T
E
L
KO
M
N
I
KA
T
e
lec
om
m
u
n
icat
ion
,
Com
p
u
t
i
n
g,
E
lec
t
r
on
ics
an
d
Cont
r
ol
Vol.
18
,
No.
3
,
J
une
2020
,
pp.
131
0
~
13
1
8
I
S
S
N:
1693
-
6930
,
a
c
c
r
e
dit
e
d
F
ir
s
t
G
r
a
de
by
Ke
me
nr
is
tekdikti
,
De
c
r
e
e
No:
21/E
/KP
T
/2018
DO
I
:
10.
12928/
T
E
L
KO
M
NI
KA
.
v18i3.
14753
1310
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
U
N
e
t
-
V
GG
16 w
ith
t
r
a
n
sf
e
r
l
e
ar
n
in
g
f
or
M
R
I
-
b
as
e
d
b
r
ai
n
t
u
m
or
s
e
gm
e
n
t
a
t
io
n
Ani
n
d
ya
Aprili
yan
t
i
P
r
avit
as
ar
i
1
,
Nur
I
r
iawan
2
,
M
awand
a
Alm
u
h
ayar
3
,
T
au
f
ik
Az
m
i
4
,
I
r
h
am
ah
5
,
Kart
ik
a
F
i
t
h
r
ias
ar
i
6
,
S
a
n
t
i
W
u
lan
P
u
r
n
am
i
7
,
Wi
d
ian
a
F
e
r
r
ias
t
u
t
i
8
1,
2,
3,
4
,
5,
6
,
7
D
ep
ar
t
men
t
o
f
St
a
t
i
s
t
i
cs
,
In
s
t
i
t
u
t
T
e
k
n
o
l
o
g
i
Sep
u
l
u
h
N
o
p
em
b
er,
In
d
o
n
e
s
i
a
1
D
ep
ar
t
men
t
o
f
St
a
t
i
s
t
i
cs
,
U
n
i
v
ers
i
t
a
s
Pad
j
a
j
aran
,
In
d
o
n
es
i
a
8
D
ep
ar
t
men
t
o
f
Rad
i
o
l
o
g
y
,
U
n
i
v
ers
i
t
a
s
A
i
r
l
an
g
g
a,
In
d
o
n
es
i
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
Aug
15
,
2019
R
e
vis
e
d
J
a
n
29
,
2020
Ac
c
e
pted
F
e
b
26
,
2020
A
b
ra
i
n
t
u
m
o
r
i
s
o
n
e
o
f
a
d
ea
d
l
y
d
i
s
ea
s
e
t
h
a
t
n
ee
d
s
h
i
g
h
acc
u
racy
i
n
i
t
s
med
i
ca
l
s
u
rg
er
y
.
Brai
n
t
u
m
o
r
d
et
ec
t
i
o
n
can
b
e
d
o
n
e
t
h
ro
u
g
h
mag
n
et
i
c
res
o
n
an
ce
i
ma
g
i
n
g
(MRI).
Imag
e
s
eg
men
t
at
i
o
n
fo
r
t
h
e
MRI
b
rai
n
t
u
m
o
r
ai
ms
t
o
s
e
p
ara
t
e
t
h
e
t
u
mo
r
area
(a
s
t
h
e
reg
i
o
n
o
f
i
n
t
eres
t
o
r
RO
I)
w
i
t
h
a
h
eal
t
h
y
b
r
a
i
n
a
n
d
p
r
o
v
i
d
e
a
c
l
e
a
r
b
o
u
n
d
a
r
y
o
f
t
h
e
t
u
m
o
r
.
T
h
i
s
s
t
u
d
y
c
l
a
s
s
i
f
i
e
s
t
h
e
R
O
I
a
n
d
n
on
-
R
O
I
u
s
i
n
g
f
u
l
l
y
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
w
i
t
h
n
e
w
a
r
c
h
i
t
e
c
t
u
r
e
,
n
a
m
e
l
y
U
N
e
t
-
V
G
G
1
6
.
T
h
i
s
m
o
d
e
l
o
r
a
r
c
h
i
t
e
c
t
u
r
e
i
s
a
h
y
b
r
i
d
o
f
U
-
N
e
t
a
n
d
V
G
G
1
6
w
i
t
h
t
r
a
n
s
f
e
r
L
e
a
r
n
i
n
g
t
o
s
i
m
p
l
i
f
y
t
h
e
U
-
N
e
t
a
r
c
h
i
t
e
c
t
u
r
e
.
T
h
i
s
m
e
t
h
o
d
h
a
s
a
h
i
g
h
a
c
c
u
r
a
c
y
o
f
a
b
o
u
t
9
6
.
1
%
i
n
t
h
e
l
e
a
r
n
i
n
g
d
a
t
a
s
e
t
.
T
h
e
v
a
l
i
d
a
t
i
o
n
i
s
d
o
n
e
b
y
c
a
l
c
u
l
a
t
i
n
g
t
h
e
c
o
r
r
e
c
t
c
l
a
s
s
i
f
i
c
a
t
i
o
n
r
a
tio
(
C
C
R
)
t
o
c
o
m
p
a
r
i
n
g
t
h
e
s
e
g
m
e
n
t
a
t
i
o
n
r
e
s
u
l
t
w
i
t
h
t
h
e
g
r
o
u
n
d
t
r
u
t
h
.
T
h
e
C
C
R
v
a
l
u
e
s
h
o
w
s
t
h
at
t
h
i
s
U
N
et
-
V
G
G
1
6
c
o
u
l
d
reco
g
n
i
ze
t
h
e
b
ra
i
n
t
u
mo
r
area
w
i
t
h
a
mean
o
f
CCR
v
a
l
u
e
i
s
ab
o
u
t
9
5
.
6
9
%
.
K
e
y
w
o
r
d
s
:
F
ull
y
c
onvolut
ion
ne
twor
k
I
mage
s
e
gmenta
ti
on
T
r
a
ns
f
e
r
l
e
a
r
ning
U
-
Ne
t
VG
G16
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e
.
C
or
r
e
s
pon
din
g
A
u
th
or
:
Nur
I
r
iaw
a
n,
De
pa
r
tm
e
nt
of
S
tatis
ti
c
s
,
I
ns
ti
tut
T
e
knologi
S
e
puluh
Nope
mber
,
Ka
mpus
I
T
S
S
ukoli
lo
-
S
ur
a
ba
ya
60111,
I
ndone
s
ia
.
E
mail:
nu
r
_i@s
tatis
ti
ka
.
it
s
.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
A
br
a
in
tum
or
is
the
15
th
de
a
dly
dis
e
a
s
e
in
I
ndone
s
ia
c
ompar
e
d
to
a
ll
types
of
c
a
nc
e
r
.
Ac
c
or
ding
to
the
W
HO
,
ther
e
we
r
e
5,
323
c
a
s
e
s
of
br
a
in
a
nd
ne
r
vous
s
ys
tem
tum
or
s
in
I
ndone
s
ia
with
4
,
229
mor
ta
li
ty
c
a
s
e
s
dur
ing
2018
[
1]
.
Due
to
thi
s
r
e
a
s
on,
a
br
a
in
tum
o
r
is
c
ons
ider
e
d
to
be
a
n
im
por
tant
top
ic.
De
tec
ti
on
of
b
r
a
in
tum
or
c
ould
be
done
unde
r
the
medic
a
l
e
quipm
e
nt
c
a
ll
e
d
magne
ti
c
r
e
s
ona
nc
e
im
a
ging
or
M
R
I
.
Ge
ne
r
a
l
Hos
pit
a
l
(
R
S
UD
)
Dr
.
S
oe
tom
o
S
u
r
a
ba
ya
pr
ovides
r
a
diol
ogica
l
e
xa
mi
na
ti
on
s
e
r
vice
s
u
s
ing
M
R
I
1
.
5
T
e
s
la
a
nd
3
T
e
s
la.
T
he
gr
e
a
ter
the
T
e
s
la
number
,
the
be
tt
e
r
the
im
a
ge
qua
li
ty,
but
it
’
s
a
ls
o
mor
e
c
os
tl
y.
T
he
M
R
I
1.
5
T
e
s
la
is
a
f
a
vor
a
ble
s
e
r
vice
,
s
ince
it
ha
s
a
mi
nim
um
c
os
t
a
nd
a
ls
o
c
ove
r
e
d
by
the
I
ndone
s
ian
gove
r
nment's
s
oc
ial
s
e
c
ur
it
y
(
B
P
J
S
)
.
How
e
ve
r
,
be
tt
e
r
im
a
ge
qua
li
ty
is
ne
e
de
d
in
medic
a
l
t
r
e
a
tm
e
nt.
T
he
im
a
ge
s
e
gmenta
ti
on
is
one
of
methodology
th
a
t
c
ould
pr
ovide
be
tt
e
r
s
ight
of
br
a
in
tum
or
s
by
s
e
pa
r
a
ti
ng
the
tum
or
a
r
e
a
(
a
s
the
r
e
gion
of
int
e
r
e
s
t
or
R
OI
)
with
the
he
a
lt
hy
br
a
in
a
nd
pr
ovide
a
c
lea
r
boun
da
r
y
of
the
tum
or
.
T
he
c
lea
r
bounda
r
y
he
lps
the
medic
a
l
tr
e
a
tm
e
nt,
e
s
pe
c
ially
in
s
ur
ge
r
y,
to
r
unning
s
a
f
e
ly
without
da
maging
he
a
lt
hy
pa
r
ts
of
the
br
a
in
.
