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.
1
8
,
No.
3
,
J
une
2020
,
pp.
1522
~
1528
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
.
v1
8
i
3
.
14839
1522
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
Pu
lmon
a
r
y r
o
n
t
g
e
n
c
la
ssi
f
ic
at
io
n
t
o d
e
t
e
c
t
p
n
e
u
m
on
ia
d
is
e
a
se
u
si
n
g c
o
n
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
or
k
s
Z
u
h
e
r
m
an
Rus
t
am
,
Rivan
P
r
at
a
m
a
Yud
a
,
Ham
im
ah
Alat
as
,
Chelvian
Ar
oe
f
D
ep
ar
t
men
t
o
f
Mah
ema
t
i
c
s
,
U
n
i
v
er
s
i
t
y
o
f
I
n
d
o
n
e
s
i
a,
In
d
o
n
e
s
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
14
,
2020
Ac
c
e
pted
F
e
b
24
,
2020
E
v
er
y
o
rg
a
n
i
s
m
i
s
k
n
o
w
n
t
o
h
a
v
e
d
i
ffere
n
t
s
t
ru
c
t
u
ra
l
an
d
b
i
o
l
o
g
i
cal
s
y
s
t
em,
s
p
ec
i
fi
ca
l
l
y
i
n
h
u
ma
n
i
mmu
n
i
t
y
.
If
t
h
e
i
mmu
n
e
s
y
s
t
em
w
eak
e
n
s
,
t
h
e
b
o
d
y
i
s
s
u
s
cep
t
i
b
l
e
t
o
d
i
s
eas
e
es
p
eci
a
l
l
y
p
n
eu
m
o
n
i
a
d
i
s
eas
e.
Pn
eu
mo
n
i
a
d
i
s
ea
s
e
i
s
cau
s
e
d
b
y
t
h
e
b
ac
t
eri
u
m
St
re
p
t
o
c
o
cc
u
s
p
n
e
u
mo
n
i
a
,
an
d
acco
r
d
i
n
g
t
o
t
h
e
W
o
rl
d
H
eal
t
h
O
rg
a
n
i
za
t
i
o
n
(
W
H
O
),
i
t
i
s
i
d
e
n
t
i
fi
e
d
a
s
t
h
e
l
ea
d
i
n
g
ca
u
s
e
o
f
d
eat
h
i
n
ch
i
l
d
ren
w
o
rl
d
w
i
d
e,
w
h
i
c
h
i
s
a
b
o
u
t
1
6
%
,
fo
r
t
h
o
s
e
u
n
d
er
t
h
e
a
g
e
o
f
5
.
Mean
w
h
i
l
e,
s
o
meo
n
e
w
h
o
i
s
p
re
d
i
c
t
ed
t
o
h
av
e
p
n
e
u
mo
n
i
a
b
y
a
d
o
ct
o
r
i
s
reco
mmen
d
ed
fo
r
a
n
X
-
ra
y
.
Co
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ral
n
et
w
o
r
k
s
(CN
N
s
)
i
s
an
accu
rat
e
met
h
o
d
t
o
h
el
p
t
h
e
d
o
ct
o
r's
p
re
d
i
c
t
ed
co
rre
c
t
l
y
.
CN
N
s
i
s
d
i
v
i
d
ed
i
n
t
o
t
w
o
i
mp
o
rt
a
n
t
p
ar
t
s
,
feat
u
re
ex
t
rac
t
i
o
n
l
a
y
er
(co
n
v
o
l
u
t
i
o
n
a
l
l
ay
er
an
d
p
o
o
l
i
n
g
l
ay
e
r
)
an
d
f
u
l
l
y
c
o
n
n
e
ct
e
d
l
a
y
er
.
CN
N
s
me
t
h
o
d
i
s
c
o
mmo
n
l
y
u
s
ed
fo
r
i
ma
g
e
d
a
t
a
cl
a
s
s
i
fi
ca
t
i
o
n
.
T
h
eref
o
re,
CN
N
s
i
s
s
u
i
t
a
b
l
e
t
o
c
l
as
s
i
f
y
p
n
e
u
mo
n
i
a
b
as
e
d
o
n
l
u
n
g
X
-
ray
i
n
o
r
d
er
t
o
o
b
t
ai
n
accu
r
at
e
p
red
i
c
t
i
o
n
res
u
l
t
s
.
A
n
d
t
h
en
,
t
h
e
res
u
l
t
s
ca
n
b
e
s
ee
n
b
as
e
d
o
n
t
h
e
g
rap
h
o
f
t
h
e
acc
u
ra
cy
v
al
u
e
an
d
t
h
e
l
o
s
s
v
al
u
e.
W
h
e
n
CN
N
s
met
h
o
d
ap
p
l
i
ed
o
n
t
h
e
d
at
as
e
t
,
an
accu
racy
r
a
t
e
o
f
9
7
%
w
as
o
b
t
a
i
n
e
d
.
B
a
s
ed
o
n
a
c
c
u
ra
c
y
r
a
t
e
,
i
t
s
h
o
w
s
t
h
a
t
C
N
N
s
c
a
n
b
e
a
p
p
l
i
e
d
t
o
i
m
ag
e
d
at
a
(
e
s
p
e
ci
a
l
l
y
l
u
n
g
X
-
r
a
y
)
f
o
r
c
l
a
s
s
i
fi
c
a
t
i
o
n
o
f
p
n
eu
m
o
n
i
a
d
i
s
e
as
e
.
K
e
y
w
o
r
d
s
:
C
las
s
if
ica
ti
on
C
NN
s
P
ne
umoni
a
X
-
r
a
y
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
:
Z
uhe
r
man
R
us
tam,
De
pa
r
tm
e
nt
of
M
a
he
matics
,
Unive
r
s
it
y
of
I
ndone
s
ia,
De
pok,
16424,
I
ndone
s
ia.
E
mail:
r
us
tam@ui.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
Huma
ns
a
r
e
li
ving
c
r
e
a
tur
e
with
the
pos
s
e
s
s
ion
of
a
mi
nd
a
nd
he
a
r
t
,
whic
h
s
e
r
ve
s
a
s
the
dis
ti
nguis
hing
f
a
c
tor
f
r
om
other
c
r
e
a
tur
e
s
[
1]
.
M
e
a
nwhile,
the
body
a
na
tom
y
indi
c
a
tes
the
pr
e
s
e
nc
e
of
numer
ous
s
ys
tems
,
or
ga
ns
a
nd
ti
s
s
ue
s
,
one
of
whic
h
is
that
f
or
im
muni
ty
[
1]
,
a
nd
while
talking
a
bou
t
it
ther
e
is
a
c
e
r
tain
r
e
lations
hip
wi
th
he
a
lt
h
.
F
ur
ther
mo
r
e
,
it
is
be
tt
e
r
known
a
s
a
s
ys
tem
that
e
xis
ts
to
wor
k
in
the
dir
e
c
ti
on
of
p
r
otec
ti
on
f
r
om
outs
ide
inf
luenc
e
s
,
e
nc
ompas
s
ing
vir
us
e
s
or
ba
c
ter
ia
[
1]
.
T
he
r
e
f
or
e
,
i
ns
tanc
e
s
whe
r
e
thi
s
de
f
e
ns
e
is
una
ble
to
f
uf
il
it
s
f
unc
ti
on,
t
he
r
e
is
a
higher
tende
nc
y
of
be
ing
a
f
f
e
c
ted
by
dis
e
a
s
e
s
,
due
to
pa
thogen
invas
ion.
