o
g
r
ap
h
y
Dee
p
lear
n
in
g
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Nik
u
n
j H
.
Do
m
ad
i
y
a
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
E
n
g
i
n
ee
r
in
g
,
L
.
D.
C
o
l
leg
e
o
f
E
n
g
i
n
ee
r
in
g
Ah
m
ed
ab
ad
,
Gu
jar
at,
I
n
d
ia
E
m
ail: d
o
m
ad
iy
a
n
ik
u
n
j0
0
2
@
g
m
ail.
co
m
,
n
ik
u
n
j
d
o
m
ad
i
y
a.
c
e@
ld
ce
.
ac
.
in
1.
I
NT
RO
D
UCT
I
O
N
R
etin
al
o
p
tical
co
h
er
en
ce
to
m
o
g
r
ap
h
y
(
OC
T
)
im
ag
in
g
p
lay
s
a
v
ital
r
o
le
in
an
aly
zin
g
th
e
lay
er
s
o
f
th
e
r
etin
a
with
o
u
t
an
y
s
u
r
g
ical
tr
ea
tm
en
t
in
th
e
ey
e.
I
t
h
elp
s
id
en
tify
d
if
f
e
r
en
t
a
n
o
m
alies
an
d
th
eir
p
r
o
g
r
ess
io
n
,
i
f
an
y
.
I
n
o
p
h
th
alm
o
lo
g
y
,
th
e
u
s
e
o
f
r
etin
al
OC
T
im
ag
in
g
p
lay
s
a
cr
u
ci
al
r
o
le
in
s
h
ap
i
n
g
tr
ea
tm
en
t
d
ec
is
io
n
s
an
d
k
ee
p
i
n
g
a
watc
h
o
n
th
e
s
u
cc
ess
o
f
th
er
ap
ies
s
u
ch
as
laser
tr
ea
tm
en
t,
ey
e
in
jectio
n
s
,
an
d
s
u
r
g
er
ies.
C
lass
if
y
in
g
r
et
in
al
OC
T
im
ag
es
with
th
e
h
elp
o
f
m
o
d
er
n
c
o
m
p
u
ter
ized
tech
n
iq
u
es
ca
n
b
e
lev
er
ag
ed
i
n
ter
m
s
o
f
s
u
g
g
esti
v
e
s
y
s
tem
s
to
th
e
r
etin
a
s
p
ec
ialis
t.
T
h
e
ab
ilit
y
o
f
au
to
m
ated
im
a
g
e
ca
teg
o
r
izatio
n
is
to
p
r
o
v
id
e
r
e
liab
le
an
d
im
p
ar
tial
ev
alu
atio
n
s
o
f
r
etin
al
i
m
ag
es,
m
in
im
izi
n
g
h
u
m
an
e
r
r
o
r
an
d
d
iag
n
o
s
tic
in
co
n
s
is
ten
cies.
C
o
n
v
en
tio
n
al
m
ac
h
in
e
lear
n
i
n
g
alg
o
r
ith
m
s
ta
k
e
th
e
tex
tu
r
e
f
ea
tu
r
es
o
f
th
e
im
ag
es
as
an
i
n
p
u
t
th
r
o
u
g
h
wh
ich
lo
ca
l
o
r
g
lo
b
al
in
f
o
r
m
atio
n
o
f
th
e
im
a
g
e
ca
n
b
e
r
ep
r
esen
ted
.
On
ce
th
e
f
ea
tu
r
e
v
ec
to
r
is
im
p
lem
en
ted
b
y
d
if
f
er
en
t f
ea
t
u
r
e
ex
tr
ac
tio
n
tech
n
i
q
u
es: L
o
ca
l
b
in
ar
y
p
atter
n
(
L
B
P)
[
1
]
,
h
i
s
to
g
r
am
o
f
o
r
ien
te
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t d
ee
p
lea
r
n
in
g
a
p
p
r
o
a
ch
fo
r
r
etin
a
l O
C
T ima
g
e
cla
s
s
ifica
tio
n
:
A
C
N
N
w
ith
…
(
P
a
r
th
R
.
Da
ve
)
415
g
r
ad
ien
ts
(
HOG
)
[
2
]
,
g
r
ay
-
le
v
el
co
-
o
cc
u
r
r
en
ce
m
atr
ix
(
GL
C
M)
[
3
]
,
ed
g
e
d
etec
tio
n
,
a
n
d
s
h
ap
e
d
etec
tio
n
,
th
e
v
ec
to
r
is
p
r
o
v
id
e
d
to
an
ar
ti
f
icial
n
eu
r
al
n
etwo
r
k
(
ANN)
,
s
u
p
p
o
r
t
v
ec
t
o
r
m
ac
h
in
e
(
SVM)
[
4
]
o
r
o
th
er
s
u
p
er
v
is
ed
alg
o
r
ith
m
wh
ich
ev
en
tu
ally
class
if
y
th
e
i
m
a
g
e
in
th
e
r
e
q
u
ir
e
d
class
.
T
h
e
lag
g
i
n
g
p
ar
t
o
f
co
n
v
en
tio
n
al
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
s
is
th
at
th
e
im
ag
es
ar
e
d
escr
ib
ed
b
y
s
o
m
e
f
e
atu
r
e
an
d
it
is
n
o
t
lear
n
in
g
th
e
f
ea
tu
r
es
its
elf
,
h
e
n
ce
m
ay
lea
d
to
lo
s
s
y
in
f
o
r
m
a
tio
n
r
ep
r
esen
tatio
n
an
d
en
d
u
p
with
less
ac
cu
r
ac
y
o
f
th
e
alg
o
r
ith
m
.
Dee
p
lear
n
in
g
h
as
attr
ac
ted
s
i
g
n
if
ican
t
in
ter
est
f
r
o
m
r
esear
ch
er
s
in
co
m
p
u
ter
v
is
io
n
b
ec
au
s
e
o
f
its
ca
p
ab
ilit
y
to
i
n
d
ep
e
n
d
en
tly
lear
n
in
tr
icate
d
etails
an
d
f
ea
tu
r
es
f
r
o
m
in
p
u
t
im
ag
es.
C
NN
h
as
ch
an
g
ed
th
e
m
o
d
u
s
o
p
er
an
d
i o
f
s
h
allo
w
an
d
co
n
v
e
n
tio
n
al
n
etwo
r
k
s
,
an
d
it
ex
ce
ed
s
th
eir
p
er
f
o
r
m
an
c
e
o
v
er
th
em
in
p
o
ts
o
f
asp
ec
ts
.
Dee
p
lear
n
in
g
is
n
’
t c
o
n
f
in
ed
to
ju
s
t c
lass
if
icatio
n
task
s
—
it e
x
ten
d
s
its
ap
p
licatio
n
s
to
a
wid
e
r
an
g
e
o
f
ar
ea
s
,
in
clu
d
in
g
o
b
ject
d
etec
tio
n
,
m
e
d
ical
im
ag
in
g
,
v
id
eo
an
aly
s
is
,
s
p
ee
ch
r
ec
o
g
n
itio
n
,
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
,
an
d
b
ey
o
n
d
.
I
n
r
ec
en
t
tim
e,
f
o
r
th
e
f
ield
o
f
co
m
p
u
ter
v
is
io
n
,
m
an
y
p
r
etr
a
in
ed
m
o
d
els:
Alex
Net
[
5
]
,
V
GG1
6
[
6
]
,
VGG1
9
[
6
]
,
R
esn
et5
0
[
7
]
,
a
n
d
Go
o
g
leNe
t
[
8
]
.
a
r
e
a
v
ailab
l
e
,
wh
ich
ar
e
ac
tu
ally
tr
ain
ed
o
n
v
er
y
lar
g
e
im
a
g
e
d
atasets
to
lear
n
th
e
d
i
v
er
s
ity
.
T
h
ese
m
o
d
els
o
f
ten
c
o
n
tain
a
lar
g
e
a
n
d
d
iv
er
s
e
s
et
o
f
tr
ain
ab
le
p
ar
am
ete
r
s
,
wh
ich
ca
n
b
e
u
tili
ze
d
f
o
r
f
e
atu
r
e
ex
tr
ac
tio
n
f
r
o
m
a
s
p
ec
i
f
ic
im
ag
e
d
ataset
b
y
elim
in
a
tin
g
th
e
f
in
al
f
u
lly
co
n
n
ec
ted
la
y
er
(
s
)
.
C
NN
lear
n
s
th
e
s
p
a
tial
in
f
o
r
m
atio
n
f
r
o
m
in
p
u
t
im
ag
e
o
r
d
ata
with
th
e
h
elp
o
f
a
lay
er
ed
co
n
v
o
l
u
tio
n
al
an
d
p
o
o
lin
g
ar
c
h
itectu
r
e
wh
ich
th
en
f
i
n
ally
g
ets
co
n
n
ec
ted
to
th
e
d
en
s
e
lay
er
f
o
r
th
e
en
d
p
o
i
n
t
class
if
icatio
n
.
T
h
e
d
en
s
e
s
ec
tio
n
m
ay
co
n
s
is
t
o
f
o
n
e
o
r
m
o
r
e
f
u
lly
co
n
n
ec
ted
h
id
d
en
u
n
its
p
r
ec
e
d
in
g
th
e
f
in
al
o
u
tp
u
t
co
m
p
o
n
en
t.
T
h
is
o
u
tp
u
t
co
m
p
o
n
en
t
co
n
tain
s
n
eu
r
o
n
s
co
r
r
esp
o
n
d
in
g
to
t
h
e
n
u
m
b
er
o
f
class
es
in
th
e
d
ataset,
en
ab
lin
g
class
if
icatio
n
.
T
h
e
r
e
ar
e
m
an
y
ac
tiv
ati
o
n
f
u
n
ctio
n
s
,
p
ar
a
m
eter
s
,
h
y
p
er
p
ar
am
eter
s
an
d
m
ec
h
an
is
m
s
av
aila
b
le
to
h
an
d
le
th
e
u
n
d
er
f
itti
n
g
o
r
o
v
er
f
itti
n
g
o
f
th
e
m
o
d
el
an
d
m
ak
e
it
m
o
r
e
g
en
er
alis
e
d
to
war
d
s
th
e
test
d
ataset.
T
o
f
u
r
th
er
en
h
an
ce
th
e
in
ter
p
r
etab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
C
NN
m
o
d
el
th
r
o
u
g
h
ex
p
lain
ab
le
AI
,
class
ac
tiv
atio
n
m
ap
p
in
g
(
C
AM
)
[
9
]
was
em
p
lo
y
ed
.
T
h
i
s
m
eth
o
d
en
ab
les
v
is
u
aliza
tio
n
o
f
th
e
r
eg
io
n
s
with
in
r
etin
al
OC
T
im
ag
es
th
at
co
n
tr
ib
u
te
m
o
s
t
s
ig
n
if
ican
tly
to
class
if
icatio
n
o
u
tco
m
es.
B
y
g
en
er
atin
g
C
AM
s
at
d
if
f
er
e
n
t
co
n
v
o
l
u
tio
n
al
lay
e
r
s
,
th
e
f
ea
tu
r
e
lear
n
in
g
p
r
o
ce
s
s
o
f
th
e
m
o
d
el
ca
n
b
e
a
n
aly
ze
d
in
a
clin
ically
m
ea
n
in
g
f
u
l m
a
n
n
er
.
T
h
is
p
ap
er
is
d
iv
id
ed
in
t
o
f
iv
e
s
ec
tio
n
s
.
Sectio
n
2
co
v
er
s
th
e
R
elate
d
W
o
r
k
.
T
h
e
Pro
p
o
s
ed
Ap
p
r
o
ac
h
is
co
v
er
ed
in
s
ec
tio
n
3
.