T
his
s
tudy
tr
i
e
s
to
s
e
gment
the
M
R
I
br
a
in
tum
o
r
to
give
a
be
tt
e
r
s
ight
of
the
M
R
I
im
a
ge
f
r
o
m
a
1
.
5
T
e
s
la
mac
hine.
M
a
ny
types
o
f
r
e
s
e
a
r
c
h
ha
d
be
e
n
de
ve
loped
f
o
r
im
a
ge
s
e
gmenta
ti
on.
S
e
ve
r
a
l
methods
us
e
c
lus
ter
ing
a
s
the
ba
s
is
of
modeling,
while
other
s
us
e
c
las
s
if
ica
ti
on.
T
he
a
im
is
to
ge
t
the
be
s
t
model
that
c
ould
r
e
c
ognize
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
UN
e
t
-
V
GG
16
w
it
h
tr
ans
fer
lear
ning
for
M
R
I
-
bas
e
d
br
ain
tumor
…
(
A
nindya
A
pr
il
iyanti
P
r
av
it
as
ar
i
)
1311
the
tum
or
a
r
e
a
mor
e
pr
e
c
is
e
ly.
T
he
pr
e
vious
s
tudi
e
s
us
e
th
e
c
lu
s
ter
ing
a
s
the
ba
s
is
of
s
e
gment
a
ti
on
a
r
e
pe
r
f
or
med
by
[
2]
whic
h
us
e
s
the
ge
ne
ti
c
a
lgor
i
th
m
a
nd
[
3
]
whic
h
e
mpl
oy
fu
z
z
y
c
lus
ter
ing
,
Ots
u
m
e
thod
a
nd
K
-
mea
ns
c
lus
ter
to
s
e
gment
the
ve
hi
c
le
im
a
g
e
.
T
he
model
-
ba
s
e
d
c
l
us
ter
ing
is
pe
r
f
or
med
by
[
4
-
6]
in
the
f
o
r
m
of
a
F
ini
te
mi
xtu
r
e
model
to
s
e
gment
the
M
R
I
b
r
a
in
tum
or
im
a
ge
.
Di
f
f
e
r
e
nt
f
r
om
the
pr
e
vious
s
tudi
e
s
,
thi
s
s
tudy
us
e
s
the
c
la
s
s
if
ica
ti
on
method
with
Ne
ur
a
l
Ne
twor
k.
Ne
ur
a
l
Ne
twor
k
is
us
e
d
s
in
c
e
it
c
a
n
a
da
pt
the
wor
king
of
human
ne
ur
ons
in
r
e
c
ognizing
im
a
ge
s
.
P
r
e
vious
s
tudi
e
s
done
unde
r
the
NN
a
p
pr
oa
c
h
is
pe
r
f
or
med
by
[
7
-
9]
that
us
e
d
the
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
C
NN
)
a
nd
De
e
p
L
e
a
r
ning
f
o
r
im
a
ge
s
e
gmenta
ti
on.
T
his
s
tudy
us
e
s
the
F
ull
y
C
onvolu
ti
ona
l
Ne
twor
k
(
F
C
N)
s
ince
it
ha
s
g
r
e
a
t
pe
r
f
or
m
a
nc
e
f
or
s
e
mantic
s
e
gmenta
ti
on
[
10,
11]
.
T
he
F
C
N
us
e
d
the
U
-
N
e
t
a
r
c
hit
e
c
tur
e
by
the
r
e
c
omm
e
nda
ti
on
of
[
12]
.
T
he
U
-
Ne
t
model
c
a
n
a
c
hiev
e
ve
r
y
good
pe
r
f
or
manc
e
on
ve
r
y
d
if
f
e
r
e
nt
biom
e
dica
l
s
e
gmenta
ti
on
a
ppli
c
a
ti
ons
.
T
his
pa
pe
r
us
e
s
a
r
e
a
l
da
tas
e
t
f
r
om
Ge
ne
r
a
l
Hos
pit
a
l
(
R
S
UD
)
Dr
.
S
oe
tom
o
.
T
he
li
mi
ted
number
o
f
br
a
in
tum
or
pa
ti
e
nts
who
c
ome
to
Dr
.
S
oe
tom
o
,
make
s
the
number
of
da
tas
e
ts
a
na
lyze
d
a
ls
o
li
mi
ted.
T
he
U
-
Ne
t
will
be
hybr
idi
z
e
d
with
VG
G16
a
r
c
hit
e
c
tur
e
a
s
it
s
e
nc
ode
r
(
c
ont
r
a
c
ti
ng)
laye
r
[
13]
to
s
im
pli
f
y
the
a
r
c
hit
e
c
tur
e
a
nd
o
ve
r
c
ome
the
pr
oblem
o
f
s
mall
number
da
ta
c
las
s
if
ica
ti
on.
On
the
other
ha
nd
,
the
c
ompl
e
xit
y
of
U
-
Ne
t
of
te
n
s
pe
nds
a
lot
of
ti
me
in
it
s
e
xe
c
uti
on
a
nd
is
gr
e
a
tl
y
a
f
f
e
c
ted
whe
n
the
c
omput
e
r
s
pe
c
if
ica
ti
ons
a
r
e
inade
qua
te.
I
n
view
of
thes
e
s
hor
tcomings
a
nd
to
give
the
s
tage
of
the
a
r
t
of
thi
s
pa
pe
r
we
pe
r
f
or
m
the
tr
a
ns
f
e
r
lea
r
ning
i
n
the
pa
r
t
of
tr
a
ini
ng
da
ta
in
the
hyb
r
id
of
U
-
Ne
t
a
nd
VG
G16.
T
he
na
me
of
th
is
pr
opos
e
d
model
or
a
r
c
hit
e
c
tur
e
is
UN
e
t
-
VG
G16
with
T
r
a
ns
f
e
r
L
e
a
r
ning.
M
or
e
ove
r
,
we
tr
y
to
c
ompar
e
the
UN
e
t
-
VG
G16
with
s
e
ve
r
a
l
s
c
e
na
r
ios
of
the
U
-
Ne
t
model
a
nd
de
c
ide
the
be
s
t
model
wit
h
the
va
lue
o
f
los
s
a
nd
a
c
c
ur
a
c
y.
T
he
c
or
r
e
c
t
c
las
s
if
ica
ti
on
r
a
ti
o
(
CCR)
will
be
c
a
lcula
ted
to
c
ompar
e
the
s
e
g
menta
ti
on
r
e
s
ult
s
f
r
om
the
be
s
t
mode
l
with
the
gr
o
und
tr
uth
a
s
the
mea
s
ur
e
of
E
va
luation.
All
the
da
ta
a
nd
g
r
ound
t
r
uth
us
e
d
in
thi
s
s
tudy
a
r
e
pa
s
s
e
d
by
the
medic
a
l
a
ppr
ova
l
f
r
om
Dr
.
S
oe
tom
o
S
ur
a
ba
ya
.
2.
RE
S
E
AR
CH
M
E
T
HO
D
2
.
1.
F
u
ll
y
c
on
vol
u
t
ion
al
n
e
t
wor
k
:
U
-
n
e
t
wit
h
t
r
an
s
f
e
r
lear
n
in
g
F
ull
y
c
onvolut
ional
ne
twor
k
(
F
C
N)
is
one
of
the
d
e
e
p
lea
r
ning
us
e
d
in
s
e
mantic
s
e
gmenta
ti
on.
F
C
N
pr
opos
e
d
by
[
14]
that
e
xa
mi
ne
s
pixel
-
to
-
pixel
mapping
a
nd
us
e
d
the
gr
ound
tr
uth
to
de
ter
mi
ne
t
he
pixel
c
las
s
.
T
he
F
C
N
is
the
de
ve
lopm
e
nt
o
f
the
c
las
s
ica
l
c
onvolut
ion
a
l
ne
ur
a
l
ne
twor
k
(
C
NN
)
.
C
NN
c
ons
is
ts
of
c
onvolut
ion
,
poo
li
ng,
a
nd
f
ull
y
-
c
onne
c
ted
a
s
th
e
main
laye
r
s
,
whic
h
in
the
F
C
N,
the
f
ull
y
c
onne
c
ted
laye
r
is
r
e
plac
e
d
by
the
c
onvolut
ion
laye
r
.
F
C
N,
ther
e
f
or
e
,
c
a
n
c
las
s
if
y
e
ve
r
y
pixel
in
the
im
a
ge
a
nd
gi
ve
them
t
he
a
bil
it
y
to
make
pr
e
dictions
on
a
r
bit
r
a
r
y
-
s
ize
d
input
s
.
T
he
input
da
ta
is
a
n
im
a
ge
that
c
ons
is
ts
of
thr
e
e
laye
r
s
na
mely
he
ight
(
ℎ
)
,
width
(
)
,
a
nd
de
pth
(
)
.
T
he
ℎ
a
nd
e
xplaine
d
the
s
pa
ti
a
l
dim
e
ns
ion,
while
is
the
f
e
a
tur
e
or
c
ha
nne
l
dim
e
ns
ion.
T
he
im
a
g
e
f
ir
s
t
laye
r
ha
s
ℎ
×
dim
e
ns
ion
a
nd
c
olor
c
ha
nne
l
(
=
1
whe
n
it
c
ontains
only
gr
a
ys
c
a
le
int
e
ns
it
y,
or
=
3
whe
n
it
ha
s
r
e
d,
gr
e
e
n,
a
nd
blue
(
R
GB
)
int
e
ns
it
ies
)
.
