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
.
2
,
A
p
r
il
2
0
2
6
,
p
p
.
1
4
5
1
~
1
4
6
3
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
2
.
p
p
1
4
5
1
-
1
4
6
3
1451
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Exponentia
l long
sho
rt
-
ter
m memo
ry
wit
h L
ev
y
f
lig
h
t
o
ptimiza
tion for
l
ung
nodu
le clas
sif
ica
tion
K
a
lib
a
G
o
wt
ha
m
i,
K
a
m
a
la
ka
nn
a
n J
a
y
a
s
ee
la
n
S
c
h
o
o
l
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
E
n
g
i
n
e
e
r
i
n
g
a
n
d
I
n
f
o
r
ma
t
i
o
n
S
y
st
e
m,
V
e
l
l
o
r
e
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
,
V
e
l
l
o
r
e
,
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
J
u
l
18
,
2
0
2
5
R
ev
is
ed
J
an
9
,
2
0
2
6
Acc
ep
ted
J
an
25
,
2
0
2
6
Lu
n
g
c
a
n
c
e
r,
wh
ich
c
o
m
m
o
n
ly
a
p
p
e
a
rs
a
s
lu
n
g
n
o
d
u
les
is
a
d
e
a
d
ly
t
y
p
e
o
f
c
a
n
c
e
r
th
a
t
d
e
v
e
lo
p
s
i
n
a
l
u
n
g
.
E
a
rly
d
e
tec
ti
o
n
o
f
l
u
n
g
c
a
n
c
e
r
is
c
rit
ica
l
a
n
d
c
h
a
ll
e
n
g
i
n
g
tas
k
d
u
e
to
p
re
se
n
c
e
o
f
o
v
e
rlap
p
in
g
stru
c
t
u
re
s,
wh
ic
h
m
a
k
e
it
c
h
a
ll
e
n
g
i
n
g
t
o
d
iffere
n
ti
a
te
th
e
b
e
n
ig
n
a
n
d
m
a
li
g
n
a
n
t
re
g
io
n
s.
T
h
i
s
re
se
a
rc
h
p
ro
p
o
se
s
lo
n
g
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
(LS
T
M
)
wit
h
e
x
p
o
n
e
n
ti
a
l
li
n
e
a
r
u
n
i
t
(EL
U)
m
e
th
o
d
fo
r
th
e
c
las
sifica
ti
o
n
o
f
d
iffere
n
t
c
las
se
s
o
f
l
u
n
g
n
o
d
u
les
.
T
h
e
h
y
p
e
r
p
a
ra
m
e
ters
o
f
th
e
LS
TM
n
e
two
rk
a
re
o
p
ti
m
ize
d
u
sin
g
th
e
d
e
v
e
lo
p
e
d
d
y
n
a
m
ic
Lev
y
fli
g
h
t
–
Arc
h
ime
d
e
s
o
p
ti
m
iza
ti
o
n
a
lg
o
r
it
h
m
(D
LF
-
AO
A),
wh
ich
e
ffe
c
ti
v
e
ly
id
e
n
ti
fies
th
e
o
p
ti
m
a
l
p
a
ra
m
e
ters
fo
r
c
las
sific
a
ti
o
n
.
T
h
e
Re
sN
e
t
-
1
8
m
e
th
o
d
is
u
se
d
fo
r
th
e
e
x
trac
ti
o
n
o
f
h
ig
h
-
lev
e
l
f
e
a
tu
re
s
to
d
iffere
n
ti
a
te
v
a
rio
u
s
c
las
se
s
o
f
l
u
n
g
n
o
d
u
les
.
F
u
rth
e
rm
o
re
,
Ba
y
e
sia
n
a
c
ti
v
e
c
o
n
to
u
r
(BAC)
is
e
m
p
l
o
y
e
d
f
o
r
th
e
se
g
m
e
n
tatio
n
o
f
ima
g
e
s
a
s
c
o
n
tai
n
in
g
c
a
n
c
e
ro
u
s
a
n
d
n
o
n
-
c
a
n
c
e
ro
u
s
re
g
io
n
s
o
f
l
u
n
g
n
o
d
u
les
.
T
h
e
LS
TM
with
EL
U
m
e
th
o
d
a
c
h
iev
e
s
9
8
.
5
6
%
a
c
c
u
ra
c
y
,
9
7
.
5
4
%
se
n
siti
v
it
y
,
9
8
.
2
2
%
sp
e
c
ifi
c
it
y
,
96
.
9
3
%
p
re
c
isio
n
,
9
6
.
3
3
%
F
1
-
sc
o
re
,
a
n
d
1
.
4
4
e
rr
o
r
ra
te
in
IQ
-
OTH/NCCD
lu
n
g
c
a
n
c
e
r
d
a
tas
e
t.
K
ey
w
o
r
d
s
:
A
r
c
h
i
m
e
d
e
s
o
p
t
i
m
i
z
a
t
i
o
n
a
l
g
o
r
i
t
h
m
B
ay
esian
ac
tiv
e
co
n
to
u
r
Dy
n
am
ic
L
ev
y
f
lig
h
t
E
x
p
o
n
e
n
tial lin
ea
r
u
n
it
L
o
n
g
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
L
u
n
g
n
o
d
u
le
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
:
Kalib
a
Go
wth
am
i
Sch
o
o
l o
f
C
o
m
p
u
ter
Scien
ce
E
n
g
in
ee
r
in
g
an
d
I
n
f
o
r
m
atio
n
Sy
s
tem
,
Vello
r
e
I
n
s
titu
te
o
f
T
ec
h
n
o
lo
g
y
Vello
r
e,
I
n
d
ia
E
m
ail:
g
o
wth
am
i.k
2
0
2
2
@
v
its
tu
d
en
t.a
c.
in
1.
I
NT
RO
D
UCT
I
O
N
L
u
n
g
ca
n
ce
r
is
a
d
ea
d
lies
t
ty
p
e
o
f
ca
n
ce
r
wo
r
ld
wid
e
.
Ho
wev
er
,
ea
r
ly
d
etec
tio
n
o
f
lu
n
g
ca
n
ce
r
ca
n
s
ig
n
if
ican
tly
im
p
r
o
v
e
s
u
r
v
iv
a
l
r
ate
[
1
]
.
T
h
e
m
alig
n
an
t
(
ca
n
ce
r
o
u
s
)
an
d
b
en
ig
n
(
n
o
n
-
ca
n
ce
r
o
u
s
)
p
u
lm
o
n
ar
y
n
o
d
u
les ar
e
s
m
all
g
r
o
wth
o
f
ce
lls
in
th
e
lu
n
g
s
[
2
]
.
T
h
e
d
etec
tio
n
o
f
m
alig
n
a
n
t lu
n
g
n
o
d
u
le
s
in
th
e
in
itial p
h
ase
is
ess
en
tial
f
o
r
ef
f
ec
tiv
e
p
r
o
g
n
o
s
is
[
3
]
.
I
n
t
h
e
in
itial
p
h
ase,
ca
n
ce
r
o
u
s
lu
n
g
n
o
d
u
les
ca
n
ap
p
ea
r
s
im
ilar
to
non
-
ca
n
ce
r
o
u
s
n
o
d
u
les,
b
u
t
r
eq
u
ir
e
d
if
f
e
r
en
tial
d
iag
n
o
s
is
b
ased
o
n
s
u
b
tle
m
o
r
p
h
o
lo
g
ical
ch
an
g
es,
p
o
s
itio
n
s
an
d
clin
ical
b
i
o
m
ar
k
e
r
s
[
4
]
–
[
6
]
.
On
e
o
f
th
e
m
o
s
t
ch
alle
n
g
in
g
asp
ec
ts
is
to
ca
lcu
late
th
e
p
r
o
b
ab
ilit
y
o
f
m
alig
n
an
cy
in
ea
r
ly
-
s
tag
e
ca
n
ce
r
o
u
s
lu
n
g
n
o
d
u
les
[
7
]
.
Var
io
u
s
d
iag
n
o
s
tics
m
eth
o
d
s
,
in
clu
d
in
g
co
m
p
u
ted
to
m
o
g
r
a
p
h
y
(
C
T
)
s
ca
n
a
n
aly
s
is
an
d
p
o
s
itro
n
em
is
s
io
n
to
m
o
g
r
ap
h
y
(
PET
)
h
a
v
e
b
e
en
u
tili
ze
d
b
y
p
h
y
s
ician
s
f
o
r
th
e
in
itial
d
iag
n
o
s
is
o
f
m
alig
n
an
t
lu
n
g
n
o
d
u
les
[
8
]
.
Ad
d
itio
n
ally
,
m
a
n
y
in
v
asiv
e
tech
n
iq
u
es
s
u
ch
as
s
u
r
g
er
ies
o
r
b
io
p
s
ies
ar
e
u
tili
ze
d
b
y
h
ea
lth
ca
r
e
p
r
ac
titi
o
n
er
s
f
o
r
d
if
f
er
e
n
tiatin
g
b
etwe
en
m
a
lig
n
an
t
an
d
b
en
ig
n
lu
n
g
n
o
d
u
les
[
9
]
.
Giv
en
th
e
s
en
s
itiv
ity
o
f
f
r
ag
ile
o
r
g
an
s
lik
e
th
e
lu
n
g
s
,
th
ese
in
v
asiv
e
tech
n
iq
u
es
ca
r
r
y
h
ig
h
r
is
k
s
an
d
in
cr
ea
s
e
p
atien
t
an
x
iety
.
T
h
e
m
o
s
t
s
u
itab
le
tech
n
iq
u
e
f
o
r
in
v
esti
g
atin
g
lu
n
g
d
is
ea
s
es
is
th
r
o
u
g
h
C
T
im
ag
es
[
1
0
]
.
R
esear
ch
er
s
h
av
e
d
ev
elo
p
ed
m
an
y
s
eg
m
en
tatio
n
tech
n
iq
u
e
s
to
ass
is
t
r
ad
io
lo
g
is
ts
in
d
etec
tin
g
lu
n
g
ca
n
ce
r
o
u
s
tu
m
o
r
s
[
1
1
]
.
T
h
e
l
u
n
g
ca
n
ce
r
s
eg
m
en
tatio
n
tec
h
n
iq
u
es
ar
e
s
ep
ar
ated
in
to
t
wo
m
ain
ca
teg
o
r
ies:
class
ical
alg
o
r
ith
m
s
an
d
d
ee
p
lear
n
i
n
g
(
DL
)
alg
o
r
ith
m
s
[
1
2
]
.
C
lass
ical
alg
o
r
ith
m
s
p
r
im
ar
ily
f
o
cu
s
o
n
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
.
2
,
Ap
r
il
20
26
:
1
4
5
1
-
1
4
6
3
1452
in
ten
s
ity
-
b
ased
tech
n
iq
u
es
s
u
ch
as
r
eg
io
n
g
r
o
wth
,
ad
ap
tiv
e
th
r
esh
o
l
d
,
m
o
r
p
h
o
lo
g
ical
tech
n
iq
u
e
,
ac
tiv
e
-
co
n
to
u
r
m
eth
o
d
an
d
s
h
ap
e
an
aly
s
is
[
1
3
]
.
Ho
wev
er
,
t
h
ese
tech
n
iq
u
es
ar
e
n
o
t
ef
f
ec
ti
v
e
in
d
if
f
e
r
en
tiatin
g
th
e
tu
m
o
r
s
izes
an
d
ar
e
n
o
t
s
u
itab
le
f
o
r
th
e
s
eg
m
en
tatio
n
o
f
lu
n
g
tu
m
o
r
s
[
1
4
]
.
A
d
d
itio
n
all
y
,
wh
en
tu
m
o
r
s
a
r
e
in
teg
r
ated
with
v
a
r
io
u
s
o
r
g
an
s
,
th
e
p
er
f
o
r
m
an
ce
o
f
tu
m
o
r
s
eg
m
en
tatio
n
tech
n
iq
u
es
is
af
f
ec
ted
,
r
esu
ltin
g
in
a
lo
wer
lev
el
o
f
au
to
m
atio
n
[
1
5
]
.
He
n
ce
,
class
ical
m
eth
o
d
s
h
av
e
b
ee
n
r
e
p
lace
d
with
DL
alg
o
r
ith
m
s
f
o
r
co
u
n
ter
in
g
n
u
m
er
o
u
s
is
s
u
es
in
im
ag
e
r
ec
o
g
n
itio
n
[
1
6
]
.
T
h
e
DL
alg
o
r
ith
m
s
ef
f
ec
tiv
el
y
ex
tr
ac
t
s
ig
n
if
ican
t
f
ea
tu
r
es,
with
o
u
t th
e
n
ee
d
f
o
r
h
u
m
an
i
n
ter
v
en
tio
n
[
1
7
]
,
[
1
8
]
.