T
hus
,
the
in
it
iation
o
f
a
bn
or
mal
c
ondit
ion
in
the
body
or
mi
nd,
whic
h
is
c
a
pa
ble
of
c
a
us
ing
dis
c
omf
or
t,
dys
f
unc
ti
on
or
d
if
f
iculty
to
th
e
pe
r
s
on,
is
c
a
ll
e
d
i
ll
ne
s
s
[
1]
.
F
ur
the
r
mor
e
,
dis
e
a
s
e
s
oc
c
ur
i
n
numer
ous
types
,
r
a
nging
f
r
om
in
f
e
c
ti
ous
,
non
-
c
omm
unica
ble,
mi
ld,
moder
a
te
a
nd
s
e
ve
r
e
,
a
nd
a
ls
o,
the
r
e
is
a
tende
nc
y
to
a
tt
a
c
k
a
ll
a
ge
s
a
nd
gr
oups
,
inl
uding
toddl
e
r
s
,
c
hil
dr
e
n,
a
dult
s
a
nd
e
lder
ly
.
I
n
a
ddit
ion,
they
a
r
e
numer
ous
,
c
a
us
ing
ve
r
y
ter
r
ibl
e
c
onv
e
r
s
a
ti
on
a
m
ong
the
pe
ople
a
s
we
ll
a
s
the
bigges
t
a
nd
mos
t
s
igni
giant
ini
ti
a
tor
of
de
a
th.
T
he
r
e
f
or
e
,
d
is
e
a
s
e
is
s
ti
ll
a
s
e
r
i
ous
pr
oblem
to
be
de
a
lt
with
in
e
ve
r
y
c
ount
r
y,
i
nc
ludi
ng
I
ndone
s
ia.
T
his
is
indi
c
a
ted
by
the
inc
r
e
a
s
ing
mor
talit
y
r
a
te,
ba
s
e
d
on
the
late
s
t
global
he
a
lt
h
in
de
x,
it
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
P
ulmonar
y
r
ontgen
c
las
s
if
ication
to
de
tec
t
pne
umo
nia
…
(
Z
uhe
r
man
R
us
tam
)
1523
r
a
nke
d
101
out
of
149
c
ountr
ies
a
c
c
or
ding
to
the
2017
r
e
por
t
o
f
the
lega
tum
p
r
os
pe
r
it
y
index
.
How
e
ve
r
,
ther
e
is
a
lwa
ys
a
c
onti
nuous
s
tr
ive
a
nd
e
xpe
r
im
e
nt
to
identif
y
the
r
ight
tr
e
a
tm
e
nt
f
or
c
ur
e
s
.
One
of
the
da
nge
r
ous
dis
e
a
s
e
s
c
a
us
ing
de
a
th
wor
ldwide,
e
s
pe
c
ially
in
I
ndone
s
ia
is
pne
umoni
a
[
2]
,
c
omm
only
known
a
s
we
t
lung,
whic
hi
is
inf
e
c
ti
ous
a
nd
a
tt
a
c
ks
the
lung
s
,
he
nc
e
,
the
a
ir
s
a
c
s
be
c
ome
inf
lame
d
a
nd
s
wollen.
T
his
pos
s
e
s
s
e
s
the
a
bil
it
y
to
a
tt
a
c
k
a
ll
wa
lks
of
li
f
e
,
a
lt
ho
ugh
mor
talit
y
is
mor
e
domi
na
nt
in
inf
a
nts
a
nd
toddl
e
r
s
.
Ac
c
or
ding
to
the
W
or
ld
He
a
lt
h
Or
ga
niza
ti
on
(
W
HO
)
it
is
the
lea
ding
c
a
us
e
of
c
hil
dr
e
n
(
inf
a
nts
a
nd
toddl
e
r
s
)
mor
talit
y
pe
r
w
e
e
k,
whic
h
is
a
bout
16%
o
f
de
a
ths
a
t
the
a
ge
o
f
unde
r
the
a
ge
of
5
.
B
a
s
e
d
on
the
s
tate
ment
of
the
he
a
d
of
s
ub
dir
e
c
tor
a
te
(
Ka
s
ubdit
)
o
f
a
c
ute
r
e
s
pir
a
tor
y
i
nf
e
c
ti
on
(
I
S
P
A)
,
I
ndone
s
ia
is
r
a
nke
d
10th
in
the
wor
ld
,
in
c
a
s
e
s
of
pne
umoni
a
kil
ler
vir
us
a
tt
a
c
ks
,
with
15.
5
%
a
t
2015,
or
the
oc
c
ur
r
e
nc
e
o
f
a
bout
554
,
650
c
a
s
e
s
,
a
nd
it
s
c
a
us
e
is
of
ten
thr
ough
va
r
ious
mea
ns
,
inl
uding
v
ir
us
e
s
or
ba
c
ter
ia.
How
e
ve
r
,
it
is
ge
ne
r
a
ll
y
ins
ti
ga
ted
by
S
tr
e
ptococ
c
us
pne
umoni
a
,
whic
h
is
the
major
c
a
us
e
i
n
inf
a
nts
,
a
t
a
bout
50%
,
while
the
r
e
maining
is
due
to
vir
us
e
s
a
nd
s
im
il
a
r
ba
c
ter
ia
[
1,
2]
.
B
a
s
e
d
on
thi
s
,
it
pos
s
ibl
e
to
s
a
y
that
pne
umoni
a
is
not
a
n
indi
s
c
r
im
inate
dis
e
a
s
e
,
a
nd
it
a
ls
o
r
e
quir
e
s
a
ppr
opr
iate
tr
e
a
tm
e
nt
to
r
e
duc
e
the
a
mount
[
2]
.
T
he
r
e
f
or
e
,
on
ins
tanc
e
s
whe
r
e
ther
e
is
a
n
oc
c
ur
r
e
nc
e
of
s
ympt
oms
,
including
s
hor
tnes
s
of
br
e
a
th
,
c
ough
with
phlegm
or
f
e
ve
r
that
doe
s
not
he
a
l
,
it
is
r
e
c
o
mm
e
nde
d
to
pe
r
f
or
m
a
he
a
lt
h
c
he
c
k,
invol
ving
the
doc
tor
.
S
ubs
e
que
ntl
y,
ther
e
is
a
c
onduc
ti
on
of
phys
ica
l
a
nd
other
s
uppor
ti
ng
e
xa
mi
na
ti
ons
,
to
de
te
r
mi
ne
the
pa
ti
e
nt
's
c
ondit
ion,
a
nd
one
of
whic
h
is
to
pe
r
f
or
m
a
n
X
-
r
a
y
of
the
lungs
.
T
his
pr
omot
e
s
the
doc
tor
to
pr
ovide
the
r
ight
diagnos
is
a
s
to
whe
ther
or
not
pne
umoni
a
is
pr
e
s
e
nt,
while
c
lar
if
ica
ti
on
a
nd
a
s
s
is
tanc
e
in
c
la
s
s
if
ying
h
e
a
lt
hy
a
nd
inf
e
c
ted
lungs
,
pr
ovidi
ng
pr
e
c
is
e
e
xplana
ti
ons
a
nd
pr
e
dictions
,
r
e
qui
r
e
s
a
n
a
c
c
ur
a
te
method
.
T
he
a
c
c
ur
a
te
method
that
c
a
n
be
us
e
d
is
(
C
NN
s
)
,
be
c
a
us
e
thi
s
method
is
c
omm
only
us
e
d
f
or
im
a
ge
da
ta
c
las
s
if
ica
ti
on
[
3]
.