S
uppos
e
d
we
ha
ve
input
da
ta
ve
c
tor
,
in
the
loca
ti
on
(
,
)
o
f
a
pa
r
ti
c
ular
laye
r
.
W
e
a
ls
o
c
a
n
c
a
lcula
te
the
ve
c
tor
output
y
with
the
f
o
ll
owing
f
or
mul
a
[
14
]
:
y
=
(
{
+
,
+
}
,
0
≤
,
≤
)
(
1)
whe
r
e
is
the
ke
r
ne
l
s
ize
,
is
the
s
ubs
a
mpl
ing
f
a
c
tor
,
a
nd
s
pe
c
if
ies
the
type
of
laye
r
us
e
d:
a
mat
r
ix
mul
ti
pli
c
a
ti
on
f
or
c
onvolut
ion
or
a
ve
r
a
ge
pooli
ng
,
a
s
pa
ti
a
l
max
f
or
max
poo
li
ng,
o
r
other
types
o
f
laye
r
s
.
Applying
thi
s
laye
r
c
a
n
s
igni
f
ica
ntl
y
r
e
duc
e
the
nu
mber
of
pa
r
a
mete
r
s
of
the
ne
twor
k.
M
or
e
ove
r
,
the
ne
tw
or
k
c
a
n
lea
r
n
the
c
or
r
e
lation
be
twe
e
n
ne
i
ghbo
r
hood
pixels
[
15]
.
F
igur
e
1
is
a
s
tr
uc
tur
a
l
model
in
t
he
F
C
N
method.
I
n
F
igu
r
e
1,
we
c
ould
s
e
e
that
F
C
N
c
a
n
e
f
f
icie
ntl
y
lea
r
n
to
make
de
ns
e
pr
e
dictions
f
or
pe
r
-
pixel
tas
ks
li
ke
s
e
mantic
s
e
gmenta
ti
on
[
14]
.
S
e
ve
r
a
l
a
r
c
hit
e
c
tur
e
s
a
r
e
buil
ding
unde
r
F
C
N,
th
is
s
tudy
us
e
d
the
U
-
Ne
t
a
r
c
hit
e
c
tur
e
that
f
ir
s
tl
y
r
e
c
omm
e
nde
d
by
[
12]
,
whic
h
is
g
r
e
a
t
f
or
bio
medic
a
l
im
a
ge
s
e
gmenta
ti
on.
I
n
F
igur
e
2
,
we
c
a
n
s
e
e
the
vis
ua
li
z
a
ti
on
of
U
-
Ne
t
a
r
c
hit
e
c
tur
e
.
T
he
U
-
s
ha
pe
of
a
r
c
hit
e
c
tur
e
is
the
r
e
a
s
on
be
hind
it
s
na
me.
T
he
letter
U
c
ontains
two
pa
ths
.
T
he
le
f
t
pa
th
is
c
a
ll
e
d
the
e
nc
ode
r
(
c
ontr
a
c
ti
ng
laye
r
)
a
nd
the
r
ight
pa
th
is
the
de
c
ode
r
(
e
xpa
nding
laye
r
)
.
T
he
e
nc
ode
r
is
a
ne
twor
k
in
whic
h
the
output
is
the
f
e
a
tur
e
map/
ve
c
tor
th
a
t
holds
the
inf
or
mation
r
e
pr
e
s
e
nti
ng
the
input
.
T
he
de
c
o
de
r
whic
h
ha
s
the
s
a
me
s
tr
uc
tur
e
but
in
the
oppo
s
it
e
wa
y,
is
a
ne
twor
k
that
take
s
f
e
a
tu
r
e
maps
f
r
om
the
e
n
c
ode
r
a
nd
p
r
ovides
a
s
im
il
a
r
matc
h
o
f
the
a
c
tual
input
or
int
e
nde
d
output
.
T
he
p
r
oc
e
s
s
in
the
e
nc
ode
r
pa
th
i
s
r
e
duc
ing
the
s
ize
input
matr
ix
by
inc
r
e
a
s
ing
the
number
of
the
f
e
a
tur
e
maps
,
while
in
the
de
c
ode
r
pa
th
i
s
r
e
tur
ning
the
matr
ix
to
it
s
or
igi
na
l
s
ize
by
mi
nim
izing
the
number
of
the
f
e
a
tur
e
maps
.
T
he
s
e
gmenta
ti
o
n
r
e
s
ult
s
,
ther
e
f
or
e
,
c
a
n
be
c
ompar
e
d
with
the
gr
o
und
tr
uth
in
e
ve
r
y
pixel
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
131
0
-
13
1
8
1312
F
igur
e
1.
S
t
r
uc
tur
e
o
f
f
ul
ly
c
onvolut
ional
ne
twor
k
(
s
our
c
e
d
f
r
om
[
14]
)
F
igur
e
2
.
T
he
U
-
Ne
t
a
r
c
hit
e
c
tur
e
(
e
xa
mpl
e
f
or
32x32
pixels
in
the
lowe
s
t
r
e
s
olut
ion,
s
our
c
e
d
f
r
om
[
12
]
)
U
-
Ne
t
tr
a
ns
mi
ts
the
f
e
a
tur
e
map
of
e
a
c
h
c
ontr
a
c
t
pa
th
leve
l
a
t
the
a
na
log
leve
l
of
the
e
xtens
ion
pa
th
s
o
that
the
c
las
s
if
ier
c
a
n
c
ons
ider
f
e
a
tur
e
s
o
f
va
r
yi
ng
s
c
a
les
a
nd
c
ompl
e
xit
ies
to
make
i
ts
de
c
is
ion.
T
he
r
e
f
or
e
,
U
-
Ne
t
c
a
n
lea
r
n
f
r
om
r
e
latively
s
mall
tr
a
ini
ng
s
e
t
s
[
16]
.
S
ince
it
s
c
ompl
e
xit
ies
,
the
U
-
Ne
t
a
r
c
hit
e
c
tur
e
of
ten
s
pe
nds
a
lot
of
ti
me
on
it
s
e
xe
c
uti
on.
I
n
or
de
r
to
de
a
l
with
the
pr
oblem,
thi
s
s
tudy
tr
ies
to
hybr
id
t
he
U
-
Ne
t
a
r
c
hit
e
c
tur
e
with
T
r
a
ns
f
e
r
l
e
a
r
n
ing.
T
r
a
ns
f
e
r
l
e
a
r
ning
is
a
n
a
ppr
oa
c
h
in
whic
h
p
r
e
-
tr
a
ined
models
a
r
e
us
e
d
a
s
a
s
tar
ti
ng
point
f
or
c
omput
e
r
vis
ion
a
nd
langua
ge
pr
oc
e
s
s
ing
tas
ks
,
s
ince
the
de
ve
lopm
e
nt
of
ne
ur
a
l
ne
twor
k
models
f
or
thes
e
pr
oblems
a
nd
due
to
the
e
nor
mous
lea
ps
in
qua
li
f
ica
ti
ons
r
e
quir
e
s
e
xtens
ive
c
omput
ing
a
nd
ti
me
r
e
s
our
c
e
s
.
T
he
a
im
of
T
r
a
ns
f
e
r
l
e
a
r
ning
is
to
im
pr
ove
le
a
r
ning
in
the
ta
r
ge
t
tas
k
by
us
ing
the
knowle
d
ge
f
r
om
the
s
our
c
e
tas
k.
T
r
a
ns
f
e
r
l
e
a
r
ning
is
a
n
e
f
f
e
c
ti
ve
tec
hnique
f
or
r
e
duc
ing
tr
a
ini
ng
ti
me
[
17
]
.
T
his
te
c
hnique
may
be
r
e
late
d
to
the
de
ve
lopm
e
nt
o
f
de
e
p
lea
r
nin
g
models
f
or
im
a
ge
c
las
s
if
ica
ti
on
pr
oblems
.
M
or
e
ove
r
,
to
s
im
pli
f
ying
the
U
-
Ne
t
a
r
c
hit
e
c
tur
e
,
s
e
ve
r
a
l
a
r
c
hit
e
c
tur
e
s
f
r
om
C
NN
ha
ve
be
e
n
c
ons
ider
e
d
to
hybr
id
with
the
U
-
Ne
t,
s
uc
h
a
s
L
e
Ne
t
[
18]
,
Ale
xNe
t
[
19]
,
Z
F
Ne
t
[
20]
a
nd
VG
G
-
Ne
t
[
21]
.
T
he
VG
G
-
Ne
t
c
onf
ir
ms
that
a
s
maller
ke
r
ne
l
s
ize
a
nd
a
de
e
p
C
NN
c
a
n
im
p
r
ove
model
pe
r
f
o
r
manc
e
.
T
he
a
r
c
hit
e
c
tur
e
of
VG
G
-
Ne
t
a
s
s
hown
in
F
igur
e
3
is
qui
te
s
im
il
a
r
to
the
U
-
Ne
t,
ther
e
f
or
e
,
th
is
s
tud
y
c
hoos
e
s
VGG
-
Ne
t
to
r
e
plac
e
the
e
nc
ode
r
pa
th
of
U
-
Ne
t
a
s
the
hybr
id
be
twe
e
n
thes
e
two
powe
r
f
ul
a
r
c
hit
e
c
tur
e
s
.
T
he
ne
w
a
r
c
hit
e
c
tur
e
will
be
dis
c
us
s
e
d
in
r
e
s
ult
s
a
nd
a
na
lys
is
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
UN
e
t
-
V
GG
16
w
it
h
tr
ans
fer
lear
ning
for
M
R
I
-
bas
e
d
br
ain
tumor
…
(
A
nindya
A
pr
il
iyanti
P
r
av
it
as
ar
i
)
1313
F
igur
e
3
.
