I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
5
,
O
c
to
be
r
2025
, pp.
4235
~
4249
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
5
.pp
4235
-
4249
4235
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
G
r
i
d
gr
ap
h
c
on
vol
u
t
i
o
n
al
n
e
t
w
or
k
-
c
yc
l
i
c
al
l
e
ar
n
i
n
g r
at
e
E
f
f
i
c
i
e
n
t
N
e
t
f
or
l
i
ve
r
t
u
m
or
se
gm
e
n
t
at
i
on
c
l
ass
i
f
i
c
at
i
on
S
an
gi
N
ar
as
im
h
u
lu
, C
h
D
V
S
u
b
b
a R
ao
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
a
nd E
ngi
ne
e
r
i
ng, S
r
i
V
e
nka
t
e
s
w
a
r
a
U
ni
ve
r
s
i
t
y
C
ol
l
e
ge
of
E
ngi
ne
e
r
i
ng, T
i
r
upa
t
hi
,
I
ndi
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
J
a
n
4
,
2025
R
e
vi
s
e
d
J
un
29
,
2025
A
c
c
e
pt
e
d
J
ul
13
,
2025
Liver
tumors
are
identified
in
computed
tomograp
hy
(CT)
images,
wh
ich
are
crucial
for
accurate
disease
diagnosis
and
treatment
planning
as
they
enable
clear
delineation
of
tumors.
Hence,
it
is
vital
in
the
field
of
medical
radiology
to
segment
and
classify
CT
images
of
liv
er
tumor
s
effec
tively.
However,
liver
tumor
locations
are
not
captured
accurately
at
the
bou
ndaries
in
terms
of
size
and
depth
within
the
liver
due
to
downsampled
i
mages,
leading
to
reduce
d
segmentation
and
classifica
tion
results.
This
re
search
proposes
a
grid
-
graph
convolutional
network
-
based
cyclical
learnin
g
rate
EfficientNet
(GG
CN
-
CLREN)
to
accurately
segment
and
classify
liver
tumors.
GGCN
addresse
s
inaccur
ate
liver
tumor
segmentation
due
to
dow
n
sampled
images,
which
capture
spatial
relationships
effective
ly
and
preserve
tumor
boundaries
as
well
as
depth
information.
For
classifi
cation,
CLREN
optimizes
classification
by
adjusting
the
learning
rate,
which
enhances
convergence
and
accuracy.
Th
erefore,
GGCN
-
CLREN
e
nsures
enhanced
segmentat
ion
and
classifi
cation
by
addressin
g
size
and
depth
inaccur
acies
. Golden
sine gray wolf op
timiz
ation
(GSGWO) selects
th
e most
appropriat
e
features
effectively
.
The
GGCN
-
CLREN
achieves
comme
ndable
accuracies
of
99.80%
and
99.96%,
respectivel
y,
for
the
LiTS17
and
C
HAOS
datasets
when
compared
to
the
existing
techniques:
enhanced
swim
transforme
r
network
with
adversa
rial
propagatio
n
(APESTNet)
and
adding
inception modu
le
-
UNet (AIM
-
UNet).
K
e
y
w
o
r
d
s
:
C
om
put
e
d t
om
ogr
a
phy
C
yc
li
c
a
l
le
a
r
ni
ng r
a
te
E
f
f
ic
ie
nt
N
e
t
da
ta
a
ugm
e
nt
a
ti
on
G
r
i
d
-
g
r
a
p
h
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
R
e
s
N
e
xt
50
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
S
a
ngi
N
a
r
a
s
im
hul
u
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
S
c
ie
nc
e
a
nd
E
ngi
ne
e
r
in
g, S
r
i
V
e
nka
te
s
w
a
r
a
U
ni
ve
r
s
it
y C
ol
le
ge
of
E
ngi
ne
e
r
in
g
T
ir
upa
th
i,
I
ndi
a
E
m
a
il
:
na
r
a
s
im
hul
u.s
a
ngi
@
gm
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
L
iv
e
r
tu
m
or
s
a
r
e
one
of
th
e
p
r
im
a
r
y
a
nd
le
th
a
l
f
or
m
s
o
f
c
a
nc
e
r
a
ll
ove
r
th
e
w
or
ld
,
c
a
us
in
g
a
la
r
ge
num
be
r
of
de
a
th
s
e
ve
r
y
ye
a
r
.
P
r
im
a
r
y
li
ve
r
c
a
n
c
e
r
s
a
r
e
of
te
n
c
a
us
e
d
by
c
ir
r
hos
is
,
r
e
s
ul
ti
ng
f
r
om
he
pa
ti
ti
s
B
or
C
,
a
lc
ohol
c
ons
um
pt
io
n,
or
f
a
tt
y
l
iv
e
r
di
s
e
a
s
e
[
1]
.
N
um
e
r
ou
s
im
a
gi
ng
te
s
ts
,
s
uc
h
a
s
ul
tr
a
s
ound,
c
om
put
e
r
to
m
ogr
a
phy
(
C
T
)
,
a
nd
m
a
gn
e
ti
c
r
e
s
on
a
nc
e
im
a
gi
ng
(
M
R
I
)
,
a
s
s
is
t
in
di
a
gno
s
in
g
c
ir
r
hos
is
.
A
m
ong
th
e
s
e
,
C
T
is
th
e
pr
im
a
r
y
m
e
th
od
us
e
d
f
or
di
a
gnos
is
.
C
T
pr
ovi
de
s
c
om
pr
e
he
ns
iv
e
c
r
os
s
-
s
e
c
ti
ona
l
a
bdome
n
im
a
ge
s
th
a
t
e
na
bl
e
it
,
in
c
lu
s
iv
e
of
a
ll
te
s
ts
[
2]
,
[
3
]
.
T
hi
s
is
be
c
a
us
e
th
e
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
in
C
T
im
a
ge
s
he
lp
s
di
s
ti
ngui
s
h
th
e
tu
m
or
r
e
gi
on
f
r
om
li
ve
r
pa
r
e
nc
hym
a
[
4]
.
H
e
n
c
e
,
it
is
s
ig
ni
f
ic
a
nt
in
th
e
m
e
di
c
a
l
r
a
di
ol
ogy
f
ie
ld
f
or
s
e
gm
e
nt
in
g
C
T
im
a
ge
s
of
li
ve
r
tu
m
or
s
a
c
c
ur
a
te
ly
[
5
]
.
G
ly
c
oge
n
s
to
r
a
ge
,
r
e
gul
a
ti
on
of
hor
m
one
pr
oduc
ti
on,
a
nd
r
e
d
bl
ood
c
e
ll
(
R
B
C
)
de
gr
a
da
ti
on a
r
e
va
r
io
us
m
e
ta
bol
ic
pr
oc
e
s
s
e
s
c
a
r
r
ie
d
out
in
th
e
li
ve
r
[
6]
.
A
nnot
a
ti
ng
li
ve
r
tu
m
or
s
f
r
om
a
la
r
ge
num
be
r
of
a
bdomi
na
l
im
a
ge
s
is
ti
m
e
-
c
ons
um
in
g
a
nd
la
bor
io
us
,
r
e
qui
r
in
g
m
e
di
c
a
l
e
xpe
r
ti
s
e
.
M
or
e
ov
e
r
,
pa
r
ti
a
l
vol
um
e
e
f
f
e
c
t
a
n
d
lo
w
-
dos
e
a
r
ti
f
a
c
ts
in
m
e
di
c
a
l
im
a
gi
ng
m
a
ke
it
e
ve
n m
or
e
c
ha
ll
e
ngi
ng t
o de
li
ne
a
te
a
c
c
ur
a
te
l
e
s
io
n bound
a
r
ie
s
, r
e
s
ul
ti
ng i
n i
nt
r
a
-
r
a
te
r
va
r
ia
ti
ons
[
7]
, [
8
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
4235
-
4249
4236
S
e
gm
e
nt
a
ti
on
is
e
s
s
e
nt
ia
l
f
or
pos
t
-
in
te
r
ve
nt
io
na
l
tr
a
c
ki
ng
of
a
bl
a
te
d
li
ve
r
ti
s
s
ue
,
a
s
it
he
lp
s
a
s
s
e
s
s
ne
ga
ti
ve
ti
s
s
u
e
m
a
r
gi
ns
a
nd
a
ll
ow
s
c
li
ni
c
ia
n
s
to
e
v
a
lu
a
te
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
th
e
pr
oc
e
s
s
[
9]
.
A
ddi
ti
ona
ll
y,
li
ve
r
tu
m
or
s
e
gm
e
nt
a
ti
on
pe
r
m
it
s
s
tr
uc
tu
r
a
l
a
na
ly
s
e
s
,
li
ke
tu
m
or
vol
um
e
e
s
ti
m
a
ti
on,
w
hi
c
h
is
s
ig
ni
f
ic
a
nt
in
f
ol
lo
w
-
up
di
a
gnos
is
,
im
a
ge
-
dr
iv
e
n
s
ur
ge
r
y,
a
nd
th
e
r
a
py
[
10]
.
A
c
c
ur
a
te
s
e
gm
e
nt
a
ti
on
e
na
bl
e
s
th
e
e
va
lu
a
ti
on
of
vol
um
e
-
ba
s
e
d
qua
nt
it
a
ti
ve
da
ta
,
in
c
lu
di
ng
th
e
te
xt
u
r
a
l
f
e
a
tu
r
e
s
,
w
hi
c
h
he
lp
in
li
ve
r
th
e
r
a
py
pl
a
nni
ng
a
nd
ge
ne
r
a
te
a
m
or
e
c
ons
is
te
nt
he
pa
ti
c
tu
m
or
c
la
s
s
if
ic
a
ti
on,
th
e
r
a
pe
ut
ic
r
e
s
pons
e
c
la
s
s
if
ic
a
ti
on,
a
nd
pa
ti
e
nt
s
ur
vi
va
l
pr
e
di
c
ti
on
[
11]
.
T
he
c
a
te
gor
iz
a
ti
on
of
s
e
gm
e
nt
a
ti
on
te
c
hni
que
s
is
of
te
n
s
ubj
e
c
ti
ve
,
a
s
th
e
y
a
r
e
c
a
te
gor
iz
e
d
de
pe
ndi
ng
on
th
e
e
xt
e
nt
of
hum
a
n
in
te
r
ve
nt
io
n
or
m
e
th
odol
ogy.
T
he
m
e
th
odol
ogy
-
ba
s
e
d
c
la
s
s
if
ic
a
ti
on
in
c
lu
de
s
m
ode
l
-
ba
s
e
d
a
ppr
oa
c
he
s
ba
s
e
d
on
s
ta
ti
s
ti
c
a
l
s
ha
p
e
,
a
c
ti
ve
c
ont
our
s
,
gr
a
ph
c
ut
s
,
a
nd
r
e
gi
on
gr
ow
in
g
[
12]
,
[
13]
.
M
ode
l
-
ba
s
e
d
te
c
hni
que
s
a
r
e
in
c
li
ne
d
to
a
c
hi
e
ve
be
tt
e
r
s
e
gm
e
nt
a
ti
on
p
e
r
f
or
m
a
nc
e
th
a
n
in
te
ns
it
y
-
ba
s
e
d
te
c
hni
que
s
du
e
to
th
e
ir
m
a
th
e
m
a
ti
c
a
l
a
nd
a
c
c
ur
a
te
s
ta
ti
s
ti
c
a
l
m
ode
li
ng
th
a
t
c
a
pt
ur
e
s
th
e
r
e
gi
on
of
in
te
r
e
s
t
(
R
oI
)
[
14]
,
[
15]
.
H
ow
e
ve
r
,
th
e
lo
c
a
ti
ons
of
l
iv
e
r
tu
m
or
s
a
r
e
not
a
c
c
ur
a
te
ly
c
a
pt
ur
e
d
a
t
th
e
bounda
r
ie
s
f
or
s
iz
e
a
nd
de
pt
h
du
e
to
dow
ns
a
m
pl
e
d
im
a
ge
s
,
f
ur
th
e
r
le
a
di
ng
to
r
e
duc
e
d
s
e
gm
e
nt
a
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y.
T
o
ove
r
c
om
e
th
is
is
s
u
e
,
gr
id
-
gr
a
ph
c
onvolut
io
na
l
ne
twor
k
-
ba
s
e
d
c
yc
li
c
a
l
le
a
r
ni
ng
r
a
te
E
f
f
ic
ie
nt
N
e
t
(
G
G
C
N
-
C
L
R
E
N
)
is
pr
opos
e
d
to
a
c
c
ur
a
t
e
ly
s
e
gm
e
nt
a
nd
c
la
s
s
if
y
li
ve
r
tu
m
or
s
by
le
ve
r
a
gi
ng
gr
a
ph
c
onvolut
io
na
l
ne
twor
k
(
G
C
N
)
a
nd
dyna
m
ic
le
a
r
ni
n
g
r
a
te
a
dj
us
tm
e
nt
s
,
w
hi
c
h
e
n
s
ur
e
a
c
c
ur
a
te
de
li
ne
a
ti
on a
nd t
um
or
c
la
s
s
if
ic
a
ti
on.
T
he
m
a
in
c
ont
r
ib
ut
io
ns
t
o l
iv
e
r
t
um
or
s
e
gm
e
nt
a
ti
on a
r
e
e
xpl
a
in
e
d a
s
f
ol
lo
w
s
:
‒
G
G
C
N
le
a
r
ns
th
e
s
tr
uc
tu
r
a
l
da
ta
by
r
e
pr
e
s
e
nt
in
g
th
e
im
a
ge
a
s
a
gr
a
ph,
w
hi
c
h
m
ode
ls
th
e
r
e
la
ti
on
s
hi
ps
a
m
ong r
e
gi
ons
or
ne
ig
hbor
hood pixe
ls
m
or
e
e
f
f
e
c
ti
ve
ly
, l
e
a
di
ng t
o a
c
c
ur
a
te
s
e
gm
e
nt
a
ti
on.