Sectio
n
4
co
n
tain
s
E
x
p
er
im
en
tal
Setu
p
an
d
R
esu
lts
an
d
f
in
ally
th
e
last
s
ec
tio
n
co
v
er
s
th
e
C
o
n
clu
s
io
n
an
d
F
u
tu
r
e
W
o
r
k
.
2.
R
E
L
AT
E
D
WO
RK
Me
d
ical
im
ag
e
p
r
o
ce
s
s
in
g
h
as
b
ec
o
m
e
a
v
er
y
g
o
o
d
s
u
g
g
esti
v
e
s
y
s
tem
in
r
ec
en
t
er
a
d
u
e
t
o
its
ca
p
ac
ity
to
g
en
er
ate
m
o
r
e
ac
c
u
r
ate
an
d
less
er
r
o
r
p
r
u
n
e
r
es
u
lts
.
Ov
er
tim
e,
a
v
ar
iety
o
f
tr
ad
itio
n
al
as
well
a
s
cu
ttin
g
-
ed
g
e
m
eth
o
d
s
h
av
e
b
e
en
ap
p
lied
f
o
r
t
h
e
class
if
icatio
n
o
f
r
eti
n
al
OC
T
im
ag
es.
Srin
iv
asan
et
a
l.
[
1
0
]
a
p
p
lie
d
im
ag
e
d
en
o
is
in
g
,
r
etin
al
c
u
r
v
atu
r
e
c
o
r
r
ec
tio
n
,
an
d
r
e
g
i
o
n
-
f
o
c
u
s
ed
cr
o
p
p
in
g
b
ef
o
r
e
ex
tr
ac
tin
g
f
e
atu
r
e
v
ec
to
r
s
u
s
in
g
HOG
d
escr
ip
to
r
s
[
2
]
f
o
r
r
etin
al
OC
T
im
ag
es.
E
ac
h
b
lo
c
k
’
s
d
escr
ip
to
r
v
ec
to
r
n
o
r
m
alize
d
as
with
a
s
m
all
co
n
s
tan
t.
T
h
e
v
alu
es
o
f
th
e
v
ec
to
r
s
wer
e
ca
p
p
ed
an
d
r
en
o
r
m
alize
d
.
T
h
e
f
in
al
f
ea
tu
r
e
v
ec
to
r
co
m
p
r
is
ed
n
o
r
m
alize
d
h
is
to
g
r
am
s
f
r
o
m
all
th
e
b
lo
ck
s
.
Fo
r
m
u
lti
class
class
if
icatio
n
,
th
ey
h
av
e
u
s
ed
S
VM
[
4
]
class
if
ier
in
o
n
e
v
s
o
n
e
m
eth
o
d
.
I
t
co
n
s
is
ts
o
f
s
u
ch
th
r
ee
lin
ea
r
SVMs
as
th
e
d
ataset
co
n
tain
s
th
r
ee
class
es:
AM
D,
DM
E
an
d
No
r
m
al.
T
h
e
r
etin
al
im
ag
e
d
ataset
f
o
r
th
e
s
aid
p
u
r
p
o
s
e
was c
r
ea
ted
lo
ca
lly
b
y
ta
k
in
g
t
h
e
im
ag
es o
f
v
ar
io
u
s
p
atien
ts
with
r
etin
al
d
is
o
r
d
er
s
.
L
iu
et
a
l.
[
1
1
]
p
r
o
p
o
s
ed
an
o
th
er
class
ical
m
ac
h
in
e
lear
n
in
g
m
eth
o
d
to
d
iag
n
o
s
e
m
ac
u
lar
p
ath
o
lo
g
ies
in
r
etin
al
OC
T
im
ag
es.
T
h
ey
u
s
ed
im
ag
e
alig
n
m
en
t
a
n
d
co
n
s
tr
u
ctio
n
o
f
g
l
o
b
al
an
d
lo
ca
l
f
ea
tu
r
es
o
f
t
h
e
im
ag
es.
T
h
e
g
lo
b
al
d
escr
ip
to
r
u
tili
ze
s
a
m
u
ltis
ca
le
s
p
atial
p
y
r
am
id
[
1
2
]
,
wh
ile
t
h
e
lo
ca
l
d
e
s
cr
ip
to
r
em
p
lo
y
s
a
PC
A
-
b
ased
r
ed
u
ce
d
L
B
P
h
is
to
g
r
am
[
1
]
.
T
h
ey
r
ec
o
r
d
ed
g
o
o
d
ac
cu
r
ac
y
with
th
e
h
elp
o
f
a
SVM
an
d
lo
ca
lly
cr
ea
ted
d
ataset
f
o
r
a
n
o
m
alies:
MH
,
m
ac
u
lar
ed
em
a
(
ME
)
a
n
d
AM
D.
Als
aih
et
a
l.
[
1
3
]
,
d
ev
el
o
p
ed
y
et
an
o
th
er
co
n
v
en
ti
o
n
al
ap
p
r
o
ac
h
u
s
in
g
f
ea
tu
r
e
ex
tr
ac
tio
n
t
ec
h
n
iq
u
es:
HOG
[
2
]
an
d
L
B
P
[
1
]
.
T
o
r
ed
u
ce
th
e
d
im
en
s
io
n
ality
o
f
t
h
e
f
ea
tu
r
e
v
ec
to
r
,
p
r
in
cip
al
c
o
m
p
o
n
en
t
an
aly
s
is
(
PC
A)
was a
p
p
lied
an
d
af
ter
t
h
at
SVM
[
4
]
was a
p
p
lied
f
o
r
e
n
d
class
if
icatio
n
o
n
lo
ca
l
d
ataset.
On
th
e
o
th
er
s
id
e,
m
a
n
y
cu
ttin
g
-
ed
g
e
tech
n
i
q
u
es
h
a
v
e
b
ee
n
im
p
lem
en
ted
to
class
if
y
t
h
e
r
et
in
al
OC
T
im
ag
es.
Hu
an
g
et
a
l.
[
1
4
]
im
p
lem
en
ted
a
lay
er
-
g
u
id
e
d
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
(
C
NN)
.
T
h
ey
p
r
o
p
o
s
ed
d
if
f
er
en
t
n
etwo
r
k
s
to
ex
tr
ac
t
t
h
e
in
f
o
r
m
atio
n
f
r
o
m
th
e
r
etin
al
lay
er
s
an
d
th
en
it
is
p
r
o
v
id
ed
to
th
e
f
in
al
d
ee
p
lear
n
in
g
n
etwo
r
k
to
class
if
y
th
e
im
ag
es.
B
y
co
n
ce
n
tr
atin
g
o
n
th
e
r
etin
al
lay
er
-
s
p
ec
if
ic
in
f
o
r
m
atio
n
with
th
e
h
elp
o
f
s
o
m
e
tr
a
n
s
f
er
lear
n
in
g
ap
p
r
o
ac
h
,
th
ey
ac
h
iev
e
d
b
e
tter
ac
cu
r
ac
y
o
n
OC
T
2
0
1
7
[
1
5
]
an
d
HUCM
[
1
4
]
r
etin
al
im
ag
e
d
atasets
with
f
o
u
r
class
es:
C
NV,
DM
E
,
DR
U
S
E
N
,
an
d
n
o
r
m
al.
An
o
th
er
a
p
p
r
o
ac
h
with
d
e
ep
l
ea
r
n
in
g
was
im
p
lem
e
n
ted
b
y
Kim
an
d
T
r
a
n
[
1
6
]
,
wh
ich
p
r
o
p
o
s
es
th
e
im
p
lem
en
tatio
n
o
f
two
b
in
ar
y
m
o
d
els.
B
ef
o
r
e
a
p
p
ly
i
n
g
th
e
ac
tu
al
class
if
icatio
n
,
th
ey
p
er
f
o
r
m
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
1
,
Ma
r
ch
20
26
:
4
1
4
-
4
2
7
416
p
r
ep
r
o
ce
s
s
in
g
,
ar
ea
o
f
in
ter
est
s
eg
m
en
tatio
n
u
s
in
g
U
-
n
et
,
an
d
r
o
tatio
n
v
ia
h
is
to
g
r
am
o
r
ien
t
atio
n
to
ex
tr
ac
t
th
e
d
ee
p
in
f
o
r
m
atio
n
f
r
o
m
t
h
e
im
ag
es
o
f
OC
T
2
0
1
7
[
1
5
]
d
ataset.
T
h
ey
ac
h
iev
e
d
b
etter
ac
cu
r
a
cy
co
m
p
ar
ed
to
th
e
ap
p
r
o
ac
h
d
er
i
v
ed
[
1
4
]
.
Diao
et
a
l.
[
1
7
]
p
r
o
p
o
s
ed
d
e
ep
lear
n
in
g
m
o
d
els
to
class
if
y
AM
D
f
r
o
m
r
etin
al
OC
T
i
m
ag
es.
T
h
is
ap
p
r
o
ac
h
en
co
m
p
ass
es
two
n
o
v
el
m
o
d
els:
C
M
-
C
NN
an
d
C
AM
-
Un
et.
C
M
-
C
NN
im
p
r
o
v
es
th
e
class
if
icatio
n
p
r
o
ce
s
s
b
y
p
er
f
o
r
m
in
g
s
eg
m
e
n
tatio
n
v
ia
s
o
m
e
m
ea
ns
,
an
d
C
AM
-
UNe
t
en
h
an
ce
s
th
e
s
eg
m
en
tatio
n
task
b
y
in
teg
r
atin
g
class
ac
tiv
atio
n
m
a
p
s
.
T
h
e
ap
p
r
o
ac
h
ac
h
iev
ed
g
o
o
d
ac
cu
r
ac
y
f
o
r
th
e
tar
g
eted
w
o
r
k
.
Hass
an
et
a
l.
[
1
8
]
im
p
lem
e
n
ted
a
b
len
d
e
d
a
p
p
r
o
ac
h
o
f
d
ee
p
lear
n
in
g
an
d
class
ical
m
ac
h
in
e
lear
n
in
g
to
p
er
f
o
r
m
au
to
m
ate
d
class
if
icatio
n
o
f
r
etin
al
OC
T
im
ag
es
u
s
in
g
OC
T
2
0
1
7
[
1
5
]
d
ata
s
et.
T
h
e
en
h
an
ce
d
o
p
tical
co
h
er
e
n
ce
to
m
o
g
r
a
p
h
y
(
E
OC
T
)
m
o
d
el
p
r
esen
ted
in
t
h
is
s
tu
d
y
d
em
o
n
s
tr
ates
a
s
ig
n
i
f
ican
t
ad
v
a
n
ce
m
en
t
b
y
co
m
b
i
n
in
g
d
ee
p
lear
n
in
g
(
R
esNet
-
5
0
[
7
]
)
with
m
ac
h
in
e
lear
n
in
g
(
R
an
d
o
m
Fo
r
est)
an
d
o
p
tim
izin
g
with
d
u
al
SGD
an
d
Ad
am
o
p
tim
ize
r
s
.
T
h
is
m
o
d
el
ac
h
ie
v
es st
ate
-
of
-
th
e
-
a
r
t a
cc
u
r
ac
y
.
Pau
l
et
a
l.
[
1
9
]
,
th
e
r
esear
c
h
er
s
in
tr
o
d
u
ce
d
a
m
o
d
el
ca
lled
OC
T
x
to
class
if
y
r
etin
al
OC
T
im
ag
es
in
to
f
o
u
r
ca
teg
o
r
ies:
d
iab
etic
m
ac
u
lar
ed
e
m
a
(D
ME
)
,
c
h
o
r
o
id
al
n
eo
v
ascu
lar
izatio
n
(
C
NV)
,
Dr
u
s
en
,
an
d
N
o
r
m
al
R
etin
a.