T
he
VG
G16
a
r
c
hit
e
c
tur
e
(
a
da
pted
f
r
om
[
21]
)
T
his
s
tudy
will
us
e
the
los
s
a
nd
a
c
c
ur
a
c
y
to
mea
s
ur
e
model
pe
r
f
o
r
manc
e
.
L
os
s
is
the
va
lue
of
e
r
r
or
be
twe
e
n
pr
e
dicte
d
a
nd
a
c
tual
va
lue.
T
he
c
a
s
e
with
two
c
las
s
e
s
in
mac
hine
lea
r
ni
ng
us
e
s
the
binar
y
c
r
os
s
-
e
ntr
opy
los
s
f
unc
ti
on
to
c
a
lcula
te
the
va
lue
of
los
s
o
r
e
r
r
or
[
22]
.
T
he
binar
y
c
r
os
s
-
e
ntr
opy
is
pr
ovid
e
d
by
(
2)
.
(
)
=
−
1
∑
log
(
(
)
)
=
1
+
(
1
−
)
log
(
1
−
(
)
)
,
(
2)
whe
r
e
is
the
number
o
f
da
ta
,
is
the
c
las
s
of
c
las
s
if
ica
ti
on
whic
h
ha
s
the
va
lue
o
f
0
o
r
1
,
a
nd
(
)
is
the
pr
oba
bil
it
y
o
f
.
Ac
c
ur
a
c
y
is
the
c
los
e
ne
s
s
be
t
we
e
n
pr
e
dicte
d
a
nd
a
c
tual
va
lue.
I
n
or
de
r
to
de
te
r
mi
ne
the
a
c
c
ur
a
c
y,
we
us
e
the
c
onf
us
ion
mat
r
ix
a
s
in
T
a
ble
1
[
22]
.
T
he
f
o
r
mul
a
f
or
c
a
lcula
ti
ng
the
a
c
c
ur
a
c
y
is
s
hown
by
(
3)
.
a
c
c
ur
a
c
y
=
+
+
+
+
.
(
3)
T
a
ble
1.
C
onf
us
ion
m
a
tr
ix
A
c
tu
a
l
V
a
lu
e
P
r
e
di
c
te
d V
a
lu
e
T
r
ue
F
a
ls
e
T
r
ue
(
T
r
ue
P
os
it
iv
e
)
(
F
a
ls
e
N
e
ga
ti
ve
)
F
a
ls
e
(
F
a
ls
e
P
os
it
iv
e
)
(
T
r
ue
N
e
ga
ti
ve
)
2
.
2.
T
r
ain
in
g
an
d
o
p
t
im
izat
ion
T
he
opti
mi
z
a
ti
on
us
e
d
in
thi
s
pa
pe
r
is
Ada
pti
ve
M
oment
E
s
ti
mator
(
Ada
m)
to
e
s
ti
mate
the
pa
r
a
mete
r
s
.
Ada
m
wa
s
a
c
ombi
na
ti
on
of
tw
o
method’
s
Ada
Gr
a
d
a
nd
R
M
S
P
r
op
[
23]
.
Ada
m
uti
li
z
e
s
two
mom
e
nts
’
va
r
iable
a
s
the
f
ir
s
t
mom
e
nt
(
the
mea
n)
a
nd
va
r
iable
a
s
the
s
e
c
ond
mo
ment
(
the
unc
e
nter
e
d
va
r
ianc
e
)
of
the
gr
a
dients
r
e
s
pe
c
ti
ve
ly.
Give
n
the
hype
r
-
pa
r
a
mete
r
0
≤
1
<
1
a
nd
0
≤
2
<
1
a
nd
pe
r
f
or
m
a
n
e
xpone
nti
a
ll
y
-
we
ight
e
d
movi
ng
a
ve
r
a
ge
(
E
W
M
A)
,
the
Ada
m
e
s
ti
mator
c
a
n
be
c
ons
tr
uc
ted
in
Algor
it
hm
1
a
s
f
oll
ows
[
24
]
Algorithm 1.
Adam Estimator
a.
Initialize
the
value
of
1
,
2
∈
[
0
,
1
)
,
0
,
0
,
learning
rate
,
and
stochastic
objective
function
(
(
)
)
with parameters
.
b.
Calculate
the gradient of
(
)
with following formula
=
∇
(
−
1
)
.
(4)
c.
Update
the first and second moment with equation (5) and (6)
=
1
−
1
+
(
1
−
1
)
,
(5)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
131
0
-
13
1
8
1314
=
2
−
1
+
(
1
−
2
)
⊙
(6)
d.
Calculate
the bias correction with (7) and (8)
̂
=
1
−
1
,
(7)
̂
=
1
−
2
.
(8)
e.
Re
-
adjust
the
learning
rate
for
each
element
in
the
model
parameters
using
el
ement
operations with bias
-
corrected variables
̂
and
̂
as in following formula
′
=
̂
√
̂
+
,
(9)
where
is the learning rate and
is error criterion set to
1
0
−
8
.
f.
Update the parameter
as the (10)
=
−
1
−
′
,
(10)
g.
Do step 2 to step 6 until the parameter
is convergence.
As
in
Algor
it
hm
1
,
Ada
m
us
e
d
s
tocha
s
ti
c
gr
a
dient
de
s
c
e
nt
to
maintain
a
s
ingl
e
lea
r
ning
r
a
te
f
or
a
ll
we
ight
upda
tes
.
T
his
ha
s
e
ns
ur
e
d
that
the
lea
r
ning
r
a
te
doe
s
not
c
ha
nge
dur
ing
the
tr
a
ini
ng
pr
oc
e
s
s
.
T
he
lea
r
ning
r
a
te
is
maintaine
d
f
or
e
a
c
h
ne
twor
k
we
igh
t
(
pa
r
a
mete
r
)
a
nd
a
djus
ted
s
e
pa
r
a
tely
a
s
lea
r
ning
unf
olds
.
B
e
s
ides
,
the
a
lgor
it
hm
c
a
lcula
tes
a
n
e
xp
one
nti
a
l
movi
ng
a
ve
r
a
ge
of
the
gr
a
dient
a
nd
the
s
qua
r
e
d
gr
a
dient,
with
hype
r
-
pa
r
a
mete
r
s
1
a
nd
2
to
c
ontr
ol
th
e
de
c
a
y
r
a
tes
.
T
he
in
it
ializa
ti
on
is
s
e
t
f
o
r
lea
r
ning
r
a
te
=
0
.
0001
,
1
=
0
.
9
,
a
nd
2
=
0
.
999
.
T
he
los
s
f
unc
ti
on
us
e
d
binar
y
c
r
os
s
-
e
ntr
opy
a
s
a
s
e
t
up
binar
y
c
las
s
if
ica
ti
on.
T
he
binar
y
c
r
os
s
-
e
ntr
opy
wa
s
a
c
ombi
na
ti
on
of
s
igm
oid
a
c
ti
va
ti
on
a
nd
c
r
os
s
-
e
ntr
opy
los
s
.
I
n
thi
s
s
tudy,
we
divi
de
da
ta
int
o
thr
e
e
pa
r
ti
t
ions
.
W
e
do
the
s
a
mpl
ing
without
r
e
pla
c
e
ment
to
ge
t
the
r
a
ndomi
z
e
d
da
ta
f
o
r
t
r
a
ini
ng,
va
li
da
ti
on
,
a
nd
t
e
s
ti
ng.
T
he
thr
e
e
s
e
ts
a
r
e
divi
de
d
us
ing
a
r
a
ti
o
of
80:10:
10.
T
he
a
mount
o
f
da
ta
f
or
e
a
c
h
s
e
t
is
a
s
F
igur
e
4
.
F
igur
e
4.
Divis
ion
o
f
the
da
ta
2
.
3
.
E
valu
at
ion
C
or
r
e
c
t
C
las
s
if
ica
ti
on
r
a
ti
o
(
CCR
)
is
c
a
lcula
ted
a
s
a
mea
s
ur
e
of
e
va
luation.
CCR
va
lue
is
obtaine
d
to
f
ind
out
whe
ther
the
R
OI
f
r
o
m
s
e
gmenta
ti
on
r
e
s
ult
s
in
a
c
c
or
da
nc
e
with
the
g
r
ound
t
r
uth.
T
he
g
r
e
a
ter
the
CCR
va
lue,
the
be
tt
e
r
the
s
e
gmenta
ti
on
r
e
s
ult
a
nd
vice
v
e
r
s
a
.
As
s
hown
in
(
11
)
is
the
f
or
mul
a
of
CCR
[
25
]
.
=
∑
|
∩
|
|
|
2
=
1
,
(
11)
whe
r
e
is
gr
ound
tr
uth
f
or
non
-
R
OI
(
=
1
)
a
nd
R
OI
(
=
2
)
.
f
or
=
1
de
f
ines
pixel
s
e
gmente
d
ba
s
e
d
on
the
model
a
s
non
-
R
OI
a
r
e
a
,
mea
nwhile
f
or
=
2
de
s
c
r
ibes
pixel
s
e
gmente
d
a
s
R
OI
a
nd
=
⋃
2
=
1
.
3.
RE
S
UL
T
S
A
ND
AN
AL
YSI
S
3.
1.
T
h
e
p
r
op
os
e
d
ar
c
h
it
e
c
t
u
r
e
U
-
Ne
t
is
one
o
f
C
NN
a
r
c
hit
e
c
tur
e
that
is
us
e
d
s
pe
c
if
ica
ll
y
f
or
im
a
ge
s
e
gmenta
ti
on.
