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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
, pp.
4630
~
4641
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
13
.i
4
.pp
4630
-
4641
4630
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
3D
vi
su
al
i
z
at
i
on
d
i
ag
n
ost
i
c
s f
or
l
u
n
g
c
an
c
e
r
d
e
t
e
c
t
i
on
R
an
a M
. M
ah
m
ou
d
1
,2
, M
os
t
a
f
a E
lg
e
n
d
y
1
, M
oh
am
e
d
T
ah
a
1
1
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
, F
a
c
ul
t
y of
C
om
put
e
r
a
nd A
r
t
i
f
i
c
i
a
l
I
nt
e
l
l
i
ge
nc
e
, B
e
nha
U
ni
ve
r
s
i
t
y,
B
e
nha
, E
gypt
2
F
a
c
ul
t
y
of
C
om
put
e
r
s
a
nd I
nf
or
m
a
t
i
on T
e
c
hnol
ogy
,
T
he
E
gypt
i
a
n E
-
l
e
a
r
ni
ng
U
ni
ve
r
s
i
t
y,
G
i
z
a
, E
gypt
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
O
c
t
11
,
2023
R
e
vi
s
e
d
F
e
b
22
,
2024
A
c
c
e
pt
e
d
M
a
r
21
,
2024
Lung
cancer
is
the
primary
contributor
to
cancer
-
related
deaths
gl
obally,
accountin
g
for
approxim
ately
2
milli
on
new
diagnoses
and
resulti
ng
i
n
1.76
million
deaths
yearly.
Early
detection
can
improve
survival,
and
compu
terized
tomography
(
CT
)
scans
are
a
precise
imaging
technique
to
diagnose
lung
cancer. However,
analyzin
g hundred
s of 2D
CT
slices
is chall
enging a
nd can
cause
false
alarms.
3D
visualization
of
lung
nodules
can
aid
clinicians
in
detection
and
diagnosis.
The
MobileNet
model
integrates
multi
-
vi
e
w
and
multi
-
scale
nodule
features
using
depthwise
separable
convolutional
layers.
These
layers
split
standard
convolutions
into
depthwise
and
pointwise
convolut
ions
to
reduce
computat
ional
cost.
Finall
y,
the
3D
pulmon
ary
n
odular
models
were
created
using
a
ray
-
casting
volum
e
rendering
app
roach.
Contraste
d
to
other
cutting
-
edge
deep
neural
networks,
this
factor
ization
enables
MobileN
et
to
achieve
a
much
lower
computat
ional
cost
while
maintaining
a
decent
degree
of
accura
cy.
The
proposed
approac
h
was
tested
on
a
dataset
comprising
986
nodules
from
lung
image
database
cons
ortium
(
LIDC
)
.
Experiment
findings
reveal
that
MobileNet
achieves
outstanding
performance
in
segmenting
the
LIDC
dataset
with
an
accuracy
of
93.3%.
Conclusion:
The
study
demonstrates
that
the
MobileNet
detects
and
se
gments
lung
nodules
somewhat
better
than
other
older
technologies.
As
a
res
ult,
the
proposed system proposes an automated 3D lung ca
ncer tumor visualization
.
K
e
y
w
o
r
d
s
:
3
D
r
e
c
ons
tr
uc
ti
on
C
om
put
e
d t
om
ogr
a
phy
D
e
e
p l
e
a
r
ni
ng
L
ung nodule de
te
c
ti
on
L
ung nodule s
e
gm
e
nt
a
ti
on
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
:
R
a
na
M
.
M
a
hm
oud
F
a
c
ul
ty
of
C
om
put
e
r
s
a
nd I
nf
or
m
a
ti
on T
e
c
hnol
ogy, T
he
E
gypt
ia
n E
-
le
a
r
ni
ng U
ni
ve
r
s
it
y
33 E
l
M
e
s
a
ha
S
tr
e
e
t,
D
okki
,
G
iz
a
, E
gypt
E
m
a
il
:
r
m
oha
m
e
da
li
@
e
e
lu
.e
du.e
g
1.
I
N
T
R
O
D
U
C
T
I
O
N
L
ung
c
a
nc
e
r
or
ig
in
a
te
s
in
th
e
lu
ngs
,
w
hi
c
h
a
r
e
vi
ta
l
or
ga
ns
f
or
br
e
a
th
in
g.
I
t
s
ta
nds
a
s
th
e
f
or
e
m
os
t
c
ont
r
ib
ut
or
to
gl
oba
l
c
a
nc
e
r
-
r
e
la
te
d
f
a
ta
li
ti
e
s
.
L
ung
c
a
nc
e
r
is
c
la
s
s
if
ie
d
in
to
two
ty
pe
s
:
s
m
a
ll
-
c
e
ll
lu
ng
c
a
nc
e
r
(
S
C
L
C
)
a
nd
non
-
s
m
a
ll
c
e
ll
lu
ng
c
a
nc
e
r
(
N
S
C
L
C
)
.
S
C
L
C
c
ons
is
ts
of
c
a
nc
e
r
c
e
ll
s
th
a
t
a
r
e
s
m
a
ll
a
nd
vi
e
w
e
d
unde
r
a
m
ic
r
os
c
op
e
.
I
n
c
ont
r
a
s
t,
in
S
C
L
C
,
th
e
c
a
nc
e
r
c
e
ll
s
a
r
e
m
or
e
e
xt
e
n
s
iv
e
.
G
e
ne
r
a
ll
y,
S
C
L
C
pr
ogr
e
s
s
e
s
f
a
s
te
r
t
ha
n N
S
C
L
C
.
N
S
C
L
C
r
e
pr
e
s
e
nt
s
t
he
pr
e
va
il
in
g m
a
ni
f
e
s
ta
ti
on of
t
he
di
s
e
a
s
e
. I
t
ha
pp
e
n
s
w
he
n a
bnor
m
a
l
c
e
ll
s
in
th
e
lu
ng
s
gr
ow
a
nd
di
vi
de
unc
ont
r
ol
le
d,
c
a
u
s
in
g
a
tu
m
o
r
or
c
a
nc
e
r
th
a
t
be
gi
n
s
in
th
e
lu
ng
s
,
w
hi
c
h
a
r
e
th
e
or
ga
ns
r
e
s
pons
ib
le
f
or
br
e
a
th
in
g
th
a
t
c
a
n
in
te
r
f
e
r
e
w
it
h
no
r
m
a
l
lu
ng
f
unc
ti
on.