‒
G
r
a
y
w
ol
f
opt
im
iz
a
ti
on
(
G
W
O
)
in
te
gr
a
te
s
th
e
e
f
f
e
c
ti
ve
e
xpl
o
r
a
ti
on
c
a
pa
bi
li
ti
e
s
w
it
h
s
e
a
r
c
h
di
v
e
r
s
it
y
e
nha
nc
e
m
e
nt
pr
ovi
de
d
by
th
e
gol
de
n
s
in
e
s
tr
a
te
gy
,
w
hi
c
h
a
s
s
is
t
s
in
n
a
vi
ga
ti
ng
th
e
s
e
a
r
c
h
s
p
a
c
e
e
f
f
e
c
ti
ve
ly
t
o i
de
nt
if
y a
ppr
opr
ia
te
f
e
a
tu
r
e
s
f
or
c
la
s
s
if
ic
a
ti
on.
‒
C
yc
li
c
a
l
le
a
r
ni
ng
r
a
te
(
C
L
R
)
a
dj
u
s
ts
th
e
le
a
r
ni
ng
r
a
te
dyn
a
m
ic
a
ll
y
dur
in
g
tr
a
in
in
g,
w
hi
c
h
f
ur
th
e
r
in
c
r
e
a
s
e
s
m
od
e
l
c
onve
r
ge
nc
e
a
nd
ge
ne
r
a
li
z
a
ti
on.
T
hi
s
a
ppr
oa
c
h
a
s
s
is
ts
E
f
f
ic
ie
nt
N
e
t
in
e
f
f
e
c
ti
ve
ly
le
a
r
ni
ng
c
om
pl
e
x
f
e
a
tu
r
e
s
,
w
hi
c
h
im
pr
ove
s
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y
a
nd
r
obus
tn
e
s
s
.
B
y
pe
r
f
or
m
in
g
a
ll
th
e
s
e
pr
oc
e
s
s
e
s
, t
he
pr
opos
e
d a
ppr
oa
c
h
a
c
hi
e
ve
s
be
tt
e
r
pe
r
f
or
m
a
nc
e
i
n l
iv
e
r
t
um
or
s
.
T
he
r
e
s
e
a
r
c
h
pa
p
e
r
is
or
ga
ni
s
e
d
a
s
f
ol
lo
w
s
:
s
e
c
ti
on
2
de
ta
il
s
th
e
li
te
r
a
tu
r
e
r
e
vi
e
w
of
e
xi
s
ti
ng
te
c
hni
que
s
.
S
e
c
ti
on
3
pr
e
s
e
nt
s
d
e
ta
il
e
d
in
f
or
m
a
ti
on
a
bout
th
e
p
r
opos
e
d
m
e
th
odol
ogy
.
S
e
c
ti
on
4
a
na
ly
z
e
s
th
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
of
th
e
e
xi
s
ti
ng
t
e
c
hni
que
s
a
nd
pr
opos
e
d
m
e
th
odol
ogy
.
S
e
c
ti
on
5
pr
ovi
de
s
th
e
c
onc
lu
s
io
n
of
t
hi
s
r
e
s
e
a
r
c
h pa
pe
r
.
2.
L
I
T
E
R
A
T
U
R
E
S
U
R
V
E
Y
T
he
r
e
la
te
d
w
or
ks
of
li
v
e
r
tu
m
or
s
e
gm
e
nt
a
ti
on
ba
s
e
d
on
C
T
im
a
ge
s
a
r
e
br
ie
f
ly
e
xpl
a
in
e
d
in
th
is
s
e
c
ti
on,
a
lo
ng
w
it
h
th
e
ir
be
ne
f
it
s
a
nd
li
m
it
a
ti
ons
.
T
he
s
e
m
e
th
ods
im
pr
ove
a
c
c
ur
a
c
y
by
c
a
pt
ur
in
g
de
ta
il
e
d
tu
m
or
bounda
r
ie
s
a
nd
le
ve
r
a
gi
ng
s
pa
ti
a
l
in
f
or
m
a
ti
on.
A
s
a
r
e
s
ul
t,
th
e
y
c
ont
r
ib
ut
e
to
i
m
pr
ove
d
di
a
gnos
is
a
nd
tr
e
a
tm
e
nt
pl
a
nni
ng i
n m
e
di
c
a
l
im
a
gi
ng.
W
a
ng
e
t
al
.
[
1
6]
s
ug
ge
s
te
d
a
n
E
f
f
i
c
i
e
n
tN
e
t
B
4
,
a
tt
e
nt
io
n
g
a
t
e
,
a
nd
r
e
s
id
u
a
l
l
e
a
r
ni
ng
(
E
A
R
-
U
N
e
t)
a
pp
r
o
a
c
h
to
a
tt
a
i
n
a
ut
o
m
a
ti
c
a
nd a
c
c
ur
a
t
e
s
e
gm
e
nt
a
t
io
n of
li
ve
r
t
um
or
s
.
I
ni
t
ia
ll
y
,
E
f
f
i
c
i
e
nt
B
4 w
a
s
pe
r
f
or
m
e
d a
s
th
e
e
n
c
o
de
r
f
or
e
xt
r
a
c
ti
n
g
m
or
e
f
e
a
t
ur
e
s
dur
i
ng
t
he
e
n
c
o
di
n
g
p
h
a
s
e
.
T
he
n,
a
n
a
tt
e
nt
io
n
ga
te
w
a
s
a
p
pl
i
e
d
i
n
th
e
s
k
ip
c
o
nn
e
c
ti
o
n
t
o
r
e
m
o
v
e
i
na
ppr
o
pr
i
a
t
e
r
e
g
io
n
s
a
nd
hi
g
hl
i
ght
s
pe
c
if
i
c
r
e
g
io
ns
.
A
t
l
a
s
t
,
d
e
c
od
e
r
c
onv
ol
u
ti
o
n
i
n
U
N
e
t
w
a
s
r
e
pl
a
c
e
d
w
i
th
a
r
e
s
id
ua
l
b
lo
c
k
to
r
e
d
uc
e
th
e
v
a
ni
s
hi
ng
gr
a
di
e
nt
i
s
s
u
e
,
w
hi
c
h
e
nh
a
n
c
e
d
t
h
e
c
on
ve
r
g
e
n
c
e
s
p
e
e
d.
H
o
w
e
ve
r
,
E
f
f
i
c
i
e
nt
N
e
t
B
4
s
tr
u
ggl
e
d
to
m
a
na
g
e
h
e
t
e
r
o
ge
ne
ou
s
t
um
or
t
e
xt
ur
e
s
du
e
t
o
th
e
va
r
i
a
b
il
i
ty
of
tu
m
or
a
p
pe
a
r
a
n
c
e
t
h
a
t
c
ha
ll
e
ng
e
d
t
he
m
o
de
l’
s
a
bi
li
t
y t
o
g
e
n
e
r
a
li
z
e
e
f
f
e
c
t
iv
e
ly
.
D
i
e
t
al
.
[
17
]
im
pl
e
m
e
nt
e
d
a
n
a
ut
om
a
te
d
a
ppr
oa
c
h
ba
s
e
d
on
hi
e
r
a
r
c
hi
c
a
l
it
e
r
a
ti
ve
s
upe
r
pi
xe
ls
a
nd
lo
c
a
l
s
ta
ti
s
ti
c
a
l
f
e
a
tu
r
e
s
to
s
e
gm
e
nt
li
ve
r
tu
m
or
s
.
I
ni
ti
a
ll
y,
3D
U
N
e
t
w
a
s
us
e
d
f
or
e
xt
r
a
c
ti
ng
li
ve
r
r
e
gi
ons
,
a
nd
a
hi
e
r
a
r
c
hi
c
a
l
s
upe
r
pi
xe
l
a
ppr
oa
c
h
w
a
s
a
ppl
ie
d
to
de
te
c
t
tu
m
or
bounda
r
ie
s
a
c
c
ur
a
te
ly
.
E
a
c
h
pi
xe
l
in
th
e
li
ve
r
r
e
gi
on
w
a
s
th
e
n
c
a
te
gor
iz
e
d
in
to
non
-
tu
m
or
or
tu
m
or
us
in
g
a
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
(
S
V
M
)
.
A
E
uc
li
de
a
n
di
s
ta
nc
e
vot
in
g
a
ppr
oa
c
h
w
a
s
e
m
pl
oye
d,
w
hi
c
h
in
c
or
por
a
te
d
s
upe
r
pi
xe
l
s
e
gm
e
nt
a
ti
on
a
nd
pi
xe
l
-
w
is
e
c
la
s
s
if
ic
a
ti
on
to
e
f
f
e
c
ti
ve
ly
a
nd
a
ut
om
a
ti
c
a
ll
y
id
e
nt
if
y
tu
m
or
r
e
gi
ons
.
N
e
ve
r
th
e
le
s
s
,
th
is
a
ppr
oa
c
h
di
d
not
c
a
pt
ur
e
c
om
pl
e
x
s
pa
ti
a
l
r
e
la
ti
ons
hi
p
s
a
nd
va
r
ia
ti
ons
w
it
hi
n
tu
m
or
s
,
a
s
it
r
e
li
e
d
on
pr
e
de
f
in
e
d
lo
c
a
l
s
tr
uc
tu
r
e
s
,
r
e
s
ul
ti
ng i
n i
na
c
c
ur
a
te
r
e
s
ul
ts
.
M
a
nj
u
na
th
a
n
d
K
w
a
di
ki
[
1
8]
p
r
e
s
e
n
te
d
a
m
o
di
f
i
e
d
r
e
s
i
d
ua
l
U
N
e
t
(
R
e
s
U
N
e
t
)
b
a
s
e
d
o
n
a
c
onv
ol
ut
io
n
a
l
ne
ur
a
l
ne
t
w
o
r
k
(
C
N
N
)
to
s
e
g
m
e
n
t
th
e
l
iv
e
r
f
r
om
C
T
i
m
a
ge
s
a
n
d
le
s
i
ons
f
r
o
m
s
e
gm
e
nt
e
d
le
ve
r
p
o
r
t
io
ns
.
I
n
th
e
pr
e
-
pr
oc
e
s
s
i
ng
pha
s
e
,
e
a
c
h
i
m
a
g
e
w
a
s
r
e
s
iz
e
d,
a
nd
a
no
r
m
a
l
iz
a
t
io
n
te
c
hn
iq
ue
w
a
s
a
p
pl
ie
d
to
e
ve
r
y
i
m
a
g
e
to
ob
ta
in
a
va
l
ue
be
t
w
e
e
n
z
e
r
o
a
nd
one
.
T
h
e
pr
e
s
e
nt
e
d
a
pp
r
oa
c
h
r
e
p
r
e
s
e
nt
e
d
th
e
a
bi
li
ty
t
o
s
e
g
m
e
n
t
t
he
l
iv
e
r
a
c
c
u
r
a
t
e
ly
by
a
ut
om
a
te
d
t
um
o
r
s
e
gm
e
nt
a
ti
on.
H
ow
e
ve
r
,
th
e
m
od
i
f
ie
d
R
e
s
U
N
e
t
s
t
r
u
gg
le
d
w
i
th
s
ig
ni
f
ic
a
nt
i
nt
e
r
-
pa
ti
e
nt
va
r
ia
bi
li
ty
in
li
ve
r
s
ha
pe
a
nd
le
s
i
on
a
ppe
a
r
a
nc
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
G
r
id
gr
aph c
onv
ol
ut
io
nal
ne
tw
o
r
k
-
c
y
c
li
c
al
l
e
a
r
ni
ng r
at
e
E
ff
ic
ie
nt
N
e
t
fo
r
l
iv
e
r
t
um
o
r
…
(
Sangi N
a
r
as
imhulu)
4237
be
c
a
us
e
i
ts
le
a
r
ne
d
f
e
a
tu
r
e
s
di
d
n
ot
ge
ne
r
a
l
iz
e
w
e
l
l
a
c
r
o
s
s
di
f
f
e
r
e
n
t
p
a
th
ol
og
ie
s
,
w
h
ic
h
r
e
s
u
lt
e
d
in
in
a
c
c
ur
a
te
s
e
g
m
e
n
ta
t
io
n.
L
v
e
t
al
.
[
19]
in
tr
oduc
e
d
de
e
p
s
upe
r
vi
s
io
n
(
D
S
)
a
nd
a
tr
ous
in
c
e
pt
io
n
(
A
I
)
w
it
h
c
ondi
ti
ona
l
r
a
ndom
f
ie
ld
(
C
R
F
)
to
s
e
gm
e
nt
li
ve
r
r
e
gi
ons
.
I
ni
ti
a
ll
y,
th
e
e
nc
ode
r
’
s
s
ta
nda
r
d
c
onvolut
io
n
w
a
s
r
e
pl
a
c
e
d
by
a
r
e
s
id
ua
l
bl
oc
k,
w
hi
c
h
e
nha
nc
e
d
th
e
n
e
twor
k’
s
de
pt
h.
T
he
n,
th
e
A
I
a
ppr
oa
c
h
w
a
s
a
ppl
ie
d
to
in
te
r
c
onne
c
t
th
e
e
nc
ode
r
a
nd de
c
ode
r
bl
oc
ks
, w
hi
c
h e
na
bl
e
d t
o obta
in
t
he
m
ul
ti
-
s
c
a
le
f
e
a
tu
r
e
s
. T
he
C
R
F
w
a
s
pe
r
f
or
m
e
d t
o e
nha
nc
e
t
he
a
dj
a
c
e
nt
da
ta
’
s
la
b
e
li
ng
de
vi
a
ti
on,
w
hi
c
h
le
d
to
th
e
r
e
f
in
e
m
e
nt
of
ove
r
a
ll
li
ve
r
bounda
r
ie
s
.