T
he
c
ompl
e
xit
y
of
U
-
Ne
t
(
thi
s
r
e
s
e
a
r
c
h
ha
s
31,
031
,
685
pa
r
a
mete
r
s
)
im
pa
c
ts
the
t
im
e
of
e
xe
c
uti
on
a
nd
in
s
e
ve
r
a
l
c
omput
e
r
s
with
the
li
m
i
ted
s
pe
c
if
ica
ti
on,
the
U
-
Ne
t
a
r
c
hit
e
c
t
ur
e
c
a
nnot
be
r
un.
T
o
ove
r
c
ome
thi
s
pr
oblem,
we
pr
opos
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
UN
e
t
-
V
GG
16
w
it
h
tr
ans
fer
lear
ning
for
M
R
I
-
bas
e
d
br
ain
tumor
…
(
A
nindya
A
pr
il
iyanti
P
r
av
it
as
ar
i
)
1315
a
ne
w
model
that
r
e
duc
e
s
the
laye
r
a
nd
pa
r
a
mete
r
of
U
-
Ne
t
by
c
ombi
ning
it
with
a
nother
a
r
c
hit
e
c
tur
e
na
mely
VG
G16.
T
he
c
hoice
of
VG
G16
is
be
c
a
us
e
of
i
ts
s
im
il
a
r
i
ty
to
U
-
Ne
t's
c
ontr
a
c
ted
laye
r
a
nd
it
s
number
of
pa
r
a
mete
r
s
is
a
ls
o
les
s
than
U
-
Ne
t.
I
n
a
ddit
io
n,
VG
G16
a
l
r
e
a
dy
ha
s
we
ight
s
f
r
om
pa
r
a
mete
r
s
that
a
r
e
e
a
s
il
y
a
c
c
e
s
s
e
d,
s
o
we
e
mpl
oy
thes
e
we
ight
s
to
thi
s
ne
w
model.
S
e
ve
r
a
l
models
us
e
d
f
or
s
e
gmenta
ti
on
f
r
e
que
ntl
y
c
ons
is
t
of
the
c
ontr
a
c
ti
ng
laye
r
a
nd
the
e
xpa
ns
ion
laye
r
.
I
n
th
is
s
tudy,
we
modi
f
ied
the
VG
G16
to
r
e
s
e
mbl
e
the
U
-
Ne
t
a
r
c
hit
e
c
tur
e
by
a
dding
a
n
e
xpa
ns
i
ve
laye
r
c
ons
is
ti
ng
of
s
e
ve
r
a
l
ups
a
mpl
ing
laye
r
s
a
nd
c
onvo
lut
ion
laye
r
s
a
t
the
e
nd
of
the
VG
G16
a
r
c
hi
tec
tur
e
.
T
his
is
done
unti
l
the
ove
r
a
ll
a
r
c
hit
e
c
tur
e
of
the
model
is
s
ymm
e
tr
ica
l
a
nd
r
e
s
e
mbl
e
s
the
s
ha
pe
of
the
l
e
tt
e
r
U.
T
he
r
e
f
or
e
,
in
the
a
r
c
hit
e
c
tur
e
of
the
UN
e
t
-
VG
G16
model,
we
will
ha
ve
a
c
ontr
a
c
ti
ng
laye
r
,
whic
h
is
the
VG
G16,
a
nd
the
e
xpa
ns
ion
laye
r
that
will
be
a
dde
d
late
r
.
W
it
h
thi
s
ne
w
a
r
c
hit
e
c
tur
e
,
the
pa
r
a
me
ter
s
will
be
r
e
duc
e
d
to
17
,
040,
001
with
t
r
a
inable
pa
r
a
mete
r
s
is
a
bout
2,
324
,
353.
T
he
M
R
I
b
r
a
in
tum
o
r
im
a
ge
will
be
tr
a
ined
us
ing
the
UN
e
t
-
VG
G16
model
with
the
T
r
a
ns
f
e
r
L
e
a
r
ning
method.
T
his
m
e
thod
f
r
e
e
z
e
s
the
c
ontr
a
c
ti
on
laye
r
in
UN
e
t
-
VG
G16
s
o
that
the
we
ight
e
d
laye
r
is
not
upda
ted
whe
n
e
xe
c
uti
ng
tr
a
ini
ng
da
ta.
I
ns
tea
d,
we
us
e
the
we
ight
of
the
c
onvolut
ion
laye
r
of
the
VG
G16
model.
T
he
goa
l
is
to
r
e
duc
e
the
c
omput
ing
pr
oc
e
s
s
a
nd
s
pe
e
d
up
the
t
r
a
ini
ng
ti
me
of
the
model
.
F
igur
e
5
s
hows
the
a
r
c
hit
e
c
tur
e
of
the
pr
opos
e
d
model,
na
mely
UN
e
t
-
VG
G16
with
T
r
a
ns
f
e
r
L
e
a
r
ning,
while
F
igur
e
6
is
the
pr
oc
e
s
s
vis
ua
li
z
a
ti
on
of
im
a
ge
s
e
gmenta
ti
on
unde
r
the
ne
w
pr
opos
e
d
a
r
c
hit
e
c
tur
e
.
F
igur
e
5.
T
he
a
r
c
hit
e
c
t
ur
e
o
f
UN
e
t
-
VG
G16
with
tr
a
ns
f
e
r
lea
r
ning
F
igur
e
6.
T
he
s
e
gmenta
ti
on
pr
oc
e
s
s
unde
r
UN
e
t
-
VG
G16
with
tr
a
ns
f
e
r
lea
r
ning
3.
2
.
Choos
in
g
t
h
e
b
e
s
t
m
od
e
l
I
n
thi
s
s
tudy,
we
tr
y
to
c
ompar
e
the
p
r
opos
e
d
mod
e
l
with
the
p
r
e
vious
s
tate
of
the
a
r
t
model.
U
-
Ne
t
a
r
c
hit
e
c
tur
e
is
de
ve
loped
with
va
r
ious
s
c
e
na
r
ios
s
o
that
s
e
ve
r
a
l
a
lt
e
r
na
ti
ve
models
a
r
e
obtaine
d
w
hich
will
then
be
c
ompar
e
d
f
o
r
a
c
c
ur
a
c
y
be
twe
e
n
one
a
n
othe
r
a
nd
with
the
p
r
opos
e
d
model
.
T
he
s
e
mod
if
ica
ti
on
s
c
e
na
r
ios
of
U
-
Ne
t
a
r
e
done
s
ince
the
or
igi
na
l
of
U
-
Ne
t
model
c
a
nnot
r
un
in
ou
r
c
ompu
ter
with
the
s
pe
c
if
ica
ti
on
of
P
r
oc
e
s
s
or
I
n
tel
C
or
e
i7,
32
GB
R
AM
,
128GB
S
S
D,
a
nd
without
G
P
U
a
nd
VR
AM
.
T
a
ble
2
is
a
br
e
a
kdown
of
the
numbe
r
of
c
onvolut
i
on
laye
r
s
a
nd
c
onvolut
ion
blocks
in
e
a
c
h
U
-
Ne
t
s
c
e
na
r
io.
All
the
models
buil
d
unde
r
the
P
ython
pr
og
r
a
mm
i
ng
langua
ge
with
T
e
ns
or
f
low,
Ke
r
a
s
,
a
nd
NumP
y
li
br
a
r
ies
.
I
n
the
tr
a
ini
ng
pr
oc
e
s
s
,
we
us
e
100
e
poc
h
s
a
nd
e
xe
c
uted
in
a
c
o
mput
e
r
with
p
r
oc
e
s
s
or
I
ntel
C
or
e
i7,
32GB
R
AM
,
128GB
S
S
D
,
a
nd
without
GPU
a
nd
VR
AM
.
E
a
c
h
e
poc
h
of
the
p
r
opos
e
d
mod
e
l
take
s
80
mi
nutes
of
c
omput
e
r
pr
oc
e
s
s
ing.
I
t
is
long
e
r
than
the
f
our
s
c
e
na
r
ios
of
U
-
Ne
t
s
ince
the
number
of
pa
r
a
mete
r
s
of
the
p
r
opos
e
d
mod
e
l
is
mo
r
e
than
t
he
modi
f
ied
U
-
Ne
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
131
0
-
13
1
8
1316
T
a
ble
2.
C
onvolut
ion
s
c
e
na
r
io
f
o
r
U
-
Ne
t
S
c
e
na
r
io
N
umbe
r
of
C
onv la
ye
r
C
onv bloc
k
P
a
r
a
me
te
r
T
im
e
/e
poc
h
M
ode
l
1
1 C
onv @
bl
oc
k
5
bl
oc
k
3,929,985
23 mi
n
M
ode
l
2
1 C
onv @
bl
oc
k
4
bl
oc
k
980,097
20 mi
n
M
ode
l
3
2 C
onv @
bl
oc
k
5
bl
oc
k
7,862,113
34 mi
n
M
ode
l
4
2 C
onv @
bl
oc
k
4
bl
oc
k
1,962,337
32 mi
n
F
igur
e
7
de
mons
tr
a
tes
a
lea
r
ning
c
u
r
ve
of
f
ou
r
U
-
Ne
t
s
c
e
na
r
ios
c
ompar
e
s
with
the
UN
e
t
-
VG
G1
6
s
e
gmenta
ti
on
model
with
T
r
a
ns
f
e
r
L
e
a
r
ning.