L
ung
c
a
nc
e
r
is
m
os
t
of
te
n
c
a
us
e
d
by
s
m
oki
ng;
how
e
ve
r
,
it
c
a
n
a
ls
o
be
c
a
us
e
d
by
ot
he
r
e
nvi
r
onm
e
nt
a
l
f
a
c
to
r
s
s
uc
h
a
s
a
ir
pol
lu
ti
on
a
nd
r
a
don
ga
s
e
xpos
ur
e
.
T
oba
c
c
o
us
e
is
th
e
pr
im
a
r
y
c
a
us
e
of
lu
ng
c
a
nc
e
r
,
a
c
c
ount
in
g
f
or
a
r
ound
80
%
of
a
ll
oc
c
ur
r
e
nc
e
s
[
1]
.
C
om
m
on
r
is
ks
in
c
lu
de
be
in
g
s
ubj
e
c
te
d
to
a
ir
pol
lu
ti
on,
s
e
c
ondha
nd
s
m
oke
,
r
a
don
ga
s
,
a
nd
s
pe
c
if
ic
c
he
m
ic
a
ls
a
nd
s
ubs
ta
n
c
e
s
e
nc
ount
e
r
e
d
in
c
e
r
ta
in
oc
c
up
a
ti
ona
l
s
e
tt
in
gs
.
A
c
c
or
di
ng
to
th
e
W
or
ld
H
e
a
lt
h
O
r
ga
ni
z
a
ti
on,
lu
ng
c
a
nc
e
r
w
il
l
a
c
c
ount
f
or
a
r
ound
1.8
m
i
ll
io
n
de
a
th
s
gl
oba
ll
y
in
2020,
a
c
c
ount
in
g
f
or
ne
a
r
ly
18%
of
a
ll
c
a
nc
e
r
f
a
ta
li
ti
e
s
, a
s
s
how
n i
n F
ig
ur
e
1
[
2
]
. A
c
c
or
di
n
g t
o da
ta
f
r
om
2019, 324,949
pa
ti
e
nt
s
i
n E
gypt
w
e
r
e
unde
r
goi
ng t
r
e
a
tm
e
nt
f
or
m
a
li
gna
nt
n
e
opl
a
s
m
s
a
t
th
e
e
xpe
ns
e
of
t
he
s
ta
te
[
3]
. A
s
p
e
r
t
he
D
e
c
e
m
be
r
2020
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
3D
v
is
ual
iz
at
io
n di
agnos
ti
c
s
f
o
r
l
ung c
anc
e
r
de
te
c
ti
on
(
R
ana M
. M
ahm
oud
)
4631
e
s
ti
m
a
te
s
f
r
om
t
he
G
lo
ba
l
C
a
nc
e
r
O
bs
e
r
va
to
r
y (
G
L
O
B
O
C
A
N
)
, t
he
m
os
t
pr
e
va
le
nt
c
a
nc
e
r
s
i
n E
gypt
, w
it
h a
5
-
ye
a
r
pr
e
va
le
nc
e
a
c
r
os
s
a
ll
a
g
e
gr
oups
,
in
c
lu
de
br
e
a
s
t,
lu
ng,
c
o
lo
r
e
c
ta
l,
pr
os
ta
te
,
s
to
m
a
c
h,
li
ve
r
,
a
nd
c
e
r
vi
c
a
l
c
a
nc
e
r
, t
ot
a
li
ng 8,879,843 c
a
s
e
s
f
or
a
ll
c
a
nc
e
r
s
, a
s
de
pi
c
te
d i
n F
ig
ur
e
2
[
4]
.
F
ig
ur
e
1. S
ta
ti
s
ti
c
s
f
or
c
a
nc
e
r
s
on t
he
gl
obe
[
2]
F
ig
ur
e
2. S
ta
ti
s
ti
c
s
pr
e
va
le
nt
c
a
nc
e
r
s
i
n E
gypt
[
4]
T
he
pr
im
a
r
y goa
l
now
a
da
ys
i
s
t
o de
te
c
t
a
nd f
or
e
c
a
s
t
c
a
nc
e
r
t
o be
gi
n t
r
e
a
tm
e
nt
a
s
qui
c
kl
y a
s
pos
s
ib
le
[
5]
.
S
c
r
e
e
ni
ng
a
nd
in
c
id
e
nt
a
l
f
in
di
ngs
a
r
e
th
e
two
m
os
t
pr
e
va
le
nt
m
e
th
ods
in
id
e
nt
if
yi
ng
lu
ng
c
a
nc
e
r
.
S
im
il
a
r
to
br
e
a
s
t
c
a
nc
e
r
s
c
r
e
e
ni
ng,
th
e
pr
im
a
r
y
a
ppr
oa
c
h
f
or
de
te
c
ti
ng
lu
ng
c
a
nc
e
r
s
houl
d
be
th
r
ough
s
c
r
e
e
ni
ng
[
6]
.
D
e
s
pi
te
th
is
,
m
os
t
n
a
ti
ons
,
in
c
lu
di
ng
th
e
U
K
,
la
c
k
a
s
c
r
e
e
ni
ng
pr
ogr
a
m
,
a
nd
m
os
t
c
a
s
e
s
a
r
e
de
te
c
te
d
by
c
oi
nc
id
e
nc
e
. L
ung nodule
s
a
r
e
di
s
c
ov
e
r
e
d une
xpe
c
te
dl
y w
hi
le
e
xa
m
in
in
g a
di
f
f
e
r
e
nt
or
ga
n f
r
om
t
he
one
be
in
g
e
xa
m
in
e
d.
N
odul
e
s
,
f
or
e
xa
m
pl
e
,
m
a
y
be
di
s
c
o
ve
r
e
d
dur
in
g
a
c
o
m
put
e
r
iz
e
d
to
m
ogr
a
phy
(
CT
)
s
c
a
n
of
th
e
he
a
r
t
or
li
ve
r
.
T
he
m
a
in
c
ha
ll
e
nge
is
th
a
t
r
a
di
ol
ogi
s
ts
la
c
k
th
e
ne
c
e
s
s
a
r
y
s
ki
ll
s
to
di
s
ti
ngui
s
h
be
twe
e
n
be
ni
gn
a
nd
m
a
li
gna
nt
nodule
s
[
7]
.
T
he
s
e
pa
ti
e
nt
s
s
houl
d
be
s
e
nt
to
a
pul
m
onol
ogi
s
t,
w
ho
w
il
l
pe
r
f
or
m
va
r
io
us
te
s
ts
to
c
he
c
k f
or
m
a
li
gna
nt
c
e
ll
s
a
nd
r
ul
e
out
a
ddi
ti
ona
l
m
e
di
c
a
l
c
ondi
ti
ons
. I
n t
hi
s
c
a
s
e
, c
li
ni
c
ia
ns
or
de
r
a
n
X
-
r
a
y o
f
th
e
lu
ngs
a
nd
lo
ok
f
or
s
ig
ns
of
a
tu
m
or
,
s
c
a
r
r
in
g,
or
a
b
ui
ld
up
of
f
lu
id
be
c
a
us
e
it
m
a
y
in
di
c
a
te
a
s
u
s
pi
c
io
us
lu
m
p or
nodule
[
8]
.