T
o
f
ig
h
i
et
a
l.
[
1
9
]
in
tr
o
d
u
ce
d
a
Mo
b
ileNetV2
-
s
tack
ed
g
ated
r
ec
u
r
r
e
n
t
u
n
it
(
SGR
U)
,
a
n
o
v
el
tr
an
s
f
er
lear
n
in
g
-
e
n
ab
led
p
r
ed
icto
r
to
lu
n
g
ca
n
ce
r
class
if
icatio
n
.
T
h
e
in
tr
o
d
u
ce
d
m
eth
o
d
was
em
p
lo
y
ed
f
o
r
th
e
au
to
m
atic
ex
tr
ac
tio
n
o
f
f
ea
tu
r
es
f
r
o
m
lu
n
g
C
T
im
ag
es.
Ho
wev
er
,
in
tr
o
d
u
ce
d
m
eth
o
d
d
i
d
n
o
t
r
em
o
v
e
n
o
is
e
f
r
o
m
th
e
im
ag
es
wh
ich
m
in
i
m
ized
its
clas
s
if
icatio
n
p
er
f
o
r
m
an
ce
.
Ma
et
a
l
.
[
2
0
]
s
u
g
g
ested
a
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
f
r
am
ewo
r
k
an
d
Go
o
g
L
eNe
t
with
a
d
ap
tiv
e
lay
er
s
(
Go
o
g
L
eNe
t
-
AL
)
to
lu
n
g
ca
n
ce
r
d
etec
tio
n
.
T
h
e
s
u
g
g
ested
m
eth
o
d
en
h
an
ce
d
t
h
e
o
v
e
r
all
ca
p
ab
ilit
y
o
f
th
e
m
o
d
el
to
ca
p
tu
r
e
th
e
r
elev
an
t
f
ea
tu
r
es
with
in
tr
icate
p
atter
n
s
.
Ho
we
v
er
,
th
e
s
u
g
g
ested
m
eth
o
d
f
ailed
to
a
d
d
r
ess
th
e
is
s
u
e
o
f
v
a
n
is
h
in
g
g
r
ad
ien
ts
in
th
e
tr
ain
in
g
p
h
ase,
th
er
e
b
y
m
in
im
izin
g
class
if
icatio
n
p
e
r
f
o
r
m
an
ce
.
Ku
m
ar
an
et
a
l.
[
2
1
]
i
n
teg
r
ated
DL
-
b
ase
d
alg
o
r
ith
m
with
p
r
e
-
tr
ain
ed
m
eth
o
d
s
lik
e
R
esNet
-
5
0
,
I
n
ce
p
tio
n
V3
,
a
n
d
VGG
-
1
6
f
o
r
en
h
an
ce
d
lu
n
g
ca
n
ce
r
d
iag
n
o
s
tic
ac
cu
r
ac
y
.
T
h
is
m
o
d
el
m
ax
im
ized
th
e
ca
p
a
b
ilit
y
to
d
is
ce
r
n
s
u
b
tle
p
atter
n
s
in
d
i
f
f
er
en
t lu
n
g
f
ea
tu
r
es,
b
u
t th
e
m
o
d
el
p
ar
am
ete
r
s
wer
e
n
o
t su
f
f
icien
tly
o
p
tim
ized
f
o
r
class
if
icatio
n
.
Sab
za
lian
et
a
l
.
[
2
2
]
im
p
lem
e
n
ted
a
b
id
ir
ec
tio
n
al
r
ec
u
r
r
e
n
t
n
eu
r
a
l
n
etwo
r
k
(
R
NN)
f
o
r
ac
cu
r
ate
d
iag
n
o
s
is
o
f
lu
n
g
ca
n
ce
r
s
.
An
en
h
an
ce
d
f
o
r
m
o
f
th
e
E
b
o
la
o
p
tim
izatio
n
s
ea
r
ch
alg
o
r
ith
m
was
em
p
lo
y
e
d
in
a
s
etu
p
to
m
in
im
ize
ex
ec
u
ti
o
n
c
o
s
ts
b
y
elim
in
atin
g
th
e
r
eq
u
ir
em
e
n
t
f
o
r
ex
h
a
u
s
tiv
e
s
ea
r
ch
tech
n
iq
u
es.
Ho
wev
er
,
th
e
im
p
lem
en
ted
m
eth
o
d
f
ai
led
to
s
eg
m
en
t
th
e
r
eg
io
n
s
o
f
lu
n
g
n
o
d
u
les
ef
f
ec
tiv
ely
,
m
in
im
izin
g
its
ca
p
ab
il
ity
to
d
if
f
er
e
n
tiate
ca
n
ce
r
o
u
s
an
d
n
o
n
-
ca
n
ce
r
o
u
s
r
eg
io
n
s
.
R
az
a
et
a
l.
[
2
3
]
p
r
esen
ted
a
n
o
v
el
tr
an
s
f
er
lear
n
in
g
-
b
ased
p
r
ed
icto
r
k
n
o
wn
as
L
u
n
g
-
E
f
f
Net
t
o
class
if
y
lu
n
g
ca
n
ce
r
s
.
T
h
e
p
r
esen
ted
m
eth
o
d
d
ev
elo
p
e
d
o
n
th
e
s
tr
u
ctu
r
e
o
f
E
f
f
icie
n
tNet
with
f
u
r
th
er
m
o
d
if
icatio
n
s
th
at
in
clu
d
ed
th
e
to
p
lay
er
s
f
o
r
class
if
icatio
n
.
Ho
wev
er
,
th
e
p
r
esen
ted
m
et
h
o
d
d
i
d
n
o
t
ex
tr
ac
t
h
ig
h
-
lev
el
f
ea
tu
r
es n
ec
ess
ar
y
t
o
id
en
tify
t
h
e
ca
n
ce
r
o
u
s
an
d
n
o
n
-
ca
n
ce
r
o
u
s
r
e
g
io
n
s
.
On
an
aly
s
is
o
f
th
e
ex
is
tin
g
alg
o
r
ith
m
s
,
t
h
e
f
o
llo
win
g
lim
itatio
n
s
h
av
e
b
ee
n
n
o
ted
:
f
ailu
r
e
to
r
em
o
v
e
n
o
is
e
f
r
o
m
im
a
g
es,
th
e
in
ab
il
ity
to
ad
d
r
ess
th
e
v
a
n
is
h
in
g
g
r
ad
ien
t
is
s
u
e
d
u
r
i
n
g
th
e
t
r
ain
in
g
p
h
ase,
lack
o
f
p
r
o
p
er
o
p
tim
izatio
n
o
f
m
o
d
el
p
ar
am
eter
s
,
in
ef
f
ec
tiv
e
s
eg
m
en
tatio
n
o
f
lu
n
g
n
o
d
u
le
s
,
an
d
in
s
u
f
f
icien
t
ex
tr
ac
tio
n
o
f
h
ig
h
-
lev
el
f
ea
t
u
r
es.
T
o
m
itig
ate
th
ese
lim
itat
io
n
s
,
th
is
ar
ticle
u
tili
ze
s
th
e
W
ien
er
f
ilter
an
d
co
n
tr
ast
lim
ited
ad
ap
tiv
e
h
is
to
g
r
am
eq
u
aliza
tio
n
(
C
L
AHE
)
a
s
p
r
e
-
p
r
o
ce
s
s
in
g
tech
n
iq
u
es
to
r
em
o
v
e
n
o
is
e
an
d
en
h
an
ce
im
ag
e
c
o
n
tr
ast.
T
h
e
n
,
s
eg
m
en
tatio
n
p
r
o
ce
s
s
is
p
er
f
o
r
m
e
d
u
s
in
g
B
ay
esian
ac
tiv
e
co
n
to
u
r
(
B
AC
)
tech
n
iq
u
es,
wh
ich
ef
f
ec
tiv
el
y
s
eg
m
en
t
ca
n
ce
r
o
u
s
f
r
o
m
n
o
n
-
ca
n
ce
r
o
u
s
r
eg
io
n
s
.
T
h
e
n
,
th
e
C
NN
-
b
ased
p
r
e
-
tr
ain
ed
m
o
d
el
is
em
p
lo
y
e
d
f
o
r
th
e
ex
tr
ac
tio
n
o
f
b
o
th
l
o
w
-
lev
el
an
d
h
ig
h
-
lev
el
attr
i
b
u
tes
to
ac
cu
r
ately
id
en
tify
a
n
d
d
if
f
er
e
n
tiate
v
ar
i
o
u
s
r
eg
i
o
n
s
with
in
th
e
im
a
g
e.
Fin
ally
,
class
if
icatio
n
is
ca
r
r
ied
o
u
t
u
s
in
g
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
with
th
e
ex
p
o
n
e
n
tial
lin
ea
r
u
n
it
(
E
L
U)
m
eth
o
d
,
wh
ic
h
ef
f
ec
t
iv
ely
class
if
ies
lu
n
g
n
o
d
u
le
ty
p
es.
I
n
ad
d
itio
n
,
th
e
h
y
p
er
p
a
r
am
eter
s
o
f
th
e
L
STM
ar
e
o
p
tim
ized
u
s
in
g
th
e
d
y
n
am
ic
L
ev
y
f
lig
h
t
(
DL
F)
with
Ar
ch
im
e
d
es
o
p
tim
izatio
n
alg
o
r
ith
m
(
AOA)
,
wh
ich
id
en
tifie
s
t
h
e
o
p
tim
al
p
ar
am
eter
s
f
o
r
th
e
class
if
icatio
n
p
r
o
ce
s
s
.
T
h
ese
DL
-
b
ased
alg
o
r
ith
m
s
en
h
an
ce
th
e
ac
cu
r
ac
y
o
f
d
is
ea
s
e
d
etec
tio
n
in
m
e
d
ical
f
ield
.
I
n
th
is
ar
ticle,
th
e
DL
F
m
eth
o
d
in
tr
o
d
u
ce
s
r
an
d
o
m
lar
g
e
ju
m
p
s
b
ased
o
n
L
ev
y
d
is
tr
ib
u
t
io
n
,
wh
ich
en
h
an
ce
s
th
e
alg
o
r
ith
m
'
s
ab
ili
ty
to
escap
e
s
u
b
o
p
tim
al
s
o
lu
ti
o
n
s
d
u
r
in
g
h
y
p
er
p
ar
am
eter
tu
n
in
g
.
Un
lik
e
o
th
e
r
o
p
tim
izatio
n
al
g
o
r
ith
m
s
,
s
u
ch
as
w
h
ale
o
p
tim
izatio
n
alg
o
r
it
h
m
(
W
OA
)
,
g
en
etic
alg
o
r
ith
m
(
GA
)
,
o
r
p
ar
ticle
s
war
m
o
p
tim
izatio
n
(
PSO
)
,
wh
ich
r
el
y
o
n
f
ix
ed
s
tep
s
izes,
th
e
DL
F
alg
o
r
ith
m
d
y
n
am
ically
a
d
ju
s
ts
h
y
p
er
p
ar
am
eter
s
b
y
b
alan
cin
g
g
lo
b
al
ex
p
lo
r
atio
n
an
d
f
in
e
-
tu
n
ed
lo
ca
l
s
ea
r
ch
.
T
h
e
AO
A
en
s
u
r
es
ef
f
ec
tiv
e
ex
p
lo
itatio
n
,
wh
ile
DL
F
ac
ce
l
er
ates
co
n
v
e
r
g
en
ce
b
y
im
p
r
o
v
in
g
s
ea
r
ch
d
iv
e
r
s
ity
.
Un
lik
e
e
x
is
tin
g
lu
n
g
n
o
d
u
le
class
if
icatio
n
m
o
d
els
wh
ich
ad
d
r
ess
s
eg
m
en
tatio
n
,
f
ea
tu
r
e
ex
tr
ac
tio
n
o
r
class
if
icatio
n
,
th
is
wo
r
k
in
tr
o
d
u
ce
d
m
o
d
el
wh
ic
h
in
te
g
r
ates
B
AC
-
b
ased
s
eg
m
en
tatio
n
,
d
ee
p
r
esid
u
al
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
s
eq
u
en
ce
-
awa
r
e
L
STM
class
if
ica
tio
n
o
p
tim
ized
b
y
DL
F
-
AOA
.
T
h
e
p
r
o
p
o
s
ed
DL
F
-
AOA
en
s
u
r
es
ad
ap
tiv
e
ex
p
lo
r
atio
n
an
d
ex
p
lo
itatio
n
in
h
y
p
er
p
a
r
am
e
ter
tu
n
in
g
.