F
i
gur
e
7
(
a
)
s
hows
the
c
ompa
r
is
on
of
the
los
s
va
lue
of
the
U
-
Ne
t
a
nd
the
pr
opos
e
d
model
,
while
F
igur
e
7
(
b
)
is
a
c
ompar
is
on
of
it
s
a
c
c
ur
a
c
y.
B
a
s
e
d
on
thi
s
f
igur
e
,
it
c
a
n
be
s
e
e
n
that
dur
ing
the
tr
a
ini
ng
pr
oc
e
s
s
with
100
e
poc
hs
,
the
pr
opos
e
d
model
pr
ovides
a
mi
ni
mu
m
los
s
va
lue
a
nd
maximum
a
c
c
ur
a
c
y
c
ompar
e
d
to
the
f
our
U
-
Ne
t
s
c
e
na
r
ios
.
M
or
e
ove
r
,
the
los
s
a
nd
a
c
c
ur
a
c
y
of
the
pr
opos
e
d
model
a
r
e
f
a
s
ter
c
onve
r
ge
nt
a
nd
s
t
a
ble
with
s
moot
he
r
moveme
nts
c
ompar
e
d
to
the
los
s
a
nd
a
c
c
ur
a
c
y
of
U
-
Ne
t
models
that
a
r
e
s
ti
ll
volatil
e
dur
ing
the
t
r
a
ini
ng
p
r
oc
e
s
s
.
T
a
ble
3
s
hows
the
ove
r
a
ll
pe
r
f
or
manc
e
of
e
a
c
h
model.
(
a
)
(
b)
F
igur
e
7.
(
a
)
L
os
s
a
nd
(
b)
a
c
c
ur
a
c
y
f
r
om
the
f
ou
r
U
-
Ne
t
s
c
e
na
r
ios
c
ompar
e
to
UN
e
t
-
VG
G16
with
tr
a
ns
f
e
r
lea
r
ning
F
r
om
T
a
ble
3
a
nd
F
igur
e
7
,
the
mi
nim
um
los
s
a
nd
maximum
a
c
c
ur
a
c
y
a
r
e
r
e
a
c
he
d
by
the
p
r
opos
e
d
model.
I
ts
los
s
va
lue
is
0
.
054
a
nd
the
a
c
c
ur
a
c
y
v
a
lue
is
0.
961
.
B
a
s
e
d
on
thes
e
r
e
s
ult
s
,
the
p
r
opos
e
d
model
gives
be
tt
e
r
pe
r
f
or
manc
e
than
the
f
our
s
c
e
na
r
ios
of
U
-
Ne
t.
S
ince
the
be
s
t
model
be
longs
to
UN
e
t
-
VG
G16
with
T
r
a
ns
f
e
r
L
e
a
r
ning,
f
or
the
ne
xt
a
na
lys
is
,
we
will
us
e
the
model
a
s
the
ba
s
is
o
f
M
R
I
br
a
in
tum
o
r
s
e
gmenta
ti
on.
3.
2
.
S
e
gm
e
n
t
at
ion
r
e
s
u
lt
s
b
as
e
d
on
t
h
e
b
e
s
t
m
od
e
l
T
he
s
e
gmenta
ti
on
is
done
to
the
16
-
tes
ti
ng
da
ta
unde
r
UN
e
t
-
VG
G16
with
T
r
a
ns
f
e
r
L
e
a
r
ning.
T
he
vis
ua
li
z
a
ti
on
of
s
e
gmenta
ti
on
r
e
s
ult
s
is
de
s
c
r
i
be
d
in
F
igu
r
e
8
.
T
his
f
igu
r
e
ha
s
s
hown
that
s
e
gmenta
ti
on
r
e
s
ult
s
of
s
a
mpl
e
s
e
que
nc
e
c
ould
r
e
c
ognize
the
tu
mor
a
r
e
a
a
s
R
OI
in
va
r
ious
tum
o
r
s
ize
a
nd
loca
ti
on,
bo
th
on
the
lef
t
or
r
ight
of
the
br
a
in
.
F
or
the
tes
ti
ng
d
a
ta,
we
c
a
lcula
te
the
CCR
a
s
the
mea
s
ur
e
o
f
e
va
luation
a
nd
the
s
umm
a
r
y
is
s
hown
in
T
a
ble
4.
T
he
s
e
gmenta
ti
on
r
e
s
ult
s
a
r
e
a
ppr
oa
c
hing
gr
ound
t
r
uth
ve
r
y
we
ll
s
ince
the
a
ll
CCR
va
lue
a
bove
90%
.
T
he
C
CR
gr
a
nd
mea
n
f
or
a
ll
tes
ti
ng
da
ta
is
r
e
a
c
hing
95.
69
%
.
T
a
ble
3.
T
he
pe
r
f
or
manc
e
c
ompar
is
on
of
e
a
c
h
mo
de
l
M
ode
l
c
ompa
r
is
on
L
os
s
A
c
c
ur
a
c
y
U
-
N
e
t
:
M
ode
l
1
0.124
0.942
M
ode
l
2
0.083
0.951
M
ode
l
3
0.085
0.953
M
ode
l
4
0.244
0.938
U
N
e
t
-
VGG
16 w
it
h T
L
0.054
0.961
T
a
ble
4.
T
he
s
umm
a
r
y
of
CCR
f
o
r
tes
ti
ng
da
ta
T
e
s
ti
ng
numbe
r
CCR
T
e
s
ti
ng
numbe
r
CCR
1
0.957184
9
0.957489
2
0.970047
10
0.965225
3
0.916245
11
0.961166
4
0.935806
12
0.905029
5
0.982773
13
0.933807
6
0.981598
14
0.928635
7
0.979904
15
0.985748
8
0.974136
16
0.975601
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
UN
e
t
-
V
GG
16
w
it
h
tr
ans
fer
lear
ning
for
M
R
I
-
bas
e
d
br
ain
tumor
…
(
A
nindya
A
pr
il
iyanti
P
r
av
it
as
ar
i
)
1317
I
nput
s
l
i
c
e
gr
ou
nd
t
r
ut
h
s
e
g
m
e
n
t
.
r
e
s
ul
t
I
nput
s
l
i
c
e
gr
ou
nd
t
r
ut
h
s
e
g
m
e
n
t
.
r
e
s
ul
t
D
a
ta
te
s
t
1
D
a
ta
te
s
t
9
D
a
ta
te
s
t
2
D
a
ta
te
s
t
10
D
a
ta
te
s
t
3
D
a
ta
te
s
t
11
D
a
ta
te
s
t
4
D
a
ta
te
s
t
12
D
a
ta
te
s
t
5
D
a
ta
te
s
t
13
D
a
ta
te
s
t
6
D
a
ta
te
s
t
14
D
a
ta
te
s
t
7
D
a
ta
te
s
t
15
D
a
ta
te
s
t
8
D
a
ta
te
s
t
16
]
F
igur
e
8.
T
he
s
e
gmenta
ti
on
r
e
s
ult
s
of
16
tes
ti
ng
da
ta
4.
CONC
L
USI
ON
B
a
s
e
d
on
the
r
e
s
ult
s
a
nd
a
na
lys
i
s
we
c
a
n
c
on
c
lude
that
the
pr
opos
e
d
model
na
mely
UN
e
t
-
VG
G16
with
T
r
a
ns
f
e
r
L
e
a
r
ning
is
r
unning
we
ll
on
the
c
omput
e
r
with
a
pr
oc
e
s
s
or
of
I
ntel
C
or
e
i7,
32G
B
R
AM
,
128GB
S
S
D,
a
nd
without
GPU
a
nd
VR
AM
.
T
he
pr
opos
e
d
model
ha
s
gr
e
a
t
pe
r
f
or
manc
e
c
ompar
e
d
to
the
U
-
Ne
t
model
(
in
f
our
s
c
e
na
r
ios
)
s
ince
it
ha
s
the
mi
nim
um
va
lue
of
los
s
a
nd
maximum
va
lue
of
a
c
c
ur
a
c
y.
T
he
s
e
gmenta
ti
on
r
e
s
ult
s
unde
r
the
pr
opos
e
d
mode
l
tend
to
a
ppr
oa
c
h
the
R
OI
tar
ge
t
of
e
a
c
h
br
a
in
tu
mor
M
R
I
im
a
ge
ve
r
y
we
ll
.
T
he
r
e
s
ult
s
of
s
e
gmenta
ti
on
f
r
om
tes
ti
ng
da
ta
we
r
e
obtaine
d
by
C
C
R
va
lue
o
f
95.
69
%
.
F
or
f
utu
r
e
r
e
s
e
a
r
c
h,
the
dif
f
e
r
e
nt
a
r
c
hit
e
c
tur
e
or
c
onvolut
ional
block
s
c
e
na
r
io
c
ould
be
obtaine
d
to
ge
t
mor
e
a
lt
e
r
na
ti
ve
models
.
Not
to
mention
that
the
op
ti
m
um
e
poc
h
is
s
ti
ll
de
mande
d
to
ge
t
the
opti
mal
c
o
mput
ing
ti
me
f
or
buil
ding
the
t
r
a
ini
ng
model
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
131
0
-
13
1
8
1318
AC
KNOWL
E
DGE
M
E
NT
S
T
he
a
uthor
s
a
r
e
gr
a
tef
u
l
to
the
Dir
e
c
tor
a
te
f
o
r
R
e
s
e
a
r
c
h
a
nd
C
omm
unit
y
S
e
r
vice
(
DR
P
M
)
M
ini
s
tr
y
of
R
e
s
e
a
r
c
h,
T
e
c
hnology
a
nd
Highe
r
E
duc
a
ti
on
I
ndone
s
ia
whic
h
s
uppor
ts
thi
s
r
e
s
e
a
r
c
h
unde
r
P
T
r
e
s
e
a
r
c
h
gr
a
nt
no.