T
h
eir
ap
p
r
o
ac
h
was
an
en
h
an
ce
d
en
s
em
b
le
m
o
d
el
wh
ich
is
co
m
b
in
in
g
m
u
l
tip
le
d
ee
p
lear
n
in
g
tech
n
iq
u
es
to
im
p
r
o
v
e
class
if
icatio
n
ac
cu
r
ac
y
.
H
o
wev
er
,
s
in
ce
th
e
m
o
d
el
was
d
ev
elo
p
ed
an
d
ev
alu
ated
in
a
s
p
ec
if
ic
o
r
lim
ited
d
ataset,
it
s
p
er
f
o
r
m
an
ce
m
ay
n
o
t
b
e
ef
f
ec
tiv
ely
ca
r
r
ied
o
v
er
to
r
ea
l
-
w
o
r
ld
clin
ical
s
ettin
g
s
.
T
h
is
lim
itatio
n
in
g
en
er
aliza
ti
o
n
co
u
l
d
af
f
ec
t its
d
ep
e
n
d
ab
ili
ty
wh
en
u
s
ed
o
n
d
iv
er
s
e
d
atasets
.
Yan
g
et
a
l.
[
2
0
]
,
r
esear
ch
er
s
b
u
ilt
a
C
NN
-
b
ased
m
o
d
el
to
class
if
y
ag
e
-
r
elate
d
m
ac
u
lar
d
e
g
en
er
atio
n
(
AM
D)
,
DM
E
,
an
d
No
r
m
al
R
etin
a
u
s
in
g
an
O
C
T
d
ataset.
B
y
in
co
r
p
o
r
atin
g
p
r
e
-
tr
ain
ed
I
m
ag
eNe
t
weig
h
ts
,
th
ey
s
ig
n
if
ican
tly
im
p
r
o
v
ed
th
e
m
o
d
el’
s
ac
cu
r
ac
y
s
ig
n
if
i
ca
n
tly
f
r
o
m
6
8
.
1
7
%
to
9
2
.
8
9
%.
T
h
eir
m
eth
o
d
u
tili
ze
d
an
en
s
em
b
le
o
f
th
r
ee
d
is
tin
ct
C
N
N
m
o
d
els,
ea
ch
e
n
h
an
ce
d
with
p
r
e
-
tr
ai
n
ed
weig
h
ts
an
d
f
in
e
-
tu
n
e
d
p
ar
am
eter
s
,
en
s
u
r
in
g
m
o
r
e
p
r
e
cise c
lass
if
icatio
n
o
f
r
etin
al
im
ag
es in
to
th
eir
r
esp
ec
tiv
e
ca
te
g
o
r
ies.
Stan
o
jev
i
et
a
l.
[
2
1
]
,
th
e
r
esear
ch
er
s
p
r
o
p
o
s
ed
a
d
ee
p
lear
n
in
g
-
b
ased
class
if
icatio
n
o
f
r
etin
al
d
is
ea
s
es
u
s
in
g
OC
T
im
ag
es.
I
t
ev
alu
at
es
C
NN
ar
ch
itectu
r
es
wh
ich
in
clu
d
es
Alex
Net,
VGG,
I
n
ce
p
tio
n
,
an
d
R
esNet.
T
h
e
I
n
ce
p
tio
n
1
m
o
d
el
th
at
tr
a
in
ed
with
R
MSp
r
o
p
o
p
tim
izer
,
ac
h
iev
e
d
th
e
h
i
g
h
est
ac
cu
r
ac
y
o
f
9
5
.
5
3
%,
a
lo
n
g
with
an
F1
-
s
co
r
e
o
f
0
.
9
3
6
8
7
.
T
h
e
s
tu
d
y
s
h
o
ws
th
at
I
n
ce
p
tio
n
-
b
ased
m
o
d
els
o
u
tp
er
f
o
r
m
o
th
er
s
in
ac
cu
r
ately
class
if
y
in
g
th
e
im
ag
es.
T
h
e
c
o
m
p
r
eh
e
n
s
iv
e
co
m
p
a
r
is
io
n
o
f
th
e
m
ac
h
in
e
lear
n
in
g
b
ased
an
d
d
ee
p
lear
n
in
g
b
ased
ap
p
r
o
ac
h
es f
o
r
th
e
OC
T
im
ag
e
class
if
ic
atio
n
is
d
ep
icted
in
T
ab
le
1
.
T
ab
le
1
.
Su
m
m
a
r
y
o
f
r
etin
al
O
C
T
im
ag
e
class
if
icatio
n
tech
n
iq
u
es
A
u
t
h
o
r
(
s)
M
e
t
h
o
d
t
y
p
e
D
a
t
a
s
e
t
A
p
p
r
o
a
c
h
e
s
D
r
a
w
b
a
c
k
Li
u
e
t
a
l
.
[
11
]
C
l
a
s
si
c
a
l
M
L
Lo
c
a
l
S
V
M
w
i
t
h
g
l
o
b
a
l
+
l
o
c
a
l
f
e
a
t
u
r
e
s
S
e
n
s
i
t
i
v
e
t
o
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
a
n
d
a
l
i
g
n
me
n
t
e
r
r
o
r
s
A
l
sai
h
e
t
a
l
.
[1
3
]
C
l
a
s
si
c
a
l
M
L
Lo
c
a
l
H
O
G
,
LB
P
,
P
C
A
w
i
t
h
S
V
M
D
e
p
e
n
d
e
n
c
e
o
n
ma
n
u
a
l
f
e
a
t
u
r
e
e
n
g
i
n
e
e
r
i
n
g
H
u
a
n
g
e
t
a
l
.
[
14
]
D
e
e
p
Le
a
r
n
i
n
g
O
C
T2
0
1
7
,
H
U
C
M
La
y
e
r
-
g
u
i
d
e
d
C
N
N
w
i
t
h
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
R
e
q
u
i
r
e
s
p
r
e
-
se
g
m
e
n
t
a
t
i
o
n
o
f
r
e
t
i
n
a
l
l
a
y
e
r
s
K
i
m
a
n
d
Tr
a
n
[
16
]
D
e
e
p
Le
a
r
n
i
n
g
O
C
T2
0
1
7
U
-
N
e
t
c
o
m
b
i
n
e
d
w
i
t
h
i
ma
g
e
r
o
t
a
t
i
o
n
a
n
d
b
i
n
a
r
y
c
l
a
ss
i
f
i
e
r
s
H
i
g
h
c
o
m
p
u
t
a
t
i
o
n
a
l
c
o
s
t
,
c
o
m
p
l
e
x
p
r
e
-
p
r
o
c
e
ssi
n
g
D
i
a
o
e
t
a
l
.
[
17
]
D
e
e
p
Le
a
r
n
i
n
g
N
o
t
s
p
e
c
i
f
i
e
d
CM
-
C
N
N
a
n
d
C
A
M
-
U
n
e
t
a
r
c
h
i
t
e
c
t
u
r
e
s
S
e
p
a
r
a
t
e
m
o
d
e
l
s
f
o
r
s
e
g
m
e
n
t
a
t
i
o
n
a
n
d
c
l
a
ss
i
f
i
c
a
t
i
o
n
H
a
ssan
e
t
a
l
.
[
18
]
H
y
b
r
i
d
(
D
L
+
M
L)
O
C
T2
0
1
7
R
e
sN
e
t
-
5
0
f
e
a
t
u
r
e
s
+
R
a
n
d
o
m
F
o
r
e
st
(
EO
C
T
a
p
p
r
o
a
c
h
C
o
m
p
l
e
x
h
y
b
r
i
d
f
r
a
m
e
w
o
r
k
,
h
i
g
h
e
r
t
r
a
i
n
i
n
g
c
o
m
p
l
e
x
i
t
y
P
a
u
l
e
t
a
l
.
[1
9
]
En
se
mb
l
e
D
L
O
C
Tx
,
O
C
T2
0
1
7
(
l
i
mi
t
e
d
)
M
u
l
t
i
p
l
e
D
L
m
o
d
e
l
s;
g
o
o
d
b
u
t
l
i
m
i
t
e
d
g
e
n
e
r
a
l
i
z
a
t
i
o
n
Li
mi
t
e
d
g
e
n
e
r
a
l
i
z
a
-
t
i
o
n
d
u
e
t
o
smal
l
d
a
t
a
se
t
Y
a
n
g
e
t
a
l
.
[
20
]
C
N
N
En
se
mb
l
e
O
C
T2
0
1
7
A
c
c
u
r
a
c
y
b
o
o
st
i
n
g
t
h
r
o
u
g
h
e
n
s
e
mb
l
e
l
e
a
r
n
i
n
g
H
i
g
h
m
e
mo
r
y
a
n
d
c
o
m
p
u
t
a
t
i
o
n
a
l
re
-
q
u
i
r
e
m
e
n
t
s
S
t
a
n
o
j
e
v
i
ć
e
t
a
l
.
[
21
]
D
e
e
p
Le
a
r
n
i
n
g
O
C
T2
0
1
7
I
n
c
e
p
t
i
o
n
1
m
o
d
e
l
p
e
r
f
o
r
m
e
d
b
e
s
t
a
m
o
n
g
t
e
st
e
d
mo
d
e
l
s
C
o
m
p
u
t
a
t
i
o
n
a
l
l
y
i
n
t
e
n
si
v
e
,
l
e
ss
su
i
t
e
d
f
o
r
r
e
a
l
-
t
i
me
d
e
p
l
o
y
me
n
t
W
e
p
r
o
p
o
s
e
d
a
l
i
g
h
t
we
i
g
h
t
C
N
N
m
o
d
e
l
d
e
s
i
g
n
e
d
t
o
b
a
l
a
n
ce
d
i
a
g
n
o
s
t
i
c
a
c
c
u
r
ac
y
w
i
t
h
c
o
m
p
u
t
a
t
i
o
n
al
e
f
f
i
c
i
e
n
c
y
i
n
r
e
t
i
n
a
l
O
C
T
cl
a
s
s
i
f
i
c
at
i
o
n
.
U
n
l
i
k
e
t
r
a
d
i
t
i
o
n
a
l
m
e
t
h
o
d
s
t
h
a
t
r
e
l
y
o
n
m
a
n
u
a
l
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
,
o
u
r
e
n
d
-
to
-
e
n
d
t
r
a
i
n
a
b
l
e
m
o
d
e
l
l
e
a
r
n
s
i
m
p
o
r
t
a
n
t
p
a
t
t
e
r
n
s
d
i
r
e
c
tl
y
f
r
o
m
r
a
w
i
m
a
g
e
s
,
r
e
d
u
c
i
n
g
h
u
m
a
n
b
i
a
s
.
T
h
e
a
r
c
h
i
t
e
c
t
u
r
e
u
s
es
j
u
s
t
s
i
x
l
a
y
e
r
s
a
n
d
i
n
c
o
r
p
o
r
a
t
e
s
h
y
b
r
i
d
p
o
o
l
i
n
g
-
m
a
x
p
o
o
l
i
n
g
t
o
c
a
p
t
u
r
e
k
e
y
f
e
a
t
u
r
e
s
a
n
d
a
v
e
r
a
g
e
p
o
o
l
i
n
g
t
o
r
e
t
ai
n
s
p
ati
a
l
d
e
t
ai
ls
-
r
e
s
u
lt
i
n
g
i
n
a
c
o
m
p
a
c
t
y
e
t
p
o
we
r
f
u
l
m
o
d
e
l
.