E
ve
n
m
in
ut
e
lu
ng
a
bnor
m
a
li
ti
e
s
th
a
t
a
n
X
-
r
a
y
m
ig
ht
m
is
s
c
a
n
be
de
te
c
te
d
by
a
C
T
s
c
a
n.
A
bi
ops
y
m
a
y
be
e
xe
c
ut
e
d vi
a
br
onc
hos
c
opy,
a
m
ong othe
r
t
e
c
hni
que
s
. T
he
ph
ys
ic
ia
n e
xa
m
in
e
s
a
be
r
r
a
nt
r
e
gi
ons
of
t
he
l
ung
s
by
in
s
e
r
ti
ng
a
li
t
in
s
tr
um
e
nt
dow
n
th
e
pa
ti
e
nt
'
s
e
s
oph
a
gus
a
nd
in
to
th
e
lu
ngs
.
M
e
di
a
s
ti
nos
c
opy
i
s
a
n
a
lt
e
r
na
ti
ve
te
c
hni
que
th
a
t
c
ol
le
c
ts
ly
m
ph
node
ti
s
s
u
e
s
a
m
pl
e
s
th
r
ough
th
e
in
s
e
r
ti
on
of
s
ur
gi
c
a
l
in
s
tr
um
e
nt
s
th
r
ough
a
n
in
c
is
io
n
m
a
de
a
t
th
e
ba
s
e
of
th
e
ne
c
k
a
nd
be
hi
nd
th
e
br
e
a
s
tb
one
.
A
n
a
lt
e
r
na
ti
ve
m
e
th
od
is
a
ne
e
dl
in
g
bi
ops
y,
dur
in
g
w
hi
c
h
th
e
phys
ic
ia
n
in
s
e
r
ts
a
ne
e
dl
e
in
to
th
e
lu
ng
ti
s
s
ue
vi
a
th
e
c
he
s
t
w
a
ll
a
nd
ut
il
iz
e
s
X
-
r
a
y
or
C
T
im
a
gi
ng
to
id
e
nt
if
y
a
bnor
m
a
l
c
e
ll
s
.
A
n
a
ddi
ti
ona
l
pos
s
ib
il
it
y
is
th
e
c
ol
le
c
ti
on
of
a
bi
ops
y
s
a
m
pl
e
f
r
om
ly
m
ph
node
s
or
ot
he
r
m
e
ta
s
ta
s
e
s
of
t
he
m
a
li
gna
nc
y, i
nc
lu
di
ng t
he
l
iv
e
r
[
9]
.
U
ti
li
z
in
g
th
r
e
e
-
di
m
e
ns
io
na
l
vi
s
ua
li
z
a
ti
on
di
a
gnos
ti
c
s
in
lu
ng
c
a
n
c
e
r
de
te
c
ti
on
e
na
bl
e
s
r
a
di
ol
ogi
s
t
s
a
nd
onc
ol
ogi
s
ts
be
tt
e
r
to
c
om
pr
e
he
nd
lu
ng
le
s
io
ns
'
s
p
a
ti
a
l
di
s
tr
ib
ut
io
n
a
nd
pr
ope
r
ti
e
s
.
T
hi
s
a
id
s
in
pr
e
c
i
s
e
di
a
gno
s
is
,
tr
e
a
tm
e
nt
s
tr
a
te
gi
z
in
g,
a
nd
ongoing
m
oni
to
r
in
g
of
lu
ng
c
a
nc
e
r
p
a
ti
e
nt
s
c
a
n
r
e
s
ul
t
in
im
pr
ove
d
pa
ti
e
nt
pr
ognos
e
s
ove
r
t
im
e
[
10
]
. T
he
a
ppl
ic
a
ti
on of
t
h
r
e
e
-
di
m
e
ns
io
na
l
vi
s
ua
li
z
a
ti
on dia
gnos
ti
c
s
i
n l
ung c
a
nc
e
r
de
te
c
ti
on a
ll
ow
s
r
a
di
ol
ogi
s
ts
a
nd
onc
ol
ogi
s
ts
to
ha
ve
a
m
or
e
c
om
pr
e
he
ns
iv
e
un
de
r
s
ta
ndi
ng
of
lu
ng
le
s
io
ns
'
s
pa
ti
a
l
di
s
tr
ib
ut
io
n
a
nd
c
ha
r
a
c
te
r
is
ti
c
s
.
T
hi
s
c
a
n
a
s
s
i
s
t
in
a
c
c
ur
a
te
di
a
gno
s
is
,
tr
e
a
t
m
e
nt
pl
a
nni
ng,
a
nd
m
oni
to
r
in
g
of
lu
ng
c
a
nc
e
r
pa
ti
e
nt
s
,
le
a
di
ng
to
im
pr
ove
d
pa
ti
e
nt
out
c
om
e
s
.
V
ol
um
e
tr
ic
r
e
c
ons
tr
uc
ti
on
is
p
e
r
f
or
m
e
d
in
2D
us
in
g
th
e
M
obi
le
N
e
t
a
ppr
oa
c
h
by
us
in
g
th
e
r
e
s
ul
t
a
s
b
e
gi
nni
ng
da
ta
a
nd
pr
opa
ga
ti
ng
th
e
r
e
s
ul
t.
I
t
is
a
c
c
om
pl
is
he
d
by
de
pl
oyi
ng ma
c
hi
ne
l
e
a
r
ni
ng on ma
s
s
iv
e
he
a
lt
hc
a
r
e
i
m
a
g
e
c
ol
le
c
ti
ons
.
L
iu
e
t
al
.
[
11]
p
r
opos
e
d a
da
ta
-
dr
iv
e
n m
e
th
od known a
s
t
he
c
a
s
c
a
de
d dua
l
-
pa
th
w
a
y r
e
s
id
ua
l
ne
twor
k
(
C
D
P
-
R
e
s
N
e
t)
,
le
ve
r
a
gi
ng
R
e
s
N
e
t
to
im
pr
ove
th
e
s
e
gm
e
nt
a
ti
on
of
lu
ng
nodule
s
in
C
T
im
a
ge
s
.