Mo
r
e
o
v
er
,
c
o
m
b
in
atio
n
o
f
E
L
U
in
L
STM
m
o
d
el
ef
f
icien
tly
o
v
er
co
m
es
v
a
n
is
h
in
g
g
r
a
d
ien
t
p
r
o
b
lem
s
lead
s
to
s
tab
le
tr
ain
in
g
an
d
e
n
h
an
ce
d
d
is
cr
im
in
a
tio
n
am
o
n
g
b
en
ig
n
,
m
alig
n
an
t a
n
d
n
o
r
m
al
n
o
d
u
les
.
T
h
e
cr
u
cial
co
n
tr
ib
u
tio
n
s
o
f
t
h
is
ar
ticle
ar
e
o
u
tlin
ed
as f
o
llo
ws
:
i)
T
h
e
B
AC
m
eth
o
d
is
u
s
ed
f
o
r
s
eg
m
en
tatio
n
to
d
is
tin
g
u
is
h
th
e
ca
n
ce
r
o
u
s
f
r
o
m
n
o
n
-
ca
n
ce
r
o
u
s
r
eg
io
n
s
i
n
lu
n
g
n
o
d
u
les.
ii)
T
h
e
L
STM
n
etwo
r
k
with
E
L
U
ac
tiv
atio
n
f
u
n
ctio
n
is
em
p
l
o
y
ed
f
o
r
class
if
y
in
g
d
if
f
e
r
en
t
ty
p
es
o
f
lu
n
g
n
o
d
u
les,
m
itig
atin
g
th
e
v
a
n
is
h
in
g
g
r
a
d
ien
t iss
u
e
an
d
en
h
an
ci
n
g
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
iii)
T
h
e
o
p
tim
al
p
a
r
am
eter
s
o
f
t
h
e
L
STM
n
etwo
r
k
f
o
r
lu
n
g
n
o
d
u
le
class
if
icatio
n
ar
e
id
e
n
tifie
d
u
s
in
g
th
e
DL
F
–
AOA,
wh
ich
h
elp
s
im
p
r
o
v
e
class
if
icatio
n
ac
cu
r
ac
y
.
T
h
is
r
esear
ch
p
ap
e
r
is
f
u
r
t
h
er
o
r
g
an
ized
as
f
o
llo
ws:
s
ec
tio
n
2
d
etails
p
r
o
p
o
s
ed
tech
n
iq
u
e
.
Sectio
n
3
p
r
o
v
id
es th
e
r
esu
lts
an
d
a
d
is
c
u
s
s
io
n
o
f
th
e
p
r
o
p
o
s
ed
tech
n
i
q
u
e
. F
in
ally
,
s
ec
tio
n
4
c
o
n
clu
d
e
s
th
e
r
esear
ch
.
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
E
xp
o
n
en
tia
l l
o
n
g
s
h
o
r
t
-
term me
mo
r
y
w
ith
Levy
flig
h
t o
p
timiz
a
tio
n
fo
r
lu
n
g
n
o
d
u
le
…
(
K
a
lib
a
Go
w
th
a
mi
)
1453
2.
M
E
T
H
O
D
An
ef
f
icien
t
DL
-
b
ased
m
o
d
el
is
d
ev
elo
p
ed
f
o
r
class
if
icatio
n
o
f
v
ar
io
u
s
class
es
o
f
lu
n
g
n
o
d
u
les.
T
h
e
IQ
-
OT
H/NC
C
D
lu
n
g
ca
n
ce
r
d
ataset
with
im
ag
es
p
r
e
-
p
r
o
ce
s
s
ed
u
s
in
g
a
wien
er
f
ilter
f
o
r
n
o
is
e
r
em
o
v
al
a
n
d
C
L
AHE
u
s
ed
f
o
r
en
h
an
cin
g
i
m
ag
e
q
u
ality
is
u
s
ed
in
th
is
r
esear
ch
.
Fu
r
th
er
m
o
r
e,
B
AC
-
b
ased
s
eg
m
en
tatio
n
is
p
er
f
o
r
m
ed
to
s
eg
m
en
t
t
h
e
im
ag
es
in
to
p
ix
els
f
o
r
t
h
e
id
en
ti
f
icatio
n
o
f
ca
n
ce
r
o
u
s
an
d
n
o
n
-
ca
n
ce
r
o
u
s
r
eg
io
n
s
.
R
esNet
-
1
8
b
ased
f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
el
is
u
s
ed
to
ca
p
tu
r
e
lo
w
-
lev
el
an
d
h
ig
h
-
lev
el
f
ea
tu
r
es
f
o
r
d
if
f
er
en
tiatin
g
class
es.
Fin
all
y
,
th
e
L
STM
with
E
L
U
m
et
h
o
d
is
d
ev
elo
p
ed
to
ac
c
u
r
ate
ly
class
if
y
d
if
f
er
en
t
class
es
o
f
lu
n
g
n
o
d
u
les,
b
esid
es
th
e
o
p
tim
al
p
ar
am
eter
s
id
e
n
tifie
d
u
s
in
g
DL
F
with
AOA.
Fig
u
r
e
1
illu
s
tr
ates
th
e
p
r
o
ce
s
s
o
f
lu
n
g
n
o
d
u
les cla
s
s
if
icatio
n
.
Fig
u
r
e
1
.
Pro
ce
s
s
o
f
lu
n
g
n
o
d
u
le
class
if
icatio
n
2
.
1
.
Da
t
a
s
et
IQ
-
OT
H/NC
C
D
lu
n
g
ca
n
ce
r
d
ataset
[
2
4
]
u
tili
ze
d
i
n
th
is
s
tu
d
y
class
if
ies
im
ag
e
d
ata
in
to
th
r
ee
ca
teg
o
r
ies:
b
en
ig
n
,
m
alig
n
an
t
,
an
d
n
o
r
m
al.
T
h
e
im
a
g
es
ex
h
ib
it
v
ar
io
u
s
d
if
f
er
en
ce
s
in
b
o
th
d
im
e
n
s
io
n
s
an
d
lab
els,
wh
ich
r
ef
lect
lay
er
ed
c
o
m
p
lex
ity
.
T
h
e
s
tan
d
ar
d
im
ag
e
s
ize
in
a
d
ataset
is
5
1
2
×5
1
2
p
ix
els.
T
h
e
d
ataset
in
clu
d
es
1
2
0
b
e
n
ig
n
s
am
p
les,
ea
ch
with
a
s
ize
o
f
5
1
2
×5
1
2
p
ix
els.
Fo
r
th
e
m
alig
n
an
t
c
lass
,
th
er
e
ar
e
5
6
1
s
am
p
les:
5
0
1
ar
e
5
1
2
×
5
1
2
p
ix
els,
3
1
ar
e
5
1
2
×
6
2
3
p
i
x
els,
2
8
ar
e
5
1
2
×8
0
1
p
ix
els,
an
d
o
n
e
is
4
0
4
×5
1
1
p
ix
els.
T
h
e
n
o
r
m
al
class
co
n
tain
s
4
1
6
s
am
p
les,
with
4
1
5
b
ein
g
5
1
2
×5
1
2
p
ix
els
an
d
o
n
e
b
ein
g
3
3
1
×5
0
6
p
ix
els.
T
h
e
d
is
tr
ib
u
tio
n
o
f
a
d
ataset
is
s
h
o
wn
in
T
ab
le
1
an
d
Fig
u
r
e
2
illu
s
tr
ates sam
p
le
im
ag
es f
r
o
m
t
h
e
d
ataset.
T
ab
le
1
.
Data
s
et
d
is
tr
ib
u
tio
n
Ty
p
e
o
f
c
l
a
s
ses
N
u
mb
e
r
o
f
sam
p
l
e
s
B
e
n
i
g
n
1
2
0
M
a
l
i
g
n
a
n
t
5
6
1
N
o
r
mal
4
1
6
Fig
u
r
e
2
.
Sam
p
le
im
a
g
es in
th
e
d
ataset
2
.
2
.
P
re
-
pro
ce
s
s
ing
T
h
e
d
ataset
in
clu
d
es
1
,
0
0
0
C
T
im
ag
es
in
.
jp
g
f
o
r
m
at,
class
if
ied
in
to
4
ca
teg
o
r
ies:
lar
g
e
ce
l
l
ca
r
cin
o
m
a,
s
q
u
am
o
u
s
ce
ll
ca
r
cin
o
m
a,
n
o
r
m
al,
an
d
ad
en
o
ca
r
cin
o
m
a.
T
h
e
im
ag
es
in
a
d
ataset
ar
e
p
r
e
-
p
r
o
ce
s
s
ed
b
y
W
ien
er
f
ilter
an
d
c
o
n
tr
ast
en
h
an
ce
m
en
t
tec
h
n
iq
u
es,
wh
ic
h
r
em
o
v
e
n
o
is
e
an
d
en
h
an
ce
im
a
g
e
co
n
tr
ast.
A
d
etailed
ex
p
lan
atio
n
o
f
th
ese
p
r
e
-
p
r
o
ce
s
s
in
g
m
eth
o
d
s
is
p
r
o
v
id
e
d
as:
i)
W
ien
er
f
ilter
:
th
e
W
ien
er
f
il
ter
co
m
p
u
tes
an
o
p
tim
u
m
t
r
ad
eo
f
f
b
etwe
en
in
v
er
s
e
f
ilter
in
g
an
d
n
o
is
e
s
m
o
o
th
in
g
.
T
h
is
f
ilter
elim
in
ates
ad
d
itio
n
al
n
o
is
e,
in
v
er
ts
b
lu
r
r
in
g
,
an
d
r
e
d
u
ce
s
o
v
e
r
all
m
ea
n
s
q
u
ar
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
.
2
,
Ap
r
il
20
26
:
1
4
5
1
-
1
4
6
3
1454
er
r
o
r
(
MSE
)
d
u
r
in
g
in
v
er
s
e
f
i
lter
in
g
an
d
n
o
is
e
s
m
o
o
th
in
g
.
T
h
e
m
ath
em
atica
l
f
o
r
m
u
la
f
o
r
th
e
f
ilter
in
g
p
r
o
ce
s
s
is
g
iv
en
in
(
1
)
.
(
,
)
=
×
(
1
,
2
)
(
1
,
2
)
(
1
,
2
)
2
(
1
,
2
)
+
(
1
,
2
)
(
1
)
ii)
C
o
n
tr
ast
en
h
an
ce
m
en
t:
th
is
p
h
ase
en
h
an
ce
s
p
e
r
ce
p
tio
n
f
o
r
f
u
r
t
h
er
p
r
o
ce
s
s
in
g
b
y
s
h
a
r
p
en
in
g
im
ag
e
b
o
r
d
er
s
an
d
m
ax
im
izin
g
ac
c
u
r
ac
y
with
a
n
em
p
h
asis
o
n
b
r
i
g
h
tn
ess
v
ar
ian
ce
b
etwe
en
t
h
e
f
o
r
eg
r
o
u
n
d
an
d
b
ac
k
g
r
o
u
n
d
.
B
esid
es,
C
L
AH
E
is
u
tili
ze
d
to
lim
it
co
n
tr
ast
am
p
lific
atio
n
f
o
r
n
o
is
e
r
em
o
v
al.
I
n
m
ed
ical
im
ag
es,
it
en
h
an
ce
s
th
e
lo
c
al
d
etails
wh
ile
m
ain
tain
in
g
s
m
o
o
th
n
ess
in
ar
ea
s
w
ith
o
u
t
s
ig
n
if
ican
t
m
o
d
if
icatio
n
s
.
T
h
e
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
es
ar
e
f
e
d
in
to
th
e
s
e
g
m
en
tatio
n
p
r
o
ce
s
s
to
s
ep
ar
at
e
th
e
d
is
ea
s
ed
an
d
n
o
n
-
d
is
ea
s
ed
r
eg
i
o
n
s
o
f
t
h
e
lu
n
g
n
o
d
u
les.
Fig
u
r
e
3
illu
s
tr
ates th
e
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
es.
Fig
u
r
e
3
.
Pre
-
p
r
o
ce
s
s
ed
im
ag
e
s
2
.
3
.
Seg
m
ent
a
t
i
o
n
T
h
e
B
AC
m
eth
o
d
is
u
s
ed
in
th
is
p
h
ase,
wh
ich
in
co
r
p
o
r
ates a
B
ay
esian
f
r
am
ewo
r
k
th
at
in
clu
d
es p
r
io
r
k
n
o
wled
g
e,
m
a
k
in
g
it
r
o
b
u
s
t
f
o
r
o
v
er
lap
p
in
g
lu
n
g
s
tr
u
ctu
r
e
s
an
d
v
ar
y
in
g
tu
m
o
r
s
h
ap
es.
T
h
is
is
p
ar
ticu
lar
ly
b
en
ef
icial
wh
en
a
n
n
o
tated
d
at
a
is
lim
ited
,
as
it
is
r
esil
ien
t
t
o
n
o
is
e,
p
r
eser
v
es
ed
g
es,
a
n
d
p
er
f
o
r
m
s
well
with
s
m
all
d
ataset
s
.