T
o
im
p
r
o
v
e
t
r
a
i
n
i
n
g
,
w
e
a
p
p
l
i
e
d
e
x
p
o
n
e
n
t
i
a
l
l
e
a
r
n
i
n
g
r
at
e
d
e
c
a
y
,
w
h
i
c
h
a
l
l
o
ws
f
as
t
e
r
co
n
v
e
r
g
e
n
c
e
a
n
d
f
i
n
e
r
t
u
n
i
n
g
,
u
n
l
i
k
e
f
i
x
e
d
l
e
a
r
n
i
n
g
r
a
t
e
s
.
Da
t
a
a
u
g
m
e
n
t
a
ti
o
n
f
u
r
t
h
e
r
e
n
h
a
n
c
e
s
m
o
d
e
l
p
e
r
f
o
r
m
a
n
c
e
b
y
a
d
d
r
e
s
s
i
n
g
c
l
as
s
i
m
b
a
l
a
n
c
e
a
n
d
r
e
d
u
c
i
n
g
o
v
e
r
f
i
t
t
i
n
g
.
W
i
t
h
o
u
t
r
e
l
y
i
n
g
o
n
c
o
m
p
l
e
x
p
r
e
p
r
o
c
e
s
s
i
n
g
o
r
h
e
a
v
y
t
r
a
n
s
f
e
r
l
e
a
r
n
i
n
g
,
o
u
r
a
p
p
r
o
a
c
h
o
f
f
e
r
s
s
t
r
o
ng
g
e
n
e
r
a
l
i
z
a
ti
o
n
a
c
r
o
s
s
v
a
r
i
e
d
i
m
a
g
i
n
g
c
o
n
d
i
t
i
o
n
s
.
T
h
is
m
a
k
es
i
t
w
e
l
l
-
s
u
i
t
e
d
f
o
r
p
r
a
c
ti
c
a
l
cl
i
n
i
ca
l
u
s
e
,
es
p
e
c
ia
l
l
y
in
l
o
w
-
r
e
s
o
u
r
c
e
s
e
t
ti
n
g
s
o
r
m
o
b
i
le
d
i
a
g
n
o
s
t
i
c
t
o
o
l
s
,
w
h
e
r
e
s
p
e
e
d
,
e
f
f
i
c
i
e
n
c
y
,
a
n
d
a
c
c
u
r
a
c
y
a
r
e
es
s
e
n
t
i
a
l
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t d
ee
p
lea
r
n
in
g
a
p
p
r
o
a
ch
fo
r
r
etin
a
l O
C
T ima
g
e
cla
s
s
ifica
tio
n
:
A
C
N
N
w
ith
…
(
P
a
r
th
R
.
Da
ve
)
417
3.
P
RO
P
O
SE
D
WO
RK
T
h
e
p
r
o
p
o
s
ed
a
p
p
r
o
ac
h
e
x
p
l
o
r
es
d
ata
au
g
m
en
tatio
n
,
ex
p
o
n
en
tial
lear
n
in
g
r
ate
d
ec
ay
an
d
C
NN
tech
n
iq
u
es.
T
h
e
u
s
e
o
f
ex
p
o
n
e
n
tial
lear
n
in
g
r
ate
d
ec
ay
h
elp
s
in
m
ak
in
g
th
e
lear
n
in
g
f
aster
in
th
e
in
itial
s
tag
e
s
wh
ile
in
th
e
later
s
tag
es
h
elp
s
to
g
et
n
ea
r
m
in
im
a
b
y
tak
in
g
s
m
all
s
tep
s
o
r
ju
m
p
s
.
I
m
ag
e
ex
p
an
s
io
n
tech
n
iq
u
es
allo
w
a
s
in
g
le
im
a
g
e
to
b
e
r
ep
r
esen
ted
in
v
ar
io
u
s
f
o
r
m
s
b
y
m
o
d
if
y
in
g
attr
ib
u
tes
lik
e
o
r
ie
n
tatio
n
,
s
h
ea
r
r
an
g
e,
a
n
d
z
o
o
m
r
a
t
i
o
.
I
t
h
e
l
p
s
t
o
g
e
t
g
r
e
a
t
e
r
d
e
t
a
il
s
f
o
r
t
h
e
t
r
a
i
n
i
n
g
p
u
r
p
o
s
e
a
s
t
h
e
t
r
ai
n
i
n
g
d
a
t
as
e
t
g
e
t
s
e
v
o
l
v
e
d
b
y
i
t
.
T
h
e
au
g
m
en
tatio
n
m
ay
also
h
elp
to
id
en
tify
an
d
d
if
f
er
en
tiate
s
o
m
e
o
v
er
lap
p
in
g
f
ea
tu
r
es
o
f
th
e
im
ag
es
o
f
d
if
f
er
e
n
t
class
es;
p
o
ten
tially
in
cr
ea
s
in
g
th
e
ac
cu
r
a
cy
o
f
th
e
m
o
d
el.
T
h
e
e
x
p
an
d
e
d
d
ataset
is
f
ed
i
n
to
th
e
C
NN
f
o
r
tr
ain
in
g
,
allo
win
g
it
to
ca
p
tu
r
e
f
in
e
-
tu
n
ed
im
a
g
e
f
ea
tu
r
es
th
r
o
u
g
h
its
co
n
v
o
l
u
tio
n
al
lay
er
s
.
T
h
e
f
in
al
s
tag
e
o
f
t
h
e
m
o
d
el
i
n
clu
d
es
a
d
e
n
s
e
lay
er
co
n
tain
in
g
1
,
0
2
4
n
e
u
r
o
n
s
,
wh
ich
f
u
n
ctio
n
s
as
a
h
id
d
en
u
n
it,
wh
ich
is
u
ltima
tely
lin
k
ed
to
4
o
u
tp
u
t
n
e
u
r
o
n
s
co
r
r
esp
o
n
d
in
g
to
th
e
f
o
u
r
d
is
tin
ct
class
e
s
in
th
e
OC
T
2
0
1
7
[
1
5
]
d
ataset.
T
h
e
co
m
p
lete
wo
r
k
f
lo
w
o
f
th
e
p
r
o
p
o
s
ed
r
etin
al
OC
T
class
if
icatio
n
m
o
d
el
is
o
u
tlin
ed
in
Alg
o
r
ith
m
1
.
A
d
etailed
in
ter
p
r
etatio
n
o
f
e
ac
h
s
tep
o
f
Alg
o
r
ith
m
1
is
ex
p
lain
ed
in
t
h
e
s
u
b
s
eq
u
e
n
t
s
ec
tio
n
s
to
e
n
s
u
r
e
t
h
e
co
m
p
lete
an
d
clea
r
u
n
d
er
s
tan
d
in
g
o
f
o
u
r
m
eth
o
d
o
lo
g
y
.
Alg
o
r
ith
m
1
: Pr
o
p
o
s
ed
C
NN
m
o
d
el
f
o
r
r
etin
al
OC
T
class
if
icatio
n
Require:
OCT2017 dataset with four retinal classes
Ensure:
Classified disease label
Step 1: Data Preprocessing
−
Resize
images to 200 × 200, normalize pixel values
Step 2: Data Augmentation
−
Perform shear, zoom, and horizontal flip transformations
Step 3: CNN Model Training
−
Initialize convolutional layers with hybrid pooling
−
Use ReLU activation and dropout to prevent over
fitting
−
Optimize model using Adam with exponential learning rate decay
Step 4: Model Evaluation
−
Test on unseen images and compute accuracy, confusion matrix
−
Generate ROC curves for performance validation
Step 5: Deployment
−
Save lightweight model for
real
-
time clinical usage
3
.
1
.
Da
t
a
prepa
ra
t
i
o
n
T
h
e
OC
T
2
0
1
7
d
ataset
[
1
5
]
c
o
n
s
is
ts
o
f
4
d
if
f
er
en
t
class
es:
C
NV,
DM
E
,
Dr
u
s
en
an
d
No
r
m
al.
T
h
e
tr
ain
in
g
d
ataset
co
n
tain
s
8
3
,
4
8
4
im
ag
es
d
iv
id
ed
s
u
ch
t
h
at
C
NV
co
n
tain
s
3
7
,
2
0
5
im
a
g
es,
DM
E
h
as
1
1
,
3
4
8
,
Dr
u
s
en
h
as
8
,
6
1
6
an
d
f
in
ally
No
r
m
al
co
n
tain
s
2
6
,
3
1
5
im
ag
es.
T
h
e
im
ag
es
in
th
e
tr
ain
in
g
d
ataset
ar
e
h
av
in
g
v
ar
y
in
g
p
i
x
el
d
im
e
n
s
io
n
ality
a
n
d
also
co
n
tain
s
o
m
e
n
o
is
e
as we
ll.
T
h
e
DPI
(
d
o
ts
p
er
in
c
h
)
o
f
all
th
e
im
a
g
es
is
9
6
.
On
t
h
e
o
th
er
h
an
d
,
th
e
te
s
t
d
ataset
co
n
tain
s
a
to
tal
o
f
9
6
8
im
a
g
es,
2
4
2
im
ag
es
ea
ch
p
er
class
.
Fo
r
th
e
v
alid
atio
n
p
u
r
p
o
s
e,
a
to
tal
o
f
3
2
im
ag
es,
8
e
ac
h
p
er
class
i
s
g
iv
en
.
T
h
e
Fig
u
r
e
1
r
ep
r
esen
ts
th
e
s
am
p
le
im
ag
es
f
o
r
ea
ch
class
in
th
e
d
ataset.
Fig
u
r
e
1
.
Sam
p
le
OC
T
im
ag
es
f
r
o
m
OC
T
2
0
1
7
d
ataset
3
.
2
.
Da
t
a
a
ug
m
ent
a
t
io
n
Data
au
g
m
en
tatio
n
is
a
tec
h
n
iq
u
e
th
r
o
u
g
h
wh
ic
h
n
ew
im
a
g
es
ca
n
b
e
g
e
n
er
ated
b
y
ap
p
l
y
in
g
s
o
m
e
attr
ib
u
tes
to
th
e
tr
ain
i
n
g
d
ataset
im
ag
es.
T
h
e
o
v
er
lap
p
in
g
f
e
atu
r
es
o
f
d
if
f
er
en
t
class
es
m
u
s
t
b
e
lear
n
t
b
y
th
e
C
NN
m
o
d
el
an
d
th
at
ca
n
b
e
e
n
h
an
ce
d
u
p
to
s
o
m
e
lev
el
b
y
th
e
d
ata
a
u
g
m
e
n
tatio
n
m
ec
h
a
n
is
m
as
it
p
r
o
v
id
es
m
o
r
e
in
f
o
r
m
atio
n
in
d
if
f
er
e
n
t
way
s
.
T
h
er
ef
o
r
e,
t
h
e
d
ata
a
u
g
m
en
tatio
n
is
ap
p
lied
to
all
t
h
e
im
ag
es
in
to
t
h
e
tr
ain
in
g
d
ataset.
As
m
en
tio
n
ed
b
y
Dav
e
an
d
Pan
d
y
a
[
2
2
]
,
d
ata
au
g
m
en
tatio
n
ca
n
lead
to
im
p
r
o
v
ed
ac
cu
r
ac
y
o
f
th
e
m
o
d
el.
T
h
e
attr
ib
u
tes
wh
ich
ar
e
u
s
ed
f
o
r
d
ata
au
g
m
en
tatio
n
in
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
ar
e
d
escr
ib
ed
in
T
ab
le
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
1
,
Ma
r
ch
20
26
:
4
1
4
-
4
2
7
418
Me
d
ical
im
ag
es
o
f
ten
h
a
v
e
class
im
b
alan
ce
s
an
d
lim
ited
v
ar
iatio
n
in
o
r
ien
tatio
n
,
b
r
i
g
h
tn
ess
,
o
r
s
tr
u
ctu
r
e
-
f
ac
to
r
s
cr
itical
f
o
r
ac
cu
r
ate
class
if
icatio
n
.