T
hus
C
NN
s
is
s
uit
a
ble
to
c
las
s
if
y
pn
e
umoni
a
dis
e
a
s
e
ba
s
e
d
on
lung
X
-
r
a
y
in
or
de
r
t
o
obtain
a
c
c
ur
a
te
pr
e
diction
r
e
s
ult
s
.
T
his
is
e
videnc
e
d
by
the
pr
e
vious
s
tudi
e
s
us
ing
C
NN
s
s
uc
h
a
s
I
mage
Ne
t
C
las
s
if
ica
t
ion
[
4]
,
Doc
ument
P
r
e
pr
oc
e
s
s
ing
[
5]
,
F
a
c
e
De
tec
ti
on
[
6]
,
I
mage
S
e
gmenta
ti
on
[
7
]
,
e
tc.
Not
only
that,
r
e
s
e
a
r
c
h
a
bout
pne
umoni
a
us
ing
c
he
s
t
r
a
diog
r
a
phs
or
X
-
r
a
y
ha
s
be
e
n
done
be
f
or
e
by
us
ing
s
e
ve
r
a
l
mac
hine
lea
r
ning
methods
s
uc
h
a
s
k
-
n
e
a
r
e
s
t
ne
ig
hbor
(
KN
N)
,
s
uppor
t
ve
c
tor
mac
hines
(
S
VM
)
a
nd
Na
ïve
B
a
ye
s
[
8]
.
T
he
r
e
f
or
e
,
in
thi
s
s
tudy
C
NN
s
method
wil
l
be
us
e
d
to
c
las
s
if
y
lung
X
-
r
a
y
to
de
tec
t
pne
umoni
a
dis
e
a
s
e
.
2.
RE
S
E
AR
CH
M
E
T
HO
D
2.
1.
Da
t
as
e
t
Da
tas
e
t
f
r
om
pa
ti
e
nt
lung
X
-
r
a
y
we
r
e
us
e
d
in
thi
s
e
xpe
r
im
e
nt,
obtaine
d
f
r
om
W
omen
a
nd
C
hil
dr
e
n's
M
e
dica
l
C
e
nter
,
Gua
ngz
hou.
T
he
number
of
im
a
ge
us
e
d
we
r
e
330,
whe
r
e
150
im
a
ge
s
a
s
he
a
lt
hy
c
a
tegor
y,
112
im
a
ge
s
a
s
ba
c
ter
ial
pne
umoni
a
,
a
nd
68
im
a
g
e
s
a
s
vir
a
l
p
ne
umoni
a
.
F
igur
e
s
1
-
3
a
r
e
s
ome
e
xa
mpl
e
s
of
the
pa
ti
e
nt's
X
-
r
a
y
tes
t
f
r
om
e
a
c
h
c
a
tegor
y
.
B
a
s
e
d
on
the
f
igur
e
s
,
to
f
ind
ou
t
the
dif
f
e
r
e
nc
e
,
X
-
r
a
y
a
f
f
e
c
ted
by
pne
umoni
a
mor
e
f
a
de
the
c
olo
r
o
f
the
lung
than
nor
mal.
And
to
tr
y
da
tas
e
t
int
o
pr
og
r
a
m,
in
e
a
c
h
c
a
t
e
gor
y
80%
of
a
ll
im
a
ge
s
we
r
e
us
e
d
a
s
t
r
a
ini
ng
da
ta
a
nd
2
0%
f
o
r
tes
ti
ng.
F
igur
e
1.
Nor
mal
l
ungs
F
igur
e
2.
B
a
c
ter
ial
pne
umoni
a
F
igur
e
3.
Vir
a
l
p
ne
umoni
a
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.
1
8
,
No
.
3
,
J
une
2020:
1522
-
1528
1524
2.
2.
Co
n
volu
t
io
n
al
n
e
u
r
al
n
e
t
wor
k
s
T
he
method
us
e
d
in
thi
s
e
xpe
r
im
e
nt
is
(
C
NN
s
)
,
w
hich
is
one
type
of
De
e
p
Ne
ur
a
l
Ne
twor
ks
that
is
a
r
e
s
ult
of
the
de
ve
lopm
e
nt
of
the
mul
ti
laye
r
pe
r
c
e
ptr
on
(
M
L
P
)
[
3,
7]
.
T
he
dis
ti
nguis
h
be
twe
e
n
C
NN
s
a
nd
M
L
P
is
c
a
n
r
e
taining
the
c
a
pa
bil
it
y
of
be
ing
us
e
d
in
the
de
tec
ti
on
a
nd
r
e
c
ognit
ion
of
objec
ts
in
a
n
i
mage
that
a
r
e
not
jus
t
in
the
mi
ddle
but
M
L
P
c
a
nnot
be
li
k
e
that.
C
NN
s
will
ge
t
be
tt
e
r
r
e
s
ult
s
than
ne
ur
a
l
n
e
twor
ks
(
NN
s
)
be
c
a
us
e
of
the
a
ddit
ion
o
f
one
laye
r
to
C
NN
s
a
f
ter
the
input
is
c
a
ll
e
d
c
onvolut
ional
la
ye
r
.
T
he
r
e
f
or
e
,
it
s
de
ve
lopm
e
nt
is
C
NN
s
,
whic
h
a
r
e
f
o
r
ms
of
ne
ur
a
l
ne
twor
k
c
omm
only
us
e
d
in
i
mage
da
ta.
T
he
c
o
mponent
in
C
NN
s
a
r
e
not
much
dif
f
e
r
e
nt
f
r
om
NN
s
,
us
u
a
ll
y
c
ons
is
t
of
ne
ur
ons
with
we
ight
,
bias
a
nd
a
c
ti
va
ti
on
f
unc
ti
on
[
3]
.
C
NN
s
divi
de
d
int
o
2
im
por
tant
pa
r
ts
,
the
f
ir
s
t
pa
r
t
is
f
e
a
tur
e
e
xtr
a
c
ti
on
(
c
onvolut
ional
l
a
ye
r
a
nd
pooli
ng
laye
r
)
a
nd
the
s
e
c
ond
is
f
ull
y
-
c
onne
c
ted
laye
r
[
3
,
7,
9]
.
I
ll
us
tr
a
ti
on
o
f
C
NN
s
c
a
n
be
s
e
e
n
in
F
i
gur
e
4
.
F
igur
e
4.
I
l
lus
tr
a
ti
on
of
c
onvolut
ional
ne
u
r
a
l
ne
tw
or
k
2.
2.
1.
F
e
a
t
u
r
e
e
xt
r
ac
t
ion
layer
F
e
a
tur
e
e
xtr
a
c
ti
on
laye
r
"
e
nc
ode
s
"
a
n
im
a
ge
int
o
f
e
a
tur
e
s
that
a
r
e
in
the
f
or
m
o
f
r
e
pr
e
s
e
ntative
number
s
(
f
e
a
tur
e
e
xt
r
a
c
ti
on
)
[
3]
.