H
ow
e
ve
r
,
th
is
a
ppr
oa
c
h r
e
s
ul
te
d i
n ove
r
or
unde
r
-
s
e
gm
e
nt
a
ti
on e
r
r
or
s
f
or
l
e
ve
l
tu
m
or
s
s
it
ua
te
d a
t
th
e
bounda
r
ie
s
due
t
o t
he
ir
c
om
pl
e
x s
pa
ti
a
l
r
e
la
ti
ons
hi
ps
a
nd va
r
ia
ti
ons
.
P
ope
s
c
u
e
t
al
.
[
20]
de
ve
lo
p
e
d
a
dva
nc
e
d
C
N
N
te
c
hni
que
s
to
s
e
g
m
e
nt
th
e
li
ve
r
a
nd
he
pa
ti
c
tu
m
or
s
by
in
c
or
por
a
ti
ng
f
our
e
f
f
e
c
ti
ve
ne
ur
a
l
ne
twor
ks
li
ke
R
e
s
N
e
X
t1
01,
R
e
s
N
e
t1
52,
D
e
ns
e
N
e
t2
01,
a
nd
I
nc
e
pt
io
nV
3.
G
lo
ba
l
s
e
gm
e
nt
a
ti
on
w
a
s
pe
r
f
or
m
e
d
by
tr
a
in
in
g
s
e
pa
r
a
te
i
ndi
vi
dua
l
c
la
s
s
if
ie
r
s
a
nd
th
e
n
in
te
gr
a
ti
ng
it
s
de
c
is
io
n
in
to
a
uni
f
ie
d
s
ys
te
m
.
T
he
im
a
ge
s
unde
r
w
e
nt
a
pos
t
-
pr
oc
e
s
s
in
g
pr
oc
e
s
s
th
a
t
e
f
f
e
c
ti
ve
ly
r
e
m
ove
d
a
r
ti
f
a
c
ts
a
f
te
r
s
e
gm
e
nt
a
ti
on
ba
s
e
d
on
th
e
ne
ur
a
l
ne
twor
ks
.
H
ow
e
ve
r
,
th
e
C
N
N
s
uf
f
e
r
e
d
f
r
om
li
m
it
e
d
ge
ne
r
a
li
z
a
ti
on
a
c
r
os
s
di
ve
r
s
e
im
a
gi
ng
c
ondi
ti
on
s
be
c
a
u
s
e
of
va
r
ia
ti
ons
in
im
a
ge
qua
li
ty
a
nd
a
n
a
to
m
ic
a
l
di
f
f
e
r
e
nc
e
s
a
m
ong pa
ti
e
nt
s
.
B
a
la
s
ubr
a
m
a
ni
a
n
e
t
al
.
[
21]
pr
e
s
e
nt
e
d
a
n
e
nha
nc
e
d
s
w
im
T
r
a
ns
f
or
m
e
r
ne
twor
k
w
it
h
a
dv
e
r
s
a
r
ia
l
pr
opa
ga
ti
on
(
A
P
E
S
T
N
e
t)
to
s
e
gm
e
nt
a
nd
c
la
s
s
if
y
li
ve
r
tu
m
or
s
.
M
e
di
um
f
il
te
r
in
g
a
nd
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
w
e
r
e
ut
il
iz
e
d i
n t
he
pr
e
-
pr
oc
e
s
s
in
g pha
s
e
, w
hi
c
h i
m
pr
ove
d t
he
i
nput
i
m
a
ge
s
. T
he
e
nha
nc
e
d m
a
s
k r
e
gi
on
C
N
N
(R
-
C
N
N
)
w
a
s
a
ppl
ie
d
f
or
s
e
gm
e
nt
in
g
th
e
li
ve
r
tu
m
or
,
a
nd
A
P
E
S
T
N
e
t
w
a
s
e
m
pl
oye
d
to
c
a
te
gor
iz
e
th
e
li
ve
r
tu
m
or
.
O
ve
r
f
it
ti
ng
pr
obl
e
m
s
w
e
r
e
s
ol
ve
d
us
in
g
th
e
s
w
im
T
r
a
ns
f
or
m
e
r
m
ode
l
by
c
ons
tr
uc
ti
ng
a
dve
r
s
a
r
ia
l
pr
opa
ga
ti
on
in
th
e
c
l
a
s
s
if
ie
r
.
H
ow
e
ve
r
,
A
P
E
S
T
N
e
t
f
a
c
e
d
c
ha
ll
e
nge
s
in
a
da
pt
in
g
to
di
f
f
e
r
e
nt
tu
m
or
a
ppe
a
r
a
nc
e
s
a
nd
a
na
to
m
ic
a
l
va
r
ia
ti
ons
,
a
s
a
r
e
s
ul
t
of
c
a
p
tu
r
in
g
s
ubt
le
f
e
a
tu
r
e
s
s
pe
c
if
ic
to
di
f
f
e
r
e
nt
tu
m
or
t
ype
s
.
K
ol
li
e
t
al
.
[
22]
i
m
pl
e
m
e
nt
e
d
a
n
im
pr
ove
d
pr
oba
bi
li
s
ti
c
ne
ur
a
l
ne
twor
k
a
nd
B
a
ye
s
ia
n
opt
im
iz
a
ti
on
(
I
P
N
N
-
B
O
)
t
o
c
la
s
s
if
y l
iv
e
r
t
um
or
s
. T
he
f
ul
ly
a
ut
om
a
te
d a
ppr
o
a
c
h w
a
s
us
e
d t
o s
e
pa
r
a
t
e
m
a
li
gna
nc
ie
s
a
nd t
he
li
ve
r
f
r
om
C
T
s
c
a
ns
.
O
pt
im
a
l
hype
r
pa
r
a
m
e
te
r
tu
ni
ng
w
a
s
a
ut
o
m
a
ti
c
a
ll
y
a
ppl
ie
d
by
ut
il
iz
in
g
B
O
w
it
h
I
P
N
N
a
ppr
oa
c
h,
w
hi
c
h
a
c
hi
e
ve
d
a
c
c
ur
a
te
s
e
gm
e
nt
a
ti
on
a
nd
c
la
s
s
i
f
ic
a
ti
on
r
e
s
ul
ts
.
H
ow
e
v
e
r
,
th
e
im
pl
e
m
e
nt
e
d
a
ppr
oa
c
h
f
a
c
e
d
s
tr
uggl
e
s
w
it
h
in
tr
ic
a
te
in
te
r
a
c
ti
on
s
a
m
ong
tu
m
or
c
ha
r
a
c
te
r
is
ti
c
s
du
e
to
c
om
pl
e
x
a
nd
non
-
li
ne
a
r
r
e
la
ti
ons
hi
ps
a
m
ong va
r
io
us
f
e
a
tu
r
e
s
.
Ö
z
c
a
n
e
t
al
.
[
23]
s
ugge
s
te
d
a
ddi
ng
in
c
e
pt
io
n
m
odul
e
-
U
N
e
t
(
A
I
M
-
U
N
e
t)
by
in
te
gr
a
ti
ng
U
N
e
t
a
nd
I
nc
e
pt
io
nV
3
to
s
e
gm
e
nt
li
ve
r
tu
m
or
s
.
D
a
ta
a
ugm
e
nt
a
ti
on,
im
a
ge
r
ot
a
ti
on,
r
e
s
iz
in
g,
a
nd
s
li
c
in
g
w
e
r
e
ut
il
iz
e
d
in
th
e
pr
e
-
pr
oc
e
s
s
in
g
s
ta
ge
to
in
c
r
e
a
s
e
th
e
d
a
ta
s
e
t
s
iz
e
a
nd
r
e
s
iz
e
th
e
im
a
ge
s
.
A
I
M
-
U
N
e
t
w
a
s
de
v
e
lo
pe
d
by
pl
a
c
in
g
c
onvolut
io
na
l
la
ye
r
s
of
va
r
io
us
f
il
te
r
s
iz
e
s
on
a
s
ki
p
c
onne
c
ti
on.
T
he
s
ugg
e
s
te
d
a
ppr
oa
c
h
pr
ovi
de
d
be
tt
e
r
pe
r
f
or
m
a
nc
e
s
by
pr
oc
e
s
s
in
g
e
dge
da
ta
a
nd
m
or
phol
ogy
f
e
a
tu
r
e
s
to
a
gr
e
a
te
r
e
xt
e
nt
.
H
ow
e
ve
r
,
A
I
M
-
U
N
e
t
s
uf
f
e
r
e
d
f
r
om
in
c
r
e
a
s
e
d
m
e
m
or
y
us
a
g
e
a
nd
pot
e
nt
ia
l
ov
e
r
f
it
ti
ng
due
to
a
l
a
r
ge
num
be
r
of
p
a
r
a
m
e
te
r
s
in
tr
oduc
e
d by I
nc
e
pt
io
n m
odul
e
s
, w
hi
c
h i
m
pa
c
te
d t
he
m
ode
l’
s
pe
r
f
or
m
a
nc
e
.
X
ia
e
t
al
.
[
24
]
de
ve
lo
pe
d a
m
ul
ti
vi
e
w
in
f
or
m
a
ti
on
f
us
io
n a
nd C
R
F
t
o s
e
gm
e
nt
l
iv
e
r
t
um
or
s
. I
n
it
ia
ll
y,
th
e
dua
l
s
e
lf
-
a
tt
e
nt
io
n
(
D
S
A
)
a
ppr
oa
c
h
w
a
s
e
m
pl
oye
d
to
de
t
e
r
m
in
e
th
e
s
ig
ni
f
ic
a
nt
s
pa
ti
a
l
s
tr
uc
tu
r
e
s
a
nd
pa
tt
e
r
ns
,
a
s
w
e
ll
a
s
c
a
pt
ur
e
r
e
la
ti
on
s
hi
ps
a
m
ong
va
r
io
us
f
e
a
tu
r
e
di
m
e
ns
io
ns
a
nd
c
ha
nne
l
s
.
A
li
ght
w
e
ig
ht
3D
ne
twor
k
w
a
s
c
ons
tr
uc
te
d
to
c
om
bi
ne
s
e
gm
e
nt
a
ti
on
r
e
s
ul
ts
f
r
om
di
f
f
e
r
e
nt
vi
e
w
s
a
nd
pr
oduc
e
a
3D
out
c
om
e
.
A
t
la
s
t,
C
R
F
w
a
s
ge
n
e
r
a
te
d
f
or
3D
s
e
gm
e
nt
a
ti
on
r
e
f
in
e
m
e
nt
,
w
hi
c
h
e
li
m
in
a
te
d
ove
r
-
s
e
gm
e
nt
e
d
e
r
r
or
s
a
nd
e
nha
nc
e
d
s
e
gm
e
nt
a
ti
on
a
c
c
ur
a
c
y.
N
e
ve
r
th
e
le
s
s
,
th
e
de
v
e
lo
pe
d
a
ppr
oa
c
h
w
a
s
a
ps
e
udo
-
3D
te
c
hni
que
th
a
t
e
xt
r
a
c
te
d 3D
f
e
a
tu
r
e
da
ta
by c
om
bi
ni
ng 2D s
e
gm
e
nt
a
ti
on outc
o
m
e
s
f
r
om
va
r
io
us
pe
r
s
pe
c
ti
ve
s
, w
hi
c
h r
e
s
ul
te
d
in
t
he
l
os
s
of
c
e
r
ta
in
s
ubt
le
pa
tt
e
r
ns
.
X
ie
e
t
al
.
[
25]
in
tr
oduc
e
d
a
m
ul
ti
-
s
c
a
le
c
ont
e
xt
in
te
gr
a
ti
on
ne
twor
k
(
M
C
I
-
N
e
t)
to
s
e
gm
e
nt
li
ve
r
im
a
ge
s
.
T
he
r
e
s
id
ua
l
a
ppr
oa
c
h
w
a
s
c
on
s
tr
uc
te
d
to
a
voi
d
ne
twor
k
de
gr
a
da
ti
on.
T
he
m
ul
ti
-
s
c
a
le
c
ont
e
xt
e
xt
r
a
c
ti
on
m
odul
e
w
a
s
de
pl
oye
d
by
in
te
gr
a
ti
ng
hyb
r
id
di
la
te
d
c
onvolut
io
ns
to
c
a
pt
ur
e
de
e
pe
r
a
nd
b
r
oa
de
r
f
e
a
tu
r
e
s
a
t
di
f
f
e
r
e
nt
s
c
a
le
s
.
A
bound
a
r
y
c
or
r
e
c
ti
on
bl
oc
k
w
a
s
ge
ne
r
a
te
d,
w
hi
c
h
e
nha
nc
e
d
th
e
lo
c
a
li
z
a
ti
on
c
a
pa
bi
li
ty
of
bounda
r
y
in
f
or
m
a
ti
on.
H
ow
e
ve
r
,
th
e
2D
C
N
N
w
a
s
ut
il
iz
e
d
f
or
s
e
gm
e
nt
in
g
3D
m
e
di
c
a
l
im
a
ge
s
,
w
hi
c
h l
e
d t
o t
he
l
os
s
of
s
p
a
ti
a
l
da
ta
, t
he
r
e
by a
f
f
e
c
ti
ng s
e
gm
e
nt
a
ti
on a
c
c
ur
a
c
y.
K
ha
n
e
t
al
.