T
ec
h
n
iq
u
es
lik
e
s
h
ea
r
in
g
,
zo
o
m
in
g
,
an
d
f
lip
p
in
g
in
tr
o
d
u
ce
s
y
n
t
h
etic
d
iv
er
s
ity
,
h
elp
in
g
th
e
m
o
d
el
r
ec
o
g
n
iz
e
k
ey
f
ea
tu
r
es,
esp
ec
ially
in
u
n
d
er
r
e
p
r
esen
ted
class
es
lik
e
Dr
u
s
en
.
T
h
is
r
ed
u
ce
s
o
v
er
f
itti
n
g
an
d
b
o
o
s
ts
th
e
m
o
d
el’
s
r
o
b
u
s
tn
ess
wh
en
ap
p
lied
to
r
ea
l
-
wo
r
ld
clin
ical
d
ata.
Au
g
m
en
tatio
n
en
s
u
r
es
th
e
m
o
d
el
lear
n
s
f
r
o
m
a
b
r
o
a
d
er
r
an
g
e
o
f
e
x
am
p
les,
im
p
r
o
v
in
g
its
r
eliab
ilit
y
an
d
d
ia
g
n
o
s
tic
p
er
f
o
r
m
an
ce
.
T
ab
le
2
.
Data
au
g
m
en
tatio
n
attr
ib
u
tes
A
t
t
r
i
b
u
t
e
n
a
me
A
t
t
r
i
b
u
t
e
v
a
l
u
e
r
e
sca
l
e
1
.
/
2
5
5
sh
e
a
r
r
a
n
g
e
0
.
3
z
o
o
m ra
n
g
e
0
.
3
h
o
r
i
z
o
n
t
a
l
f
l
i
p
Tr
u
e
Ta
r
g
e
t
S
i
z
e
(
p
x
)
2
0
0
x
2
0
0
3.
3
.
E
x
po
nentia
l
lea
rning
ra
t
e
deca
y
L
ea
r
n
in
g
r
ate
is
a
cr
u
cial
h
y
p
er
-
p
ar
am
eter
i
n
d
ee
p
lear
n
in
g
o
r
m
ac
h
in
e
lear
n
in
g
.
I
t
h
an
d
le
s
h
o
w
f
ast
o
r
g
r
ad
u
al
lear
n
in
g
h
ap
p
en
s
wh
ile
tr
ain
in
g
.
I
n
m
ath
em
atica
l
ter
m
s
it
d
er
iv
es
th
e
s
ize
o
f
th
e
ju
m
p
o
n
th
e
er
r
o
r
s
u
r
f
ac
e
to
b
e
tak
en
to
r
ea
c
h
u
p
to
th
e
m
in
im
a.
T
h
er
e
ca
n
b
e
a
tr
ad
e
o
f
f
if
we
p
u
t
th
e
lea
r
n
in
g
r
ate
s
o
h
ig
h
;
lead
in
g
to
f
ast
lear
n
in
g
in
t
h
e
in
itial
s
tag
es
wh
ile
g
r
ea
te
r
o
s
cillatio
n
in
later
o
n
s
tag
es
g
e
n
er
atin
g
c
h
an
ce
s
to
m
is
s
th
e
ac
tu
al
m
in
im
a.
On
t
h
e
o
th
er
h
an
d
,
if
we
p
u
t
th
e
lear
n
in
g
r
ate
v
er
y
s
m
all
th
en
v
er
y
g
r
ad
u
al
s
tep
s
wo
u
ld
b
e
ca
r
r
ied
o
u
t to
war
d
s
m
in
im
a
o
n
t
h
e
er
r
o
r
s
u
r
f
ac
e;
i
n
cr
ea
s
in
g
th
e
tim
e
o
f
lear
n
i
n
g
at
a
g
r
ea
ter
s
p
ac
e.
=
×
(
)
(
/
)
(
1
)
T
o
co
m
b
at
th
is
s
itu
atio
n
,
th
e
p
r
o
p
o
s
ed
wo
r
k
h
as
u
s
ed
ex
p
o
n
en
tial
lear
n
in
g
r
ate
d
ec
ay
wh
ich
u
s
es
lar
g
er
ju
m
p
s
in
th
e
i
n
itial
s
et
o
f
tr
ain
in
g
an
d
th
e
n
g
r
ad
u
ally
d
ec
r
ea
s
in
g
th
e
s
ize
o
f
th
e
ju
m
p
to
war
d
s
th
e
ac
t
u
al
m
in
im
a,
m
ak
in
g
v
er
y
s
tab
le
u
p
d
ates
in
later
s
t
ag
es
o
f
lear
n
in
g
.
T
h
e
m
ath
em
atica
l
in
ter
p
o
l
atio
n
o
f
th
e
s
am
e
is
g
iv
en
b
y
(
1
)
.
T
h
is
im
p
lem
en
ta
tio
n
u
ltima
tely
h
elp
s
in
f
aster
lear
n
in
g
in
in
itial
s
tag
es
an
d
av
o
id
s
o
v
er
s
h
o
o
tin
g
an
d
o
s
cillatio
n
s
ar
o
u
n
d
ac
tu
al
m
in
im
a
o
f
th
e
er
r
o
r
s
u
r
f
ac
e.
T
ab
le
3
lis
ts
th
e
p
ar
a
m
eter
s
u
s
ed
f
o
r
th
e
ex
p
o
n
e
n
tial
d
ec
r
ea
s
e
in
th
e
le
ar
n
in
g
r
ate.
B
ased
o
n
th
ese
v
a
lu
es
an
d
(
1
)
,
T
a
b
le
4
s
h
o
ws
h
o
w
th
e
lear
n
i
n
g
r
ate
ch
an
g
es st
ep
b
y
s
tep
d
u
r
in
g
tr
ain
in
g
.
T
ab
le
3
.
Par
am
eter
v
alu
es f
o
r
ex
p
o
n
e
n
tial le
ar
n
in
g
r
ate
d
ec
ay
P
a
r
a
me
t
e
r
n
a
m
e
P
a
r
a
me
t
e
r
v
a
l
u
e
i
n
i
t
i
a
l
l
e
a
r
n
i
n
g
r
a
t
e
0
.
0
0
1
d
e
c
a
y
st
e
p
s
1
0
0
0
0
d
e
c
a
y
r
a
t
e
0
.
9
T
ab
le
4
.
Up
d
ates in
lear
n
in
g
r
ate
u
s
in
g
p
ar
am
eter
s
o
f
T
ab
le
3
S
t
e
p
s
Eq
u
a
t
i
o
n
N
e
w
V
a
l
u
e
o
f
LR
1
0
0
0
0
LR
=
0
.
0
0
1
x
(
0
.
9
)
1
0
.
0
0
0
9
2
0
0
0
0
LR
=
0
.
0
0
1
x
(
0
.
9
)
2
0
.
0
0
0
8
1
3
0
0
0
0
LR
=
0
.
0
0
1
x
(
0
.
9
)
3
0
.
0
0
0
7
2
9
3.
4
.
Co
nv
o
lutio
na
l
neura
l net
wo
rk
s
t
ruct
ure
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
C
NN
m
o
d
el
u
s
ed
in
o
u
r
ap
p
r
o
ac
h
is
d
escr
ib
ed
in
Fig
u
r
e
2
.
T
h
e
m
o
d
el
co
m
p
r
is
es
th
e
co
m
b
in
atio
n
o
f
m
ax
an
d
av
er
ag
e
p
o
o
lin
g
b
y
wh
ic
h
r
o
b
u
s
t
f
ea
tu
r
es
o
f
th
e
im
ag
es
ca
n
b
e
ex
tr
ac
ted
.
T
h
e
f
ir
s
t
co
n
v
o
lu
tio
n
al
lay
er
ap
p
li
es
3
2
f
ilter
s
o
f
3
×3
s
ize
o
n
th
e
o
r
ig
in
al
in
p
u
t
im
a
g
e
o
f
s
ize
2
0
0
×2
0
0
p
ix
els
with
a
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
[
2
3
]
as
a
n
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
e
f
ir
s
t
lay
er
u
s
es
th
e
m
ax
p
o
o
lin
g
o
f
s
ize
2
×
2
to
g
r
ab
th
e
m
a
x
im
u
m
p
ix
el
v
a
lu
e
an
d
to
r
e
d
u
ce
th
e
f
ea
t
u
r
e
m
ap
s
ize.
T
h
e
s
ec
o
n
d
lay
e
r
co
n
s
is
ts
o
f
1
6
f
ilter
s
o
f
s
ize
3
×3
a
n
d
a
v
er
ag
e
p
o
o
lin
g
o
f
s
ize
2
×
2
.
Av
er
ag
e
p
o
o
lin
g
g
ath
er
s
th
e
av
er
a
g
e
p
ix
el
v
al
u
e
o
f
its
s
u
r
r
o
u
n
d
in
g
to
n
o
t
m
is
s
d
etailed
in
f
o
r
m
ati
o
n
.
T
h
e
ess
en
ce
o
f
m
ax
im
u
m
an
d
av
er
ag
e
p
o
o
lin
g
is
r
ep
r
e
s
en
ted
in
Fig
u
r
e
2
.
T
h
e
s
ec
o
n
d
,
th
ir
d
,
f
o
u
r
th
,
a
n
d
f
if
th
lay
er
s
co
n
tain
3
2
,
6
4
,
6
4
,
an
d
1
2
8
k
er
n
els,
r
esp
ec
tiv
e
ly
,
ea
ch
with
a
s
ize
o
f
3
×3
.
T
h
ese
ar
e
f
o
llo
wed
b
y
p
o
o
li
n
g
o
p
er
atio
n
s
in
t
h
e
o
r
d
er
:
av
er
a
g
e
(
2
×2
)
,
m
ax
(
2
×
2
)
,
m
ax
(
2
×2
)
,
an
d
av
er
ag
e
(
2
×2
)
.
(
)
=
(
)
1
≤
≤
,
1
≤
≤
(
2
)
(
)
=
1
2
∑
∑
=
1
=
1
(
3
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t d
ee
p
lea
r
n
in
g
a
p
p
r
o
a
ch
fo
r
r
etin
a
l O
C
T ima
g
e
cla
s
s
ifica
tio
n
:
A
C
N
N
w
ith
…
(
P
a
r
th
R
.
Da
ve
)
419
T
h
e
m
ath
em
atica
l
in
ter
p
o
latio
n
f
o
r
th
e
m
a
x
im
u
m
an
d
av
er
a
g
e
p
o
o
lin
g
is
d
escr
ib
ed
b
y
(
2
)
an
d
(
3
)
,
r
esp
ec
tiv
ely
,
wh
er
e
x
ij
r
ep
r
ese
n
ts
th
e
v
alu
es
o
f
th
e
in
p
u
t
f
ea
tu
r
e
m
ap
with
in
th
e
p
o
o
lin
g
win
d
o
w
an
d
K
is
th
e
s
ize
o
f
th
e
p
o
o
lin
g
win
d
o
w.
H
y
b
r
id
p
o
o
lin
g
b
alan
ce
s
s
p
ec
if
i
city
(
b
y
m
a
x
p
o
o
lin
g
)
an
d
r
o
b
u
s
tn
ess
(
b
y
av
er
ag
e
p
o
o
lin
g
)
,
en
s
u
r
i
n
g
th
at
th
e
m
o
d
el
ca
p
tu
r
es
b
o
th
d
is
cr
im
in
ativ
e
f
ea
tu
r
es
an
d
s
tr
u
ctu
r
a
l
co
h
er
en
ce
.