C
o
n
s
id
er
N
C
T
im
ag
e
f
r
am
es,
wh
er
e
a
lu
n
g
c
o
n
to
u
r
o
f
a
n
^th
im
ag
e
is
r
ep
r
esen
ted
b
y
th
e
s
tate
v
ec
to
r
.
E
ac
h
f
r
am
e
co
r
r
esp
o
n
d
s
to
a
s
p
atially
s
liced
im
ag
e
f
r
o
m
to
p
t
o
b
o
tto
m
.
State
v
ec
to
r
is
th
en
ev
o
lv
ed
u
s
in
g
(
2
)
.
=
(
−
1
,
)
(
2
)
W
h
er
e
r
e
p
r
esen
ts
a
tr
a
n
s
itio
n
f
u
n
ctio
n
c
o
r
r
esp
o
n
d
in
g
to
th
e
ch
an
g
e
o
f
lu
n
g
co
n
t
o
u
r
f
r
o
m
(
n
−
1
)
th
im
ag
e.
No
is
e
ter
m
r
ep
r
esen
ts
ze
r
o
-
m
ea
n
,
in
d
ep
e
n
d
en
t
wh
ite
-
n
o
is
e,
u
n
af
f
ec
ted
b
y
b
o
th
p
r
ev
i
o
u
s
an
d
p
r
esen
t
s
tates.
I
n
B
ay
esian
m
eth
o
d
,
m
ea
s
u
r
em
en
t
v
ec
to
r
is
r
elev
an
c
e
to
s
tate
v
ec
to
r
th
r
o
u
g
h
o
b
s
e
r
v
atio
n
e
q
u
atio
n
.
T
h
e
m
ea
s
u
r
em
en
t
v
ec
to
r
is
r
e
p
r
esen
ted
as
.
T
h
e
o
b
s
er
v
atio
n
eq
u
atio
n
is
g
iv
e
n
in
(
3
)
.
=
ℎ
(
,
)
(
3
)
W
h
er
e
ℎ
r
e
p
r
esen
ts
m
ea
s
u
r
e
m
en
t
f
u
n
ctio
n
an
d
r
ep
r
ese
n
ts
n
o
is
e
wh
ich
is
a
ze
r
o
-
m
ea
n
s
eq
u
e
n
ce
,
in
d
ep
en
d
en
t
wh
ite
n
o
is
e,
u
n
r
elate
d
to
b
o
th
p
r
e
v
io
u
s
an
d
cu
r
r
en
t
s
tates,
as
well
as
s
y
s
tem
n
o
is
e.
I
n
itial
p
r
o
b
a
b
ilit
y
d
en
s
ity
f
u
n
ctio
n
(
PDF)
is
r
ep
r
esen
ted
as
(
1
|
0
)
=
(
1
)
.
T
h
e
av
ailab
le
d
ata
in
n
th
f
r
am
e
is
s
et
to
a
m
ea
s
u
r
em
en
t v
ec
to
r
.
T
h
e
o
b
jectiv
e
is
to
d
ev
elo
p
a
PDF o
f
s
tate
v
ec
to
r
,
b
ased
o
n
all
av
ailab
le
d
ata,
wh
ich
is
ac
q
u
ir
ed
r
ec
u
r
s
iv
ely
in
two
p
h
ases
o
f
u
p
d
ate
an
d
p
r
ed
ictio
n
.
R
eq
u
ir
e
d
PDF
an
d
p
r
ev
io
u
s
PDF
o
f
th
e
s
tate
v
ec
to
r
is
attain
ed
u
s
in
g
(
4
)
an
d
(
5
)
.
(
|
−
1
)
=
∫
(
|
−
1
)
(
−
1
|
−
1
)
−
1
(
4
)
Sin
ce
th
e
(
|
−
1
)
=
(
)
.
(
|
−
1
)
=
∫
(
−
(
−
1
,
)
)
(
)
(
5
)
W
h
er
e
δ
(
.
)
r
ep
r
esen
ts
a
d
ir
ac
d
elt
a
f
u
n
ctio
n
.
I
n
n
th
f
r
am
e,
r
ep
r
esen
ts
th
e
m
ea
s
u
r
em
en
t
v
ec
to
r
p
r
esen
ted
an
d
u
tili
ze
d
f
o
r
u
p
d
atin
g
th
e
p
r
ev
io
u
s
PDF
f
o
r
p
o
s
ter
io
r
PDF
in
ac
co
r
d
an
ce
with
th
e
B
ay
es’
r
u
le,
as
m
ath
em
atica
lly
f
o
r
m
u
lated
in
(
6
)
.
T
h
e
m
ath
em
atica
l
f
o
r
m
u
l
a
f
o
r
d
en
o
m
in
ato
r
n
o
r
m
alizin
g
is
g
iv
e
n
as
in
(
7
)
.
T
h
e
m
ath
em
atica
l f
o
r
m
u
la
f
o
r
PDF co
n
d
itio
n
is
r
ep
r
esen
ted
t
h
r
o
u
g
h
m
ea
s
u
r
e
m
en
t f
u
n
ctio
n
,
as g
iv
en
in
(
8
)
.
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
E
xp
o
n
en
tia
l l
o
n
g
s
h
o
r
t
-
term me
mo
r
y
w
ith
Levy
flig
h
t o
p
timiz
a
tio
n
fo
r
lu
n
g
n
o
d
u
le
…
(
K
a
lib
a
Go
w
th
a
mi
)
1455
(
|
)
=
(
|
)
(
|
−
1
)
(
|
−
1
)
(
6
)
(
|
−
1
)
=
∫
(
|
)
(
|
)
(
7
)
(
|
)
=
∫
(
−
ℎ
(
,
)
)
(
)
(
8
)
W
h
er
e
r
ep
r
esen
ts
th
e
m
ea
s
u
r
em
en
t
v
ec
to
r
u
tili
ze
d
to
m
o
d
if
y
th
e
p
r
ed
icte
d
p
r
ev
i
o
u
s
PDF
f
o
r
o
b
tain
in
g
th
e
s
u
b
s
eq
u
en
tly
r
e
q
u
ir
ed
PDF
o
f
a
s
tate
v
ec
to
r
.
Nex
t,
an
u
p
d
ated
PDF
is
u
tili
ze
d
f
o
r
a
p
r
ev
io
u
s
PDF
in
+
1
th
f
r
am
e.
T
h
e
r
ec
u
r
r
en
t
r
el
atio
n
s
o
f
eq
u
atio
n
s
estab
lis
h
a
g
en
er
al
s
o
lu
tio
n
f
o
r
B
ay
esian
r
ec
u
r
s
iv
e
ev
alu
atio
n
is
s
u
es.
2
.
4
.
F
e
a
t
ure
ex
t
r
a
ct
io
n
T
h
e
s
eg
m
en
ted
im
ag
es
ar
e
in
p
u
t
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
,
wh
er
e
m
ea
n
in
g
f
u
l
a
n
d
d
ee
p
f
ea
t
u
r
es
o
f
th
e
lu
n
g
n
o
d
u
les
ar
e
ex
tr
ac
ted
u
s
in
g
th
e
R
esNet
-
1
8
m
et
h
o
d
.
R
esNet
-
1
8
lev
er
ag
es
r
esid
u
al
c
o
n
n
ec
tio
n
s
,
w
h
ich
h
elp
m
itig
ate
th
e
v
an
is
h
in
g
g
r
ad
ien
t
is
s
u
e
an
d
en
h
a
n
ce
d
ee
p
f
ea
tu
r
e
lear
n
in
g
.
R
esNet
is
a
r
esid
u
al
n
etwo
r
k
with
an
1
8
-
lay
e
r
C
NN
p
r
e
-
tr
ain
ed
m
o
d
el.
I
n
th
is
s
tu
d
y
,
R
esNet
-
1
8
u
tili
ze
s
3
×3
f
ilter
s
with
a
s
tr
id
e
an
d
p
ad
d
in
g
o
f
1
,
an
av
er
ag
e
p
o
o
l
in
g
lay
er
with
a
1
×1
f
ilter
,
1
f
u
lly
co
n
n
ec
ted
lay
er
,
an
d
a
f
i
n
al
S
o
f
t
M
ax
lay
er
.
T
h
e
R
esNet
-
1
8
in
clu
d
es
f
o
u
r
m
ajo
r
p
h
ases
wh
er
e
ev
er
y
p
h
ase
co
n
tain
s
a
m
u
lti
-
r
esid
u
al
b
lo
ck
.
T
h
ese
b
lo
ck
s
u
tili
ze
s
k
ip
co
n
n
ec
tio
n
s
wh
ich
allo
w
d
ata
f
o
r
b
y
p
ass
in
g
s
o
m
e
lay
er
s
,
th
er
eb
y
ass
is
tin
g
th
e
n
etwo
r
k
in
lear
n
in
g
d
ee
p
f
ea
t
u
r
es
b
y
m
itig
atin
g
v
an
is
h
in
g
g
r
ad
ien
ts
,
p
r
o
v
i
n
g
a
d
v
an
tag
eo
u
s
in
id
e
n
tify
in
g
th
e
r
elev
an
t
s
tr
u
ctu
r
es.
Af
ter
ev
er
y
r
esid
u
al
b
lo
ck
,
th
e
p
o
o
lin
g
la
y
er
m
in
im
izes
th
e
s
p
atial
d
im
en
s
io
n
s
o
f
f
ea
tu
r
e
s
an
d
co
n
ce
n
tr
ates
o
n
th
e
s
ig
n
if
ica
n
t
d
ata,
m
a
k
i
n
g
th
e
n
etwo
r
k
m
o
r
e
r
o
b
u
s
t
to
s
p
atial
d
if
f
er
en
ce
s
in
th
e
in
p
u
t.
Af
ter
f
in
al
co
n
v
o
l
u
tio
n
al
b
l
o
ck
,
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
is
em
p
lo
y
e
d
to
ef
f
icac
io
u
s
ly
o
b
tain
f
ea
tu
r
e
v
ec
to
r
s
wh
ich
d
escr
ib
es th
e
en
tire
im
ag
e
i
n
a
co
n
d
e
n
s
ed
f
o
r
m
at.
2
.
5
.
Cla
s
s
if
ica
t
io
n us
ing
lo
ng
s
ho
rt
-
t
er
m m
e
m
o
ry
wit
h
ex
po
nentia
l linea
r
un
it
R
NNs
ar
e
k
in
d
o
f
d
ee
p
n
e
u
r
al
n
etwo
r
k
(
DNN)
wh
ich
estab
l
is
h
co
n
n
ec
tio
n
s
b
etwe
en
n
e
u
r
o
n
s
with
in
th
e
s
am
e
lay
er
,
m
a
k
in
g
th
em
s
u
itab
le
f
o
r
tim
e
-
s
er
ies
task
s
.
L
STM
is
a
v
ar
ian
t
o
f
R
NN
th
at
ad
d
r
ess
es
d
ata
r
ed
u
n
d
an
cy
b
y
i
n
co
r
p
o
r
atin
g
a
"g
ate"
ar
ch
itectu
r
e
at
s
tr
ateg
ic
p
o
s
itio
n
s
.
T
h
is
allo
ws
th
e
n
etwo
r
k
t
o
s
elec
tiv
ely
r
etain
o
r
d
is
ca
r
d
d
ata
as
it
f
lo
ws
th
r
o
u
g
h
ea
c
h
n
eu
r
o
n
.
L
STM
s
ar
e
e
f
f
ec
tiv
e
in
o
v
er
c
o
m
in
g
is
s
u
es
s
u
ch
as
v
a
n
is
h
in
g
g
r
ad
ien
ts
an
d
ex
p
l
o
d
in
g
g
r
a
d
ien
ts
,
wh
ich
co
m
m
o
n
ly
af
f
ec
t
tr
ad
itio
n
al
R
NNs
[
2
5
]
.
Mo
r
eo
v
er
,
th
ey
ex
ce
l
in
h
a
n
d
l
in
g
lo
n
g
-
te
r
m
d
ep
e
n
d
en
cies
in
th
e
d
ata.
T
h
e
an
d
is
weig
h
t
an
d
b
ias
o
f
a
in
p
u
t
g
ate,
th
e
an
d
is
weig
h
t
an
d
b
ias
o
f
a
o
u
tp
u
t
g
ate,
an
d
is
weig
h
t
an
d
b
ias
o
f
a
ce
ll
s
tates
,
an
d
is
weig
h
t
a
n
d
b
ias
o
f
f
o
r
g
et
g
ate.
T
h
e
ℎ
−
1
is
o
u
tp
u
t
o
f
p
r
ev
io
u
s
tim
e
s
tep
,
wh
ile
is
in
p
u
t
o
f
cu
r
r
en
t
tim
e
s
tep
.