[
26]
pr
e
s
e
nt
e
d
a
r
e
s
id
ua
l
m
ul
ti
-
s
c
a
l
e
U
N
e
t
(
R
M
S
-
U
N
e
t)
to
s
e
gm
e
nt
th
e
li
ve
r
a
nd
le
s
io
n
e
f
f
e
c
ti
ve
ly
.
I
ns
te
a
d
of
u
s
in
g
va
r
io
us
ke
r
ne
l
s
iz
e
s
,
a
m
ul
ti
-
s
c
a
le
c
ont
e
xt
la
ye
r
w
it
h
di
f
f
e
r
e
nt
di
la
ti
on
r
a
te
s
w
a
s
a
ppl
ie
d
to
e
nha
nc
e
th
e
uni
que
a
nd
va
lu
a
bl
e
d
a
ta
f
r
om
e
ve
r
y
la
ye
r
.
R
e
s
id
ua
l
bl
oc
k
s
w
e
r
e
in
c
lu
de
d
to
c
om
pe
ns
a
te
f
or
t
r
a
in
in
g l
os
s
be
c
a
us
e
of
t
he
i
nc
r
e
a
s
e
i
n t
he
a
m
o
unt
of
c
onvolut
io
n l
a
ye
r
s
. B
a
tc
h nor
m
a
li
z
a
ti
on
w
a
s
a
ppl
ie
d
in
R
M
S
-
U
N
e
t
to
e
nha
nc
e
le
a
r
ni
ng
w
it
hout
a
ny
lo
s
s
of
va
lu
a
bl
e
in
f
or
m
a
ti
on.
H
ow
e
ve
r
,
R
M
S
-
U
N
e
t
s
tr
uggl
e
d
w
it
h
tr
a
in
in
g
s
ta
bi
li
ty
due
to
c
om
pl
e
x
in
te
r
a
c
ti
ons
a
m
ong
r
e
s
id
ua
l
a
nd
m
ul
ti
-
s
c
a
l
e
c
om
pone
nt
s
, w
hi
c
h t
r
ig
ge
r
e
d c
onve
r
ge
nc
e
pr
obl
e
m
s
.
K
us
hn
u
r
e
e
t
al
.
[
2
7
]
s
ug
ge
s
te
d
a
l
ig
h
twe
ig
ht
m
u
lt
i
-
le
v
e
l
m
ul
ti
s
c
a
le
ne
tw
o
r
k
w
it
h
a
de
e
p
r
e
s
i
dua
l
a
pp
r
o
a
c
h
(
L
iM
-
N
e
t
)
to
s
e
g
m
e
n
t
th
e
l
iv
e
r
t
um
or
us
in
g
C
T
im
a
ge
s
.
T
he
c
o
m
p
ut
a
ti
ona
ll
y
e
f
f
e
c
ti
v
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
4235
-
4249
4238
pr
e
-
a
c
ti
va
t
e
d
R
e
s
2
N
e
t
w
it
h
c
ha
nne
l
-
w
is
e
a
t
te
n
ti
o
n
(
P
A
R
C
A
)
w
a
s
e
m
pl
oye
d
in
th
e
U
N
e
t
f
o
r
e
x
t
r
a
c
t
in
g
m
ul
ti
-
s
c
a
le
f
in
e
-
g
r
a
in
e
d
f
e
a
tu
r
e
s
to
e
n
ha
n
c
e
de
c
o
de
r
c
o
m
pe
te
nc
e
.
T
he
ne
tw
o
r
k
w
a
s
op
ti
m
iz
e
d
by
ut
i
li
z
i
ng
a
c
us
t
om
l
os
s
f
u
nc
t
io
n
th
a
t
e
f
f
e
c
t
iv
e
ly
m
a
na
ge
d
c
la
s
s
im
ba
l
a
nc
e
s
.
N
on
e
th
e
le
s
s
,
L
iM
-
N
e
t
e
xhi
bi
te
d
i
na
c
c
ur
a
te
s
e
gm
e
nt
a
t
io
n
be
c
a
us
e
it
s
l
ig
ht
w
e
ig
ht
de
s
ig
n
di
d
n
ot
c
a
p
tu
r
e
th
e
c
o
m
p
le
x
tu
m
o
r
in
f
o
r
m
a
ti
on
e
f
f
e
c
t
iv
e
ly
.
O
u
e
t
al
.
[
28]
im
pl
e
m
e
nt
e
d
a
R
e
s
T
r
a
ns
U
N
e
t
th
a
t
in
te
gr
a
te
d
U
N
e
t
a
nd
T
r
a
ns
f
or
m
e
r
f
or
li
ve
r
tu
m
o
r
s
e
gm
e
nt
a
ti
on.
T
he
s
qu
e
e
z
e
a
nd
e
x
c
it
a
ti
on
(
S
E
)
bl
oc
k
w
e
r
e
in
c
or
por
a
te
d
in
U
N
e
t
to
c
a
pt
ur
e
m
or
e
s
a
li
e
nt
im
a
ge
f
e
a
tu
r
e
s
.
A
dua
l
-
pa
th
a
ppr
oa
c
h
w
a
s
e
m
pl
oye
d
in
th
e
e
nc
odi
ng
s
tr
uc
tu
r
e
,
w
he
r
e
f
e
a
tu
r
e
s
w
e
r
e
s
e
pa
r
a
te
ly
e
xt
r
a
c
te
d
ut
il
iz
in
g
bot
h
C
N
N
a
nd
T
r
a
n
s
f
or
m
e
r
.
A
n
e
f
f
ic
ie
nt
f
e
a
tu
r
e
e
nha
nc
e
m
e
nt
uni
t
w
a
s
c
ons
tr
uc
te
d
f
or
tr
a
ns
f
e
r
r
in
g
th
e
gl
oba
l
e
xt
r
a
c
te
d
f
e
a
tu
r
e
s
by
t
he
T
r
a
n
s
f
or
m
e
r
to
C
N
N
,
w
hi
c
h
e
nha
nc
e
d
it
s
pe
r
f
or
m
a
nc
e
.
N
e
ve
r
th
e
le
s
s
,
R
e
s
T
r
a
ns
U
N
e
t
ha
d
di
f
f
ic
ul
ti
e
s
w
it
h
tr
a
in
in
g
s
ta
bi
li
ty
a
nd
c
onve
r
ge
n
c
e
du
e
to
th
e
in
te
gr
a
ti
on
of
c
onvolut
io
n
a
nd
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
s
,
w
hi
c
h
s
tr
uggl
e
d
to
e
f
f
e
c
ti
ve
ly
le
a
r
n
a
nd
opt
im
iz
e
it
s
pa
r
a
m
e
te
r
s
.
F
r
om
th
e
ove
r
a
ll
a
na
ly
s
is
,
th
e
e
xi
s
ti
ng
m
e
th
ods
a
r
e
s
e
e
n
to
ha
ve
li
m
it
a
ti
ons
a
s
f
ol
lo
w
s
:
ove
r
or
unde
r
-
s
e
gm
e
nt
a
ti
on
e
r
r
or
s
,
s
tr
uggl
e
s
in
ge
ne
r
a
li
z
a
ti
on
a
bi
li
ty
,
i
nt
e
r
pr
e
ta
bi
li
ty
,
a
nd
tr
a
ns
pa
r
e
nc
y.
M
or
e
ove
r
,
it
doe
s
not
c
a
pt
ur
e
a
c
c
ur
a
te
bound
a
r
ie
s
in
te
r
m
s
of
s
iz
e
a
nd
d
e
pt
h.
I
n
or
de
r
to
a
ddr
e
s
s
th
i
s
is
s
ue
,
G
G
C
N
-
C
L
R
E
f
f
ic
ie
nt
N
e
t
is
pr
opos
e
d t
o
a
c
c
ur
a
te
ly
s
e
gm
e
nt
a
nd c
la
s
s
if
y l
iv
e
r
t
um
or
s
by
us
in
g
G
C
N
a
nd dyna
m
ic
l
e
a
r
ni
ng
r
a
te
a
dj
us
tm
e
nt
s
, w
hi
c
h e
n
a
bl
e
a
c
c
ur
a
te
de
li
ne
a
ti
on a
nd t
um
or
c
la
s
s
if
ic
a
ti
on.
3.
P
R
O
P
O
S
E
D
M
E
T
H
O
D
O
L
O
G
Y
I
n
th
is
r
e
s
e
a
r
c
h,
G
G
C
N
-
C
L
R
E
N
is
p
r
opos
e
d
f
or
s
e
gm
e
nt
in
g
a
nd
c
la
s
s
if
yi
ng
li
ve
r
tu
m
or
s
f
r
om
C
T
im
a
ge
s
.
L
iT
S
17
a
nd
C
H
A
O
S
a
r
e
th
e
two
s
ta
nd
a
r
d
be
nc
hm
a
r
k
da
ta
s
e
ts
us
e
d
to
d
e
te
r
m
in
e
th
e
pe
r
f
or
m
a
nc
e
of
th
e
pr
opos
e
d
a
ppr
oa
c
h.
A
m
e
di
a
n
f
il
te
r
a
nd
d
a
ta
a
ugm
e
nt
a
t
io
n
a
r
e
us
e
d
in
th
e
pr
e
-
pr
oc
e
s
s
in
g
pha
s
e
to
r
e
m
ove
noi
s
e
a
nd
in
c
r
e
a
s
e
th
e
im
a
ge
s
iz
e
.
F
ur
th
e
r
,
th
e
G
G
C
N
i
s
us
e
d
to
s
e
gm
e
nt
im
a
g
e
s
w
hi
le
R
e
s
N
e
xt
50
is
a
ppl
ie
d
to
e
xt
r
a
c
t
th
e
f
e
a
tu
r
e
s
.
T
he
n,
th
e
gol
de
n
s
in
e
gr
a
y
w
ol
f
opt
im
iz
a
ti
on
(
G
S
G
W
O
)
is
e
s
ta
bl
is
he
d
to
s
e
le
c
t
th
e
e
xt
r
a
c
te
d
f
e
a
tu
r
e
s
.
A
t
la
s
t,
C
L
R
E
N
c
la
s
s
if
ie
s
th
e
li
ve
r
tu
m
or
s
a
c
c
ur
a
te
ly
.
F
ig
ur
e
1
in
di
c
a
te
s
a
bl
oc
k
di
a
gr
a
m
f
or
t
he
pr
opos
e
d a
ppr
oa
c
h.
F
ig
ur
e
1. B
lo
c
k di
a
gr
a
m
f
or
t
he
pr
opos
e
d a
ppr
oa
c
h
3.1.
D
at
as
e
t
s
I
n
th
is
r
e
s
e
a
r
c
h,
L
iT
S
17
[
29
]
a
nd
C
H
A
O
S
[
30
]
da
ta
s
e
ts
a
r
e
us
e
d
f
or
li
ve
r
tu
m
or
s
e
gm
e
nt
a
ti
on.
T
he
s
e
two
da
ta
s
e
ts
a
r
e
ba
s
e
d
on
C
T
s
c
a
ns
,
w
hi
c
h
ha
ve
m
a
n
y
s
li
c
e
s
iz
e
s
,
s
p
a
c
in
g,
a
nd
th
ic
kne
s
s
.
A
br
ie
f
de
s
c
r
ip
ti
on of
t
he
s
e
da
ta
s
e
ts
i
s
e
xpl
a
in
e
d
a
s
f
ol
lo
w
s
.
3.1.1.
L
iT
S
17
I
t
c
ont
a
in
s
a
tr
a
in
in
g
s
e
t
w
it
h
131
C
T
s
c
a
ns
a
nd
a
te
s
t
s
e
t
w
it
h
70
C
T
s
c
a
n
s
.
E
ve
r
y
C
T
c
ont
a
in
s
va
r
yi
ng
s
li
c
e
s
,
r
a
ngi
ng
f
r
om
42
to
1026,
w
it
h
512×
512
-
pi
xe
l
r
e
s
ol
ut
io
n
a
nd
a
s
li
c
e
th
ic
kne
s
s
r
a
ngi
ng
f
r
om
0.45
to
6.0
m
m
.
T
he
vol
um
e
of
th
e
131
C
T
s
c
a
ns
i
s
di
vi
de
d
r
a
n
dom
ly
in
to
two
pa
r
ts
:
30
c
a
s
e
s
f
or
te
s
ti
ng
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
G
r
id
gr
aph c
onv
ol
ut
io
nal
ne
tw
o
r
k
-
c
y
c
li
c
al
l
e
a
r
ni
ng r
at
e
E
ff
ic
ie
nt
N
e
t
fo
r
l
iv
e
r
t
um
o
r
…
(
Sangi N
a
r
as
imhulu)
4239
101
c
a
s
e
s
a
s
tr
a
in
in
g
da
ta
.
F
ig
ur
e
2
r
e
pr
e
s
e
nt
s
a
s
a
m
pl
e
im
a
ge
f
r
om
th
e
L
iT
S
17
da
ta
s
e
t,
w
it
h
F
ig
ur
e
2(
a
)
r
e
pr
e
s
e
nt
in
g t
he
or
ig
in
a
l
im
a
ge
a
nd F
ig
ur
e
2(
b)
r
e
pr
e
s
e
nt
in
g t
he
m
a
s
k i
m
a
ge
.
(
a
)
(
b)
F
ig
ur
e
2. S
a
m
pl
e
i
m
a
ge
s
f
or
L
iT
S
17 da
ta
s
e
t
of
(
a
)
or
ig
in
a
l
im
a
ge
a
nd (
b)
t
r
ue
m
a
s
k
3.1.2.