944/P
KS/
I
T
S
/2019
.
RE
F
E
RE
NC
E
S
[1
]
G
l
o
b
o
can
,
“
In
d
o
n
es
ia
S
o
u
r
ce:
G
l
o
b
o
c
a
n
2
0
1
8
,
”
2
0
1
9
.
[
O
n
l
i
n
e].
A
v
ai
b
l
e
:
h
t
t
p
:
/
/
g
co
.
i
arc.
fr/
t
o
d
a
y
/
d
at
a
/
fac
t
s
h
e
et
s
/
p
o
p
u
l
at
i
o
n
s
/
3
6
0
-
I
n
d
o
n
e
s
i
a
-
fac
t
-
s
h
eet
s
.
p
d
f
.
[2
]
V
.
J
ai
s
w
a
l
,
V
.
Sh
arma,
S.
V
arma.
,
“A
n
Imp
l
eme
n
t
a
t
i
o
n
o
f
N
o
v
e
l
G
en
et
i
c
-
Ba
s
ed
C
l
u
s
t
er
i
n
g
A
l
g
o
ri
t
h
m
fo
r
C
o
l
o
r
Imag
e
Seg
men
t
at
i
o
n
,”
TE
LKO
M
NIK
A
Tel
eco
m
m
u
n
i
c
a
t
i
o
n
Co
m
p
u
t
i
n
g
E
l
ec
t
r
o
n
i
cs
a
n
d
Co
n
t
r
o
l
,
v
o
l
1
7
,
n
o
3
,
p
p
1
4
6
1
-
1
4
6
7
,
J
u
n
e
2
0
1
9
.
[3
]
P.
B.
Prak
o
s
o
,
Y
.
Sar
i
.
,
“V
e
h
i
c
l
e
D
e
t
ect
i
o
n
u
s
i
n
g
Back
g
r
o
u
n
d
S
u
b
t
rac
t
i
o
n
a
n
d
C
l
u
s
t
er
i
n
g
A
l
g
o
r
i
t
h
m
s
,”
TE
LKO
M
NIK
A
Tel
ec
o
m
m
u
n
i
c
a
t
i
o
n
Co
m
p
u
t
i
n
g
E
l
ec
t
r
o
n
i
c
s
a
n
d
Co
n
t
r
o
l
,
v
o
l
1
7
,
n
o
3
,
p
p
1
3
9
3
-
1
3
9
8
,
J
u
n
e
2
0
1
9
.
[4
]
N
.
Iri
aw
a
n
,
A
.
A
.
Prav
i
t
as
ar
i
,
K
.
Fi
t
h
ri
a
s
ari
,
Irh
ama
h
,
S.
W
.
Pu
rn
ami
,
W
.
Ferri
a
s
t
u
t
i
.
,
“Co
mp
ara
t
i
v
e
St
u
d
y
o
f
Brai
n
T
u
m
o
r
Seg
me
n
t
a
t
i
o
n
u
s
i
n
g
D
i
ffere
n
t
Se
g
me
n
t
a
t
i
o
n
T
ec
h
n
i
q
u
es
i
n
H
an
d
l
i
n
g
N
o
i
s
e
,”
2
0
1
8
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
er
e
n
ce
o
n
Co
m
p
u
t
e
r
E
n
g
i
n
ee
r
i
n
g
,
Net
w
o
r
k
a
n
d
I
n
t
e
l
l
i
g
e
n
t
M
u
l
t
i
m
e
d
i
a
(C
E
NIM
),
Su
ra
b
ay
a,
In
d
o
n
es
i
a,
p
p
.
2
8
9
-
2
9
3
,
2
0
1
8
.
[5
]
A
.
A
.
Prav
i
t
a
s
ar
i
,
M.
A
.
I.
Safa,
N
.
Iri
aw
an
,
Irh
ama
h
,
K
.
Fi
t
h
r
i
as
a
ri
,
S.
W
.
Pu
rn
ami
,
W
.
Ferri
a
s
t
u
t
i
,
“MRI
-
Bas
e
d
Brai
n
T
u
m
o
r
Seg
men
t
at
i
o
n
u
s
i
n
g
Mo
d
i
fie
d
St
ab
l
e
St
u
d
e
n
t
’s
t
fro
m
Bu
rr
Mi
x
t
u
re
Mo
d
el
w
i
t
h
Bay
es
i
an
A
p
p
ro
a
c
h
,”
M
a
l
a
y
s
i
a
n
Jo
u
r
n
a
l
o
f
M
a
t
h
e
m
a
t
i
c
a
l
S
ci
e
n
ces
,
v
o
l
1
3
,
n
o
3
,
p
p
.
2
9
7
-
3
1
0
,
2
0
1
9
.
[6
]
A
.
A
.
Prav
i
t
a
s
ari
,
N
.
I.
N
i
rma
l
as
ar
i
,
N
.
Iri
a
w
an
,
Ir
h
ama
h
,
K
.
Fi
t
h
r
i
as
ar
i
,
S.
W
.
Pu
rn
am
i
,
W
.
Ferr
i
as
t
u
t
i
,
“Bay
e
s
i
an
Sp
at
i
al
l
y
co
n
s
t
rai
n
ed
Fern
a
n
d
e
z
-
St
ee
l
Sk
ew
N
o
rmal
M
i
x
t
u
re
mo
d
el
fo
r
MRI
-
b
as
e
d
Brai
n
T
u
mo
r
Se
g
men
t
at
i
o
n
,
”
In
A
IP
C
o
n
f
er
e
n
ce
P
r
o
ceed
i
n
g
s
,
v
o
l
.
2
1
9
4
,
n
o
.
1
,
D
e
ce
mb
er
2
0
1
9
.
[7
]
M.
Pat
h
ak
,
N
.
Sri
n
i
v
a
s
u
,
V
.
Bai
rag
i
,
“E
ffect
i
v
e
S
eg
men
t
at
i
o
n
o
f
Scl
era,
Iri
s
,
an
d
Pu
p
i
l
i
n
N
o
i
s
y
E
y
e
Imag
es
,
”
TE
LK
O
M
NIK
A
Tel
e
co
m
m
u
n
i
ca
t
i
o
n
C
o
m
p
u
t
i
n
g
E
l
ec
t
r
o
n
i
cs
a
n
d
Co
n
t
r
o
l
,
v
o
l
1
7
,
n
o
5
,
p
p
2
3
4
6
-
2
3
5
4
,
O
ct
o
b
er
2
0
1
9
.
[
8
]
I
.
B
.
K
.
S
u
d
i
a
t
m
i
k
a
,
F
.
R
a
h
m
a
n
,
T
r
i
s
n
o
,
S
u
y
o
t
o
,
“
I
m
a
g
e
F
o
r
g
e
r
y
D
e
t
e
c
t
i
o
n
U
s
i
n
g
E
r
r
o
r
L
e
v
e
l
A
n
a
l
y
s
i
s
a
n
d
D
e
e
p
L
e
a
r
n
i
n
g
,
”
T
E
L
K
O
M
N
I
K
A
T
e
l
e
c
o
m
m
u
n
i
c
a
t
i
o
n
C
o
m
p
u
t
i
n
g
E
l
e
c
t
r
o
n
i
c
s
a
n
d
C
o
n
t
r
o
l
,
v
o
l
1
7
,
n
o
2
,
p
p
6
5
3
-
6
5
9
,
A
p
r
i
l
2
0
1
9
.
[
9
]
E
.
G
i
b
s
o
n
,
W
.
L
i
,
C
.
H
.
S
u
d
r
e
,
L
.
F
i
d
o
n
,
D
.
S
h
a
k
i
r
,
G
.
W
a
n
g
,
Z
.
E
a
t
o
n
-
R
o
s
e
n
,
R
.
G
r
a
y
,
T
.
D
o
e
l
,
Y
.
H
u
,
T
.
W
h
y
n
t
i
e
,
P
.
N
a
c
h
e
v
,
D
.
C
.
B
a
r
r
a
t
t
,
S
.
O
u
r
s
e
l
i
n
,
M
.
J
.
C
a
r
d
o
s
o
,
T
.
V
e
r
c
a
u
t
e
r
e
n
.
,
“
N
i
f
t
y
n
e
t
:
a
D
e
e
p
-
L
e
a
r
n
i
n
g
P
l
a
t
f
o
r
m
f
o
r
M
e
d
i
c
a
l
I
m
a
g
i
n
g
,
”
C
o
m
p
u
t
e
r
M
e
t
h
o
d
s
a
n
d
P
r
o
g
r
a
m
s
i
n
B
i
o
m
e
d
i
c
i
n
e
,
v
o
l
.
1
5
8
,
p
p
.
1
1
3
-
1
2
2
,
M
a
y
2
0
1
8
.
[1
0
]
P.
V
.
T
ran
,
“A
F
u
l
l
y
C
o
n
v
o
l
u
t
i
o
n
a
l
N
e
u
ral
N
et
w
o
r
k
fo
r
Card
i
ac
Se
g
men
t
at
i
o
n
i
n
Sh
o
rt
-
A
x
i
s
MRI
,
”
arX
i
v
p
re
p
ri
n
t
arX
i
v
:
1
6
0
4
.
0
0
4
9
4
,
2
0
1
6
.
[1
1
]
Y
.
L
i
,
H
.