He
nc
e
,
C
NN
s
is
tec
hnica
ll
y
a
n
a
r
c
hit
e
c
tur
e
e
nc
ompas
s
ing
s
e
ve
r
a
l
s
tage
s
,
a
nd
e
a
c
h
input
a
nd
ou
tput
c
ons
is
ts
of
numer
ous
a
r
r
a
ys
,
ter
med
f
e
a
tur
e
maps
,
whi
le
the
e
xtr
a
c
ti
on
laye
r
indi
vidually
c
ompr
is
e
s
of
two
im
po
r
tant
pa
r
ts
,
in
c
ludi
ng
1)
c
onvolut
ional
laye
r
,
whic
h
is
major
a
nd
mos
t
im
por
tant
f
or
us
e
,
c
los
e
ly
f
oll
owe
d
by
2
)
the
p
ooli
ng
laye
r
,
c
ombi
na
t
ion
us
e
d
to
e
xtr
a
c
t
the
a
ve
r
a
ge
or
maximum
va
lue
on
the
pixel
,
a
nd
e
a
c
h
input
pos
s
e
s
s
e
s
a
dif
f
e
r
e
nt
volum
e
,
a
nd
is
r
e
p
r
e
s
e
nted
ba
s
e
d
on
de
pth,
he
ight
a
nd
width
[
4
,
9
-
11
]
.
T
he
r
e
f
o
r
e
,
the
r
e
s
ult
s
obtaine
d
include
the
va
r
ied
a
mount
due
to
the
f
a
c
t
that
it
is
inf
luenc
e
d
by
the
f
il
ter
ing
of
the
pr
e
vious
laye
r
s
,
a
nd
a
ls
o
the
a
mount
o
f
f
il
ter
s
us
e
d.
a.
C
onvolut
ional
laye
r
T
his
is
a
s
tage
that
us
e
s
a
c
onvolut
ion
pr
oc
e
s
s
t
hr
ough
the
uti
li
z
a
ti
on
o
f
f
il
ter
s
ins
ide,
whic
h
ha
s
s
im
il
a
r
s
ize
a
s
that
of
a
n
im
a
ge
in
ge
ne
r
a
l,
includi
ng
the
length
a
nd
he
ight
(
pixels
)
,
a
s
we
ll
a
s
width
[
12,
13]
,
a
s
il
lus
tr
a
ted
i
n
F
igur
e
5.
B
a
s
e
d
on
F
igur
e
5
,
a
ll
pa
r
ts
of
the
im
a
ge
a
r
e
s
hif
ted
by
the
f
i
lt
e
r
s
us
e
d,
a
nd
f
r
o
m
e
a
c
h
pos
it
ion,
number
s
a
r
e
ge
ne
r
a
ted,
whic
h
a
r
e
t
he
r
e
s
ult
s
of
"
do
t
pr
oduc
t"
be
twe
e
n
the
im
a
ge
in
puts
a
nd
f
il
ter
va
lues
[
14
-
16]
.
T
he
r
e
f
or
e
,
the
r
e
s
ult
s
a
r
e
ter
med
a
c
ti
va
ti
on/f
e
a
tur
e
maps
a
s
in
the
pur
ple
pa
r
t.
F
igur
e
5.
I
l
lus
tr
a
ti
on
of
c
onvolut
ional
laye
r
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
P
ulmonar
y
r
ontgen
c
las
s
if
ication
to
de
tec
t
pne
umo
nia
…
(
Z
uhe
r
man
R
us
tam
)
1525
b.
S
t
r
ide
S
tr
ide
is
a
pa
r
a
mete
r
that
de
ter
mi
ne
s
the
a
mount
of
s
hif
ts
in
the
e
nti
r
e
im
a
ge
[
3,
17]
,
a
nd
ins
tanc
e
s
whe
r
e
the
va
lue
us
e
d
wa
s
1,
is
a
n
indi
c
a
ti
on
that
the
c
onvolut
io
l
f
il
ter
s
hif
ted
ho
r
izonta
ll
y
f
ir
s
t
by
1
p
ixel,
f
oll
owe
d
by
1
f
o
r
the
ve
r
ti
c
a
l.
How
e
ve
r
,
the
s
e
lec
ti
on
of
s
tep
numbe
r
s
or
s
tr
ide
ha
s
no
r
ules
,
a
nd
mor
e
inf
or
mation
is
obtaine
d
f
r
om
a
n
input
,
only
whi
le
us
ing
a
s
maller
s
tr
ide.
M
e
a
nwhile,
if
the
s
e
le
c
ted
one
be
c
omes
s
maller
,
the
dot
pr
oduc
t
c
a
lcula
ti
on
be
tw
e
e
n
the
input
a
nd
the
va
lue
o
f
the
f
il
ter
is
pe
r
f
or
m
e
d
mor
e
than
th
e
c
hoice
of
a
lar
ge
r
s
ize
while
a
s
mall
s
tr
ide
doe
s
not
ne
c
e
s
s
a
r
il
y
lea
d
to
a
good
pe
r
f
o
r
manc
e
[
18]
.
c
.
Z
e
r
o
pa
dding
Z
e
r
o
P
a
dding
is
a
pa
r
a
mete
r
that
is
c
a
pa
ble
o
f
de
ter
mi
ning
the
number
of
pixels
that
c
ontain
the
va
lue
0,
a
dde
d
on
e
a
c
h
s
ide
of
the
im
a
ge
in
put
[
18,
19]
,
whic
h
is
us
e
d
with
a
n
a
im
of
manipulating
the
output
dim
e
ns
ions
of
the
c
onvolut
ional
laye
r
(
f
e
a
tur
e
map
).
d.
An
ove
r
view
of
f
e
a
tur
e
e
xt
r
a
c
ti
on
laye
r
I
n
a
c
c
or
da
nc
e
with
F
igu
r
e
6
,
it
is
s
e
e
n
that
the
f
ir
s
t
laye
r
in
the
f
e
a
tur
e
e
xt
r
a
c
ti
on
is
a
n
input
im
a
ge
,
with
a
s
ize
of
5x5x3,
including
2
f
il
ter
s
,
a
nd
f
or
e
xa
mpl
e
the
us
e
o
f
2
s
tr
ides
a
nd
a
s
ingl
e
z
e
r
o
pa
dding.
T
he
r
e
f
or
e
,
the
5x5x3
mea
ns
5
pixels
in
lengt
h,
5
pixels
in
he
ight
a
nd
3
piec
e
s
in
thi
c
kn
e
s
s
,
a
nd
the
il
lus
tr
a
ti
on
a
s
s
hown
in
F
igur
e
6
indi
c
a
tes
that
the
dim
e
ns
ions
of
the
a
c
tual
input
a
r
e
5x5,
on
i
ns
tanc
e
s
whe
r
e
c
onvolut
ion
is
c
onduc
ted
with
a
3x3
f
il
ter
,
a
nd
a
s
tr
ide
of
2,
thus
,
a
f
e
a
tur
e
map
with
2x2
s
ize
is
obtaine
d.
M
e
a
nwhile,
if
z
e
r
o
pa
dding
is
a
dde
d
a
s
much
a
s
1,
a
3x3
s
ize
is
obtaine
d,
a
nd
mo
r
e
inf
o
r
mation
is
f
ur
ther
ge
ne
r
a
ted.
C
ons
ider
the
il
lus
tr
a
ti
on
a
s
s
hown
in
F
igur
e
6.
T
o
c
a
lcula
te
the
dim
e
ns
ions
of
a
f
e
a
tur
e
map,
the
(
1)
is
us
e
d
[
3,
20]
:
=
−
+
2
2
+
1
(
1)
whe
r
e
,
=
long
or
high
of
input
=
long
o
r
high
of
f
il
ter
=
a
s
a
pa
dding
=
a
s
a
s
tr
ide
f
or
e
xa
mpl
e
,
ba
s
e
d
on
F
igu
r
e
3,
with
W
=
5
N
=
3
P
=
1
S
=
2
,
the
output
dim
e
ns
ion
obtaine
d
is
3
.