I
n
th
e
p
r
o
p
o
s
ed
C
NN
m
o
d
el,
a
co
m
b
in
atio
n
o
f
h
y
b
r
i
d
p
o
o
lin
g
is
u
tili
ze
d
to
en
h
an
ce
p
er
f
o
r
m
a
n
ce
.
Ma
th
em
ati
ca
lly
,
h
y
b
r
id
p
o
o
lin
g
co
m
b
i
n
es
m
ax
-
p
o
o
lin
g
an
d
av
e
r
ag
e
-
p
o
o
lin
g
o
u
tp
u
ts
as
(
4
)
,
wh
er
e
α
∈
[
0
,
1
]
co
n
tr
o
ls
th
e
tr
ad
e
-
o
f
f
b
etwe
en
f
ea
tu
r
e
s
h
ar
p
n
ess
an
d
s
m
o
o
th
in
g
.
T
h
is
m
ec
h
an
is
m
p
r
eser
v
es
s
alien
t
f
ea
tu
r
es
wh
ile
m
ain
tain
in
g
g
lo
b
al
co
n
tex
t,
o
f
f
er
in
g
a
m
o
r
e
b
alan
ce
d
f
ea
t
u
r
e
r
e
p
r
esen
tatio
n
th
an
in
d
iv
id
u
al
p
o
o
lin
g
tec
h
n
iq
u
es.
(
)
=
×
(
)
+
(
1
−
)
×
(
)
(
4
)
T
h
e
C
NN
s
tr
u
ctu
r
e
u
tili
ze
s
t
h
e
R
eL
U
[
2
3
]
ac
tiv
atio
n
f
u
n
c
tio
n
ac
r
o
s
s
all
f
ea
tu
r
e
ex
t
r
ac
tio
n
lay
er
s
.
R
eL
U
[
2
3
]
h
as
th
e
p
r
o
p
e
r
ties
o
f
n
o
t
b
ein
g
s
atu
r
ated
o
v
e
r
th
e
in
p
u
t
d
ata
p
o
in
ts
an
d
is
ac
tiv
ated
o
n
a
p
o
s
itiv
e
s
et
o
f
in
p
u
ts
o
r
n
eu
r
o
n
s
.
I
n
ad
d
itio
n
to
t
h
ese
p
r
o
p
er
ties
,
R
eL
U
[
2
3
]
also
h
elp
s
s
o
lv
e
th
e
p
r
o
b
lem
o
f
v
a
n
is
h
in
g
g
r
ad
ien
ts
at
lar
g
e.
T
h
e
f
u
n
ctio
n
c
an
b
e
d
escr
ib
ed
b
y
(
5
)
,
w
h
er
e
x
d
en
o
tes
in
p
u
t
d
ata
t
o
th
e
f
u
n
ctio
n
.
R
esear
ch
an
d
r
esu
lts
h
av
e
s
h
o
w
n
th
at
R
eL
U
g
en
er
ates f
aster
o
u
t
p
u
t in
lar
g
e
an
d
co
m
p
le
x
n
etwo
r
k
s
.
Fig
u
r
e
2
.
I
ll
u
s
tr
atio
n
o
f
2
x
2
m
ax
an
d
a
v
er
ag
e
p
o
o
lin
g
T
h
e
s
ec
o
n
d
,
th
ir
d
,
f
o
u
r
t
h
,
an
d
f
if
th
lay
er
s
co
n
tain
3
2
,
6
4
,
6
4
,
an
d
1
2
8
k
e
r
n
els,
r
esp
ec
t
iv
ely
,
ea
c
h
with
a
s
ize
o
f
3
x
3
.
T
h
ese
ar
e
f
o
llo
wed
b
y
p
o
o
lin
g
o
p
e
r
atio
n
s
in
th
e
o
r
d
e
r
:
av
er
a
g
e
(
2
x
2
)
,
m
a
x
(
2
x
2
)
,
m
ax
(
2
x
2
)
,
an
d
av
er
a
g
e
(
2
x
2
)
.
T
h
e
C
NN
s
tr
u
ctu
r
e
u
tili
ze
s
th
e
R
eL
U
[
2
2
]
ac
t
iv
atio
n
f
u
n
ctio
n
ac
r
o
s
s
all
f
ea
tu
r
e
ex
tr
ac
tio
n
lay
er
s
.
R
elu
[
2
2
]
h
as
th
e
p
r
o
p
er
ties
o
f
n
o
t
b
ei
n
g
s
atu
r
ated
o
v
e
r
th
e
in
p
u
t
d
ata
p
o
in
ts
an
d
is
ac
tiv
ated
o
n
a
p
o
s
itiv
e
s
et
o
f
in
p
u
ts
o
r
n
eu
r
o
n
s
.
I
n
a
d
d
itio
n
t
o
th
ese
p
r
o
p
er
ties
,
R
elu
[
2
2
]
a
ls
o
h
elp
s
s
o
lv
e
th
e
p
r
o
b
lem
o
f
v
a
n
is
h
in
g
g
r
ad
ien
t
s
at
lar
g
e.
T
h
e
f
u
n
ctio
n
ca
n
b
e
d
escr
ib
ed
b
y
(
2
)
,
w
h
er
e
x
d
en
o
tes
in
p
u
t
d
ata
to
th
e
f
u
n
ctio
n
.
R
esear
ch
a
n
d
r
esu
lts
h
av
e
s
h
o
wn
th
at
R
elu
g
en
er
ates
f
aster
o
u
tp
u
t
in
l
ar
g
e
an
d
co
m
p
lex
n
etwo
r
k
s
.
(
)
=
(
,
0
)
(
5
)
T
h
e
last
p
o
o
lin
g
lay
er
in
t
h
e
s
tr
u
ctu
r
e
is
an
av
er
ag
e
p
o
o
l
in
g
o
f
2
×
2
o
n
1
2
8
f
ilter
s
o
f
th
e
3
×3
co
n
v
o
l
u
tio
n
al
lay
er
.
He
n
ce
,
p
er
f
o
r
m
in
g
th
e
f
latten
in
g
o
f
th
e
f
ea
tu
r
es
at
th
is
s
tag
e
wo
u
ld
h
av
e
1
×1
×1
2
8
f
ea
tu
r
es
wh
ich
ar
e
c
o
n
n
ec
te
d
to
th
e
1
0
2
4
d
en
s
e
n
eu
r
o
n
s
.
T
h
is
d
en
s
e
u
n
it
ac
ts
as
a
h
id
d
en
u
n
it
p
lace
d
ju
s
t
b
ef
o
r
e
th
e
f
i
n
al
o
u
tp
u
t
s
tag
e.
T
o
h
elp
p
r
ev
en
t
o
v
e
r
f
itti
n
g
an
d
im
p
r
o
v
e
th
e
m
o
d
el’
s
a
b
ilit
y
to
g
e
n
er
alize
,
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
3
is
ap
p
lied
.
Fin
ally
,
th
e
o
u
tp
u
t
lay
er
in
cl
u
d
es
4
n
eu
r
o
n
s
,
ea
ch
co
r
r
esp
o
n
d
in
g
to
o
n
e
o
f
th
e
f
o
u
r
class
es
in
th
e
d
ataset.
T
o
class
if
y
th
e
im
ag
es
in
to
p
r
o
p
e
r
class
es,
th
e
last
lay
er
u
s
es
th
e
s
o
f
tm
ax
f
u
n
ctio
n
,
wh
ich
wo
r
k
s
o
n
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
.
C
ateg
o
r
ical
-
cr
o
s
s
-
en
tr
o
p
y
is
u
s
ed
with
Ad
am
o
p
tim
izer
an
d
ex
p
o
n
e
n
tial
lear
n
in
g
r
ate
d
ec
a
y
to
illu
s
tr
ate
m
u
lticlas
s
clas
s
i
f
icatio
n
.
T
h
e
p
r
o
p
o
s
ed
C
NN
ar
ch
itectu
r
e
ca
n
b
e
v
is
u
alize
d
b
y
Fig
u
r
e
3
.
(
)
=
(
)
∑
(
)
=
1
(
6
)
T
h
e
s
o
f
tm
ax
f
u
n
ctio
n
is
m
ath
em
atica
lly
r
ep
r
esen
ted
b
y
(
6
)
,
with
it
s
g
r
ap
h
i
ca
l
in
ter
p
r
etatio
n
s
h
o
wn
in
Fig
u
r
e
4
.
T
h
is
f
u
n
cti
o
n
s
ca
les
th
e
o
u
tp
u
t
v
alu
es
b
etwe
en
0
an
d
1
,
en
s
u
r
in
g
th
at
th
e
t
o
tal
s
u
m
o
f
all
class
p
r
o
b
a
b
ilit
ies
eq
u
als
1
.
I
n
(
6
)
,
Z
r
ep
r
esen
ts
th
e
in
p
u
t
v
ec
to
r
t
o
th
e
o
u
t
p
u
t
la
y
er
,
an
d
th
e
in
d
ex
j
r
ef
er
s
to
ea
c
h
o
u
tp
u
t
u
n
it,
wh
er
e
j =
1
,
2
,
.
.
.
,
K.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
1
,
Ma
r
ch
20
26
:
4
1
4
-
4
2
7
420
Fig
u
r
e
3
.
Pro
p
o
s
ed
C
NN
ar
ch
i
tectu
r
e
Fig
u
r
e
4
.
So
f
tm
a
x
o
u
tp
u
t a
s
cl
ass
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
4.
E
XP
E
R
I
M
E
N
T
A
L
SE
T
UP
AND
RE
SUL
T
S
T
h
e
im
p
lem
e
n
tatio
n
was
ca
r
r
ied
o
u
t
o
n
a
lo
ca
l
s
y
s
tem
eq
u
i
p
p
ed
with
an
AM
D
R
y
ze
n
7
5
7
0
0
G
6
4
-
b
it
p
r
o
ce
s
s
o
r
,
R
ad
eo
n
Gr
a
p
h
i
cs
at
3
.
8
0
GHz
,
an
d
1
6
GB
o
f
R
AM
.
T
h
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
was
d
ev
elo
p
ed
in
Py
th
o
n
,
u
tili
zin
g
lib
r
ar
ies
s
u
ch
as
Ker
as,
T
en
s
o
r
Flo
w,
s
cik
it
-
lear
n
,
an
d
Ma
tp
lo
tlib
.
T
h
e
ap
p
r
o
ac
h
u
s
es
ca
llb
ac
k
as
th
e
ea
r
ly
s
to
p
p
in
g
m
ec
h
an
is
m
f
o
r
th
e
C
NN
tr
ai
n
in
g
b
ased
o
n
v
al
id
atio
n
u
p
d
a
te
lo
s
s
v
alu
es
an
d
a
p
atien
ce
v
alu
e
k
ep
t
at
5
.
W
e
s
to
r
e
th
e
b
est
weig
h
ts
f
o
r
t
h
e
m
o
d
el
th
at
h
as
b
ee
n
tr
ain
e
d
o
v
er
1
1
ep
o
c
h
s
.
As
m
en
tio
n
ed
in
th
e
p
r
e
v
io
u
s
s
e
ctio
n
,
a
to
tal
o
f
8
3
,
4
8
4
in
p
u
t
in
s
tan
ce
s
wer
e
u
s
ed
f
o
r
tr
ai
n
in
g
,
3
2
in
s
tan
ce
s
s
er
v
ed
as
th
e
v
alid
atio
n
s
et,
an
d
9
6
8
in
s
tan
ce
s
wer
e
allo
ca
ted
f
o
r
test
in
g
.
4.
1
.