T
h
e
d
ev
el
o
p
ed
n
ew
v
ec
t
o
r
is
d
en
o
ted
as
[
ℎ
−
1
,
]
.
T
h
e
p
r
o
ce
s
s
o
f
m
atr
ix
m
u
ltip
licatio
n
o
f
[
ℎ
−
1
,
]
v
ec
to
r
with
an
d
th
r
o
u
g
h
th
e
tan
h
ac
tiv
atio
n
f
u
n
ctio
n
is
u
s
ed
to
o
b
tai
n
th
e
in
p
u
t,
as
d
esc
r
ib
ed
in
(
9
)
.
T
h
e
weig
h
ts
an
d
p
ar
a
m
eter
ter
m
s
o
f
ea
ch
g
ate
ar
e
p
r
o
ce
s
s
ed
th
r
o
u
g
h
th
e
s
ig
m
o
i
d
f
u
n
ctio
n
to
p
r
o
d
u
c
e
th
e
f
o
r
g
et,
in
p
u
t,
an
d
o
u
tp
u
t g
ates.
T
h
e
m
ath
em
atica
l f
o
r
m
u
l
as f
o
r
th
ese
g
ates a
r
e
p
r
o
v
id
e
d
in
(
1
0
)
to
(
1
2
)
.
′
=
ℎ
(
∙
[
ℎ
−
1
,
]
+
)
(
9
)
=
(
∙
[
ℎ
−
1
,
]
+
)
(
1
0
)
=
(
∙
[
ℎ
−
1
,
]
+
)
(
1
1
)
=
(
∙
[
ℎ
−
1
,
]
+
)
(
1
2
)
B
y
m
u
ltip
ly
in
g
with
th
e
p
r
ev
io
u
s
ce
ll
s
tate
−
1
,
th
e
d
ata
to
b
e
f
o
r
g
o
tten
an
d
r
em
em
b
er
e
d
is
d
eter
m
in
ed
,
th
er
eb
y
c
o
n
tr
o
lli
n
g
−
1
.
B
y
m
u
ltip
ly
i
n
g
with
th
e
cu
r
r
e
n
t
in
p
u
t
ce
ll
s
tate
′
,
th
e
d
ata
′
r
eq
u
ir
ed
to
b
e
s
to
r
e
d
an
d
u
s
ed
is
d
eter
m
in
e
d
.
Fin
ally
,
th
e
s
ta
te
v
alu
e
o
f
t
h
e
h
i
d
d
en
lay
er
at
tim
e
is
o
b
tain
e
d
b
y
ad
d
in
g
two
p
r
o
d
u
cts,
as
s
h
o
wn
in
(
1
3
)
.
Fin
ally
,
r
esu
lt
at
tim
e
is
d
eter
m
in
ed
b
y
u
n
it
s
tate,
o
u
tp
u
t
g
ate,
an
d
tan
h
ac
tiv
atio
n
f
u
n
ctio
n
,
a
s
d
escr
ib
ed
in
(
1
4
)
.
=
×
−
1
+
×
′
(
1
3
)
ℎ
=
×
ℎ
(
)
(
1
4
)
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
.
2
,
Ap
r
il
20
26
:
1
4
5
1
-
1
4
6
3
1456
2
.
5
.
1
.
E
x
po
nentia
l
lin
ea
r
un
i
t
T
h
e
E
L
U
in
teg
r
ates
two
ac
tiv
atio
n
f
u
n
ctio
n
s
,
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
an
d
s
ig
m
o
id
.
T
h
e
ac
tiv
atio
n
f
u
n
ctio
n
is
u
s
ed
to
weig
h
th
e
in
teg
r
atio
n
o
f
s
o
u
r
c
es
o
r
r
esu
lts
ap
p
lied
in
th
e
s
u
b
s
eq
u
en
t
p
h
ase.
T
h
e
o
u
tp
u
t
o
f
th
is
f
u
n
ctio
n
f
lu
ctu
ates
b
etwe
en
0
an
d
1
wh
en
t
h
e
p
er
ce
p
tr
o
n
is
a
m
ath
em
ati
ca
l
f
u
n
ctio
n
.
E
L
U
d
if
f
er
s
f
r
o
m
R
eL
U
b
y
allo
win
g
n
eg
ativ
e
v
alu
es.
I
t
o
f
f
er
s
a
lo
wer
e
x
ec
u
tio
n
co
s
t
co
m
p
ar
ed
to
b
atch
n
o
r
m
aliza
tio
n
a
n
d
r
esu
lts
in
h
ig
h
er
av
e
r
ag
e
u
n
it
ac
tiv
atio
n
,
wh
ich
is
clo
s
er
to
0
.
T
h
e
h
y
p
er
p
ar
am
eter
co
n
tr
o
ls
th
e
s
atu
r
atio
n
v
alu
e
o
f
th
e
E
L
U
f
o
r
n
eg
ativ
e
i
n
p
u
ts
.
T
h
e
m
ath
em
atica
l
f
o
r
m
u
las
f
o
r
E
L
U
ac
tiv
atio
n
f
u
n
ctio
n
a
r
e
g
iv
e
n
in
(
1
5
)
an
d
(
1
6
)
.
(
)
=
{
,
>
0
(
(
)
−
1
)
,
≤
0
(
1
5
)
′
(
)
=
{
1
,
>
0
(
)
+
,
≤
0
(
1
6
)
C
o
m
p
ar
ed
t
o
o
th
er
ac
tiv
atio
n
f
u
n
ctio
n
s
,
th
e
E
L
U
en
h
an
ce
s
lear
n
in
g
ch
ar
ac
ter
is
tics
at
all
tim
es
an
d
ac
h
iev
es
an
av
er
a
g
e
ac
tiv
ati
o
n
clo
s
er
to
0
,
wh
ic
h
ac
ce
le
r
ates
lear
n
in
g
.
T
h
e
E
L
U
ac
ti
v
atio
n
r
ed
u
ce
s
th
e
d
if
f
er
en
ce
in
f
o
r
war
d
p
r
o
p
ag
a
tio
n
an
d
th
e
d
ata
tr
an
s
f
er
r
ed
t
o
th
e
n
ex
t
lay
er
,
e
v
en
wh
en
t
h
e
in
p
u
t
p
ar
a
m
eter
s
ar
e
lim
ited
.
I
n
t
h
is
p
ap
e
r
,
th
e
E
L
U
is
u
s
ed
in
s
tead
o
f
th
e
s
ig
m
o
id
,
o
f
f
e
r
in
g
b
o
t
h
r
o
b
u
s
tn
ess
an
d
r
ed
u
ce
d
co
m
p
lex
ity
.
T
h
is
s
im
p
lifie
s
th
e
p
r
o
ce
s
s
an
d
a
d
d
r
ess
es th
e
is
s
u
e
o
f
v
an
is
h
in
g
g
r
ad
ien
ts
.
2
.
5
.
2
.
H
y
perpa
ra
m
et
er
t
un
ing
us
ing
dy
na
m
ic
L
ev
y
f
lig
ht
wit
h
Arc
him
edes o
ptim
iza
t
io
n a
lg
o
rit
hm
I
n
th
is
ar
ticle,
th
e
DL
F in
tr
o
d
u
ce
s
r
an
d
o
m
lar
g
e
ju
m
p
s
b
ase
d
o
n
th
e
L
ev
y
d
is
tr
ib
u
tio
n
,
e
n
h
an
cin
g
th
e
ca
p
ab
ilit
y
to
escap
e
p
o
o
r
s
o
lu
tio
n
s
d
u
r
in
g
h
y
p
er
p
ar
am
eter
t
u
n
in
g
.
Un
lik
e
o
th
er
o
p
tim
izati
o
n
alg
o
r
ith
m
s
s
u
ch
as
W
OA,
GA,
o
r
PS
O,
wh
ich
r
ely
o
n
f
ix
e
d
s
tep
s
izes,
DL
F
-
AOA
d
y
n
am
ically
ad
ju
s
ts
h
y
p
er
p
ar
am
eter
s
b
y
b
alan
cin
g
g
lo
b
al
ex
p
lo
r
atio
n
a
n
d
f
i
n
e
-
tu
n
e
d
lo
c
al
s
ea
r
ch
.
T
h
e
AOA
en
s
u
r
es
ef
f
ec
tiv
e
e
x
p
l
o
itatio
n
,
wh
ile
DL
F
ac
ce
ler
ates
co
n
v
er
g
en
ce
b
y
im
p
r
o
v
in
g
s
ea
r
ch
d
iv
er
s
ity
.
AOA
is
a
p
o
p
u
latio
n
-
b
ased
alg
o
r
ith
m
,
wh
er
e
in
d
iv
id
u
als in
th
e
p
o
p
u
latio
n
a
r
e
f
o
cu
s
ed
o
n
s
p
ec
if
ic
o
b
jectiv
es.
Similar
to
o
th
er
p
o
p
u
latio
n
-
b
ased
alg
o
r
ith
m
s
,
AOA
d
r
iv
es
th
e
p
r
o
ce
s
s
with
a
m
ain
p
o
p
u
latio
n
o
f
in
d
iv
id
u
als
th
at
ar
e
s
u
b
ject
to
r
an
d
o
m
ac
ce
ler
atio
n
s
,
d
en
s
ities
,
an
d
v
o
l
u
m
es.
T
h
e
AOA
b
eg
in
s
b
y
ev
alu
atin
g
th
e
f
itn
ess
o
f
t
h
e
p
r
elim
in
ar
y
p
o
p
u
latio
n
an
d
co
n
tin
u
es
iter
atin
g
u
n
til
ter
m
i
n
atio
n
co
n
d
itio
n
is
m
et.
I
n
ev
er
y
iter
atio
n
,
v
o
l
u
m
e
an
d
d
en
s
ity
o
f
in
d
iv
id
u
als
ar
e
u
p
d
ated
.
T
h
e
in
d
iv
id
u
al'
s
ac
ce
ler
atio
n
is
r
ev
is
ed
b
ased
o
n
th
eir
in
ter
ac
tio
n
with
n
eig
h
b
o
r
in
g
in
d
iv
i
d
u
als,
an
d
th
e
u
p
d
ated
d
en
s
ity
,
v
o
lu
m
e,
an
d
ac
ce
ler
atio
n
d
ete
r
m
i
n
e
n
ew
p
o
s
itio
n
o
f
th
e
in
d
i
v
id
u
al.
Ma
th
em
atica
l
f
o
r
m
u
la
f
o
r
i
n
itializatio
n
is
p
r
o
v
id
ed
i
n
(
1
7
)
.
=
+
×
(
×
)
,
=
1
,
2
,
.
.
.
,
(
1
7
)
W
h
er
e
is
n
u
m
b
er
o
f
in
d
iv
id
u
als
in
p
o
p
u
latio
n
,
is
th
e
ℎ
i
n
d
iv
id
u
al,
r
ep
r
esen
ts
th
e
lo
wer
lim
it,
r
ep
r
esen
ts
th
e
u
p
p
er
lim
it
o
f
th
e
s
ea
r
ch
ar
ea
,
a
n
d
r
ep
r
esen
ts
th
e
d
im
en
s
io
n
v
ec
to
r
ar
b
itra
r
ily
g
iv
e
n
in
th
e
r
an
g
e
[
0
,
1
]
.
T
h
e
m
ath
em
atica
l
f
o
r
m
u
la
f
o
r
r
eset
d
en
s
ity
f
o
r
ev
er
y
ℎ
in
d
iv
id
u
al
is
g
iv
en
in
(
1
8
)
.
Fin
ally
,
th
e
in
itial
ac
ce
ler
atio
n
o
f
th
e
ℎ
p
ar
am
ete
r
a
n
d
its
m
ath
em
atica
l
f
o
r
m
u
la
a
r
e
g
iv
en
in
(
1
9
)
.
I
n
th
is
p
h
ase,
th
e
in
itial
p
o
p
u
latio
n
is
ev
alu
ated
an
d
r
elev
an
t
i
n
d
iv
id
u
als
th
at
h
av
e
o
p
tim
al
f
itn
ess
v
alu
es
ar
e
s
elec
ted
.
T
h
en
,
,
,
,
an
d
ar
e
em
p
lo
y
ed
,
wh
er
e
r
ep
r
esen
ts
th
e
o
p
tim
al
in
d
iv
id
u
al,
,
,
an
d
r
ep
r
esen
t d
en
s
ity
,
v
o
lu
m
e
,
an
d
ac
ce
ler
atio
n
r
elev
a
n
c
e
with
a
f
itn
ess
v
alu
e.