F
igur
e
6.
I
l
lus
tr
a
ti
on
of
f
e
a
tur
e
map
e
.
P
ooli
ng
l
a
ye
r
T
he
pooli
ng
laye
r
is
us
ua
ll
y
loca
ted
a
f
ter
the
c
onv
olut
ional
[
21
,
22]
,
a
nd
the
point
is
that
it
c
ons
is
ts
of
a
c
e
r
tain
s
ize
d
f
il
ter
,
a
nd
s
tr
ides
that
s
hif
ts
t
hr
oughout
the
f
e
a
tu
r
e
map
a
r
e
a
.
M
e
a
nwhile,
the
p
ooli
ng
pr
oc
e
s
s
c
omm
only
us
e
d
include
s
(
1)
max
pooli
ng
,
whe
r
e
the
us
e
of
2x2
with
s
tr
ide
2
lea
ds
to
the
maximum
va
lue
of
2x2
pixel
a
r
e
a
s
e
lec
ted
f
or
e
a
c
h
s
hif
t,
(
2
)
a
ve
r
a
ge
pooli
ng
,
whic
h
c
hoos
e
s
it
s
mea
n
va
lue,
a
nd
a
n
il
lus
tr
a
ti
on
of
pooli
ng
laye
r
(
e
s
pe
c
ially
max
pooli
ng)
a
r
e
s
hown
in
F
igur
e
7
.
F
r
om
the
il
lus
tr
a
ti
on
w
it
h
a
2x2
f
il
ter
us
e
d,
the
f
il
ter
will
move
in
the
input
a
r
e
a
wi
th
a
s
tr
ides
of
2
a
nd
take
the
maximum
va
lue
o
f
e
a
c
h
f
il
ter
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.
1
8
,
No
.
3
,
J
une
2020:
1522
-
1528
1526
moveme
nt
towa
r
ds
the
input
.
T
he
output
dim
e
n
s
ion
of
the
pooli
ng
la
ye
r
a
ls
o
us
e
s
the
s
a
me
f
o
r
mul
a
a
s
the
c
onvolut
ional,
a
nd
the
r
e
a
s
on
f
or
us
e
is
to
r
e
duc
e
the
dim
e
ns
ions
of
the
f
e
a
tur
e
map
(
down
s
a
mpl
ing)
,
thus
,
s
pe
e
ding
up
c
omput
a
ti
on,
due
to
the
pr
e
s
e
nc
e
of
f
e
we
r
pa
r
a
mete
r
s
the
r
e
quir
e
a
n
upd
a
ted
a
nd
ove
r
f
it
ti
ng
is
s
ue
s
.
F
igur
e
7.
I
l
lus
tr
a
ti
on
of
max
pool
ing
2.
2.
2.
F
u
ll
y
c
on
n
e
c
t
e
d
l
aye
r
T
he
f
e
a
tur
e
map
is
c
r
e
a
ted
ba
s
e
d
on
the
f
e
a
tur
e
e
xtr
a
c
ti
on
laye
r
,
whic
h
is
s
ti
ll
a
mul
ti
di
mens
ional
a
r
r
a
y,
he
nc
e
,
ther
e
is
ne
e
d
to
f
latten
or
r
e
s
ha
pe
i
t
in
the
ve
c
tor
f
e
a
tu
r
e
map
,
whic
h
is
us
e
d
a
s
inp
ut
f
r
om
the
f
ull
y
-
c
onne
c
ted
laye
r
[
23,
2
4]
.
I
n
a
ddit
ion
,
a
ll
that
a
c
ti
va
te
ne
ur
ons
f
r
om
the
pr
e
vious
a
r
e
li
nke
d
in
the
ne
xt
laye
r
,
a
s
s
e
e
n
in
ne
ur
a
l
ne
twor
ks
.
T
he
r
e
f
or
e
,
in
o
r
de
r
to
c
onne
c
t
pr
ope
r
ly
,
indi
vidual
a
c
ti
v
a
ti
on
of
the
pr
e
vious
ought
to
be
f
ir
s
t
c
onve
r
ted
int
o
on
e
-
dim
e
ns
ional
da
ta.
T
he
s
e
us
ua
ll
y
us
e
a
method
ter
med
mul
ti
-
pe
r
c
e
ptr
on
laye
r
,
whic
h
ha
s
a
tar
ge
t
in
pr
oc
e
s
s
ing
da
ta,
in
or
de
r
f
o
r
it
to
c
a
r
r
y
ou
t
pr
ope
r
c
las
s
if
ica
ti
on
[
25]
.
M
e
a
nwhile,
the
dif
f
e
r
e
nc
e
a
ga
ins
t
c
onvolut
ion
laye
r
s
is
that
the
ne
ur
ons
in
the
l
a
tt
e
r
a
r
e
c
onne
c
ted
to
a
s
pe
c
if
ic
input
a
r
e
a
,
while
the
f
u
ll
y
-
c
onne
c
ted
oc
c
ur
s
in
a
lm
os
t
a
ll
pa
r
ts
.
How
e
v
e
r
,
both
c
onti
nue
to
pe
r
f
o
r
m
“
dot
pr
oduc
t”
ope
r
a
ti
ons
,
thus
,
their
f
unc
ti
ons
a
r
e
not
s
igni
f
ica
ntl
y
dif
f
e
r
e
nt.
2.
3.
M
od
e
l
T
h
e
m
e
t
h
o
d
a
p
p
l
i
e
d
i
s
C
N
N
s
w
i
t
h
i
n
T
e
n
s
o
r
f
l
o
w
,
w
h
i
c
h
i
s
a
f
r
a
m
e
w
o
r
k
of
p
r
o
g
r
a
m
s
o
r
p
a
c
k
a
g
e
s
t
h
a
t
a
s
s
i
s
t
t
h
e
s
y
s
t
e
m
s
t
o
p
r
o
c
e
s
s
i
m
a
g
e
d
a
t
a
.
M
e
a
n
w
h
i
l
e
,
t
h
e
c
h
o
s
e
n
m
o
d
e
l
i
n
d
i
c
a
t
e
s
t
h
e
u
s
e
o
f
4
c
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
s
,
i
n
c
l
u
d
i
n
g
:
−
F
ir
s
t
l
a
ye
r
us
e
16
-
3
x
3
f
il
ter
s
−
S
e
c
ond
laye
r
us
e
32
-
3
x
3
f
il
ter
s
−
T
hir
d
l
a
ye
r
us
e
64
-
3
x
3
f
il
te
r
s
−
F
our
th
l
a
ye
r
us
e
128
-
3
x
3
f
il
ter
s
E
a
c
h
of
thes
e
invol
ve
max
-
pooli
ng
of
2x2
,
with
a
mount
of
s
tr
ide
be
ing
2
a
nd
input
ha
s
a
s
ize
of
28
×
28
pixels
with
a
thi
c
kne
s
s
of
1.
T
he
r
e
f
o
r
e
,
be
c
a
u
s
e
thi
s
s
tudy
will
be
c
las
s
if
ied
int
o
thr
e
e
c
las
s
e
s
,
the
c
a
tegor
ica
l
c
r
os
s
-
e
ntr
opy
los
s
f
unc
ti
on
will
b
e
us
e
d
a
nd
Ada
m
(
Ada
pti
ve
M
omentum)
a
s
a
n
o
pti
mi
z
e
r
with
a
lea
r
ning
r
a
te
o
f
0.