E
v
a
lua
t
i
o
n o
f
pro
po
s
ed
a
pp
ro
a
ch
T
h
e
p
r
o
p
o
s
ed
d
ee
p
n
eu
r
al
n
et
wo
r
k
ar
c
h
itectu
r
e
was
ca
r
ef
u
lly
tr
ain
ed
an
d
test
ed
u
s
in
g
t
h
e
OC
T
2
0
1
7
im
ag
e
d
ataset
to
en
s
u
r
e
a
r
eliab
le
an
d
co
m
p
r
eh
e
n
s
iv
e
ev
alu
a
tio
n
o
f
its
p
er
f
o
r
m
an
ce
.
T
r
ai
n
in
g
was
co
n
d
u
cted
o
v
er
1
1
e
p
o
ch
s
,
u
tili
zin
g
an
ea
r
ly
s
to
p
p
in
g
m
ec
h
an
is
m
to
p
r
ev
en
t
o
v
er
f
itti
n
g
an
d
o
p
ti
m
ize
g
en
er
aliza
tio
n
.
T
h
e
en
tire
tr
ain
in
g
p
r
o
ce
s
s
to
o
k
a
p
p
r
o
x
im
ately
1
7
0
m
in
u
te
s
,
wh
ile
test
in
g
r
e
q
u
ir
ed
ju
s
t
1
2
1
m
illi
s
ec
o
n
d
s
,
d
em
o
n
s
tr
atin
g
t
h
e
ef
f
icien
c
y
o
f
o
u
r
lig
h
tweig
h
t
a
p
p
r
o
ac
h
.
B
y
in
co
r
p
o
r
atin
g
d
ata
au
g
m
en
tatio
n
a
n
d
an
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t d
ee
p
lea
r
n
in
g
a
p
p
r
o
a
ch
fo
r
r
etin
a
l O
C
T ima
g
e
cla
s
s
ifica
tio
n
:
A
C
N
N
w
ith
…
(
P
a
r
th
R
.
Da
ve
)
421
ex
p
o
n
e
n
tial
lear
n
in
g
r
ate
d
e
ca
y
s
tr
ateg
y
,
th
e
m
o
d
el
s
u
c
ce
s
s
f
u
lly
lear
n
ed
i
n
tr
icate
r
e
tin
al
p
atter
n
s
wh
ile
m
ain
tain
in
g
s
tab
ilit
y
d
u
r
i
n
g
tr
ain
in
g
.
T
h
e
tr
ain
in
g
a
n
d
v
alid
atio
n
er
r
o
r
s
ch
an
g
ed
o
v
er
1
1
ep
o
ch
s
f
o
r
th
e
p
r
o
p
o
s
ed
C
NN
m
o
d
el
is
d
ep
icted
b
y
Fig
u
r
e
5
.
As th
e
f
ig
u
r
e
illu
s
tr
ates,
th
e
tr
ain
in
g
er
r
o
r
co
n
s
i
s
ten
tly
g
o
es d
o
wn
,
wh
ich
s
u
g
g
ests
th
at
th
e
m
o
d
el
is
lear
n
in
g
well
f
r
o
m
th
e
d
ata.
T
h
is
s
m
o
o
th
an
d
s
tead
y
tr
en
d
h
ig
h
lig
h
ts
th
e
e
f
f
ec
tiv
en
ess
o
f
u
s
in
g
ex
p
o
n
en
tial
lear
n
in
g
r
ate
d
ec
ay
an
d
d
ata
a
u
g
m
en
tatio
n
,
b
o
th
o
f
wh
ich
h
elp
ed
th
e
m
o
d
el
t
r
ain
ef
f
icien
t
ly
an
d
r
eliab
ly
.
T
h
e
clo
s
e
m
atch
b
etwe
en
th
e
tr
ain
in
g
an
d
v
alid
atio
n
cu
r
v
es
s
h
o
ws
th
at
th
e
m
o
d
el
p
er
f
o
r
m
s
well
ev
en
o
n
u
n
s
ee
n
d
ata,
d
em
o
n
s
tr
atin
g
its
o
v
er
all
r
o
b
u
s
tn
ess
.
T
h
e
f
o
llo
win
g
eq
u
atio
n
s
r
ep
r
esen
t
co
m
m
o
n
m
etr
ics
s
u
ch
as
p
r
ec
is
io
n
,
r
ec
all
an
d
F1
s
co
r
e,
u
s
ed
in
ev
alu
atio
n
o
f
class
if
ic
atio
n
m
o
d
els:
=
+
(
7
)
=
+
(
8
)
1
=
2
∗
∗
+
(
9)
Fig
u
r
e
5
.
Mo
d
el
er
r
o
r
p
r
o
g
r
ess
io
n
ac
r
o
s
s
ep
o
ch
s
f
o
r
b
o
th
tr
a
in
in
g
an
d
v
alid
atio
n
s
ets
T
h
e
co
n
f
u
s
io
n
m
atr
ix
s
h
o
w
n
in
Fig
u
r
e
6
f
u
r
th
er
illu
s
tr
ate
s
th
e
p
r
ec
is
io
n
o
f
th
e
m
o
d
el,
r
ev
ea
lin
g
m
in
im
al
m
is
class
if
icat
io
n
s
in
th
e
f
o
u
r
ca
teg
o
r
ies
o
f
th
e
r
etin
a:
C
NV,
DM
E
,
Dr
u
s
en
,
an
d
No
r
m
al.
T
h
e
p
r
o
f
icien
c
y
o
f
th
e
m
o
d
el
in
cl
ass
if
y
in
g
Dr
u
s
en
a
n
d
DM
E
,
t
wo
co
n
d
itio
n
s
th
at
o
f
ten
s
h
a
r
e
o
v
er
la
p
p
in
g
v
is
u
al
f
ea
tu
r
es,
is
p
ar
ticu
lar
ly
n
o
tab
l
e.
T
h
e
h
ig
h
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
s
co
r
e
v
alu
es
d
ep
icted
in
T
ab
le
5
r
ea
f
f
ir
m
th
e
r
eliab
ilit
y
o
f
th
e
m
o
d
el.
I
n
f
ac
t,
it
ac
h
iev
ed
an
F1
s
c
o
r
e
o
f
1
.
0
0
f
o
r
th
e
No
r
m
al
a
n
d
Dr
u
s
en
class
es,
in
d
icatin
g
f
lawless
class
if
icati
o
n
in
th
o
s
e
ca
s
es.
Desp
ite
class
im
b
alan
ce
,
esp
ec
ially
with
Dr
u
s
en
h
av
i
n
g
f
ewe
r
s
am
p
les,
au
g
m
en
tatio
n
tech
n
iq
u
es
wer
e
ap
p
lied
u
n
if
o
r
m
ly
to
all
class
es,
in
cr
ea
s
in
g
s
am
p
le
d
iv
er
s
ity
.
E
ac
h
s
am
p
le,
(
x
,
y
)
,
was
tr
an
s
f
o
r
m
ed
in
to
(
T
(
x
)
,
y
)
,
wh
er
e
T
r
ep
r
esen
t
s
r
an
d
o
m
tr
an
s
f
o
r
m
atio
n
s
lik
e
s
h
ea
r
in
g
an
d
f
lip
p
in
g
.
T
h
is
led
to
im
p
r
o
v
e
d
Dr
u
s
en
class
p
er
f
o
r
m
an
ce
with
an
F
1
-
s
co
r
e
o
f
0
.
9
9
an
d
a
r
ec
all
o
f
0
.
9
8
,
h
ig
h
lig
h
tin
g
th
e
au
g
m
en
tatio
n
’
s
r
o
le
in
ad
d
r
ess
in
g
im
b
alan
ce
an
d
e
n
h
an
cin
g
m
o
d
el
g
e
n
er
aliza
tio
n
.
T
h
ese
f
ig
u
r
es
an
d
tab
les
co
r
r
esp
o
n
d
to
th
e
n
u
m
er
ical
lab
els 0
,
1
,
2
,
an
d
3
,
r
ep
r
esen
tin
g
t
h
e
f
o
u
r
class
es: C
NV,
DM
E
,
Dr
u
s
en
,
an
d
No
r
m
al,
r
esp
ec
tiv
ely
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
1
,
Ma
r
ch
20
26
:
4
1
4
-
4
2
7
422
Fig
u
r
e
6
.
C
o
n
f
u
s
io
n
m
atr
i
x
ev
alu
atio
n
o
n
p
r
o
p
o
s
ed
m
o
d
el
f
o
r
th
e
test
d
ataset
T
ab
le
5
.
Per
f
o
r
m
an
ce
m
etr
ics o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
o
n
test
d
ataset
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
C
l
a
s
s
C
N
V
0
.
9
7
0
.
9
9
0
.
9
8
2
4
2
D
M
E
0
.
9
9
0
.
9
9
0
.
9
9
2
4
2
D
r
u
sen
1
.
0
0
0
.
9
8
0
.
9
9
2
4
2
N
o
r
mal
1
.
0
0
1
.
0
0
1
.
0
0
2
4
2
A
c
c
u
r
a
c
y
0
.
9
9
9
6
8
M
a
c
r
o
A
v
g
0
.
9
9
0
.
9
9
0
.
9
9
9
6
8
W
e
i
g
h
t
e
d
A
v
g
0
.
9
9
0
.
9
9
0
.
9
9
9
6
8
T
o
f
u
r
th
er
v
alid
ate
its
r
o
b
u
s
t
n
ess
,
th
e
r
ec
eiv
er
o
p
er
atin
g
c
h
ar
ac
ter
is
tic
(
R
O
C
)
cu
r
v
e
o
f
th
e
m
o
d
el
s
h
o
wn
in
Fig
u
r
e
7
was
p
lo
tte
d
,
r
ev
ea
lin
g
a
n
AUC
clo
s
e
to
1
,
wh
ich
in
d
icate
s
s
tr
o
n
g
p
r
e
d
ictiv
e
ca
p
ac
ity
an
d
co
n
f
id
en
ce
in
i
ts
class
if
icatio
n
s
.
C
o
m
p
ar
ed
to
p
r
ev
io
u
s
m
eth
o
d
s
,
in
clu
d
in
g
L
a
y
er
-
Gu
id
e
d
C
NN,
B
in
ar
y
C
NN,
an
d
E
OC
T
m
o
d
els,
o
u
r
ap
p
r
o
ac
h
co
n
s
is
ten
tly
o
u
tp
e
r
f
o
r
m
e
d
th
em
in
te
r
m
s
o
f
ac
c
u
r
ac
y
.
T
h
e
co
m
p
a
r
is
o
n
o
f
th
e
ex
is
tin
g
ap
p
r
o
ac
h
es with
o
u
r
p
r
o
p
o
s
ed
a
p
p
r
o
ac
h
is
s
h
o
w
n
in
T
a
b
le
6
.
T
h
e
co
m
p
a
r
ativ
e
o
v
er
v
iew
o
f
f
o
u
r
p
r
o
g
r
ess
iv
ely
r
ef
in
ed
C
NN
ar
ch
itectu
r
es
d
ev
elo
p
ed
f
o
r
r
etin
al
OC
T
im
ag
e
class
if
icatio
n
is
d
ep
icted
in
T
ab
le
7
.
T
h
e
in
itial
m
o
d
el
with
s
am
e
n
u
m
b
er
o
f
c
o
n
v
o
lu
ti
o
n
al
lay
er
s
d
ep
icted
b
y
Fig
u
r
e
4
(
M
o
d
el
A)
u
s
es
m
a
x
p
o
o
lin
g
af
ter
e
ac
h
co
n
v
o
lu
tio
n
al
lay
er
,
a
f
ix
ed
lear
n
in
g
r
ate
o
f
0
.
0
0
1
,
an
d
was
tr
ain
ed
f
o
r
1
1
ep
o
ch
s
.
I
t
g
ain
ed
a
n
ac
cu
r
ac
y
o
f
9
4
.