Q
i
,
J
.
D
ai
,
X
.
J
i
,
Y
.
W
ei
,
“F
u
l
l
y
C
o
n
v
o
l
u
t
i
o
n
a
l
In
s
t
a
n
ce
-
A
w
are
Sema
n
t
i
c
Seg
me
n
t
a
t
i
o
n
,
”
2
0
1
7
I
E
E
E
Co
n
f
er
e
n
ce
o
n
C
o
m
p
u
t
er
V
i
s
i
o
n
a
n
d
P
a
t
t
er
n
R
ec
o
g
n
i
t
i
o
n
(CV
P
R
),
H
o
n
o
l
u
l
u
,
H
I,
p
p
.
4
4
3
8
-
4
4
4
6
,
2
0
1
7
.
[1
2
]
O
.
Ro
n
n
eb
er
g
e
r,
P.
Fi
s
ch
er,
T
.
Bro
x
,
“U
-
N
e
t
:
C
o
n
v
o
l
u
t
i
o
n
a
l
N
et
w
o
r
k
s
f
o
r
Bi
o
me
d
i
ca
l
Imag
e
Se
g
men
t
at
i
o
n
,
”
M
ed
i
ca
l
Im
a
g
e
Co
m
p
u
t
i
n
g
a
n
d
Co
m
p
u
t
e
r
-
A
s
s
i
s
t
ed
In
t
er
ve
n
t
i
o
n
-
M
ICC
A
I
2
0
1
5
,
v
o
l
.
9
3
5
1
,
p
p
.
2
3
4
-
2
4
1
,
N
o
v
emb
er
2
0
1
5
.
[1
3
]
C.
Bal
ak
ri
s
h
n
a,
D
.
Sars
h
ar,
S.
So
l
t
an
i
n
e
j
ad
,
“A
u
t
o
mat
i
c
d
et
ec
t
i
o
n
o
f
l
u
me
n
an
d
med
i
a
i
n
t
h
e
IV
U
S
i
ma
g
es
u
s
i
n
g
U
-
N
e
t
w
i
t
h
V
G
G
1
6
E
n
c
o
d
er
,”
a
r
X
i
v
p
r
e
p
r
i
n
t
a
r
X
i
v:1
8
0
6
.
0
7
5
5
4
,
J
u
n
e
2
0
1
8
.
[1
4
]
J
.
L
o
n
g
,
E
.
Sh
e
l
h
amer,
T
.
D
arre
l
l
,
“Fu
l
l
y
Co
n
v
o
l
u
t
i
o
n
a
l
N
et
w
o
r
k
s
f
o
r
Seman
t
i
c
Seg
me
n
t
a
t
i
o
n
,
”
2
0
1
5
IE
E
E
Co
n
f
er
e
n
ce
o
n
C
o
m
p
u
t
er
V
i
s
i
o
n
a
n
d
P
a
t
t
er
n
R
ec
o
g
n
i
t
i
o
n
(CV
P
R
),
Bo
s
t
o
n
,
MA
,
p
p
.
3
4
3
1
-
3
4
4
0
,
2
0
1
5
.
[1
5
]
Y
.
L
e
Cu
n
,
Y
.
Ben
g
i
o
,
“Co
n
v
o
l
u
t
i
o
n
al
N
e
t
w
o
rk
s
fo
r
Imag
es
,
Sp
eech
,
an
d
T
i
me
Seri
e
s
,
”
H
a
n
d
B
r
a
i
n
Th
eo
r
y
Neu
r
a
l
Net
w
,
p
p
3
3
6
1
-
3
3
7
1
,
J
an
u
ary
1
9
9
5
.
[1
6
]
H
.
C.
Sh
i
n
,
H
.
R.
Ro
t
h
,
M.
G
ao
,
L
.
L
u
,
Z
.
X
u
,
I.
N
o
g
u
es
,
J
.
Y
ao
,
D
.
Mo
l
l
u
ra,
R.
M
Su
mmer
s
,
“D
ee
p
Co
n
v
o
l
u
t
i
o
n
a
l
N
eu
ra
l
N
et
w
o
r
k
s
f
o
r
Co
m
p
u
t
er
-
ai
d
ed
D
et
ec
t
i
o
n
:
C
N
N
A
rc
h
i
t
ect
u
re
s
,
D
at
a
s
et
C
h
aract
er
i
s
t
i
c
s
,
an
d
T
ran
s
fer
L
earn
i
n
g
,
”
in
IE
E
E
T
r
a
n
s
a
ct
i
o
n
s
o
n
M
e
d
i
c
a
l
Im
a
g
i
n
g
,
v
o
l
.
3
5
,
n
o
.
5
,
p
p
.
1
2
8
5
-
1
2
9
8
,
May
2
0
1
6
.
[1
7
]
F.
Ch
o
l
l
e
t
,
“
D
ee
p
L
earn
i
n
g
w
i
t
h
Py
t
h
o
n
,”
M
a
n
n
i
n
g
:
S
h
e
l
t
e
r
Is
l
a
n
d
,
2
0
1
8
.
[1
8
]
Y
.
L
ecu
n
,
L
.
Bo
t
t
o
u
,
Y
.
Ben
g
i
o
,
P.
H
affn
er,
“G
ra
d
i
en
t
-
b
as
e
d
L
earn
i
n
g
A
p
p
l
i
e
d
t
o
D
o
c
u
men
t
Reco
g
n
i
t
i
o
n
,
”
P
r
o
ceed
i
n
g
s
o
f
t
h
e
IE
E
E
,
v
o
l
.
8
6
,
n
o
.
1
1
,
p
p
.
2
2
7
8
-
2
3
2
4
,
N
o
v
.
1
9
9
8
.
[1
9
]
A
.
K
ri
z
h
ev
s
k
y
,
I.
Su
t
s
k
ev
er,
G
.
E
.
H
i
n
t
o
n
,
“Imag
en
e
t
Cl
as
s
i
fi
cat
i
o
n
w
i
t
h
D
eep
Co
n
v
o
l
u
t
i
o
n
al
N
eu
ra
l
N
et
w
o
r
k
s
,
”
i
n
A
d
va
n
ces
i
n
n
e
u
r
a
l
i
n
f
o
r
m
a
t
i
o
n
p
r
o
ces
s
i
n
g
s
ys
t
em
s
,
p
p
1
0
9
7
-
1
1
0
5
,
2
0
1
2
.
[2
0
]
M.
D
.
Z
ei
l
er,
R.
Ferg
u
s
,
“V
i
s
u
al
i
zi
n
g
an
d
U
n
d
er
s
t
a
n
d
i
n
g
Co
n
v
o
l
u
t
i
o
n
a
l
N
et
w
o
r
k
s
,
”
E
u
r
o
p
ea
n
co
n
f
e
r
en
ce
o
n
co
m
p
u
t
er
vi
s
i
o
n
,
Sp
r
i
n
g
er,
Ch
am,
v
o
l
.
8
6
8
9
,
p
p
8
1
8
-
8
3
3
,
2
0
1
4
.
[2
1
]
A
.
K
ami
l
ar
i
s
,
F.
A
Pren
afet
a
-
Bo
l
d
ú
,
“A
Rev
i
e
w
o
f
T
h
e
U
s
e
o
f
Co
n
v
o
l
u
t
i
o
n
al
N
eu
ral
N
et
w
o
r
k
s
i
n
A
g
r
i
cu
l
t
u
r
e
,
”
Th
e
Jo
u
r
n
a
l
o
f
A
g
r
i
cu
l
t
u
r
a
l
S
ci
e
n
ce
,
p
p
1
–
1
1
,
2
0
1
8
.
[2
2
]
A
.
Z
h
an
g
,
Z
.
C.
L
i
p
t
o
n
,
M.
L
i
,
A
.
J
.
Smo
l
a
,
“
D
i
v
e
i
n
t
o
D
eep
L
earn
i
n
g
,”
S
A
G
E
P
u
b
l
i
c
a
t
i
o
n
s
In
c
,
2
0
1
9
.
[2
3
]
D
.
P.
K
i
n
g
ma,
J
.
Ba
,
“A
d
am:
A
Me
t
h
o
d
fo
r
S
t
o
c
h
as
t
i
c
O
p
t
i
mi
zat
i
o
n
,
”
arX
i
v
p
rep
r
i
n
t
arX
i
v
:
1
4
1
2
.
6
9
8
0
,
2
0
1
4
.
[2
4
]
S.
Red
d
i
,
S.
K
al
e,
S.
K
u
mar,
“O
n
t
h
e
C
o
n
v
erg
e
n
ce
o
f
A
d
am
an
d
Be
y
o
n
d
,
”
arX
i
v
p
re
p
ri
n
t
ar
X
i
v
:
1
9
0
4
.
0
9
2
3
7
,
2
0
1
9
.
[2
5
]
C.
N
i
k
o
u
,
N
.
P.
G
al
at
s
an
o
s
,
A
.
C.
L
i
k
as
,
“A
Cl
as
s
-
ad
a
p
t
i
v
e
Sp
at
i
al
l
y
V
ar
i
an
t
Mi
x
t
u
re
Mo
d
e
l
fo
r
Imag
e
Seg
men
t
at
i
o
n
,
”
in
IE
E
E
T
r
a
n
s
a
ct
i
o
n
s
o
n
Im
a
g
e
P
r
o
ce
s
s
i
n
g
,
v
o
l
.
1
6
,
n
o
.
4
,
p
p
.
1
1
2
1
-
1
1
3
0
,
A
p
r
i
l
2
0
0
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.