C
H
A
O
S
T
hi
s
da
ta
s
e
t
ha
s
C
T
im
a
ge
s
w
it
h
va
r
yi
ng
s
li
c
e
s
,
r
a
ngi
ng
f
r
o
m
42
to
1
,
026,
w
it
h
512×
512
-
pi
xe
l
r
e
s
ol
ut
io
n,
a
nd
s
li
c
e
th
ic
kne
s
s
,
in
c
lu
di
ng
6
,
407
s
e
c
ti
ons
f
r
o
m
40
pa
ti
e
nt
s
.
I
t
is
c
om
pos
e
d
of
20
s
e
ts
of
unl
a
be
le
d
a
nd
20
s
e
ts
of
la
be
le
d
da
t
a
.
T
he
16
s
e
ts
of
C
T
im
a
ge
s
a
r
e
s
e
le
c
t
e
d
r
a
ndoml
y,
w
it
h
416
s
e
c
ti
ons
f
or
te
s
ti
ng
a
nd
2
,
458
s
e
c
ti
ons
f
or
tr
a
in
in
g
s
e
ts
.
T
h
e
num
be
r
of
s
li
c
e
s
r
a
nge
s
f
r
om
81
to
266,
w
it
h
512
×
512
-
pi
xe
l
r
e
s
ol
ut
io
n a
nd s
li
c
e
t
hi
c
kne
s
s
r
a
ngi
ng f
r
om
2.0
t
o 3.2
m
m
.
F
ig
u
r
e
3 i
ndi
c
a
te
s
a
s
a
m
pl
e
i
m
a
ge
f
or
t
he
C
H
O
A
S
da
ta
s
e
t.
T
a
bl
e
1
de
not
e
s
th
e
da
t
a
s
e
t
s
de
s
c
r
ip
ti
on,
a
nd
F
ig
ur
e
3
de
pi
c
ts
th
e
s
a
m
pl
e
im
a
ge
s
f
r
om
th
e
C
H
A
O
S
da
ta
s
e
ts
i
n F
ig
ur
e
3(
a
)
r
e
pr
e
s
e
nt
in
g t
he
or
ig
in
a
l
im
a
ge
a
nd F
ig
u
r
e
3(
b)
r
e
pr
e
s
e
nt
in
g t
he
t
r
ue
m
a
s
k.
(
a
)
(
b)
F
ig
ur
e
3. S
a
m
pl
e
i
m
a
ge
s
f
or
C
H
A
O
S
da
ta
s
e
t
of
(
a
)
or
ig
in
a
l
im
a
ge
a
nd (
b)
t
r
ue
m
a
s
k
T
a
bl
e
1. D
a
ta
s
e
t
de
s
c
r
ip
ti
on
D
a
t
a
s
e
t
S
l
i
c
e
S
i
z
e
of
s
l
i
c
e
S
l
i
c
e
t
hi
c
kne
s
s
(
m
m
)
S
l
i
c
e
s
pa
c
i
ng (
m
m
)
L
i
T
S
17
42~
1026
512×512
0.45~
6.0
0.55~
1.0
C
H
A
O
S
81~
266
512×512
2.0~
3.2
0.57~
0.79
3.2.
P
r
e
-
p
r
oc
e
s
s
in
g
T
he
ga
th
e
r
e
d
in
put
im
a
ge
s
a
r
e
pr
e
-
pr
oc
e
s
s
e
d
ut
il
iz
in
g
two
a
ppr
oa
c
he
s
:
m
e
di
a
n
f
il
te
r
in
g
a
nd
da
ta
a
ugm
e
nt
a
ti
on.
T
he
s
e
a
ppr
oa
c
he
s
e
nha
n
c
e
th
e
im
a
ge
qua
li
ty
a
nd
in
c
r
e
a
s
e
th
e
r
obus
tn
e
s
s
a
nd
s
e
gm
e
nt
a
ti
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
4235
-
4249
4240
pe
r
f
or
m
a
nc
e
of
th
e
m
ode
l
by
ge
ne
r
a
ti
ng
m
o
r
e
di
ve
r
s
e
tr
a
in
in
g
im
a
ge
s
in
li
ve
r
tu
m
or
,
li
ke
pa
ddi
ng,
c
r
opp
in
g,
f
li
ppi
ng, a
nd r
a
ndom r
ot
a
ti
on. A de
ta
il
e
d e
xpl
a
na
ti
on of
t
he
s
e
m
e
th
ods
is
e
xpl
a
in
e
d
a
s
f
ol
lo
w
s
.
3.2.1.
M
e
d
ia
n
f
il
t
e
r
in
g
A
m
e
di
a
n
f
il
te
r
[
31]
is
a
non
-
li
ne
a
r
di
gi
ta
l
f
il
te
r
be
c
a
us
e
it
pr
e
s
e
r
ve
s
th
e
e
dge
s
w
hi
le
r
e
m
ovi
ng
im
pul
s
e
noi
s
e
,
s
uc
h
a
s
s
a
lt
a
nd
pe
ppe
r
noi
s
e
,
f
r
om
th
e
im
a
ge
of
a
li
ve
r
tu
m
o
r
.
T
hi
s
ty
pe
of
noi
s
e
a
ppe
a
r
s
a
s
da
r
k
pi
xe
ls
or
is
ol
a
te
d
br
ig
ht
s
pot
s
,
w
hi
c
h
de
gr
a
de
th
e
im
a
ge
qua
li
ty
.
T
he
m
e
di
a
n
f
il
te
r
r
e
pl
a
c
e
s
e
a
c
h
pi
xe
l
w
it
h
th
e
m
e
di
a
n
va
lu
e
f
r
om
a
lo
c
a
l
ne
ig
hbor
hood,
w
hi
c
h
e
f
f
e
c
ti
ve
ly
s
m
oot
he
s
out
noi
s
e
-
in
duc
e
d
out
li
e
r
s
w
hi
le
pr
e
s
e
r
vi
ng
s
ig
ni
f
ic
a
nt
im
a
ge
in
f
o
r
m
a
ti
on
a
nd
e
dge
s
.
I
ni
t
ia
ll
y,
th
e
im
a
ge
’
s
m
e
di
a
n
va
lu
e
is
a
c
qui
r
e
d
t
o
r
e
a
d
th
e
pi
xe
l
va
lu
e
s
,
a
nd
th
e
n
it
is
c
om
put
e
d
by
c
hoos
in
g
th
e
m
id
dl
e
va
lu
e
to
m
odi
f
y
th
e
p
ix
e
l’
s
in
te
ns
it
y
va
lu
e
(
,
)
.
T
he
pr
oc
e
s
s
of
m
e
di
a
n
f
il
te
r
in
g
is
r
e
pr
e
s
e
nt
e
d
in
(
1)
.
H
e
r
e
,
r
e
pr
e
s
e
nt
s
th
e
ne
ig
hbor
hood
pi
xe
l
f
ix
e
d
a
t
th
e
[
,
]
lo
c
a
ti
on
in
th
e
C
T
im
a
ge
,
[
,
]
de
not
e
s
th
e
out
put
va
lu
e
of
th
e
f
il
te
r
e
d
im
a
ge
,
a
nd
[
,
]
de
te
r
m
in
e
s
th
e
in
put
v
a
lu
e
in
th
e
or
ig
in
a
l
im
a
g
e
.
T
hi
s
pr
oc
e
s
s
in
c
r
e
a
s
e
s
th
e
im
a
ge
qua
li
ty
w
hi
le
pr
e
s
e
r
vi
ng t
he
s
ig
ni
f
ic
a
nt
e
dge
i
nf
or
m
a
ti
on.
[
,
]
=
{
[
,
]
}
,
(
,
)
ϵ
ω
(
1)
3.2.2.
D
at
a
au
gm
e
n
t
at
io
n
D
a
ta
a
ugm
e
nt
a
ti
on
a
s
s
is
t
s
in
in
c
r
e
a
s
in
g
th
e
di
ve
r
s
it
y
of
tr
a
i
ni
ng
da
ta
,
w
hi
c
h
m
in
im
iz
e
s
li
m
i
te
d
a
nnot
a
te
d
im
a
ge
s
.
D
e
e
p
le
a
r
ni
ng
(
DL
)
a
ppr
oa
c
he
s
r
e
qui
r
e
a
s
ig
ni
f
ic
a
nt
ly
huge
a
m
ount
of
la
be
le
d
da
ta
f
o
r
tr
a
in
in
g.
T
o
s
ol
ve
th
is
is
s
u
e
,
da
ta
a
ugm
e
nt
a
ti
on
is
pe
r
f
or
m
e
d
to
e
nha
nc
e
th
e
a
va
il
a
bl
e
da
t
a
f
or
tr
a
in
in
g.
I
t
e
xt
e
nds
th
e
da
ta
by
ut
il
iz
in
g
va
r
io
us
a
ppr
oa
c
h
e
s
li
ke
pa
ddi
n
g,
c
r
oppi
ng,
r
a
ndom
r
ot
a
ti
on,
a
nd
hor
iz
ont
a
l
f
li
ppi
ng
to
c
r
e
a
te
di
ve
r
s
e
va
r
ia
ti
ons
of
tr
a
in
in
g
im
a
ge
s
f
or
D
L
m
e
th
od.
B
e
f
or
e
e
f
f
ic
ie
nt
ly
de
pl
oyi
ng
th
e
D
L
a
ppr
oa
c
h,
th
e
da
ta
s
iz
e
is
in
c
r
e
a
s
e
d
th
r
ough
s
ynt
he
ti
c
a
ugm
e
nt
a
ti
on
f
or
C
T
s
e
gm
e
nt
a
ti
on.
I
t
e
nha
n
c
e
s
m
od
e
l
r
obus
tn
e
s
s
by ge
ne
r
a
ti
ng dif
f
e
r
e
nt
t
r
a
ns
f
or
m
a
ti
ons
of
t
he
or
ig
in
a
l
im
a
ge
s
, w
hi
c
h l
e
a
ds
t
o be
tt
e
r
ge
ne
r
a
li
z
a
ti
on
a
nd
e
nha
nc
e
d
s
e
gm
e
nt
a
ti
on
a
c
c
ur
a
c
y.
F
ig
ur
e
s
4
a
nd
5
r
e
pr
e
s
e
nt
s
a
m
pl
e
a
ugm
e
nt
e
d
im
a
ge
s
f
or
th
e
L
iT
S
17
a
nd
C
H
A
O
S
da
ta
s
e
ts
,
w
hi
c
h
is
de
m
ons
tr
a
te
d
in
:
F
ig
ur
e
s
4
(
a
)
a
nd
5(
a
)
s
how
th
e
pa
ddi
ng,
F
ig
ur
e
s
4
(
b)
a
nd
5(
b)
s
how
th
e
c
r
oppi
ng,
F
ig
ur
e
s
4
(
c
)
a
nd
5(
c
)
s
how
th
e
r
a
ndom
r
ot
a
ti
on,
a
nd
F
ig
ur
e
s
4
(
d)
a
nd
5(
d)
s
how
t
he
hor
iz
ont
a
l
f
li
ppi
ng. T
he
pr
e
-
p
r
o
c
e
s
s
e
d da
ta
i
s
f
e
d a
s
i
nput
t
o t
he
s
e
gm
e
nt
a
ti
on
pr
oc
e
s
s
us
in
g
G
G
C
N
.
(
a
)
(
b)
(
c
)
(
d)
F
ig
ur
e
4. S
a
m
pl
e
a
ugm
e
nt
e
d i
m
a
ge
s
f
or
L
iT
S
17 da
ta
s
e
t
of
(a
)
pa
ddi
ng, (
b)
c
r
oppi
ng, (
c
)
r
a
ndom
r
ot
a
ti
on, a
nd
(
d)
hor
iz
ont
a
l
f
li
ppi
ng
F
ig
ur
e
5. S
a
m
pl
e
a
ugm
e
nt
e
d i
m
a
ge
s
f
or
C
H
A
O
S
da
ta
s
e
t
of
(
a
)
pa
ddi
ng, (
b)
c
r
oppi
ng, (
c
)
r
a
ndom
r
ot
a
ti
on,
a
nd (
d)
hor
iz
ont
a
l
f
li
ppi
ng
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
G
r
id
gr
aph c
onv
ol
ut
io
nal
ne
tw
o
r
k
-
c
y
c
li
c
al
l
e
a
r
ni
ng r
at
e
E
ff
ic
ie
nt
N
e
t
fo
r
l
iv
e
r
t
um
o
r
…
(
Sangi N
a
r
as
imhulu)
4241
3.3.
S
e
gm
e
n
t
at
io
n
A
f
te
r
pr
e
-
pr
oc
e
s
s
in
g,
G
G
C
N
is
us
e
d
to
s
e
gm
e
nt
th
e
li
ve
r
tu
m
o
r
e
f
f
e
c
ti
ve
ly
.
G
G
C
N
c
a
pt
ur
e
s
s
pa
ti
a
l
r
e
la
ti
ons
hi
ps
a
nd
de
pe
nde
n
c
ie
s
e
f
f
e
c
ti
ve
ly
th
r
ough
a
gr
a
ph
r
e
pr
e
s
e
nt
a
ti
on,
w
hi
c
h
e
ns
ur
e
s
m
ode
l
unde
r
s
ta
ndi
ng
of
th
e
va
r
io
us
c
ont
e
xt
ti
s
s
ue
s
a
nd
or
ga
n
s
.
G
G
C
N
pr
e
s
e
r
ve
s
s
tr
uc
tu
r
a
l
da
ta
by
r
e
pr
e
s
e
nt
in
g
th
e
im
a
ge
a
s
a
gr
a
ph
a
nd
b
e
tt
e
r
m
ode
ls
th
e
r
e
la
ti
ons
hi
p
s
a
m
ong
r
e
gi
ons
or
ne
ig
hbor
hood
pi
xe
l
s
,
w
hi
c
h
le
a
ds
to
m
or
e
a
c
c
ur
a
te
s
e
gm
e
nt
a
ti
on.