T
hi
s
m
ode
l
s
ubs
e
que
nt
ly
c
a
lc
ul
a
te
s
th
e
pr
oba
bi
li
ty
of
voxe
l
m
e
m
b
e
r
s
hi
p
w
it
hi
n
th
e
nodule
a
nd
pr
ovi
de
s
a
3D
vi
s
ua
li
z
a
ti
on
of
th
e
f
in
a
l
s
e
gm
e
nt
a
ti
on
out
c
om
e
.
F
our
r
a
di
ol
ogi
s
ts
e
xt
e
n
s
i
ve
ly
e
va
lu
a
te
d
th
a
t
m
e
th
od
on
a
lu
ng
im
a
ge
da
ta
ba
s
e
c
ons
or
ti
um
(
L
I
D
C
)
da
ta
s
e
t
c
ont
a
in
in
g
986
a
nnot
a
te
d
nodule
s
.
T
he
r
e
s
ul
ts
in
di
c
a
te
th
a
t
th
e
C
D
P
-
R
e
s
N
e
t
m
ode
l
out
pe
r
f
or
m
s
a
s
s
e
s
s
m
e
nt
s
c
ondu
c
te
d
by
f
our
di
f
f
e
r
e
nt
r
a
di
ol
ogi
s
ts
.
A
ddi
ti
ona
ll
y,
th
e
di
c
e
s
im
il
a
r
it
y
c
oe
f
f
ic
ie
nt
(
D
S
C
)
be
twe
e
n
C
D
P
-
R
e
s
N
e
t
a
nd
e
a
c
h
r
a
di
ol
ogi
s
t
a
ve
r
a
ge
s
82.69%
,
s
li
ght
ly
s
ur
pa
s
s
in
g
th
e
a
ve
r
a
ge
va
r
ia
bi
li
ty
obs
e
r
ve
d a
m
ong r
a
di
ol
ogi
s
ts
on t
h
e
L
I
D
C
da
ta
s
e
t,
w
hi
c
h s
ta
nd
s
a
t
82.66%
[
11]
.
L
iu
e
t
al
.
[
12]
us
e
d
th
e
c
ont
e
xt
a
tt
e
nt
io
n
ne
twor
k
(
C
A
-
N
e
t)
.
I
t
e
xt
r
a
c
ts
bot
h
c
ont
e
xt
u
a
l
f
e
a
tu
r
e
s
a
nd
nodule
s
,
e
f
f
e
c
ti
ve
ly
in
te
gr
a
ti
ng
th
e
m
dur
in
g
be
ni
gni
ty
/m
a
li
gna
nc
y
c
la
s
s
if
ic
a
ti
on.
T
o
pr
e
c
is
e
ly
c
a
pt
ur
e
0
5
0
0
0
0
0
1
0
0
0
0
0
0
1
5
0
0
0
0
0
2
0
0
0
0
0
0
2
5
0
0
0
0
0
Ca
s
es
Ca
s
es
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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4630
-
4641
4632
c
ont
e
xt
ua
l
f
e
a
tu
r
e
s
in
f
lu
e
nc
e
d
by
or
a
s
s
oc
ia
te
d
w
it
h
th
e
nod
ul
e
,
a
n
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
in
c
or
por
a
te
s
th
e
nodule
'
s
in
f
or
m
a
ti
on
a
s
a
r
e
f
e
r
e
nc
e
.
B
e
s
id
e
s
,
c
ont
e
xt
ua
l
f
e
a
t
ur
e
s
'
e
f
f
e
c
t
on
c
la
s
s
if
ic
a
ti
on
c
a
n
va
r
y
a
c
r
os
s
nodule
s
.
T
he
d
a
ta
s
c
ie
nc
e
bow
l
2017
(
D
S
B
)
is
a
da
ta
s
e
t
f
or
lu
ng
c
a
nc
e
r
pr
e
di
c
ti
on
us
e
d
f
or
te
s
ti
ng
a
nd
e
va
lu
a
ti
on. T
he
e
xp
e
r
im
e
nt
a
l
r
e
s
ul
ts
s
ho
w
t
ha
t
C
A
-
N
e
t
r
e
a
c
h
e
s
a
n a
c
c
ur
a
c
y of
83.79%
[
12]
.
M
e
i
e
t
al
.
[
13]
pr
opos
e
d
th
e
M
a
s
k
r
e
gi
on
-
ba
s
e
d
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
R
-
C
N
N
)
us
in
g
R
e
s
N
e
t
-
50
a
s
th
e
ba
c
kbone
a
nd
a
ppl
ie
d
f
e
a
tu
r
e
pyr
a
m
id
ne
tw
or
k
(
F
P
N
)
to
e
xpl
or
e
m
ul
ti
-
s
c
a
le
f
e
a
tu
r
e
m
a
ps
f
ul
ly
.
U
s
in
g
r
e
gi
on
pr
opos
a
l
ne
twor
k
(
R
P
N
)
to
pr
opos
e
c
a
ndi
d
a
te
bounding
boxe
s
.
T
he
lu
ng
nodule
a
n
a
ly
s
is
(
L
U
N
A
)
c
ha
ll
e
nge
is
te
s
te
d
a
nd
e
v
a
lu
a
te
d
on
th
e
publ
ic
ly
a
c
c
e
s
s
ib
le
L
U
N
A
16
da
ta
s
e
t
pr
ovi
de
d
a
s
pa
r
t
of
I
S
B
I
2016. E
xpe
r
im
e
nt
a
l
f
in
di
ngs
i
ndi
c
a
te
t
ha
t
M
a
s
k R
-
C
N
N
a
c
hi
e
v
e
s
a
s
e
ns
it
iv
it
y of
88.1%
[
13]
.
C
a
i
e
t
al
.
[
14]
pr
opos
e
d
a
n
a
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
lu
ng
im
a
gi
ng
a
na
ly
s
is
s
y
s
te
m
(
A
L
I
A
S
)
f
e
a
tu
r
in
g
lu
n
g
nodule
de
te
c
ti
on
a
nd s
e
gm
e
nt
a
ti
on ne
twor
k
s
.
T
he
t
hr
e
e
-
di
m
e
ns
i
ona
l
r
e
c
ti
f
ie
d l
in
e
a
r
uni
t
(
R
e
L
U
)
c
a
s
c
a
de
F
P
N
is
us
e
d
f
or
nodule
de
te
c
ti
on.
I
n
nodule
-
ba
s
e
d
a
n
a
ly
s
is
,
f
e
a
tu
r
e
s
s
uc
h
a
s
hi
s
to
gr
a
m
s
of
houns
f
ie
ld
uni
ts
(H
u
s
)
a
nd
r
a
di
oni
c
s
f
e
a
tu
r
e
s
a
r
e
e
xt
r
a
c
te
d.