0001
.
T
he
model
of
C
NN
s
is
e
xplaine
d
in
the
f
or
m
of
a
f
lowc
ha
r
t
a
s
s
hown
in
F
igur
e
8
.
F
igur
e
8.
M
ode
l
of
C
NN
s
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
P
ulmonar
y
r
ontgen
c
las
s
if
ication
to
de
tec
t
pne
umo
nia
…
(
Z
uhe
r
man
R
us
tam
)
1527
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ONS
T
his
s
tudy
us
e
d
s
of
twa
r
e
P
ython
ve
r
s
ion
2
.
1.
9
f
or
c
onvolut
ional
ne
u
r
a
l
ne
two
r
ks
.
Af
ter
r
unning
the
pr
ogr
a
m
,
a
c
c
ur
a
c
y
va
lue
a
nd
los
s
va
lue
will
be
ge
ne
r
a
ted
in
g
r
a
phica
l
f
or
m.
T
he
im
a
ge
da
ta
that
ha
s
be
e
n
obtaine
d
a
s
e
xplaine
d
in
the
da
tas
e
t
s
e
c
ti
on,
330
we
r
e
c
las
s
if
ied
int
o
3
c
las
s
e
s
,
with
the
output
a
s
a
c
c
ur
a
c
y
va
lue
a
s
s
hown
in
F
igur
e
9
a
nd
los
s
va
lue
a
s
s
hown
in
F
igu
r
e
10
.
T
he
a
c
c
ur
a
c
y
va
lue
a
nd
the
los
s
va
lue
a
r
e
us
e
d
to
s
e
e
how
c
a
pa
ble
the
C
NN
s
model
that
ha
s
be
e
n
f
or
med
to
c
las
s
if
y
pne
umoni
a
dis
e
a
s
e
.
B
a
s
e
d
on
F
igur
e
9,
the
a
c
c
ur
a
c
y
va
lue
ha
s
incr
e
a
s
e
d
in
e
poc
hs
to
0
-
25,
whic
h
mea
ns
the
model
is
tr
ying
to
memo
r
ize
da
ta.
B
ut
a
f
ter
25
-
50
e
poc
hs
t
he
a
c
c
ur
a
c
y
va
lue
s
tar
ts
to
s
tagna
nt
wi
th
t
r
a
ini
ng
a
c
c
ur
a
c
y
c
onve
r
ge
to
1.
00
a
nd
tes
t
a
c
c
ur
a
c
y
c
onve
r
ge
to
0
.
91.
D
i
f
f
e
r
e
nt
with
F
igur
e
10
,
los
s
tr
a
ini
ng
de
c
r
e
a
s
e
s
a
nd
c
onve
r
ge
s
to
z
e
r
o
while
the
los
s
tes
t
c
onti
nue
s
t
o
incr
e
a
s
e
.
F
ur
ther
mor
e
,
tes
t
a
c
c
ur
a
c
y
a
nd
tes
t
lo
s
t
s
hows
f
r
om
F
igur
e
9
a
nd
F
igu
r
e
10
,
a
s
ign
of
ove
r
f
i
tt
ing
.
B
ut
ba
s
e
d
on
the
r
e
s
ult
s
of
pr
e
vious
s
tud
ies
the
a
c
c
ur
a
c
y
va
lue
obtaine
d
in
de
tec
ti
on
of
pne
umoni
a
dis
e
a
s
e
ba
s
e
d
on
X
-
r
a
y
im
a
ge
s
us
ing
the
S
VM
method
is
77%
,
KN
N
method
is
70%
a
nd
Na
ïve
B
a
y
e
s
is
68%
[
8]
.
And
whe
n
c
ompar
e
d
with
the
r
e
s
ult
s
of
thi
s
s
tudy
or
us
ing
the
C
NN
s
method
pr
oduc
e
s
be
tt
e
r
a
c
c
ur
a
c
y
va
lues
than
thr
e
e
methods
,
whic
h
is
91%
.
F
igur
e
9.
Ac
c
ur
a
c
y
va
lue
F
igur
e
10.
L
os
s
va
lue
4.
CONC
L
USI
ON
B
a
s
e
d
on
r
e
s
ult
s
a
nd
dis
c
us
s
ions
,
the
c
onvolut
ional
ne
ur
a
l
ne
twor
ks
(
C
NN
s
)
method
is
a
ble
to
c
las
s
if
y
X
-
r
a
y
im
a
ge
s
int
o
he
a
lt
hy
lungs
a
nd
pne
umoni
a
,
due
to
ba
c
ter
ia
a
nd
vir
us
e
s
.
S
o
that
thi
s
method
is
a
ble
to
a
s
s
i
s
t
doc
tor
s
in
pr
ovidi
ng
the
a
ppr
opr
i
a
te
be
li
e
f
s
a
nd
pr
e
dictions
to
pa
ti
e
nts
.
I
n
a
ddit
ion,
it
is
e
videnc
e
d
in
t
he
e
va
luation
obtaine
d
a
f
ter
e
poc
hs
25,
f
or
tr
a
ini
ng
a
c
c
ur
a
c
y
is
0.
97
a
nd
f
o
r
tes
t
a
c
c
ur
a
c
y
is
0.
91.
T
his
e
videnc
e
d
s
hows
the
c
a
pa
bil
it
y
f
or
the
pr
ogr
a
m
f
r
om
C
NN
s
method
to
a
c
c
ur
a
tely
identif
y
the
pa
ti
e
nt's
X
-
r
a
y
tes
t
im
a
ge
s
.
How
e
ve
r
,
the
a
c
c
ur
a
c
y
r
a
te
a
t
s
ome
e
poc
hs
is
identif
ied
a
s
uns
table
,
due
to
the
f
a
c
t
that
the
a
tt
e
mpt
o
f
C
NN
s
to
memor
ize
the
da
ta
s
e
t
im
a
ge
.
T
he
li
mi
tation
of
thi
s
s
tudy
li
e
s
in
the
im
a
ge
da
tas
e
t
be
c
a
us
e
that
is
not
take
n
in
lar
g
e
qua
nti
ti
e
s
a
nd
the
pr
ogr
a
m
o
f
C
NN
s
doe
s
n’
t
us
e
a
goo
d
pr
oc
e
s
s
or
.
S
o,
f
or
the
ne
xt
r
e
s
e
a
r
c
he
r
c
a
n
modi
f
y
o
r
a
dding
high
pr
oc
e
s
s
or
to
s
pe
e
d
up
the
pr
og
r
a
m
of
C
NN
s
.
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.
1
8
,
No
.
3
,
J
une
2020:
1522
-
1528
1528
AC
KNOWL
E
DGE
M
E
NT
S
T
his
r
e
s
e
a
r
c
h
wa
s
f
ull
s
uppor
ted
f
inanc
ially
by
Un
iver
s
it
y
of
I
ndone
s
ia
,
with
a
DR
P
M
P
U
T
I
KI
Q2
2020
r
e
s
e
a
r
c
h
gr
a
nt
s
c
he
me
.
RE
F
E
RE
NC
E
S
[1
]
U
rry
L
.
A
.
,
et
al
,
“
Camp
b
el
l
b
i
o
l
o
g
y
,
”
E
l
ev
e
n
t
h
E
d
i
t
i
o
n
N
ew
Y
o
rk
:
Pears
o
n
,
p
p
.