G
G
C
N
e
m
pl
oys
a
c
ove
r
a
ge
-
a
w
a
r
e
gr
id
que
r
y
(
C
A
G
Q
)
by
le
ve
r
a
gi
ng
gr
id
s
pa
c
e
e
f
f
ic
ie
nc
y
a
nd
e
nha
n
c
in
g
s
p
a
ti
a
l
c
ove
r
a
ge
.
F
or
e
ve
r
y
poi
nt
gr
o
up
ge
ne
r
a
te
d
by C
A
G
Q
,
th
e
G
C
A
is
e
m
pl
oye
d
f
or
a
ggr
e
ga
ti
ng
f
e
a
tu
r
e
s
f
r
om
poi
nt
s
of
th
e
node
to
th
e
gr
o
up
c
e
nt
e
r
.
I
ni
ti
a
ll
y,
a
lo
c
a
l
gr
a
ph
(
,
)
is
c
ons
tr
uc
te
d,
w
he
r
e
V
is
th
e
gr
oup
c
e
nt
e
r
,
a
nd
K
is
th
e
poi
nt
of
th
e
node
ge
ne
r
a
te
d
by
C
A
G
Q
.
T
he
n,
e
ve
r
y
node
poi
nt
is
c
onne
c
te
d
to
th
e
gr
oup
c
e
nt
e
r
a
nd
de
ve
lo
ps
node
poi
nt
f
e
a
tu
r
e
s
to
̃
.
T
he
G
C
A
c
om
put
e
s
̃
a
nd
a
ggr
e
ga
te
s
e
ve
r
y
f
e
a
tu
r
e
a
s
a
c
e
nt
e
r
f
e
a
tu
r
e
de
p
e
ndi
ng
o
n
th
e
e
dge
r
e
la
ti
on
be
twe
e
n
th
e
nod
e
a
nd
th
e
c
e
nt
e
r
. T
he
m
a
th
e
m
a
ti
c
a
l
G
C
A
m
odul
e
i
s
e
xpr
e
s
s
e
d i
n (
2)
a
nd (
3)
. T
he
̃
r
e
pr
e
s
e
nt
s
t
he
node
c
ont
r
ib
ut
io
n,
de
not
e
s
th
e
node
lo
c
a
ti
on,
ℳ
de
te
r
m
in
e
s
th
e
m
ul
ti
-
la
ye
r
pe
r
c
e
pt
r
on
(
M
L
P
)
,
s
how
s
th
e
e
dge
a
tt
e
nt
io
n
f
unc
ti
on,
a
nd
e
xpl
a
in
s
th
e
a
ggr
e
ga
ti
on
f
unc
ti
on.
T
he
e
dge
a
tt
e
nt
io
n
f
unc
ti
on
is
e
m
pl
oye
d
by
us
in
g
th
e
c
e
nt
e
r
a
nd
node
to
m
ode
l
e
dge
a
tt
e
nt
io
n
a
s
a
g
e
om
e
tr
ic
r
e
la
ti
on
f
unc
ti
on.
F
ig
ur
e
6
r
e
pr
e
s
e
nt
s
th
e
G
G
C
N
a
r
c
hi
te
c
tu
r
e
.
,
̃
=
(
,
)
∗
ℳ
(
)
(
2)
̃
=
(
{
,
̃
}
,
1
,
…
,
)
(
3)
F
ig
ur
e
6. A
r
c
hi
te
c
tu
r
e
of
G
G
C
N
M
or
e
ove
r
,
th
e
f
or
m
ul
a
ti
on
di
s
r
e
ga
r
ds
th
e
unde
r
ly
in
g
c
ont
r
ib
ut
io
n
of
e
ve
r
y
node
poi
nt
f
r
om
pr
io
r
la
ye
r
s
.
T
he
c
ove
r
a
ge
w
e
ig
ht
is
de
f
in
e
d
a
s
th
e
num
be
r
of
poi
nt
s
a
ggr
e
ga
te
d
to
a
node
in
pr
io
r
la
ye
r
s
.
T
hi
s
va
lu
e
is
c
a
lc
ul
a
t
e
d
e
a
s
il
y
in
C
A
G
Q
,
a
nd
th
e
c
ove
r
a
g
e
w
e
ig
ht
is
a
s
ig
ni
f
ic
a
nt
f
e
a
tu
r
e
in
c
om
put
in
g
e
dge
a
tt
e
nt
io
n.
S
e
m
a
nt
ic
r
e
la
ti
on
is
a
not
he
r
e
s
s
e
nt
ia
l
a
s
pe
c
t
of
c
om
put
in
g
e
dge
a
tt
e
nt
io
n.
T
he
s
e
m
a
nt
ic
r
e
la
ti
on
is
e
nc
ode
d
by
ut
il
iz
in
g
f
e
a
tu
r
e
s
of
th
e
gr
oup
c
e
nt
e
r
a
nd
th
e
f
e
a
t
ur
e
’
s
node
poi
nt
,
w
hi
c
h
ne
e
d
s
a
c
e
nt
e
r
of
th
e
gr
oup
to
be
c
hos
e
n
f
r
om
th
e
node
poi
nt
s
.
T
he
gr
id
c
ont
e
xt
pool
in
g
is
e
s
ta
bl
is
he
d,
w
hi
c
h
e
xt
r
a
c
ts
th
e
c
ont
e
xt
f
e
a
tu
r
e
s
by
pool
in
g
f
r
om
e
v
e
r
y
c
ont
e
xt
poi
nt
in
C
A
G
Q
th
a
t
s
uf
f
ic
ie
nt
ly
c
ov
e
r
s
th
e
lo
c
a
l
gr
a
ph’
s
w
hol
e
gr
id
s
pa
c
e
. G
r
id
c
ont
e
xt
pooli
ng de
te
r
m
in
e
s
t
he
f
ol
lo
w
in
g be
ne
f
it
s
:
‒
m
ode
ls
th
e
vi
r
tu
a
l
gr
oup
c
e
nt
e
r
f
e
a
tu
r
e
s
th
a
t
e
n
a
bl
e
th
e
e
s
ti
m
a
ti
on
of
s
e
m
a
nt
ic
r
e
la
ti
ons
a
m
ong
node
poi
nt
s
a
nd c
e
nt
e
r
s
.
‒
E
ve
n
w
hi
le
th
e
gr
oup
c
e
nt
e
r
is
s
e
le
c
te
d
a
t
a
phys
ic
a
l
poi
nt
,
is
s
ig
ni
f
ic
a
nt
in
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
on
be
c
a
us
e
i
t
c
ov
e
r
s
m
or
e
poi
nt
s
i
n a
ne
ig
hbor
hood ins
te
a
d of
onl
y
gr
a
ph point
s
.
‒
I
n
C
A
G
Q
,
th
e
c
ont
e
xt
poi
nt
s
a
r
e
a
s
s
oc
ia
te
d
w
it
h
th
e
ir
c
e
nt
e
r
voxe
l,
a
nd
th
e
r
e
a
r
e
no
e
xt
r
a
que
r
y
ove
r
he
a
d
poi
nt
s
.
I
n
a
lo
c
a
l
gr
a
ph,
is
s
ha
r
e
d
a
m
ong
e
ve
r
y
e
d
ge
c
om
put
a
ti
on,
a
nd
th
e
pool
in
g
is
a
li
ght
w
e
ig
ht
ope
r
a
ti
on
r
e
qui
r
in
g
no
le
a
r
na
bl
e
w
e
ig
ht
a
nd
l
e
s
s
c
om
put
a
ti
ona
l
ove
r
he
a
d.
T
h
e
e
dge
a
tt
e
nt
io
n f
unc
ti
on i
s
f
or
m
ul
a
te
d i
n (
4)
.
=
(
(
,
,
)
,
(
,
)
)
(
4)
M
e
di
c
a
l
im
a
g
e
s
v
a
r
y
s
ig
ni
f
ic
a
nt
ly
in
p
a
th
ol
ogy
a
nd
a
na
to
m
y
.
T
he
G
G
C
N
s
m
a
na
ge
th
e
ir
r
e
gul
a
r
s
ha
pe
s
of
li
ve
r
tu
m
or
s
by
us
in
g
gr
a
ph
s
tr
uc
tu
r
e
s
,
w
hi
c
h
l
e
a
ds
to
e
nha
nc
e
d
a
c
c
ur
a
c
y
a
nd
r
obus
tn
e
s
s
.
A
ddi
ti
ona
ll
y,
G
G
C
N
in
c
or
po
r
a
te
s
m
ul
ti
-
s
c
a
le
da
ta
,
w
hi
c
h
ge
ne
r
a
te
s
m
or
e
da
ta
a
nd
a
c
c
ur
a
te
s
e
gm
e
nt
a
ti
on.
F
ig
ur
e
7
in
di
c
a
te
s
th
e
s
e
gm
e
nt
e
d
s
a
m
pl
e
im
a
ge
s
,
w
hi
c
h
a
r
e
s
how
n
in
F
ig
ur
e
7
(
a
)
r
e
pr
e
s
e
nt
in
g
th
e
L
iT
S
17
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
4235
-
4249
4242
da
ta
s
e
t
a
nd
F
ig
ur
e
7(
b)
r
e
pr
e
s
e
nt
in
g
th
e
C
H
A
O
S
da
ta
s
e
t.
T
h
e
n,
th
e
s
e
gm
e
nt
e
d
in
put
is
f
e
d
in
to
th
e
f
e
a
tu
r
e
e
xt
r
a
c
ti
on pr
oc
e
s
s
u
s
in
g R
e
s
N
e
xt
50.
(
a
)
(
b)
F
ig
ur
e
7. S
a
m
pl
e
f
or
s
e
gm
e
nt
e
d i
m
a
ge
s
of
(
a
)
L
iT
S
17 a
nd (
b)
C
H
A
O
S
3.4. F
e
at
u
r
e
e
xt
r
ac
t
io
n
A
f
te
r
s
e
gm
e
nt
a
ti
on,
R
e
s
N
e
xt
50
[
32]
is
pe
r
f
or
m
e
d
to
e
xt
r
a
c
t
f
e
a
tu
r
e
s
e
f
f
e
c
ti
ve
ly
f
r
om
li
ve
r
tu
m
or
s
.
R
e
s
N
e
xt
50
e
nh
a
nc
e
s
out
c
om
e
s
by
us
in
g
a
hi
ghl
y
m
odul
a
r
a
r
c
hi
te
c
tu
r
e
w
it
h
gr
oupe
d
c
onvolut
io
ns
th
a
t
im
pr
ove
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
e
f
f
ic
ie
nc
y
a
nd
m
ode
l
s
c
a
la
bi
li
ty
.
I
ts
c
a
r
di
na
li
ty
e
nha
nc
e
s
th
e
ne
twor
k’
s
a
bi
li
ty
t
o
le
a
r
n
di
ve
r
s
e
a
nd
r
ic
h
f
e
a
tu
r
e
s
,
w
hi
c
h
he
lp
in
c
a
pt
ur
in
g
th
e
li
ve
r
tu
m
or
’
s
c
om
pl
e
x
pa
tt
e
r
n.
T
hi
s
m
ode
l
s
ig
ni
f
ic
a
nt
ly
m
in
im
iz
e
s
c
om
put
a
ti
ona
l
de
m
a
nds
w
hi
le
pr
e
s
e
r
vi
ng
in
f
or
m
a
ti
on
r
ic
hne
s
s
.
T
he
1s
t
c
onvolut
io
n
la
ye
r
ut
il
iz
e
s
a
7×
7
c
onvolut
io
na
l
ke
r
ne
l
w
it
h
a
2
-
s
tr
id
e
,
s
uc
c
e
e
de
d
by
3×
3
la
y
e
r
s
of
m
a
x
-
pool
in
g
w
it
h
a
2
-
s
tr
id
e
.
T
he
s
e
c
ond
c
onvolut
io
na
l
m
odul
e
ha
s
3
ki
nds
of
c
onvolut
io
ns
:
a
1×
1
c
onvolut
io
na
l
w
it
h
128
c
ha
nne
ls
,
a
3×
3
c
onvolut
io
na
l
w
it
h
128
c
ha
nne
l
s
,
a
nd
a
1
×
1
c
onvolut
io
na
l
w
it
h
256
c
ha
nn
e
ls
.
T
he
th
ir
d
c
onvolut
io
na
l
ha
s
a
1
×
1
c
onvolut
io
na
l
w
it
h
256
c
h
a
nne
ls
,
3×
3
c
onvolut
io
na
l
w
it
h
256
c
ha
nne
l
s
s
pl
it
in
to
32
c
onvolut
io
na
l
gr
oups
,
a
nd
a
1
×
1
c
onvolut
io
na
l
w
it
h
512
c
ha
nne
ls
pr
e
pa
r
e
d
in
4
gr
oups
.
T
he
f
our
th
c
onvolut
io
na
l
ha
s
a
1
×
1
c
onvolut
io
na
l
ke
r
ne
l
w
it
h
512
c
ha
nne
l
s
,
a
3×
3
c
onvolut
io
na
l
w
it
h
512
c
ha
nn
e
ls
s
pl
it
in
to
32
c
onvolut
io
na
l
gr
oups
,
a
nd
a
1×
1
c
onvolut
io
na
l
w
it
h
1
,
024
c
ha
nne
ls
,
r
e
pe
a
te
d
in
6
gr
oups
.