T
he
s
e
f
e
a
tu
r
e
s
a
ll
ow
f
or
th
e
c
om
pa
r
is
on
of
di
s
c
r
e
pa
nc
ie
s
obs
e
r
ve
d
be
twe
e
n
m
a
li
gna
nt
a
nd
be
ni
gn
nodul
e
s
.
of
va
r
io
us
s
iz
e
s
.
T
he
A
L
I
A
S
unde
r
goe
s
te
s
ti
ng
a
nd
e
va
lu
a
ti
on
on
im
a
ge
s
obt
a
in
e
d
f
r
om
c
ol
la
bor
a
ti
ve
in
s
ti
tu
ti
ons
,
f
ol
lo
w
in
g
th
e
in
s
ti
tu
ti
ona
l
r
e
vi
e
w
boa
r
d
pr
ot
oc
ol
s
a
nd
r
e
s
pe
c
ti
ve
m
a
t
e
r
ia
l
tr
a
ns
f
e
r
a
gr
e
e
m
e
nt
s
.
T
hi
s
e
va
lu
a
ti
on
e
nc
o
m
pa
s
s
e
s
a
to
ta
l
of
8540
pul
m
ona
r
y
C
T
im
a
ge
s
f
r
om
7
,
716
pa
ti
e
nt
s
.
W
it
hi
n
th
e
te
s
ti
ng
s
e
t,
th
e
r
e
a
r
e
138
m
a
li
gna
nt
nodule
s
(
pos
it
iv
e
)
a
nd
91
be
ni
gn
nodule
s
(
ne
ga
ti
ve
)
, r
e
s
ul
ti
ng i
n a
n a
c
c
ur
a
c
y of
83.8%
[
14]
.
C
he
n
e
t
al
.
[
15]
in
tr
oduc
e
d
th
e
s
li
c
e
-
a
w
a
r
e
ne
twor
k
(
S
A
N
e
t)
f
or
th
e
de
te
c
ti
on
of
lu
ng
nodule
s
.
T
hi
s
ne
twor
k
ut
il
iz
e
s
a
in
te
gr
a
te
s
a
s
li
c
e
-
gr
oupe
d
non
-
lo
c
a
l
(
S
G
N
L
)
m
odul
e
a
nd
U
-
N
e
t
-
li
ke
s
tr
uc
tu
r
e
a
nd
R
e
s
B
lo
c
k
to
ge
ne
r
a
te
nodul
e
c
a
ndi
da
te
s
.
T
h
e
pe
r
f
or
m
a
nc
e
of
th
e
S
A
N
e
t
is
a
s
s
e
s
s
e
d
a
nd
va
li
da
te
d
u
s
in
g
th
e
pul
m
ona
r
y
nodule
da
ta
s
e
t
(
P
N
9
)
,
w
hi
c
h
c
om
pr
is
e
s
40,439
a
nnot
a
te
d
a
nd
nodule
s
8,798
th
or
a
c
ic
C
T
s
c
a
ns
.
T
h
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
s
how
t
ha
t
S
A
N
e
t
r
e
a
c
he
s
a
pr
e
c
is
io
n of
35
.92%
a
nd a
r
e
c
a
ll
of
70.20%
, w
hi
c
h m
e
a
ns
th
e
a
c
c
ur
a
c
y i
s
87%
[
15]
.
M
k
in
d
u
e
t
al
.
[
16]
pr
o
po
s
e
d
a
c
o
m
pu
te
r
-
a
i
de
d
d
e
t
e
c
ti
o
n
(
C
A
D
)
s
c
h
e
m
e
in
c
or
por
a
ti
ng
a
3
D
m
ul
ti
-
s
c
a
l
e
vi
s
io
n t
r
a
n
s
f
or
m
e
r
(
M
S
V
i
T
)
.
T
hi
s
a
r
c
hi
te
c
t
ur
e
e
m
pl
oy
s
a
l
oc
a
l
-
g
lo
b
a
l
tr
a
n
s
f
or
m
e
r
bl
o
c
k
s
tr
u
c
t
ur
e
,
w
it
h t
h
e
lo
c
a
l
tr
a
n
s
f
or
m
e
r
s
t
a
g
e
pr
o
c
e
s
s
in
g
e
a
c
h
s
c
a
l
e
pa
tc
h
i
nd
e
p
e
n
de
nt
l
y
be
f
or
e
m
e
r
gi
ng
th
e
m
in
to
th
e
gl
o
b
a
l
tr
a
n
s
f
or
m
e
r
le
v
e
l
t
o
in
c
or
por
a
t
e
m
u
lt
i
-
s
c
a
le
f
e
a
tu
r
e
s
.
T
h
e
C
A
D
s
c
h
e
m
e
pr
o
po
s
e
d
w
a
s
te
s
t
e
d
on
8
88
C
T
i
m
a
ge
s
f
r
om
th
e
publ
ic
ly
a
va
i
la
bl
e
L
U
N
A
16
d
a
t
a
s
e
t.
T
he
3
D
-
M
S
V
i
T
a
lg
or
it
hm
a
c
hi
e
v
e
d
a
n
a
c
c
ur
a
c
y
r
a
te
of
9
1.1%
[
1
6]
.
Z
ha
ng
a
nd
Z
ha
ng
[
17]
pr
opos
e
d
th
a
t
a
3D
s
e
le
c
ti
ve
k
e
r
ne
l
r
e
s
id
ua
l
ne
twor
k
(
S
K
-
R
e
s
N
e
t)
ba
s
e
d
on
th
e
s
e
le
c
ti
ve
ke
r
ne
l
ne
twor
k
a
nd
th
r
e
e
-
di
m
e
n
s
io
na
l
r
e
s
id
ua
l
ne
twor
k
is
lo
c
a
te
d.
A
3D
R
P
N
,
e
m
pl
oyi
ng
SK
-
R
e
s
N
e
t,
is
de
ve
lo
p
e
d
f
or
th
e
id
e
nt
if
ic
a
ti
on
of
lu
ng
nodule
s
,
a
c
c
om
pa
ni
e
d
by
a
m
ul
ti
-
s
c
a
le
f
e
a
tu
r
e
f
us
io
n
ne
twor
k t
a
il
or
e
d f
or
nodule
c
la
s
s
if
ic
a
ti
on.
T
he
e
f
f
e
c
ti
ve
ne
s
s
of
t
he
m
e
th
od i
s
e
xa
m
in
e
d
a
nd
a
s
s
e
s
s
e
d u
s
in
g t
he
publ
ic
ly
a
va
il
a
bl
e
L
U
N
A
16 da
ta
s
e
t.