=
,
=
(
1
8
)
=
+
×
(
×
)
(
1
9
)
i)
R
en
ew
v
o
lu
m
es
an
d
d
en
s
ities
:
d
en
s
ity
an
d
v
o
l
u
m
e
o
f
t
h
e
in
d
iv
id
u
al
in
r
e
p
etitio
n
+
1
an
d
its
m
ath
em
atica
l
f
o
r
m
u
la
is
g
iv
en
as
(
2
0
)
an
d
(
2
1
)
.
W
h
er
e
an
d
r
ep
r
esen
t
d
en
s
ity
an
d
v
o
l
u
m
e
r
elev
an
ce
.
W
ith
a
n
o
p
tim
al
i
n
d
iv
id
u
al,
d
is
tr
ib
u
ted
u
n
if
o
r
m
ly
as
th
e
r
a
n
d
o
m
n
u
m
b
er
is
r
ep
r
esen
ted
as
.
+
1
=
+
(
−
)
(
2
0
)
+
1
=
+
(
−
)
(
2
1
)
ii)
Den
s
ity
f
ac
to
r
an
d
tr
an
s
f
er
o
p
er
ato
r
:
i
n
itially
,
in
d
iv
i
d
u
al
s
co
llid
e,
an
d
o
v
er
tim
e,
th
ey
attem
p
t
to
co
n
v
er
g
e
to
war
d
a
s
y
m
m
etr
i
c
s
tate.
I
n
AOA,
th
is
p
r
o
ce
s
s
is
f
ac
ilit
ated
b
y
th
e
tr
an
s
f
e
r
o
p
er
ato
r
,
th
at
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
E
xp
o
n
en
tia
l l
o
n
g
s
h
o
r
t
-
term me
mo
r
y
w
ith
Levy
flig
h
t o
p
timiz
a
tio
n
fo
r
lu
n
g
n
o
d
u
le
…
(
K
a
lib
a
Go
w
th
a
mi
)
1457
s
h
if
ts
th
e
f
o
cu
s
f
r
o
m
ex
p
lo
r
at
io
n
to
ex
p
lo
itatio
n
.
T
h
e
m
at
h
em
atica
l
f
o
r
m
u
la
f
o
r
th
e
tr
an
s
f
er
o
p
er
ato
r
is
g
iv
en
in
(
2
2
)
.
I
n
th
e
(
2
2
)
,
is
n
u
m
b
er
o
f
iter
atio
n
s
a
n
d
is
m
ax
im
u
m
iter
atio
n
s
.
T
h
e
d
en
s
ity
d
ec
lin
in
g
f
ac
to
r
(
)
h
elp
s
th
e
AOA
in
g
lo
b
al
an
d
lo
ca
l
s
ea
r
ch
es,
an
d
its
m
ath
em
atica
l
f
o
r
m
u
la
is
g
iv
en
in
(
2
3
)
.
I
n
th
e
(
2
3
)
,
+
1
r
e
p
r
esen
ts
th
e
d
ec
lin
in
g
tim
e
w
h
ich
g
iv
es
th
e
m
o
d
el
th
e
ab
il
ity
to
co
n
v
er
g
e
in
p
ast
r
ec
o
g
n
ized
r
eg
io
n
s
.
T
h
is
r
ep
r
esen
ts
th
e
s
u
itab
le
h
an
d
lin
g
o
f
th
e
v
ar
iati
o
n
to
s
tr
ik
e
a
b
alan
ce
am
o
n
g
ex
p
lo
itatio
n
an
d
ex
p
lo
r
atio
n
p
h
ases
in
AOA
alg
o
r
ith
m
.
=
(
−
)
(
2
2
)
+
1
=
(
−
)
−
(
)
(
2
3
)
iii)
E
x
p
lo
r
atio
n
p
h
ase
(
c
o
llis
io
n
am
o
n
g
in
d
iv
id
u
als)
:
w
h
en
≤
0
.
5
,
th
er
e
is
a
co
llis
io
n
am
o
n
g
in
d
iv
id
u
als,
an
d
s
o
a
r
an
d
o
m
m
ater
ial
is
s
elec
ted
an
d
ac
ce
l
er
atio
n
o
f
th
e
p
a
r
am
eter
f
o
r
+
1
r
ep
etitio
n
is
r
ev
is
ed
,
as
s
h
o
wn
in
(
2
4
)
.
I
n
th
e
(
2
4
)
,
r
ep
r
esen
ts
d
en
s
ity
,
r
ep
r
esen
ts
th
e
v
o
l
u
m
e
,
an
d
r
ep
r
esen
ts
ac
ce
ler
atio
n
o
f
ℎ
i
n
d
iv
id
u
al.
T
h
is
is
ess
en
tial
f
o
r
r
ep
r
esen
tatio
n
as
≤
0
.
5
,
wh
ich
g
u
ar
an
tees e
x
p
lo
r
atio
n
at
tim
e
o
f
1
/3
iter
atio
n
.
+
1
=
+
+
+
1
+
+
1
(
2
4
)
iv
)
E
x
p
lo
itatio
n
s
tag
e
(
n
u
ll
co
llis
io
n
am
o
n
g
i
n
d
iv
id
u
als)
:
i
n
co
n
d
itio
n
>
0
.
5
,
th
er
e
is
n
u
ll
co
lli
s
io
n
b
etwe
en
in
d
iv
i
d
u
als,
r
ev
is
in
g
th
e
ac
ce
ler
atio
n
o
f
in
d
iv
id
u
als
f
o
r
iter
atio
n
(
+
1
)
,
as
m
ath
e
m
atica
lly
f
o
r
m
u
lated
in
(
2
5
)
.
B
y
u
s
in
g
(
2
6
)
,
ac
ce
ler
atio
n
is
r
eq
u
i
r
ed
t
o
b
e
n
o
r
m
alize
d
f
o
r
e
x
ec
u
tin
g
th
e
v
a
r
iatio
n
p
r
o
p
o
r
tio
n
.
I
n
th
e
(
2
6
)
,
=
0
.
9
an
d
=
0
.
1
r
ep
r
esen
t
th
e
n
o
r
m
aliza
tio
n
r
a
n
g
e.
T
h
e
−
+
1
r
ep
r
esen
ts
th
e
r
atio
o
f
p
h
ases
in
wh
ich
ea
ch
ag
en
t
v
ar
ies.
I
n
th
e
co
n
d
itio
n
wh
er
e
is
an
in
d
iv
id
u
al
th
at
is
af
ield
f
r
o
m
th
e
g
lo
b
al
o
p
tim
u
m
,
t
h
e
v
alu
e
o
f
ac
ce
ler
atio
n
is
h
ig
h
,
m
ea
n
in
g
th
at
in
d
iv
id
u
al
r
em
ain
s
in
ex
p
lo
itatio
n
o
r
ex
p
lo
r
atio
n
s
tag
e
an
d
p
ath
o
f
ex
p
lo
r
atio
n
tr
an
s
itio
n
s
f
r
o
m
a
s
tag
e
o
f
ex
p
lo
r
atio
n
t
o
ex
p
lo
itatio
n
.
Un
d
er
g
e
n
er
al
c
o
n
d
itio
n
s
,
an
ac
ce
ler
atio
n
c
o
m
p
o
n
e
n
t
in
itiates
with
h
ig
h
er
v
alu
e
wh
ich
d
ec
lin
es
with
tim
e.
T
h
is
h
elp
s
th
e
s
ea
r
ch
ag
en
ts
in
n
av
ig
ati
n
g
to
war
d
th
e
g
lo
b
al
o
p
tim
al
o
u
tco
m
es
an
d
awa
y
f
r
o
m
lo
ca
l
s
o
lu
tio
n
s
.
T
h
e
k
ey
o
b
s
er
v
atio
n
is
th
at
th
er
e
ar
e
a
f
ew
s
ea
r
ch
ag
en
ts
th
at
r
eq
u
ir
e
ex
tr
a
tim
e
f
o
r
th
e
e
x
p
lo
r
atio
n
p
h
ase
co
m
p
ar
e
d
to
g
en
er
al
cir
c
u
m
s
tan
ce
s
.
Hen
ce
,
th
e
s
y
m
b
io
tic
o
r
g
an
is
m
s
s
ea
r
ch
(
SOS
)
p
r
o
v
i
d
es st
ab
ilit
y
in
b
o
t
h
ex
p
lo
itatio
n
an
d
ex
p
l
o
r
atio
n
.
+
1
=
+
+
+
1
+
+
1
(
2
5
)
−
+
1
=
×
+
1
−
(
)
(
)
+
(
)
(
2
6
)
v)
Up
d
ate
p
o
s
itio
n
:
w
h
en
≤
0
.
5
,
th
e
p
o
s
itio
n
o
f
i
n
d
iv
id
u
al
f
o
r
a
n
ex
t
iter
atio
n
+
1
is
g
iv
en
as
in
(
2
7
)
.
I
n
t
h
e
(
2
7
)
,
1
r
ep
r
esen
ts
a
co
n
s
tan
t
s
et
to
2
.
Alter
n
ativ
e
ly
,
wh
en
>
0
.
5
,
in
d
iv
i
d
u
als
r
ev
is
e
th
eir
p
o
s
itio
n
s
,
as
m
ath
em
atic
ally
f
o
r
m
u
lated
in
(
2
8
)
.
I
n
th
e
(
2
8
)
,
2
r
ep
r
esen
ts
a
co
n
s
tan
t
s
et
to
6
.
,
p
r
o
p
o
r
tio
n
al
t
o
T
O,
is
d
ef
in
e
d
u
s
in
g
=
3
×
.
W
h
er
e,
3
r
ep
r
esen
ts
a
co
n
s
tan
t
s
et
to
2
.
As
in
cr
ea
s
es,
it
co
n
s
id
er
s
th
e
tim
e
lim
it
an
d
th
e
r
atio
f
r
o
m
t
h
e
o
p
tim
al
p
o
s
itio
n
.
I
t
s
tar
ts
with
s
m
all
r
atio
s
,
r
esu
ltin
g
in
a
h
ig
h
v
ar
ia
n
ce
c
o
n
ce
r
n
i
n
th
e
s
tep
-
s
ize
o
f
r
a
n
d
o
m
walk
b
etwe
en
th
e
o
p
tim
al
an
d
cu
r
r
en
t
lo
ca
tio
n
s
.
As
a
s
ea
r
c
h
p
r
o
g
r
ess
es,
th
is
r
atio
in
cr
ea
s
es,
r
e
d
u
cin
g
th
e
v
ar
ian
ce
b
etwe
en
th
e
o
p
tim
al
p
o
s
itio
n
an
d
th
e
p
r
esen
t
s
tate,
th
er
eb
y
ac
h
iev
i
n
g
an
ap
p
r
o
p
r
iate
eq
u
ilib
r
iu
m
b
etwe
en
e
x
p
lo
itatio
n
an
d
ex
p
lo
r
atio
n
.
T
h
e
f
lag
in
d
icate
s
v
ar
io
u
s
d
ir
ec
tio
n
s
o
f
m
o
tio
n
,
an
d
its
m
ath
e
m
atica
l
f
o
r
m
u
l
a
is
g
iv
e
n
in
(
2
9
)
.
I
n
t
h
e
(
2
9
)
,
=
2
×
−
4
an
d
4
r
ep
r
es
en
t
a
co
n
s
tan
t
eq
u
al
to
0
.
5
.
E
ac
h
p
ar
am
eter
is
ev
alu
ated
u
s
in
g
th
e
f
itn
ess
f
u
n
ctio
n
an
d
r
ec
o
r
d
th
e
o
p
ti
m
al
o
u
tp
u
t
o
b
tain
ed
s
o
f
ar
.
T
h
e
v
alu
es
o
f
,
,
an
d
ar
e
allo
ca
ted
.
+
1
=
+
1
×
×
−
+
1
×
×
(
−
)
(
2
7
)
+
1
=
+
1
+
+
2
×
×
−
+
1
×
×
(
×
−
)
(
2
8
)
=
{
+
1
≤
0
.
5
−
1
>
0
.
5
}
(
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
.
2
,
Ap
r
il
20
26
:
1
4
5
1
-
1
4
6
3
1458
v
i)
DL
F
with
AOA
:
i
n
th
is
s
ec
tio
n
,
d
y
n
am
ic
f
ea
tu
r
es
ar
e
em
p
lo
y
ed
f
o
r
th
e
s
o
lu
tio
n
s
in
AOA.
Me
tah
eu
r
is
tic
alg
o
r
ith
m
s
g
e
n
er
ally
c
o
n
s
is
t
o
f
two
p
r
im
ar
y
p
h
ases
:
ex
p
lo
r
atio
n
an
d
e
x
p
lo
itatio
n
,
wh
ic
h
m
ain
tain
a
n
o
p
tim
al
b
alan
ce
b
etwe
en
th
e
m
.