T
he
f
if
th
c
onvolut
io
na
l
ha
s
a
1
×
1
c
onvolut
io
na
l
ke
r
ne
l
w
it
h
1
,
024
c
ha
nn
e
ls
,
a
3×
3
c
onvolut
io
na
l
w
it
h
1
,
024
c
ha
nn
e
ls
,
w
hi
c
h s
pl
it
i
nt
o 32
c
onvolut
io
na
l
g
r
oups
, a
nd a
1×
1 c
onvolut
io
na
l
w
it
h 2
,
048 c
ha
nne
ls
, w
hi
c
h a
r
e
r
e
pe
a
te
d i
n
3
gr
oups
.
T
h
e
n,
th
e
out
c
om
e
is
g
e
ne
r
a
te
d
by
us
in
g
th
e
a
ve
r
a
ge
pool
in
g
la
ye
r
a
nd
a
f
ul
ly
c
onne
c
te
d
(
F
C
)
la
ye
r
. R
e
s
N
e
xt
50 e
xt
r
a
c
ts
2
,
048 f
e
a
tu
r
e
s
e
f
f
e
c
ti
ve
ly
a
nd i
nc
r
e
a
s
e
s
t
he
m
ode
l’
s
a
bi
li
ty
s
ig
ni
f
ic
a
nt
ly
, t
o
c
a
pt
ur
e
di
ve
r
s
e
a
nd
in
tr
ic
a
te
f
e
a
tu
r
e
s
w
it
hi
n
th
e
s
e
gm
e
nt
e
d
im
a
ge
s
.
A
ls
o,
R
e
s
N
e
xt
50’
s
de
pt
h
a
nd
r
e
s
id
ua
l
c
onne
c
ti
ons
e
na
bl
e
r
obus
t
le
a
r
ni
ng f
or
m
or
e
a
c
c
ur
a
te
a
nd r
e
li
a
b
le
t
um
or
r
e
s
ul
ts
.
3.5. F
e
at
u
r
e
s
e
le
c
t
io
n
A
f
te
r
e
x
tr
a
c
t
in
g
f
e
a
tu
r
e
s
,
t
he
G
S
G
W
O
is
e
s
ta
b
li
s
he
d
to
s
e
le
c
t
t
he
f
e
a
t
ur
e
s
f
r
o
m
th
e
L
iT
S
1
7
a
n
d
C
H
A
O
S
da
ta
s
e
ts
f
or
li
ve
r
t
um
or
s
.
T
he
f
e
a
tu
r
e
s
e
le
c
t
io
n
p
r
oc
e
s
s
is
e
s
s
e
nt
ia
l
to
in
c
r
e
a
s
e
th
e
m
ode
l’
s
pe
r
f
o
r
m
a
nc
e
b
y
m
i
ni
m
iz
in
g
d
im
e
ns
i
ona
li
ty
a
n
d
f
oc
us
i
ng
on
th
e
m
os
t
a
pp
r
o
pr
ia
te
f
e
a
tu
r
e
s
,
w
h
ic
h
r
e
d
uc
e
ove
r
f
it
ti
ng
a
nd
i
m
p
r
o
ve
in
t
e
r
pr
e
ta
bi
li
ty
.
G
W
O
e
f
f
e
c
t
iv
e
ly
e
x
pl
o
r
e
s
a
nd
e
x
pl
o
it
s
t
he
s
e
a
r
c
h
s
pa
c
e
,
le
a
d
in
g
to
o
pt
im
a
l
f
e
a
tu
r
e
s
ubs
e
ts
.
T
h
is
r
e
s
u
lt
s
in
e
n
ha
nc
e
d
a
c
c
u
r
a
c
y
by
c
h
oos
i
ng
th
e
m
os
t
a
pp
r
o
pr
ia
te
f
e
a
t
u
r
e
s
f
r
om
th
e
e
xt
r
a
c
te
d
f
e
a
tu
r
e
s
.
G
W
O
[
3
3]
im
it
a
te
s
t
he
w
ol
f
-
hu
nt
in
g
a
pp
r
oa
c
h
f
o
r
op
ti
m
iz
a
t
io
n.
T
h
is
a
p
pr
oa
c
h
s
pl
i
ts
t
he
w
o
lv
e
s
in
to
he
a
d
w
o
l
f
∝
,
,
a
n
d
,
w
h
ic
h
he
l
ps
t
he
he
a
d
w
ol
f
i
n
hu
nt
in
g
a
nd
is
r
e
s
po
ns
ib
le
f
o
r
s
e
nt
r
y
,
r
e
c
on
na
is
s
a
nc
e
,
a
lo
ng
w
it
h
a
no
th
e
r
w
ol
f
ω
.
T
he
w
o
l
ve
s
'
h
un
ti
ng
be
ha
v
io
r
is
p
r
im
a
r
i
ly
s
pl
it
i
nt
o
th
r
e
e
pha
s
e
s
:
e
nc
ir
c
le
m
e
n
t,
a
t
ta
c
k,
a
nd
p
ur
s
ui
t.
I
n
th
e
e
nc
i
r
c
le
m
e
n
t
s
ta
ge
,
th
e
G
W
O
u
pda
te
s
th
e
p
os
i
ti
o
n
by
ut
i
li
z
i
ng
(
5
)
t
o
(
8
)
.
(
+
1
)
=
(
)
−
.
|
.
(
)
−
(
)
|
(
5)
=
2
1
.
1
−
1
(
6)
=
2
2
(
7)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
G
r
id
gr
aph c
onv
ol
ut
io
nal
ne
tw
o
r
k
-
c
y
c
li
c
al
l
e
a
r
ni
ng r
at
e
E
ff
ic
ie
nt
N
e
t
fo
r
l
iv
e
r
t
um
o
r
…
(
Sangi N
a
r
as
imhulu)
4243
1
=
2
−
2
/
(
8)
W
he
r
e
,
de
not
e
th
e
gr
a
y
w
ol
f
a
nd
pr
e
y
w
ol
f
’
s
pos
it
io
n,
,
de
f
in
e
th
e
c
oe
f
f
ic
ie
nt
ve
c
to
r
s
,
1
de
te
r
m
in
e
s
th
e
c
onve
r
ge
nc
e
f
a
c
to
r
,
1
,
2
de
te
r
m
in
e
th
e
r
a
ndom
nu
m
be
r
a
m
ong
[
0,
1]
,
in
di
c
a
te
s
th
e
pr
e
s
e
nt
num
be
r
of
it
e
r
a
ti
ons
,
a
nd
r
e
pr
e
s
e
nt
s
th
e
m
a
xi
m
um
num
be
r
of
it
e
r
a
ti
ons
.
I
n
th
e
pur
s
ui
t
a
nd
a
tt
a
c
k
s
ta
ge
,
th
e
G
W
O
upda
te
s
th
e
pos
it
io
n
by
(
9)
a
nd
(
10)
.
T
h
e
1
,
2
,
a
n
d
3
de
f
in
e
th
e
lo
c
a
ti
on
upda
t
e
,
w
hi
c
h
in
f
lu
e
nc
e
s
th
e
f
a
c
to
r
s
of
,
, a
nd
w
ol
ve
s
.
{
1
=
−
1
.
|
1
.
−
2
=
−
2
.
|
2
.
−
3
=
−
3
.
|
3
.
−
(
9)
(
+
1
)
=
(
1
+
2
+
3
)
/
3
(
10)
3.5.1. Gol
d
e
n
s
in
e
al
gor
it
h
m
I
t
is
a
m
e
ta
-
he
ur
is
ti
c
a
ppr
oa
c
h
th
a
t
tr
a
ns
ve
r
s
e
s
e
a
c
h
poi
nt
o
n
a
c
ir
c
le
by
th
e
s
in
e
f
unc
ti
on
a
nd
m
in
im
iz
e
s
th
e
s
e
a
r
c
h
s
pa
c
e
by
th
e
gol
de
n
c
o
e
f
f
ic
ie
nt
,
th
e
r
e
by
e
na
bl
in
g
th
e
m
ode
l
to
a
c
hi
e
ve
hi
gh
s
e
a
r
c
h
e
f
f
ic
ie
nc
y, j
um
pi
ng out of
t
he
l
oc
a
l
opt
im
a
. T
he
gol
de
n s
in
e
a
ppr
oa
c
h’
s
pos
it
io
n upda
te
e
qua
ti
on i
s
e
xpr
e
s
s
e
d
in
(
11)
a
nd (
12)
.
T
he
′
(
+
1
)
de
f
in
e
s
t
he
m
ode
l’
s
c
onve
r
ge
nc
e
di
r
e
c
ti
on,
1
,
2
de
not
e
t
he
r
a
ndom num
be
r
s
a
m
ong
[
0
,
2
]
a
nd
[
0
,
]
.
T
he
1
a
nd
2
r
e
pr
e
s
e
nt
th
e
gol
de
n
a
lg
or
it
hm
c
o
e
f
f
ic
ie
nt
,
w
hi
le
de
te
r
m
in
e
s
th
e
num
be
r
of
gol
de
n s
e
c
ti
ons
. T
h
e
m
a
xi
m
um
a
c
c
ur
a
c
y i
s
us
e
d
a
s
a
f
it
ne
s
s
f
unc
ti
on
,
w
hi
c
h i
s
c
a
lc
ul
a
te
d i
n (
13)
.
′
(
+
1
)
=
(
+
1
)
.
|
1
|
−
2
.
1
.
|
1
.
1
−
2
.
(
+
1
)
|
(
11)
{
1
=
.
+
.
(
1
−
)
2
=
.
(
1
−
)
+
.
′
(
12)
=
(
)
(
13)
T
he
gol
de
n
s
in
e
a
ppr
oa
c
h
e
f
f
e
c
ti
ve
ly
pe
r
f
or
m
s
s
e
c
onda
r
y
po
pul
a
ti
on
c
onve
r
ge
nc
e
,
opt
im
iz
e
s
th
e
a
ppr
oa
c
h,
a
nd
in
c
r
e
a
s
e
s
th
e
a
lg
or
it
hm
’
s
s
e
a
r
c
h
a
bi
li
ty
.
I
t
s
e
le
c
ts
1
,
980
a
nd
1
,
890
f
e
a
tu
r
e
s
f
or
th
e
L
iT
S
17
a
nd
C
H
A
O
S
da
ta
s
e
t
s
.
T
h
e
G
S
G
W
O
pr
ovi
de
s
a
n
e
f
f
e
c
ti
ve
a
ppr
oa
c
h
by
in
c
or
por
a
ti
ng
lo
c
a
l
a
nd
gl
oba
l
s
e
a
r
c
h
a
bi
li
ti
e
s
,
e
ns
ur
in
g
a
c
c
ur
a
te
a
nd
opt
im
a
l
tu
m
or
de
li
ne
a
ti
on.
A
f
te
r
s
e
le
c
ti
ng
th
e
f
e
a
tu
r
e
s
,
th
e
G
G
C
N
-
C
L
R
E
N
is
pe
r
f
or
m
e
d f
or
l
iv
e
r
t
um
or
c
la
s
s
if
ic
a
ti
on.
3.6.
C
la
s
s
if
i
c
at
io
n
O
nc
e
th
e
f
e
a
tu
r
e
s
a
r
e
s
e
le
c
t
e
d
f
r
om
th
e
e
xt
r
a
c
te
d
f
e
a
tu
r
e
s
,
E
f
f
ic
ie
nt
N
e
t
c
la
s
s
if
ie
s
th
e
li
ve
r
tu
m
or
by
le
ve
r
a
gi
ng
a
s
c
a
la
bl
e
a
r
c
hi
te
c
tu
r
e
th
a
t
ba
la
n
c
e
s
th
e
w
id
th
,
de
pt
h,
a
nd
r
e
s
ol
ut
io
ns
.
I
t
is
a
C
N
N
m
ode
l
a
nd
s
c
a
li
ng
a
ppr
oa
c
h
th
a
t
a
ppl
ie
s
c
om
pounde
d
c
oe
f
f
ic
ie
nt
s
f
or
s
c
a
li
ng
th
e
di
m
e
ns
io
ns
e
ve
nl
y
in
th
e
li
ve
r
tu
m
or
s
.
E
f
f
ic
ie
nt
N
e
t
ha
s
8 m
ode
ls
be
twe
e
n B
0 a
nd B
7. A
s
t
he
numbe
r
of
m
ode
ls
i
nc
r
e
a
s
e
s
, t
he
numbe
r
of
pa
r
a
m
e
te
r
s
doe
s
not
r
is
e
s
ig
ni
f
ic
a
nt
ly
,
but
th
e
a
c
c
ur
a
c
y
de
c
r
e
a
s
e
s
r
e
m
a
r
k
a
bl
y.
T
he
us
e
of
D
L
a
ppr
oa
c
h
is
to
di
s
c
lo
s
e
m
or
e
e
f
f
e
c
ti
ve
m
ode
ls
w
it
h
s
m
a
ll
e
r
a
ppr
oa
c
he
s
.
E
f
f
ic
ie
nt
N
e
t
a
c
hi
e
ve
s
m
or
e
e
f
f
ic
ie
nt
r
e
s
ul
t
s
by
e
ve
nl
y
s
c
a
li
ng
w
id
th
,
r
e
s
ol
ut
io
n, a
nd
de
pt
h
w
he
n
s
c
a
li
ng
dow
n
th
e
a
pp
r
oa
c
h.
T
he
m
a
in
bui
ld
in
g
bl
oc
k
is
a
n
in
ve
r
te
d
M
B
C
onv
bot
tl
e
n
e
c
k
,
w
hi
c
h
is
ge
n
e
r
a
te
d
in
M
obi
le
N
e
tV2
f
or
E
f
f
ic
ie
nt
N
e
t.