9
7
5
-
9
9
8
,
2
0
1
6
.
[2
]
K
erman
y
D
.
,
Z
h
an
g
K
,
.
an
d
G
o
l
d
b
a
u
m
M
.
,
“
L
ab
el
ed
o
p
t
i
cal
c
o
h
ere
n
ce
t
o
mo
g
rap
h
y
(
O
CT
)
a
n
d
ch
e
s
t
X
-
Ray
i
mag
e
s
fo
r
cl
a
s
s
i
fi
ca
t
i
o
n
,
”
M
e
n
d
e
l
ey
D
a
t
a
,
v
o
l
.
2
,
2
0
1
8
.
[
3
]
O
’
S
h
e
a
K
.
,
a
n
d
N
a
s
h
R
.
,
“
A
n
i
n
t
r
o
d
u
c
t
i
o
n
t
o
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
,
”
a
r
X
i
v
,
P
r
e
p
r
i
n
t
a
r
x
i
v
:
1
5
1
1
.
0
8
4
5
8
v
2
,
2
0
1
5
.
[4
]
K
ri
z
h
ev
s
k
y
A
.
,
Su
t
s
k
e
v
er
I
.
an
d
H
i
n
t
o
n
G
.
E
.
,
“
Imag
eN
et
cl
as
s
i
f
i
cat
i
o
n
w
i
t
h
d
ee
p
co
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ral
n
et
w
o
r
k
s
,”
A
d
v
a
n
ce
s
Neu
r
a
l
I
n
f
o
r
m
a
t
i
o
n
P
r
o
ce
s
s
i
n
g
S
y
s
t
e
m
s
,
p
p
.
1
9
0
7
-
1
1
0
5
,
2
0
1
2
.
[5
]
Ch
el
l
ap
i
l
l
a
K
.
,
Pu
ri
S
.
,
an
d
Si
mard
P
.
Y
.
,
“
High
-
P
erfo
rman
ce
co
n
v
o
l
u
t
i
o
n
al
n
e
u
ral
n
e
t
w
o
rk
s
fo
r
d
o
cu
men
t
Pro
ces
s
i
n
g
,
”
P
r
o
c.
o
f
t
h
e
1
0
t
h
In
t
.
W
o
r
ks
h
o
p
o
n
F
r
o
n
t
i
e
r
s
i
n
H
a
n
d
w
r
i
t
i
n
g
R
eco
g
n
i
t
i
o
n
,
2
0
0
6
.
[6
]
Li
H
.
,
et
al
.
,
“
A
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ra
l
n
e
t
w
o
rk
ca
s
cad
e
f
o
r
face
d
e
t
ect
i
o
n
,
”
P
r
o
cee
d
i
n
g
s
o
f
t
h
e
IE
E
E
C
o
n
f
er
en
ce
o
n
Co
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
,
p
p
.
5
3
2
5
-
5
3
3
4
,
2
0
1
5
.
[7
]
T
u
ra
g
a
S
.
C
.
,
et
al
.,
“
Co
n
v
o
l
u
t
i
o
n
a
l
n
et
w
o
r
k
s
ca
n
l
e
arn
t
o
g
en
era
t
e
affi
n
i
t
y
g
ra
p
h
s
fo
r
i
ma
g
e
s
e
g
men
t
at
i
o
n
,
”
Neu
r
a
l
Co
m
p
u
t
a
t
i
o
n
,
v
o
l
.
22
,
n
o
.
2
,
p
p
.
5
1
1
–
5
3
8
,
2
0
1
0
.
[8
]
So
u
s
a
R
.
T
.
,
et
al
.
,
“
Co
mp
ara
t
i
v
e
p
erfo
rman
ce
an
a
l
y
s
i
s
o
f
mach
i
n
e
l
ear
n
i
n
g
cl
a
s
s
i
fi
er
s
i
n
d
et
ec
t
i
o
n
o
f
ch
i
l
d
h
o
o
d
p
n
e
u
mo
n
i
a
u
s
i
n
g
c
h
es
t
rad
i
o
g
ra
p
h
s
,
”
P
r
o
ced
i
a
C
o
m
p
u
t
e
r
S
ci
e
n
ce
,
p
p
.
2
5
7
9
-
2
5
8
2
,
2
0
1
3
.
[9
]
Ci
res
a
n
D
.
,
Mei
er
U
.
,
an
d
Sc
h
mi
d
h
u
b
er
J
.
,
“
Mu
l
t
i
-
co
l
u
mn
d
ee
p
n
e
u
ral
n
et
w
o
r
k
s
fo
r
i
mag
e
c
l
as
s
i
f
i
cat
i
o
n
,
”
P
r
o
ceed
i
n
g
s
o
f
t
h
e
IE
E
E
Co
n
f
e
r
en
ce
o
n
Co
m
p
u
t
e
r
V
i
s
i
o
n
a
n
d
P
a
t
t
er
n
R
eco
g
n
i
t
i
o
n
,
p
p
.
3
6
4
2
–
3
6
4
9
,
2
0
1
2
.
[1
0
]
Ci
res
a
n
D
.
C
.
,
et
al
.,
“H
i
g
h
p
erf
o
rman
ce
c
o
n
v
o
l
u
t
i
o
n
al
n
e
u
ral
n
e
t
w
o
rk
s
fo
r
i
ma
g
e
cl
as
s
i
f
i
ca
t
i
o
n
,
”
IJCA
I
P
r
o
ceed
i
n
g
s
-
I
n
t
e
r
n
a
t
i
o
n
a
l
Jo
i
n
t
C
o
n
f
er
e
n
ce
o
n
A
r
t
i
f
i
c
i
a
l
In
t
e
l
l
i
g
e
n
ce,
v
o
l
.
22
,
p
p
.
1
2
3
7
,
2
0
1
1
.
[1
1
]
Ci
res
a
n
D
.
C
.
,
et
al
.
,
“
Co
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ral
n
et
w
o
rk
co
mm
i
t
t
ees
f
o
r
h
a
n
d
w
ri
t
t
e
n
ch
arac
t
er
c
l
as
s
i
f
i
cat
i
o
n
,
”
IE
E
E
Co
n
f
er
e
n
ce
o
n
D
o
c
u
m
e
n
t
A
n
a
l
y
s
i
s
a
n
d
R
ec
o
g
n
i
t
i
o
n
,
p
p
.
1
1
3
5
–
1
1
3
9
,
2
0
1
1
.
[1
2
]
E
g
m
o
n
t
-
Pet
er
s
en
M
.
,
d
e
.
Ri
d
d
er
D
.
,
an
d
H
a
n
d
e
l
s
H
.
,
“
Imag
e
p
r
o
ces
s
i
n
g
w
i
t
h
n
eu
ra
l
n
e
t
w
o
rk
s
a
re
v
i
ew
,
”
P
a
t
t
e
r
n
r
ec
o
g
n
i
t
i
o
n
,
v
o
l
.
35
,
n
o
.
10
,
p
p
.
2
2
7
9
–
2
3
0
1
,
2
0
0
2
.
[1
3
]
Ji
S
.
,
et
al
.,
“
3
D
co
n
v
o
l
u
t
i
o
n
al
n
e
u
ral
n
e
t
w
o
rk
s
fo
r
h
u
man
act
i
o
n
reco
g
n
i
t
i
o
n
,
”
P
a
t
t
er
n
A
n
a
l
y
s
i
s
a
n
d
M
a
c
h
i
n
e
In
t
e
l
l
i
g
e
n
ce,
v
o
l
.