3
%,
b
u
t
s
tr
u
g
g
le
d
to
ac
cu
r
ately
class
if
y
m
in
o
r
ity
class
es
s
u
ch
as
Dr
u
s
en
.
I
n
Mo
d
el
B
,
th
e
ap
p
licatio
n
o
f
d
ata
a
u
g
m
en
tatio
n
s
ig
n
if
i
ca
n
tly
e
n
h
an
ce
d
th
e
m
o
d
el’
s
ab
ilit
y
to
g
en
er
alize
ac
r
o
s
s
v
ar
io
u
s
s
am
p
les
th
at
i
m
p
r
o
v
i
n
g
o
v
e
r
all
class
b
alan
ce
an
d
in
cr
ea
s
in
g
th
e
ac
cu
r
ac
y
to
9
6
.
1
2
%.
Mo
d
el
C
in
tr
o
d
u
ce
d
h
y
b
r
id
p
o
o
lin
g
as
th
e
p
r
o
p
o
s
ed
m
o
d
el
(
a
co
m
b
in
atio
n
o
f
Ma
x
an
d
Av
er
ag
e
p
o
o
lin
g
)
,
wh
ich
allo
wed
f
o
r
m
o
r
e
ef
f
ec
tiv
e
r
eten
ti
o
n
o
f
s
p
atial
an
d
tex
tu
r
al
d
eta
ils
,
th
er
eb
y
g
ain
in
g
th
e
ac
cu
r
ac
y
to
9
7
.
3
%.
B
u
ild
in
g
o
n
th
ese
im
p
r
o
v
em
e
n
ts
,
th
e
p
r
o
p
o
s
ed
M
o
d
el
D
b
r
o
u
g
h
t
t
o
g
eth
er
all
p
r
e
v
io
u
s
en
h
an
ce
m
e
n
ts
an
d
in
tr
o
d
u
ce
d
an
ex
p
o
n
e
n
tial
lear
n
in
g
r
ate
d
ec
ay
,
wh
ich
h
el
p
ed
t
h
e
m
o
d
el
lear
n
m
o
r
e
s
tead
ily
an
d
co
n
v
er
g
e
m
o
r
e
s
m
o
o
th
ly
d
u
r
in
g
tr
ain
in
g
an
d
f
in
ally
ac
h
iev
in
g
th
e
b
en
ch
m
a
r
k
ac
c
u
r
ac
y
o
f
9
8
.
7
5
%.
4.
2
.
E
x
pla
ina
bil
it
y
a
na
ly
s
is
us
ing
CA
M
T
o
g
et
a
clea
r
er
p
ictu
r
e
o
f
h
o
w
o
u
r
C
NN
“wo
r
k
s
”,
C
AM
[
9
]
is
im
p
o
s
ed
as
a
n
ex
p
lain
a
b
le
AI
to
o
l
(
XAI
)
to
p
ee
k
i
n
s
id
e
its
lay
er
s
.
C
AM
tu
r
n
s
ea
ch
co
n
v
o
lu
tio
n
al
lay
er
’
s
f
ea
tu
r
e
m
a
p
s
in
to
h
ea
tm
ap
s
,
s
h
o
win
g
ex
ac
tly
wh
ich
p
a
r
ts
o
f
an
OC
T
im
ag
e,
th
e
m
o
d
el
is
f
o
cu
s
in
g
o
n
at
th
at
d
ep
th
.
W
h
ile
co
m
p
ar
in
g
C
AM
s
o
u
tp
u
t
f
r
o
m
th
e
f
i
r
s
t,
m
id
d
le,
an
d
d
e
ep
e
s
t
lay
er
s
,
a
s
m
o
o
th
p
r
o
g
r
e
s
s
io
n
is
s
ee
n
:
ea
r
ly
o
n
th
e
n
etwo
r
k
s
p
o
ts
s
im
p
l
e
ed
g
es
an
d
tex
tu
r
es,
th
en
it
h
o
m
es
in
o
n
th
e
in
d
icativ
e
s
ig
n
s
o
f
d
is
ea
s
e
-
lik
e
f
lu
id
p
o
ck
ets
o
r
s
tr
u
ctu
r
al
b
r
ea
k
s
,
b
ef
o
r
e
m
ak
in
g
its
f
in
al
ca
ll.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t d
ee
p
lea
r
n
in
g
a
p
p
r
o
a
ch
fo
r
r
etin
a
l O
C
T ima
g
e
cla
s
s
ifica
tio
n
:
A
C
N
N
w
ith
…
(
P
a
r
th
R
.
Da
ve
)
423
Fig
u
r
e
7
.
R
OC
cu
r
v
e
f
o
r
th
e
p
r
o
p
o
s
ed
a
p
p
r
o
ac
h
T
ab
le
6
.
C
o
m
p
a
r
is
o
n
with
ex
i
s
tin
g
ap
p
r
o
ac
h
es
Y
e
a
r
R
e
f
e
r
e
n
c
e
s
M
e
t
h
o
d
D
a
t
a
s
e
t
A
c
c
2
0
1
9
D
i
a
z
e
t
a
l
.
[
24
]
C
N
N
O
C
T2
0
1
7
93
2
0
1
9
H
u
a
n
g
e
t
a
l
.
[1
4
]
La
y
e
r
G
u
i
d
e
d
C
N
N
O
C
T2
0
1
7
8
8
.
4
2
0
2
0
S
a
r
a
i
v
a
e
t
a
l
.
[
25
]
C
N
N
O
C
T2
0
1
7
9
3
.
3
2
0
2
1
K
i
m
a
n
d
Tr
a
n
[
16
]
B
i
n
a
r
y
C
N
N
M
o
d
e
l
1
/
M
o
d
e
l
2
(
H
e
a
v
y
mo
d
e
l
s)
O
C
T2
0
1
7
9
8
.
1
/
9
8
.
7
2
0
2
3
H
a
ssan
e
t
a
l
.
[
1
8
]
EO
C
T
M
o
d
e
l
O
C
T2
0
1
7
9
7
.
4
7
2
0
2
3
D
i
a
o
e
t
a
l
.
[
17
]
C
N
N
O
C
T2
0
1
7
9
6
.
9
3
2
0
2
3
O
p
o
k
u
e
t
a
l
.
[
26
]
C
a
p
su
l
e
n
e
t
w
o
r
k
w
i
t
h
c
o
n
t
r
a
st
l
i
mi
t
e
d
a
d
a
p
t
i
v
e
h
i
s
t
o
g
r
a
m
e
q
u
a
l
i
z
a
t
i
o
n
O
C
T2
0
1
7
9
7
.
7
2
0
2
4
S
t
a
n
o
j
e
v
i
´
c
e
t
a
l
.
[
21
]
D
e
e
p
C
N
N
O
C
T2
0
1
7
9
5
.
5
5
2
0
2
4
Y
a
n
g
e
t
a
l
.
[
20
]
En
se
mb
l
e
M
o
d
e
l
b
a
se
d
o
n
C
N
N
,
Ef
f
i
c
i
e
n
t
n
e
t
v
2
a
n
d
R
e
s
n
e
t
O
C
T2
0
1
7
9
7
.
8
9
2
0
2
5
Pr
o
p
o
sed
a
p
p
r
o
a
c
h
C
N
N
+ D
a
t
a
A
u
g
m
e
n
t
a
t
i
o
n
+
E
x
p
o
n
e
n
t
i
a
l
L
e
a
r
n
i
n
g
R
a
t
e
D
e
c
a
y
+
mi
x
t
u
r
e
o
f
ma
x
a
n
d
a
v
e
r
a
g
e
p
o
o
l
i
n
g
OC
T
2
0
1
7
9
8
.
7
5
Ta
b
le
7
.
Ab
latio
n
s
tu
d
y
o
f
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
C
AM
p
r
o
v
id
es
v
is
u
al
in
s
ig
h
t
s
in
to
th
e
r
eg
io
n
s
o
f
th
e
r
eti
n
al
OC
T
im
ag
es
th
at
m
o
s
t
s
ig
n
if
ican
tly
co
n
tr
ib
u
te
to
th
e
class
if
icatio
n
d
ec
is
io
n
s
m
a
d
e
b
y
th
e
C
NN
m
o
d
el.
T
h
is
v
is
u
aliza
tio
n
h
elp
s
en
s
u
r
e
th
at
th
e
m
o
d
el
is
f
o
cu
s
in
g
o
n
clin
ically
m
ea
n
in
g
f
u
l
r
eg
io
n
s
,
th
u
s
p
r
o
m
o
tin
g
tr
a
n
s
p
ar
en
cy
,
tr
u
s
t,
an
d
p
o
ten
tial
clin
ical
ap
p
licab
ilit
y
.
Ma
th
em
atica
lly
,
f
o
r
a
g
i
v
en
class
c,
th
e
C
AM
h
ea
tm
ap
M
c
(
x
,
y
)
is
d
ef
in
e
d
b
y
(
1
0
)
.
M
o
d
e
l
D
a
t
a
a
u
g
me
n
t
a
t
i
o
n
P
o
o
l
i
n
g
Le
a
r
n
i
n
g
r
a
t
e
A
c
c
O
b
serv
a
t
i
o
n
M
o
d
e
l
A
(
B
a
s
e
l
i
n
e
C
N
N
)
NO
M
a
x
P
o
o
l
i
n
g
F
i
x
e
d
LR
9
4
.
3
%
S
i
mp
l
e
C
N
N
w
i
t
h
st
a
n
d
a
r
d
p
o
o
l
i
n
g
a
n
d
f
i
x
e
d
LR
sh
o
w
s
l
i
m
i
t
e
d
l
e
a
r
n
i
n
g
,
e
s
p
e
c
i
a
l
l
y
o
n
mi
n
o
r
i
t
y
c
l
a
sses
l
i
k
e
D
r
u
s
e
n
M
o
d
e
l
B
Y
e
s
M
a
x
P
o
o
l
i
n
g
F
i
x
e
d
LR
9
6
.
1
2
%
A
u
g
m
e
n
t
a
t
i
o
n
i
mp
r
o
v
e
s
g
e
n
e
r
a
l
i
z
a
t
i
o
n
a
n
d
c
l
a
ss
b
a
l
a
n
c
e
,
e
s
p
e
c
i
a
l
l
y
f
o
r
u
n
d
e
r
r
e
p
r
e
s
e
n
t
e
d
c
l
a
sses.
M
o
d
e
l
C
Y
e
s
H
y
b
r
i
d
P
o
o
l
i
n
g
(
M
a
x
+
A
v
g
)
F
i
x
e
d
LR
9
7
.
3
%
A
d
d
i
t
i
o
n
o
f
h
y
b
r
i
d
p
o
o
l
i
n
g
b
o
o
st
s
p
e
r
f
o
r
m
a
n
c
e
b
y
b
e
t
t
e
r
p
r
e
ser
v
i
n
g
sp
a
t
i
a
l
a
n
d
e
d
g
e
d
e
t
a
i
l
s.
M
o
d
e
l
D
(
Pr
o
p
o
sed
M
o
d
e
l
)
Y
e
s
Hy
b
r
i
d
Po
o
l
i
n
g
(
M
a
x
+ A
v
g
)
E
x
p
o
n
e
n
t
i
a
l
LR
9
8
.
7
5
%
Fi
n
a
l
mo
d
e
l
;
e
x
p
o
n
e
n
t
i
a
l
L
R
d
e
c
a
y
st
a
b
i
l
i
z
e
s
l
e
a
r
n
i
n
g
,
l
e
a
d
i
n
g
t
o
o
p
t
i
m
a
l
c
o
n
v
e
r
g
e
n
c
e
a
n
d
b
e
st
a
c
c
u
r
a
c
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.