I
n
th
is
p
ap
e
r
,
th
e
d
y
n
a
m
ic
v
er
s
io
n
is
u
tili
ze
d
to
e
m
p
h
asize
b
o
t
h
ex
p
lo
itatio
n
an
d
e
x
p
lo
r
atio
n
,
wh
er
e
ea
ch
s
o
lu
tio
n
d
y
n
am
ic
ally
u
p
d
ates
its
p
o
s
itio
n
b
ase
d
o
n
th
e
b
est
s
o
lu
tio
n
attain
ed
d
u
r
i
n
g
th
e
o
p
tim
izatio
n
p
r
o
ce
s
s
.
T
h
e
d
y
n
am
ic
ca
n
d
id
ate
s
o
lu
tio
n
(
DC
S)
f
u
n
ctio
n
is
in
co
r
p
o
r
ated
with
th
e
L
ev
y
f
li
g
h
t
f
u
n
ctio
n
,
r
ep
lacin
g
(
2
5
)
a
n
d
(
2
6
)
.
T
h
e
u
p
d
ated
m
ath
em
atica
l
f
o
r
m
u
las
ar
e
p
r
o
v
id
ed
i
n
(
3
0
)
a
n
d
(
3
1
)
.
T
h
e
DC
S
f
u
n
ctio
n
is
im
p
lem
en
ted
to
ac
co
u
n
t
f
o
r
th
e
ef
f
ec
t
o
f
d
ec
r
ea
s
in
g
th
e
p
er
ce
n
tag
e
in
th
e
ca
n
d
id
at
e
s
o
lu
tio
n
d
u
r
in
g
ea
c
h
iter
atio
n
.
I
ts
v
alu
es a
r
e
g
iv
en
in
(
3
2
)
an
d
(
3
3
)
.
+
1
=
+
+
+
1
+
+
1
+
×
(
3
0
)
−
+
1
=
×
+
1
−
(
)
(
)
+
(
)
+
×
(
3
1
)
(
0
)
=
1
−
√
(
3
2
)
(
+
1
)
=
(
)
×
0
.
99
(
3
3
)
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
p
er
f
o
r
m
a
n
ce
o
f
DL
F
with
AOA
an
d
E
L
U
b
ased
L
STM
m
eth
o
d
is
s
im
u
lated
in
MA
T
L
AB
2
0
2
0
b
.
T
h
e
s
y
s
tem
is
co
n
f
ig
u
r
ed
with
i5
p
r
o
ce
s
s
o
r
,
W
in
d
o
ws
1
0
(
6
4
b
it)
an
d
8
GB
R
A
M.
Per
f
o
r
m
an
ce
o
f
a
d
ev
elo
p
e
d
m
o
d
el
is
ev
alu
ated
in
ter
m
s
o
f
th
e
p
er
f
o
r
m
a
n
ce
m
etr
ics
o
f
ac
cu
r
ac
y
,
s
en
s
itiv
ity
,
s
p
ec
if
icity
,
p
r
ec
is
io
n
,
F
1
-
s
co
r
e
,
an
d
e
r
r
o
r
r
ate.
I
n
T
ab
le
2
,
p
er
f
o
r
m
a
n
ce
o
f
a
d
ev
e
lo
p
ed
DL
F
wi
th
AO
A
b
as
ed
h
y
p
er
p
ar
am
e
te
r
tu
n
in
g
p
r
o
ce
s
s
is
ev
al
u
a
ted
b
a
s
ed
o
n
v
a
r
io
u
s
p
er
f
o
r
m
an
c
e
m
e
tr
i
c
s
.
T
h
e
d
if
f
e
r
en
t
o
p
ti
m
i
za
tio
n
alg
o
r
i
th
m
s
u
ti
li
ze
d
to
ev
alu
at
e
th
e
p
e
r
f
o
r
m
an
c
e
o
f
DL
F
w
i
t
h
AO
A
ar
e
,
P
S
O,
an
t
co
lo
n
y
o
p
tim
iz
at
io
n
(
A
C
O)
,
g
r
ey
wo
lf
o
p
ti
m
i
za
tio
n
alg
o
r
it
h
m
(
G
OA
)
,
an
d
p
e
li
ca
n
o
p
tim
iz
at
io
n
a
lg
o
r
ith
m
(
P
OA)
.
T
h
e
D
L
F
wi
th
A
O
A
b
as
ed
h
y
p
er
p
ar
a
m
e
te
r
tu
n
in
g
p
r
o
ce
s
s
a
tt
ain
s
9
8
.
5
6
%
ac
cu
r
a
cy
,
9
7
.
5
4
%
s
e
n
s
i
tiv
i
ty
,
9
8
.
2
2
%
s
p
e
cif
ic
it
y
,
9
6
.
9
3
%
p
r
e
ci
s
io
n
,
9
6
.
3
3
%
F
1
-
s
co
r
e
,
a
n
d
1
.
4
4
er
r
o
r
r
a
te
.
T
h
e
D
L
F
i
s
in
co
r
p
o
r
at
ed
i
n
to
tr
ad
i
t
io
n
a
l
A
OA
to
en
h
a
n
ce
ex
p
lo
r
at
io
n
b
y
in
tr
o
d
u
c
in
g
r
an
d
o
m
p
e
r
tu
r
b
a
tio
n
s
b
a
s
ed
o
n
th
e
L
ev
y
f
li
g
h
t
d
i
s
tr
ib
u
t
io
n
,
wh
ich
im
p
r
o
v
es
g
lo
b
al
s
e
ar
c
h
ca
p
ab
i
li
ti
e
s
.
R
a
th
er
th
an
d
ir
e
ct
ly
u
p
d
a
tin
g
b
a
s
ed
o
n
ac
c
el
er
a
tio
n
,
DL
F
d
y
n
am
ica
ll
y
ad
ju
s
t
s
th
e
s
t
ep
s
iz
e
,
en
ab
lin
g
f
in
e
-
tu
n
in
g
o
f
h
y
p
er
p
ar
am
e
ter
s
.
I
n
T
ab
le
3
,
p
er
f
o
r
m
an
c
e
o
f
an
E
L
U
a
ct
i
v
at
io
n
f
u
n
c
t
io
n
i
s
ev
al
u
a
ted
w
ith
v
a
r
io
u
s
p
er
f
o
r
m
an
c
e
m
et
r
i
cs
.
T
h
e
d
if
f
er
e
n
t
ac
tiv
at
io
n
f
u
n
c
ti
o
n
s
u
ti
li
z
ed
to
ev
a
lu
a
te
th
e
p
er
f
o
r
m
an
c
e
o
f
E
L
U
in
c
lu
d
e
R
eL
U,
L
ea
k
y
R
e
L
U
,
T
an
,
an
d
S
in
.
T
h
e
E
L
U
a
ct
iv
at
io
n
f
u
n
c
t
io
n
ac
h
iev
e
s
9
8
.
5
6
%
a
cc
u
r
ac
y
,
9
7
.
5
4
%
s
e
n
s
it
iv
i
ty
,
9
8
.
2
2
%
s
p
ec
if
ici
ty
,
9
6
.
9
3
%
p
r
e
ci
s
io
n
,
9
6
.
3
3
%
F1
-
s
c
o
r
e,
an
d
1
.
4
4
o
f
er
r
o
r
r
at
e.
T
ab
le
2
.
Per
f
o
r
m
an
ce
o
f
d
ev
el
o
p
ed
DL
F with
AOA
m
eth
o
d
M
e
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
Er
r
o
r
r
a
t
e
IQ
-
O
TH
/
N
C
C
D
l
u
n
g
c
a
n
c
e
r
d
a
t
a
s
e
t
PSO
9
3
.
5
8
9
4
.
2
2
9
5
.
1
1
9
3
.
3
4
9
2
.
4
8
6
.
4
1
A
C
O
9
0
.
1
1
9
3
.
4
7
9
0
.
1
1
9
1
.
3
0
8
9
.
2
2
9
.
8
8
GOA
9
5
.
2
2
9
2
.
7
7
8
9
.
7
7
9
2
.
3
8
9
2
.
0
0
4
.
7
7
P
O
A
9
6
.
4
4
9
4
.
2
7
9
5
.
0
2
9
4
.
0
6
9
3
.
8
9
3
.
5
5
D
LF
-
AOA
9
8
.
5
6
9
7
.
5
4
9
8
.
2
2
9
6
.
9
3
9
6
.
3
3
1
.
4
4
C
h
e
st
C
T
-
s
c
a
n
i
ma
g
e
d
a
t
a
s
e
t
PSO
9
4
.
5
8
9
5
.
2
2
9
6
.
1
1
9
4
.
3
4
9
3
.
4
8
7
.
4
1
A
C
O
9
1
.
1
1
9
4
.
4
7
9
1
.
1
1
9
1
.
4
0
9
0
.
2
2
9
.
9
8
GOA
9
6
.
2
2
9
3
.
7
7
9
0
.
7
7
9
5
.
3
8
9
5
.
0
0
5
.
7
7
P
O
A
9
8
.
7
4
9
8
.
5
6
9
7
.
9
3
9
9
.
7
4
9
8
.
1
6
1
.
1
2
D
LF
-
AOA
9
9
.
4
6
9
9
.
1
4
9
8
.
2
2
9
9
.
2
3
9
8
.
8
3
1
.
7
4
I
n
T
ab
le
4
,
t
h
e
p
e
r
f
o
r
m
an
ce
o
f
L
STM
with
E
L
U
m
eth
o
d
is
v
alid
ated
b
y
v
ar
io
u
s
p
e
r
f
o
r
m
a
n
ce
m
etr
ics.
T
h
e
d
if
f
er
e
n
t
class
if
ier
s
u
s
ed
to
ev
alu
ate
th
e
p
e
r
f
o
r
m
a
n
ce
o
f
L
STM
with
E
L
U
in
clu
d
e
s
p
ar
s
e
au
to
-
en
co
d
er
(
Sp
a
r
s
e
AE
)
,
tr
a
d
itio
n
al
L
STM
,
n
eu
r
al
n
etwo
r
k
(
NN)
,
an
d
DNN.
T
h
e
L
ST
M
with
E
L
U
m
eth
o
d
ac
h
iev
ed
9
8
.
5
6
%
ac
c
u
r
ac
y
,
9
7
.
5
4
%
s
en
s
itiv
ity
,
9
8
.
2
2
%
s
p
e
cif
icity
,
9
6
.
9
3
%
p
r
ec
is
io
n
,
9
6
.
3
3
%
F1
-
s
co
r
e,
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
xp
o
n
en
tia
l l
o
n
g
s
h
o
r
t
-
term me
mo
r
y
w
ith
Levy
flig
h
t o
p
timiz
a
tio
n
fo
r
lu
n
g
n
o
d
u
le
…
(
K
a
lib
a
Go
w
th
a
mi
)
1459
an
er
r
o
r
r
ate
o
f
1
.
4
4
.
I
n
T
ab
le
5
,
p
e
r
f
o
r
m
a
n
c
e
o
f
a
L
ST
M
wit
h
E
L
U
m
et
h
o
d
is
v
al
i
d
at
e
d
u
s
in
g
v
a
r
i
o
u
s
p
e
r
f
o
r
m
a
n
c
e
m
e
tr
i
cs
wit
h
d
i
f
f
er
en
t k
-
f
o
l
d
v
al
u
es
,
r
a
n
g
in
g
f
r
o
m
k
=
3
to
k
=
7
.
T
h
e
L
S
T
M
wit
h
th
e
E
L
U
m
e
th
o
d
ac
h
ie
v
es
9
8
.
5
6
%
a
cc
u
r
ac
y
,
9
7
.
5
4
%
s
e
n
s
i
ti
v
it
y
,
9
8
.
2
2
%
s
p
ec
if
i
cit
y
,
9
6
.
9
3
%
p
r
ec
is
io
n
,
9
6
.
3
3
%
F
1
-
s
c
o
r
e
,
an
d
an
er
r
o
r
r
ate
o
f
1
.
4
4
at
a
k
v
al
u
e
o
f
5
.
T
ab
le
3
.
Per
f
o
r
m
an
ce
o
f
E
L
U
ac
tiv
atio
n
f
u
n
cti
o
n
M
e
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
Er
r
o
r
r
a
t
e
IQ
-
O
TH
/
N
C
C
D
l
u
n
g
c
a
n
c
e
r
d
a
t
a
s
e
t
R
e
LU
9
1
.
4
8
8
9
.
4
7
9
0
.
5
5
8
8
.
5
7
8
7
.
6
9
8
.
5
1
Le
a
k
y
R
e
LU
9
2
.
5
8
8
2
.
7
4
8
8
.
4
7
8
3
.
9
1
8
5
.
1
2
7
.
4
1
Ta
n
8
8
.
1
2
8
5
.
7
8
8
2
.
6
9
8
0
.
3
0
7
5
.
4
8
1
1
.