T
h
e
S
K
-
R
e
s
N
e
t
a
lg
or
it
hm
a
c
hi
e
ve
d a
n a
c
c
ur
a
c
y r
a
te
of
91.75%
[
17]
.
T
hi
s
p
a
pe
r
us
e
s
th
e
M
obi
le
N
e
t
m
od
e
l
th
a
t
in
t
e
gr
a
te
s
m
ul
ti
-
vi
e
w
a
nd
m
ul
ti
-
s
c
a
le
c
ha
r
a
c
te
r
is
ti
c
s
of
va
r
io
us
nodule
s
f
r
om
C
T
im
a
ge
s
.
T
h
e
n,
th
e
th
r
e
e
-
di
m
e
ns
io
na
l
pul
m
ona
r
y
nodula
r
m
ode
ls
w
e
r
e
c
r
e
a
te
d
us
in
g
a
r
a
y
-
c
a
s
ti
ng
vol
um
e
r
e
nde
r
in
g
a
ppr
oa
c
h.
T
he
c
r
e
a
ti
on
of
th
e
th
r
e
e
-
di
m
e
ns
io
na
l
m
ode
l
w
il
l
of
f
e
r
lo
s
s
le
s
s
r
e
c
ons
tr
uc
ti
on,
pr
ovi
di
ng
a
m
or
e
r
e
a
li
s
ti
c
de
pi
c
ti
on
of
th
e
tu
m
or
c
om
pa
r
e
d
to
th
e
w
ir
e
d
m
ode
l,
ut
il
iz
in
g
th
e
th
r
e
e
-
di
m
e
ns
io
na
l
r
e
c
ons
tr
uc
ti
on
of
th
e
tu
m
or
c
e
ll
s
.
I
t
w
il
l
a
ll
o
w
th
e
phys
ic
ia
n
c
a
n
m
or
e
e
f
f
e
c
ti
ve
ly
e
xa
m
in
e
th
e
r
e
la
ti
ons
hi
p
be
twe
e
n
th
e
tu
m
or
a
nd
i
ts
s
ur
r
oundings
,
e
ve
n
be
f
or
e
s
ur
ge
r
y.
T
hi
s
pr
opos
e
d
s
ys
te
m
w
il
l
a
na
ly
z
e
lu
ng
da
m
a
ge
gl
oba
ll
y:
th
e
f
ul
l
C
T
im
a
ge
,
us
in
g
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
he
s
.
T
hi
s
r
e
s
e
a
r
c
h
ha
s
th
e
f
ol
lo
w
in
g
c
ont
r
ib
ut
io
ns
:
i
)
th
e
s
ys
te
m
s
ugge
s
t
s
a
th
r
e
e
-
di
m
e
ns
io
na
l
vi
s
ua
li
z
a
ti
on
te
c
hni
que
f
or
c
om
put
e
r
-
a
id
e
d
lu
ng
nodule
id
e
nt
if
ic
a
ti
on
ba
s
e
d
on
M
obi
le
N
e
t
a
nd
th
e
r
a
y
-
c
a
s
ti
ng
vol
um
e
r
e
nde
r
in
g
m
e
th
odol
ogy
a
nd
ii
)
c
la
s
s
if
y
pul
m
ona
r
y
nodule
s
.
T
h
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
in
di
c
a
te
th
a
t
M
obi
le
N
e
t
i
s
us
e
f
ul
f
or
s
e
gm
e
nt
in
g
a
nd
c
la
s
s
if
yi
ng l
ung nodules
.
2.
M
E
T
H
O
D
R
e
c
e
nt
s
tu
di
e
s
ha
ve
unde
r
s
c
or
e
d
s
ig
ni
f
ic
a
nt
a
dva
nc
e
m
e
nt
s
in
m
e
di
c
a
l
im
a
ge
s
e
gm
e
nt
a
ti
on
e
na
bl
e
d
by t
he
a
ppl
ic
a
ti
on of
de
e
p l
e
a
r
ni
ng me
th
odol
ogi
e
s
. T
h
e
i
m
pr
ove
m
e
nt
i
s
c
le
a
r
, e
s
p
e
c
ia
ll
y W
he
n
c
om
pa
r
in
g t
he
out
c
om
e
s
of
tr
a
di
ti
ona
l
s
e
gm
e
nt
a
ti
on
a
ppr
oa
c
h
e
s
w
it
h
th
o
s
e
of
de
e
p
le
a
r
ni
ng
te
c
hni
que
s
in
li
ve
r
de
li
ne
a
ti
on.
T
he
pr
opos
e
d
s
y
s
te
m
de
te
c
ts
lu
ng
tu
m
or
s
a
ut
om
a
t
ic
a
ll
y.
I
t
s
e
gm
e
nt
s
th
e
C
T
im
a
ge
s
of
th
e
lu
ng
to
r
e
c
ogni
z
e
.
T
he
m
e
th
od
r
e
c
ons
tr
uc
ts
th
e
tu
m
or
in
th
r
e
e
-
di
m
e
ns
io
na
l
im
a
ge
s
w
it
h
tu
m
or
vol
um
e
tr
y.
T
he
s
ys
te
m
in
c
lu
de
s
th
r
e
e
pha
s
e
s
,
a
s
s
how
n i
n F
ig
ur
e
3
.
−
T
he
pr
e
-
pr
oc
e
s
s
in
g
m
odul
e
(
P
r
M
)
pha
s
e
c
onve
r
ts
lu
ng
im
a
ge
s
f
r
om
w
or
ld
c
oor
di
na
te
s
to
im
a
g
e
c
oor
di
na
te
s
, s
pl
it
ti
ng i
m
a
ge
s
i
nt
o s
li
c
e
s
a
nd nor
m
a
li
z
in
g i
m
a
ge
s
.
−
T
he
de
te
c
ti
on modul
e
(
D
M
)
pha
s
e
de
te
c
ts
l
ung tum
or
s
.
−
T
he
th
r
e
e
-
di
m
e
ns
io
na
l
r
e
c
ons
tr
uc
ti
on
m
odul
e
(
3D
R
M
)
p
ha
s
e
vi
s
ua
li
z
e
s
th
e
lu
ng
tu
m
or
s
in
th
r
e
e
-
di
m
e
ns
io
na
l.
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
3D
v
is
ual
iz
at
io
n di
agnos
ti
c
s
f
o
r
l
ung c
anc
e
r
de
te
c
ti
on
(
R
ana M
. M
ahm
oud
)
4633
F
ig
ur
e
3.
T
he
pr
opos
e
d m
ode
l
-
ba
s
e
d
s
e
gm
e
nt
a
ti
on a
nd
d
e
te
c
ti
o
n f
r
a
m
e
w
or
k f
or
pul
m
ona
r
y nodu
le
t
hr
e
e
-
d
im
e
ns
io
na
l
vi
s
ua
li
z
a
ti
on dia
gno
s
is
2
.1.