35
,
n
o
.
1
,
p
p
.
2
2
1
–
2
3
1
,
2
0
1
3
.
[1
4
]
Farab
et
C
.
,
et
a
l
.
,
“
H
ar
d
w
are
acce
l
era
t
ed
c
o
n
v
o
l
u
t
i
o
n
a
l
n
eu
ra
l
n
e
t
w
o
rk
s
fo
r
s
y
n
t
h
et
i
c
v
i
s
i
o
n
s
y
s
t
em
s
,
”
P
r
o
cee
d
i
n
g
s
o
f
2
0
1
0
IE
E
E
In
t
er
n
a
t
i
o
n
a
l
S
ym
p
o
s
i
u
m
o
n
Ci
r
cu
i
t
s
a
n
d
S
ys
t
em
s
,
p
p
.
2
5
7
–
2
6
0
,
2
0
1
0
.
[1
5
]
H
i
n
t
o
n
G
.
E
.
,
et
al
.
,
“
Imp
ro
v
i
n
g
n
e
u
ral
n
e
t
w
o
r
k
s
b
y
p
re
v
en
t
i
n
g
co
-
a
d
ap
t
at
i
o
n
o
f
feat
u
re
d
e
t
ect
o
rs
,”
a
r
X
i
v
,
Prep
ri
n
t
arX
i
v
:
1
2
0
7
.
0
5
8
0
,
2
0
1
2
.
[1
6
]
K
arp
a
t
h
y
A
,
et
al
.
,
“
L
arg
e
s
cal
e
v
i
d
eo
cl
as
s
i
f
i
cat
i
o
n
w
i
t
h
co
n
v
o
l
u
t
i
o
n
al
n
e
u
ral
n
et
w
o
r
k
s
,
”
P
r
o
cee
d
i
n
g
s
o
f
t
h
e
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
,
pp.
1
7
2
5
–
1
7
3
2
,
2
0
1
4
.
[1
7
]
T
u
t
i
P
.
,
Iman
i
a
A
.
A
.
,
an
d
Pert
i
w
i
B
.
U
.
,
“
Imag
e
cl
a
s
s
i
fi
cat
i
o
n
o
f
g
o
l
ek
p
u
p
p
e
t
i
ma
g
es
u
s
i
n
g
co
n
v
o
l
u
t
i
o
n
al
n
e
u
ral
n
et
w
o
r
k
s
al
g
o
r
i
t
h
m
,
”
In
t
e
r
n
a
t
i
o
n
a
l
Jo
u
r
n
a
l
o
f
A
d
v
a
n
ce
s
,
v
o
l
.
11
,
p
p
.
34
-
45
,
2
0
1
9
.
[1
8
]
Mu
h
ama
d
Y
.
,
Bu
d
h
i
I
.
,
an
d
Cas
i
S
.
,
“
A
p
p
l
i
ca
t
i
o
n
o
f
t
r
an
s
fer
l
ear
n
i
n
g
u
s
i
n
g
co
n
v
o
l
u
t
i
o
n
a
l
n
eu
ra
l
n
et
w
o
r
k
met
h
o
d
fo
r
earl
y
d
et
ec
t
i
o
n
o
f
t
erry
’
s
n
a
i
l
,
”
Jo
u
r
n
a
l
o
f
P
h
y
s
i
c
s
:
Co
n
f
er
e
n
ce
S
er
i
es
,
v
o
l
.
1
2
0
1
,
2
0
1
9
.
[1
9
]
Si
mard
P
.
Y
.
,
St
ei
n
k
r
au
s
D
.
an
d
Pl
a
t
t
J
.
,
“
Bes
t
p
rac
t
i
ce
fo
r
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ral
n
et
w
o
r
k
s
a
p
p
l
i
e
d
t
o
v
i
s
u
a
l
d
o
c
u
men
t
an
al
y
s
i
s
,
”
In
t
er
n
a
t
i
o
n
a
l
C
o
n
f
er
e
n
ce
o
n
D
o
c
u
m
en
t
A
n
a
l
y
s
i
s
a
n
d
R
ec
o
g
n
i
t
i
o
n
,
p
p
.
9
5
8
-
9
6
2
,
2
0
0
3
.
[2
0
]
L
eCu
n
Y
.
,
et
al
.
,
“
G
rad
i
en
t
-
B
as
e
d
l
earn
i
n
g
ap
p
l
i
e
d
t
o
d
o
cu
me
n
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
,
p
p
.
2
2
7
8
-
2
3
2
4
,
1
9
9
8
.
[2
1
]
St
ei
n
k
ra
u
s
D
.
,
Bu
ck
I
.
an
d
Si
mard
P
.
Y
.
, “
G
PU
s
fo
r
mac
h
i
n
e
l
earn
i
n
g
al
g
o
r
i
t
h
ms
,
”
E
i
g
h
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
D
o
cu
m
en
t
A
n
a
l
ys
i
s
a
n
d
R
eco
g
n
i
t
i
o
n
,
p
p
.
1
1
1
5
-
1
1
2
0
,
2
0
0
5
.
[
2
2
]
W
a
n
g
L
.
,
e
t
a
l
.
,
“
T
r
a
i
n
i
n
g
d
e
e
p
e
r
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
s
w
i
t
h
d
e
e
p
s
u
p
e
r
v
i
s
i
o
n
,
”
a
r
X
i
v
P
r
e
p
r
i
n
t
a
r
X
i
v
:
1
5
0
5
.
0
2
4
9
6
,
2
0
1
5
.
[2
3
]
Szarv
as
M
.
,
et
a
l
.
,
“
Ped
e
s
t
r
i
an
d
et
ec
t
i
o
n
w
i
t
h
co
n
v
o
l
u
t
i
o
n
al
n
eu
ra
l
n
e
t
w
o
rk
s
,
”
In
t
el
l
i
g
en
t
V
e
h
i
c
l
e
s
S
y
m
p
o
s
i
u
m
,
IE
E
E
,
p
p
.
2
2
4
–
2
2
9
,
2
0
0
5
.
[2
4
]
Szeg
ed
y
C
.
,
T
o
s
h
ev
A
.
,
an
d
E
r
h
an
D
.
,
“
D
eep
n
e
u
r
al
n
e
t
w
o
rk
s
fo
r
o
b
j
ec
t
d
e
t
ect
i
o
n
,
”
A
d
va
n
ces
i
n
Neu
r
a
l
In
f
o
r
m
a
t
i
o
n
P
r
o
ce
s
s
i
n
g
S
y
s
t
e
m
s
,
p
p
.
2
5
5
3
–
2
5
6
1
,
2
0
1
3
.
[2
5
]
T
i
v
i
v
e
F.
H
.
C
.
,
an
d
Bo
u
zerd
o
u
m
A
.
,
“A
n
ew
c
l
as
s
o
f
co
n
v
o
l
u
t
i
o
n
al
n
eu
r
al
n
et
w
o
r
k
s
(
s
i
c
o
n
n
et
s
)
a
n
d
t
h
ei
r
ap
p
l
i
ca
t
i
o
n
o
f
face
d
e
t
ec
t
i
o
n
,
”
P
r
o
cee
d
i
n
g
s
o
f
t
h
e
In
t
er
n
a
t
i
o
n
a
l
J
o
i
n
t
Co
n
f
er
e
n
ce,
p
p
.
2
1
5
7
–
2
1
6
2
,
2
0
0
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.