8
7
S
i
n
8
0
.
2
3
8
1
.
4
5
8
2
.
7
7
8
0
.
9
7
8
0
.
4
8
1
9
.
7
6
ELU
9
8
.
5
6
9
7
.
5
4
9
8
.
2
2
9
6
.
9
3
9
6
.
3
3
1
.
4
4
C
h
e
st
C
T
-
s
c
a
n
i
ma
g
e
d
a
t
a
s
e
t
R
e
LU
9
2
.
5
4
9
0
.
2
4
9
1
.
5
7
9
1
.
5
7
9
1
.
6
9
8
.
5
1
Le
a
k
y
R
e
LU
9
3
.
5
8
8
3
.
7
4
8
9
.
4
7
9
3
.
3
1
9
3
.
1
2
7
.
4
1
Ta
n
8
9
.
1
2
8
5
.
7
8
8
2
.
6
9
8
8
.
3
0
7
6
.
4
8
1
1
.
8
7
S
i
n
8
1
.
2
3
8
2
.
4
5
8
3
.
7
7
8
0
.
9
7
8
0
.
4
8
1
9
.
7
6
ELU
9
9
.
4
6
9
9
.
1
4
9
8
.
2
2
9
9
.
2
3
9
8
.
8
3
1
.
7
4
T
ab
le
4
.
Per
f
o
r
m
an
ce
o
f
L
ST
M
with
E
L
U
m
eth
o
d
M
e
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
Er
r
o
r
r
a
t
e
IQ
-
O
TH
/
N
C
C
D
l
u
n
g
c
a
n
c
e
r
d
a
t
a
s
e
t
S
p
a
r
se
A
E
6
2
.
5
8
5
9
.
5
7
6
3
.
5
8
6
5
.
8
3
7
3
.
5
7
3
7
.
4
1
LSTM
9
6
.
5
8
8
9
.
4
1
9
3
.
6
6
8
9
.
7
3
9
0
.
0
5
3
.
4
1
NN
7
5
.
6
4
7
2
.
4
8
7
7
.
1
5
7
1
.
7
2
7
0
.
9
9
2
4
.
3
5
DNN
8
5
.
4
7
8
0
.
5
6
8
2
.
4
8
7
7
.
9
3
7
5
.
4
8
1
4
.
5
2
LSTM
w
i
t
h
ELU
9
8
.
5
6
9
7
.
5
4
9
8
.
2
2
9
6
.
9
3
9
6
.
3
3
1
.
4
4
C
h
e
st
C
T
-
S
c
a
n
i
m
a
g
e
d
a
t
a
s
e
t
S
p
a
r
se
A
E
7
3
.
5
8
6
0
.
5
7
6
3
.
7
5
7
3
.
2
4
7
3
.
5
7
3
8
.
4
1
LSTM
9
7
.
5
8
9
0
.
4
1
9
2
.
6
6
9
6
.
7
3
9
6
.
0
5
3
.
5
6
NN
7
5
.
6
4
7
2
.
2
4
7
5
.
5
6
7
5
.
7
2
7
5
.
9
9
3
4
.
1
4
DNN
8
6
.
4
7
8
1
.
5
6
8
2
.
4
8
8
5
.
9
3
8
5
.
4
8
1
2
.
1
4
LSTM
w
i
t
h
ELU
9
9
.
4
6
9
9
.
1
4
9
8
.
2
2
9
9
.
2
3
9
8
.
8
3
1
.
7
4
T
ab
le
5
.
Per
f
o
r
m
an
ce
o
f
L
ST
M
with
E
L
U
m
eth
o
d
f
o
r
K
-
f
o
l
d
v
alu
es
K
-
v
a
l
u
e
s
A
c
c
u
r
a
c
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
Er
r
o
r
r
a
t
e
IQ
-
O
TH
/
N
C
C
D
l
u
n
g
c
a
n
c
e
r
d
a
t
a
s
e
t
K
=
3
9
5
.
6
8
9
4
.
8
7
9
5
.
8
7
9
2
.
6
2
9
0
.
4
8
4
.
3
1
K
=
4
9
6
.
9
8
9
5
.
7
8
9
6
.
8
7
9
4
.
0
9
9
2
.
4
5
3
.
0
1
K
=
5
9
8
.
5
6
9
7
.
5
4
9
8
.
2
2
9
6
.
9
3
9
6
.
3
3
1
.
4
4
K
=
6
9
7
.
5
6
9
3
.
4
8
9
4
.
5
8
9
2
.
5
0
9
1
.
5
4
2
.
4
3
K
=
7
9
6
.
4
8
9
0
.
4
8
9
2
.
8
4
9
0
.
1
3
8
9
.
7
7
3
.
5
1
C
h
e
st
C
T
-
s
c
a
n
i
ma
g
e
d
a
t
a
s
e
t
K
=
3
9
4
.
4
3
9
3
.
8
7
9
4
.
8
7
9
3
.
4
2
9
3
.
4
8
4
.
3
1
K
=
4
9
5
.
9
2
9
5
.
2
4
9
6
.
5
2
9
4
.
5
3
9
4
.
3
1
3
.
0
1
K
=
5
9
9
.
4
6
9
9
.
1
4
9
8
.
2
2
9
9
.
2
3
9
8
.
8
3
1
.
7
4
K
=
6
9
7
.
5
6
9
3
.
4
8
9
4
.
5
8
9
6
.
5
0
9
6
.
4
2
2
.
4
3
K
=
7
9
6
.
4
3
9
1
.
3
2
9
2
.
3
6
9
5
.
1
3
9
4
.
7
7
3
.
5
1
Fig
u
r
e
4
r
ep
r
esen
ts
a
co
n
f
u
s
io
n
m
atr
ix
f
o
r
I
Q
-
OT
H/NC
C
D
l
u
n
g
ca
n
ce
r
d
ataset.
Fig
u
r
e
5
r
ep
r
esen
ts
a
co
n
f
u
s
io
n
m
atr
ix
f
o
r
t
h
e
ch
est
C
T
-
s
ca
n
im
ag
e
d
ata
s
et.
Fig
u
r
e
6
r
e
p
r
esen
ts
a
r
ec
eiv
er
o
p
er
atin
g
ch
ar
ac
ter
is
tic
(
R
OC
)
cu
r
v
e
f
o
r
I
Q
-
OT
H/NC
C
D
lu
n
g
ca
n
ce
r
d
ataset,
wh
ile
Fig
u
r
e
7
r
e
p
r
e
s
en
ts
a
R
OC
cu
r
v
e
f
o
r
th
e
ch
est
C
T
-
s
ca
n
im
ag
e
d
ataset.
T
ab
le
6
s
h
o
ws
th
e
ev
alu
atio
n
o
f
co
m
p
u
tatio
n
tim
e
an
d
m
em
o
r
y
u
s
ag
e
f
o
r
t
h
e
p
r
o
p
o
s
ed
al
g
o
r
ith
m
.
T
ab
le
7
r
ep
r
esen
ts
th
e
cr
o
s
s
-
d
ataset
g
en
er
aliza
tio
n
with
tr
ain
i
n
g
o
n
IQ
-
OT
H/NC
C
D
lu
n
g
ca
n
ce
r
d
ataset
an
d
test
in
g
o
n
ch
est
C
T
-
s
ca
n
im
ag
e
d
ataset.
T
ab
le
8
r
ep
r
esen
ts
th
e
cr
o
s
s
-
d
ataset
g
en
er
aliza
tio
n
with
tr
ain
in
g
o
n
ch
est
C
T
-
s
c
an
im
ag
e
d
ataset
an
d
test
in
g
o
n
I
Q
-
OT
H/NC
C
D
lu
n
g
ca
n
ce
r
d
ataset.
T
ab
les
7
an
d
8
r
ep
r
esen
ts
th
e
r
eliab
ilit
y
o
f
p
r
o
p
o
s
ed
m
o
d
el
wh
en
an
aly
zin
g
u
n
s
ee
n
d
ata
f
r
o
m
d
if
f
er
en
t
clin
ical
o
r
ig
i
n
s
.
T
ab
le
7
r
ep
r
esen
ts
th
at
L
STM
with
E
L
U
m
o
d
el
o
b
tain
s
h
ig
h
ac
cu
r
ac
y
o
f
8
0
.
4
3
%
wh
e
n
tr
ain
ed
o
n
I
Q
-
OT
H/NC
C
D
lu
n
g
ca
n
ce
r
d
ata
s
et
an
d
test
ed
o
n
ch
est
C
T
d
ata.
T
ab
le
8
s
h
o
ws
h
ig
h
ac
cu
r
ac
y
o
f
8
1
.
7
3
%,
as
v
ice
-
v
er
s
a
o
f
T
a
b
le
7
,
s
h
o
ws
m
o
d
el’
s
ca
p
ab
ilit
y
to
ca
p
tu
r
e
p
a
th
o
lo
g
ical
m
ar
k
er
s
th
an
p
ar
ticu
lar
d
ataset
f
ea
tu
r
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
.
2
,
Ap
r
il
20
26
:
1
4
5
1
-
1
4
6
3
1460
Fig
u
r
e
4
.
C
o
n
f
u
s
io
n
m
atr
i
x
f
o
r
I
Q
-
OT
H/NC
C
D
lu
n
g
ca
n
ce
r
d
ataset
Fig
u
r
e
5
.
C
o
n
f
u
s
io
n
m
atr
i
x
f
o
r
c
h
est C
T
-
s
ca
n
im
ag
e
d
ataset
Fig
u
r
e
6
.
R
OC
cu
r
v
e
f
o
r
I
Q
-
O
T
H/NC
C
D
lu
n
g
ca
n
ce
r
d
ataset
Fig
u
r
e
7
.
R
OC
cu
r
v
e
f
o
r
ch
est
CT
-
s
ca
n
im
ag
e
d
ataset
T
ab
le
6
.
E
v
alu
atio
n
o
f
co
m
p
u
t
atio
n
tim
e
an
d
m
em
o
r
y
u
s
ag
e
M
e
t
h
o
d
s
D
a
t
a
s
e
t
C
o
m
p
u
t
a
t
i
o
n
t
i
m
e
(
mi
n
)
M
e
m
o
r
y
u
s
a
g
e
(
M
B
)
S
p
a
r
se
A
E
IQ
-
O
TH
/
N
C
C
D
29
2
4
0
LSTM
30
2
2
5
NN
25
2
0
7
DNN
27
1
8
3
LSTM
w
i
t
h
ELU
30
2
4
5
S
p
a
r
se
A
E
C
h
e
st
C
T
-
s
c
a
n
28
2
4
5
LSTM
35
2
3
0
NN
30
2
1
3
DNN
32
1
9
7
LSTM
w
i
t
h
ELU
35
2
5
0
T
ab
le
7
.
C
r
o
s
s
d
ataset
g
en
er
al
izatio
n
with
tr
ain
in
g
o
n
I
Q
-
O
T
H/NC
C
D
an
d
test
in
g
o
n
c
h
es
t CT
-
s
can
M
e
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
S
p
a
r
se
A
E
7
7
.
4
1
7
7
.
2
7
7
6
.
8
4
7
6
.
4
7
7
6
.
2
2
LSTM
7
7
.
6
3
7
7
.
4
8
7
7
.
0
9
7
6
.
8
2
7
6
.
6
2
NN
7
8
.
1
6
7
7
.
6
5
7
7
.
3
4
7
7
.
0
2
7
6
.
8
3
DNN
7
8
.
6
4
7
8
.
2
5
7
8
.
0
2
7
7
.
6
2
7
7
.
4
8
LSTM
w
i
t
h
ELU
8
0
.
4
3
7
9
.
4
1
7
9
.
2
3
7
9
.
0
3
7
8
.
8
2
T
ab
le
8
.
C
r
o
s
s
d
ataset
g
en
er
al
izatio
n
with
tr
ain
in
g
o
n
ch
est
CT
-
s
ca
n
an
d
test
in
g
o
n
I
Q
-
OT
H/NC
C
D
M
e
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
S
p
a
r
se
A
E
7
7
.
8
2
7
7
.
4
3
7
7
.
0
4
7
6
.
5
8
7
6
.
3
2
LSTM
7
8
.
0
9
7
7
.
8
6
7
7
.
6
2
7
7
.
2
9
7
7
.
0
2
NN
7
8
.
6
8
7
8
.
3
6
7
8
.
0
2
7
7
.
8
5
7
7
.
5
9
DNN
7
9
.
4
8
7
9
.
1
8
7
8
.
8
5
7
8
.
6
2
7
8
.
2
9
LSTM
w
i
t
h
ELU
8
1
.
7
3
8
0
.
6
5
8
0
.
2
8
7
9
.
8
4
7
9
.
5
2
Evaluation Warning : The document was created with Spire.PDF for Python.