I
n
M
B
C
onv,
th
e
bl
o
c
ks
h
a
ve
a
la
ye
r
th
a
t
is
in
it
ia
ll
y
c
om
pr
e
s
s
e
d
a
nd
th
e
n
e
nl
a
r
ge
s
th
e
c
ha
n
ne
l.
A
m
ong
bot
tl
e
ne
c
ks
,
s
tr
a
ig
ht
c
onne
c
ti
ons
a
s
s
oc
ia
t
e
d
w
it
h
f
e
w
e
r
c
ha
nn
e
ls
c
om
pa
r
e
d
to
th
e
e
xpa
nde
d
la
y
e
r
s
a
r
e
e
m
pl
oye
d.
F
ur
th
e
r
m
or
e
,
th
is
s
tr
uc
tu
r
e
ha
s
in
-
de
pt
h
s
e
pa
r
a
bl
e
c
onvolut
io
ns
th
a
t
r
e
duc
e
th
e
c
a
lc
ul
a
ti
on
by
a
2
f
a
c
to
r
,
w
he
r
e
th
e
ke
r
ne
l
s
iz
e
de
not
e
s
th
e
he
ig
ht
a
nd
w
id
th
of
th
e
c
onvolut
io
n
w
in
dow
.
T
he
c
om
pound
c
oe
f
f
ic
ie
nt
is
ut
il
iz
e
d
f
or
s
c
a
li
ng
e
ve
nl
y
,
w
hi
c
h i
s
e
xpr
e
s
s
e
d i
n (
14)
.
ℎ
:
=
ℎ
:
=
:
=
≥
1
,
≥
1
,
≥
1
(
14
)
W
he
r
e
,
,
de
not
e
s
a
c
ons
ta
nt
th
a
t
is
c
a
lc
ul
a
te
d
by
gr
id
s
e
a
r
c
h
,
a
nd
is
de
te
r
m
in
e
d
a
s
a
us
e
r
-
de
f
in
e
d
c
oe
f
f
ic
ie
nt
th
a
t
ha
ndl
e
s
th
e
a
va
il
a
bl
e
r
e
s
our
c
e
s
to
s
c
a
le
th
e
m
ode
l.
T
he
f
lo
a
ti
ng
-
poi
nt
ope
r
a
ti
ons
pe
r
s
e
c
ond
(
F
L
O
P
S
)
a
r
e
pr
opor
ti
ona
l
to
,
2
,
2
.
C
om
put
in
g
c
o
s
ts
in
c
onvolut
io
n
ne
twor
ks
a
r
e
gr
e
a
te
r
ow
in
g
to
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
4235
-
4249
4244
c
onvolut
io
n
ope
r
a
ti
on,
w
hi
le
s
c
a
li
ng c
onvolut
io
na
l
ne
twor
ks
in
c
r
e
a
s
e
s
th
e
F
L
O
P
S
ne
twor
k
by
a
ppr
oxi
m
a
te
ly
(
,
2
,
2
)
(
,
2
,
2
)
. T
he
c
om
pound s
c
a
li
ng a
ppr
oa
c
h s
c
a
le
s
t
hi
s
m
ode
l
in
2
s
ta
ge
s
:
‒
S
ta
ge
1:
c
ons
id
e
r
in
g
th
a
t
th
e
r
e
a
r
e
m
or
e
th
a
n
two
a
va
il
a
bl
e
r
e
s
our
c
e
s
,
a
gr
id
s
e
a
r
c
h
is
pe
r
f
or
m
e
d
w
it
h
=
1
,
a
nd i
de
a
l
va
lu
e
s
a
r
e
e
s
ta
bl
i
s
he
d f
r
om
,
,
a
nd
.
‒
S
ta
ge
2:
th
e
obt
a
in
e
d
,
,
a
nd
va
lu
e
s
a
r
e
de
te
r
m
in
e
d a
s
c
ons
ta
nt
s
, a
nd
th
e
s
ta
nda
r
d
n
e
twor
k
is
s
c
a
l
e
d
up t
o obta
in
E
f
f
ic
ie
nt
N
e
t
-
B
1 t
o B
7 w
it
h va
r
yi
ng
va
lu
e
s
.
T
he
r
e
c
ti
f
ie
d
li
ne
a
r
uni
t
(
R
e
L
U
)
is
us
e
d
a
s
a
n
a
c
ti
va
ti
on
f
unc
ti
on
f
or
th
e
li
ve
r
tu
m
or
s
.
C
L
R
is
us
e
d
to
a
c
qui
r
e
th
e
opt
im
a
l
le
a
r
ni
ng
r
a
te
in
li
ve
r
tu
m
or
s
by
f
lu
c
tu
a
ti
ng
be
twe
e
n
th
e
m
a
xi
m
um
le
a
r
ni
ng
r
a
te
o
f
10
a
nd
a
ba
s
e
le
a
r
ni
ng
r
a
te
of
1e
-
8.
W
it
h
a
s
te
p
s
iz
e
of
50
a
nd
a
c
yc
le
le
ngt
h
of
100,
th
e
C
L
R
a
dj
us
ts
th
e
le
a
r
ni
ng
r
a
te
w
it
hi
n
th
is
r
a
nge
ove
r
100
it
e
r
a
ti
ons
by
ut
i
li
z
in
g
a
ba
tc
h
s
iz
e
of
32.
T
he
ove
r
f
it
t
in
g
is
s
ue
is
s
ol
ve
d
by
u
s
in
g
C
L
R
dur
in
g
tr
a
in
in
g,
a
ll
ow
in
g
th
e
m
ode
l
to
e
x
pl
or
e
a
w
id
e
r
r
a
nge
of
le
a
r
ni
ng
r
a
te
s
.
T
h
e
m
a
x
a
nd
ba
s
e
le
a
r
ni
ng
r
a
te
s
de
te
r
m
in
e
a
r
a
ng
e
bounda
r
y
w
he
r
e
th
e
r
a
te
of
le
a
r
ni
ng
is
f
lu
c
tu
a
te
d.
T
he
s
e
dyn
a
m
ic
a
dj
us
tm
e
nt
s
e
na
bl
e
th
e
m
ode
l
to
tr
a
in
m
or
e
e
f
f
e
c
ti
ve
ly
by
r
a
pi
d
c
onve
r
ge
nc
e
.
I
nc
or
por
a
ti
ng
C
L
R
a
nd
E
f
f
ic
ie
nt
N
e
t
e
nha
nc
e
s
tr
a
in
in
g
s
ta
bi
li
ty
,
opt
im
iz
e
s
le
a
r
ni
ng
r
a
te
f
or
be
tt
e
r
pe
r
f
or
m
a
nc
e
,
a
nd
obt
a
in
s
a
hi
gh
a
c
c
ur
a
c
y
w
it
h
f
e
w
e
r
r
e
s
our
c
e
s
,
w
hi
c
h
r
e
nde
r
s
th
is
a
ppr
oa
c
h
hi
ghl
y
e
f
f
e
c
ti
ve
f
o
r
th
e
c
la
s
s
if
ic
a
ti
on
of
li
ve
r
t
um
or
.
4.
E
X
P
E
R
I
M
E
N
T
A
L
R
E
S
U
L
T
S
T
he
r
e
s
ul
ts
a
nd
di
s
c
u
s
s
io
n
of
th
e
pe
r
f
or
m
a
nc
e
a
na
ly
s
is
a
r
e
pr
e
s
e
nt
e
d
in
th
is
s
e
c
ti
on.
T
he
pr
opos
e
d
a
ppr
oa
c
h
is
e
v
a
lu
a
te
d
u
s
in
g
s
of
twa
r
e
to
ol
s
:
A
na
c
ond
a
N
a
v
ig
a
to
r
3.5.2.0
(
64
-
bi
t)
,
P
yt
hon
3.10.12
w
it
h
W
in
dow
s
10
ope
r
a
ti
ng
s
ys
te
m
s
,
i5
I
nt
e
l
-
c
or
e
,
a
nd
8
G
B
R
A
M
.
T
he
f
r
a
m
e
w
or
ks
a
nd
li
br
a
r
ie
s
us
e
d
he
r
e
a
r
e
T
r
a
ns
f
or
m
e
r
s
,
T
e
n
s
or
f
lo
w
,
K
e
r
a
s
,
S
kl
e
a
r
n
f
r
a
m
e
w
or
ks
,
a
nd
M
a
tp
lo
tl
ib
li
br
a
r
y
to
pl
ot
.
T
he
pe
r
f
or
m
a
nc
e
m
e
tr
ic
s
ut
il
iz
e
d
in
th
is
r
e
s
e
a
r
c
h
a
r
e
di
c
e
s
im
il
a
r
it
y
c
oe
f
f
ic
ie
nt
(
D
S
C
)
,
a
c
c
ur
a
c
y,
r
e
c
a
ll
,
pr
e
c
i
s
io
n,
vol
um
e
tr
ic
ove
r
la
ppi
ng
e
r
r
or
(
V
O
E
)
,
a
nd
r
e
la
ti
ve
vol
um
e
di
f
f
e
r
e
nc
e
(
R
V
D
)
,
w
hi
c
h
a
r
e
e
xpr
e
s
s
e
d
in
(
15
)
to
(
21)
.
T
he
,
,
,
a
nd
r
e
pr
e
s
e
nt
tr
ue
pos
it
iv
e
,
f
a
ls
e
ne
ga
ti
ve
,
tr
ue
ne
ga
ti
ve
, a
nd
f
a
ls
e
pos
it
iv
e
,
a
nd
|
|
a
nd
|
|
in
di
c
a
te
t
he
vol
um
e
s
of
a
nd
, r
e
s
pe
c
ti
ve
ly
.
=
+
+
+
+
(
15)
=
+
(
16)
=
+
(
17)
1
−
=
2
2
+
+
(
18)
(
,
)
=
2
|
⋂
|
|
|
+
|
|
(
19)
(
,
)
=
|
|
−
|
|
|
|
(
20)
(
,
)
=
1
−
2
|
⋂
|
|
|
+
|
|
(
21)
4.1. P
e
r
f
or
m
an
c
e
an
al
ys
is
T
a
bl
e
2
pr
e
s
e
nt
s
th
e
pe
r
f
or
m
a
nc
e
a
na
ly
s
is
of
s
e
gm
e
nt
a
ti
on
m
e
th
ods
us
in
g
th
e
L
iT
S
17
a
nd
C
H
A
O
S
da
ta
s
e
ts
.
T
he
e
xi
s
ti
ng a
ppr
oa
c
he
s
of
U
N
e
t,
S
upe
r
pi
xe
l
s
e
gm
e
nt
a
ti
on, a
nd G
C
N
a
r
e
c
om
pa
r
e
d w
it
h t
he
G
G
C
N
a
ppr
oa
c
h.
W
he
n
c
om
pa
r
e
d
to
th
e
s
e
e
xi
s
ti
ng
a
ppr
oa
c
he
s
,
G
G
C
N
a
c
hi
e
ve
s
a
be
tt
e
r
D
S
C
of
98.50%
a
nd
97.95%
us
in
g
th
e
L
iT
S
17
a
nd
C
H
A
O
S
da
ta
s
e
ts
due
to
it
e
f
f
e
c
ti
ve
ly
in
te
gr
a
ti
ng
th
e
s
pa
ti
a
l
s
tr
uc
tu
r
e
of
gr
i
d
da
ta
w
it
h
G
C
N
,
w
hi
c
h
e
na
bl
e
s
m
or
e
a
c
c
ur
a
te
m
ode
li
ng
of
c
o
m
pl
e
x
tu
m
or
s
ha
pe
s
a
nd
s
pa
ti
a
l
r
e
la
ti
on
s
hi
ps
.
A
ls
o,
it
e
m
pl
oy
s
bot
h
gl
oba
l
a
nd
lo
c
a
l
c
ont
e
xt
s
,
w
hi
c
h
in
c
r
e
a
s
e
it
s
a
bi
li
ty
to
c
a
pt
ur
e
c
om
pl
e
x
bounda
r
ie
s
a
nd va
r
ia
ti
ons
.
T
a
bl
e
3
de
not
e
s
a
pe
r
f
or
m
a
nc
e
a
na
ly
s
is
of
f
e
a
tu
r
e
s
e
le
c
ti
on
m
e
th
ods
f
or
L
iT
S
17
a
nd
C
H
A
O
S
da
ta
s
e
ts
.
P
a
r
ti
c
le
s
w
a
r
m
opt
im
iz
a
ti
on
(
P
S
O
)
,
a
nt
c
ol
ony
opt
im
iz
a
ti
on
(
A
C
O
)
,
a
nd
G
W
O
a
r
e
c
om
pa
r
e
d
w
it
h
G
S
G
W
O
,
w
hi
c
h
a
c
hi
e
ve
s
a
b
e
tt
e
r
a
c
c
ur
a
c
y
of
99.80%
a
nd
99.96%
.
A
s
th
e
pr
opos
e
d
a
ppr
oa
c
h
e
nha
nc
e
s
e
xpl
oi
ta
ti
on
a
nd
e
xpl
or
a
ti
on
ba
la
nc
e
by
in
c
lu
di
ng
th
e
gol
d
e
n
s
in
e
s
tr
a
te
gy,
th
e
c
onve
r
ge
nc
e
s
pe
e
d
i
s
im
pr
ove
d,
a
lo
ngs
id
e
a
voi
di
ng
th
e
lo
c
a
l
opt
im
a
is
s
u
e
. T
hi
s
a
ppr
oa
c
h
in
te
gr
a
te
s
th
e
s
tr
e
ngt
hs
of
th
e
s
in
e
c
os
in
e
Evaluation Warning : The document was created with Spire.PDF for Python.