P
r
e
-
p
r
oc
e
s
s
in
g m
od
u
le
T
he
in
put
da
t
a
C
T
im
a
ge
s
a
r
e
r
e
pr
e
s
e
nt
e
d
by
di
vi
di
ng
th
r
e
e
-
di
m
e
ns
io
na
l
s
c
a
ns
in
to
2D
s
li
c
e
s
.
T
he
y
w
e
r
e
r
e
s
a
m
pl
e
d
a
nd
nor
m
a
li
z
e
d
to
e
a
c
h
c
om
pone
nt
.
F
in
a
ll
y,
e
a
c
h
s
li
c
e
pi
c
tu
r
e
is
c
onve
r
te
d
in
to
a
bi
na
r
y
im
a
ge
.
T
he
bl
obs
r
e
la
te
d
to
th
e
im
a
ge
'
s
bor
de
r
a
r
e
r
e
m
ove
d
a
s
a
pr
e
-
pr
oc
e
s
s
in
g s
te
p
to
r
e
m
ove
a
ny
a
r
ti
f
a
c
ts
or
noi
s
e
th
a
t
m
a
y
be
pr
e
s
e
nt
in
th
e
da
ta
.
T
hi
s
pha
s
e
a
s
s
is
t
s
in
in
c
r
e
a
s
i
ng
th
e
a
c
c
ur
a
c
y
a
nd
de
pe
nda
bi
li
ty
of
a
n
a
ly
ti
c
pr
oc
e
s
s
e
s
li
ke
s
e
gm
e
nt
a
ti
on
or
f
e
a
tu
r
e
e
xt
r
a
c
ti
on.
T
he
two
bi
gg
e
s
t
por
ti
ons
a
r
e
pr
e
s
e
r
ve
d
a
nd
ut
il
iz
e
d
to
di
vi
de
th
e
two
lu
ngs
.
H
ol
e
s
f
il
le
d
th
e
m
a
s
k,
s
e
pa
r
a
ti
ng
th
e
lu
ngs
a
nd
r
e
ve
a
li
ng
e
a
c
h
lu
ng'
s
c
onve
x
hul
l.
T
h
e
two
lu
ngs
a
r
e
j
oi
ne
d t
o ke
e
p nodule
s
f
ix
e
d t
o t
he
l
ung wa
ll
, a
s
s
how
n i
n F
ig
ur
e
4
.
F
ig
ur
e
4. P
r
e
-
pr
oc
e
s
s
in
g
m
odul
e
2.1.1. Coor
d
in
at
e
m
ap
p
in
g c
om
p
on
e
n
t
s
T
he
r
a
w
c
oor
di
na
te
s
of
m
e
di
c
a
l
im
a
ge
d
a
ta
a
r
e
e
xpr
e
s
s
e
d
in
th
e
gl
oba
l
c
oor
di
na
te
s
ys
te
m
.
I
n
(
1)
r
e
pr
e
s
e
nt
s
tr
a
ns
f
or
m
in
g
th
e
gl
oba
l
c
oor
di
na
te
f
r
a
m
e
w
or
k
to
th
e
im
a
ge
-
ba
s
e
d
c
oor
di
na
te
s
tr
uc
tu
r
e
.
F
ig
ur
e
5
vi
s
ua
ll
y de
pi
c
ts
t
hi
s
c
onve
r
s
io
n pr
oc
e
s
s
f
or
be
tt
e
r
unde
r
s
ta
ndi
ng
[
18]
.
I
m
a
ge
_c
oor
di
na
te
s
=
|
w
o
r
l
d
_
co
o
r
d
in
a
te
s
−
o
r
ig
in
|
(
1)
F
ig
ur
e
5. M
e
di
c
a
l
im
a
ge
-
to
-
im
a
ge
c
oor
di
na
te
s
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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4630
-
4641
4634
2.1.2. S
li
c
e
s
e
le
c
t
io
n
T
he
Z
-
a
xi
s
in
th
e
im
a
g
e
c
oor
di
na
te
s
y
s
te
m
e
xt
e
nds
f
r
om
lo
w
e
r
to
uppe
r
.
T
he
lo
gi
c
a
l
e
qu
a
ti
on
r
e
pr
e
s
e
nt
e
d
f
or
th
e
di
a
m
e
te
r
,
s
li
c
e
th
ic
kn
e
s
s
,
a
nd
c
oor
di
na
te
s
in
th
e
a
nnot
a
te
d
f
il
e
de
f
in
e
th
e
pos
it
io
n,
w
r
it
te
n
s
e
le
c
te
d z
,
a
nd t
he
numbe
r
of
c
hos
e
n
s
li
c
e
s
, w
r
it
te
n
a
s
i
n (
2)
[
19]
, [
20]
.
EN
=
{
T
r
ue
(
−
n
∗
s
>
=
3
)
a
n
d
(
c
−
n
<
=
s
z
)
a
n
d
(
c
+
n
>
=
s
z
)
F
a
l
s
e
o
t
he
r
wis
e
(
2)
W
he
r
e
E
N
is
e
f
f
e
c
ti
ve
nodule
s
,
d
is
di
a
m
e
te
r
,
s
is
s
li
c
e
t
h
i
c
kne
s
s
,
s
z
is
s
e
le
c
te
d
z
,
a
nd
c
is
c
oor
di
na
te
z
m
a
y
be
de
r
iv
e
d
f
r
om
th
e
a
nnot
a
te
d
f
il
e
.
E
f
f
e
c
ti
ve
nodule
s
w
e
r
e
de
f
in
e
d
a
s
nodule
s
gr
e
a
te
r
th
a
n
3
m
m
in
di
a
m
e
te
r
,
a
s
s
how
n i
n F
ig
ur
e
6
.
F
ig
ur
e
6. I
m
a
ge
c
oor
di
na
te
s
i
nt
o s
li
c
e
s
of
a
n i
m
a
ge
2.1.3.
N
or
m
al
iz
at
io
n
T
he
r
e
la
ti
ve
de
ns
it
y
of
ti
s
s
ue
s
a
nd
or
ga
ns
in
C
T
s
c
a
n
s
is
r
e
pr
e
s
e
nt
e
d
by
th
e
H
u.
T
he
H
ouns
f
ie
ld
va
lu
e
of
bone
s
i
n a
c
he
s
t
C
T
s
c
a
n i
s
l
a
r
ge
r
t
ha
n 400, wh
e
r
e
a
s
t
he
H
ou
ns
f
ie
ld
va
lu
e
of
t
he
l
ung r
a
nge
s
f
r
om
1000 Hu
to
400 Hu. T
he
f
in
a
l
p
ic
tu
r
e
pi
xe
l
va
lu
e
s
w
e
r
e
w
in
dow
e
d i
n
th
e
[
1000, 400]
r
a
nge
be
f
or
e
be
in
g nor
m
a
li
z
e
d t
o
[
0, 1]
, a
s
s
how
n i
n F
ig
ur
e
7
.
F
ig
ur
e
7. N
or
m
a
li
z
a
ti
on
2
.
2
.
D
e
t
e
c
t
io
n
m
od
u
le
T
hi
s
p
a
r
t
w
il
l
di
s
c
u
s
s
th
e
M
obi
le
N
e
t
c
or
e
la
ye
r
s
,
w
hi
c
h
a
r
e
de
pt
h
-
w
is
e
s
e
p
a
r
a
bl
e
f
il
te
r
s
.
T
he
s
ubs
e
que
nt
s
e
c
ti
on 2.2.2
e
la
bor
a
te
s
on t
he
ove
r
a
ll
s
tr
uc
tu
r
e
of
t
he
M
obi
le
N
e
t
n
e
twor
k. A
vi
s
ua
l
r
e
pr
e
s
e
nt
a
ti
on
of
t
hi
s
s
tr
uc
tu
r
e
i
s
pr
e
s
e
nt
e
d i
n F
ig
ur
e
8 f
or
be
tt
e
r
c
onc
e
pt
ua
li
z
a
ti
on.
F
ig
ur
e
8. M
obi
le
N
e
t
a
r
c
hi
te
c
tu
r
e
[
21]
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
3D
v
is
ual
iz
at
io
n di
agnos
ti
c
s
f
o
r
l
ung c
anc
e
r
de
te
c
ti
on
(
R
ana M
. M
ahm
oud
)
4635
2.2.1.
D
e
p
t
h
-
w
is
e
s
e
p
ar
ab
le
c
on
vol
u
t
io
n
A
n
e
xa
m
pl
e
of
a
f
a
c
to
r
iz
e
d
c
onvolut
io
n
is
th
e
de
pt
h
-
w
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
n,
w
hi
c
h
ta
ke
s
a
r
e
gul
a
r
c
onvolut
io
n
a
nd
s
pl
it
s
it
in
to
two:
de
pt
h
-
w
is
e
a
nd
poi
nt
w
is
e
.
A
s
s
uc
h,
th
e
M
obi
le
N
e
t
m
ode
l
is
ba
s
e
d
on
th
is
de
s
ig
n.
W
it
h
de
pt
h
-
w
is
e
c
onvolut
io
n,
M
obi
le
N
e
t
pr
oc
e
s
s
e
s
e
a
c
h
in
put
c
ha
nne
l
w
it
h
a
s
in
gl
e
f
il
te
r
.
E
le
ve
n
c
onvolut
io
ns
a
r
e
e
xe
c
ut
e
d by me
r
gi
ng t
he
r
e
s
ul
ts
of
de
pt
h
-
w
is
e
a
nd point
w
is
e
c
onvolut
io
ns
. C
om
bi
ne
s
i
nput
s
w
it
h
or
di
na
r
y
c
onvolut
io
n
f
il
te
r
s
to
pr
oduc
e
m
or
e
out
put
s
a
ll
a
t
onc
e
.
A
de
pt
h
-
w
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
n
c
r
e
a
te
s
a
f
il
te
r
in
g
la
ye
r
a
nd
a
c
om
bi
n
in
g
la
ye
r
.
C
om
put
in
g
a
nd mode
l
s
iz
e
a
r
e
bot
h
s
ig
ni
f
ic
a
nt
ly
r
e
duc
e
d
us
in
g
th
e
f
a
c
to
r
iz
a
ti
on
te
c
hni
que
.
T
he
s
qu
a
r
e
in
put
f
e
a
tu
r
e
m
a
p
F
'
s
s
pa
ti
a
l
w
id
th
a
nd
he
ig
ht
,
de
not
e
d
a
s
D
a
ta
F
r
a
m
e
(
DF
)
,
a
nd
th
e
f
e
a
tu
r
e
m
a
p
G
th
a
t
r
e
s
ul
ts
f
r
om
it
a
r
e
r
e
pr
e
s
e
nt
e
d
a
s
D
F
×
D
F
×
N
.
I
n
a
s
qua
r
e
out
put
f
e
a
tu
r
e
m
a
p,
dyna
m
ic
gr
oup
(
DG
)
de
not
e
s
th
e
s
pa
ti
a
l
e
xt
e
nt
a
nd
he
ig
h
t,
a
nd
N
th
e
num
be
r
of
ou
tp
ut
c
ha
nne
ls
(
in
put
de
pt
h)
.
O
ne
m
il
li
on
in
di
c
a
te
s
th
e
de
ns
it
y
of
in
put
c
ha
nne
ls
in
a
gi
ve
n
s
pa
ti
a
l
c
ont
e
xt
.
T
he
c
onve
nt
io
na
l
c
onvolut
io
na
l
la
ye
r
us
e
s
a
c
onvolut
io
n
ke
r
ne
l
K
w
it
h
pa
r
a
m
e
te
r
s
D
K
*
D
K
*
M
*
N
, w
he
r
e
D
K
is
th
e
s
pa
ti
a
l
di
m
e
ns
io
n of
t
he
k
e
r
ne
l,
M
i
s
t
h
e
numbe
r
of
i
nput
c
ha
nne
l
s
, a
nd
N
i
s
t
he
numbe
r
of
out
put
c
ha
nne
l
s
.
T
hi
s
i
s
in
ke
e
pi
ng
w
it
h
w
h
a
t
w
a
s
pr
e
vi
ous
ly
de
s
c
r
ib
e
d.
F
e
a
tu
r
e
m
a
ps
pr
oduc
e
d
by c
onve
nt
io
na
l
c
onvolut
io
n
w
it
h
s
tr
id
e
one
a
nd pa
ddi
ng l
ook li
ke
t
hi
s
:
,
,
=
∑
+
−
1
,
+
−
1
,
∙
,
,
,
,
,
T
he
c
om
put
a
ti
ona
l
c
o
s
t
of
s
ta
nda
r
d c
onvolut
io
ns
i
s
:
∙
∙
∙
∙
∙
I
f
th
e
c
om
put
a
ti
ona
l
c
os
t
de
p
e
nds
on
th
e
s
iz
e
of
th
e
f
e
a
tu
r
e
m
a
p,
ke
r
ne
l,
in
put
c
ha
nne
l
s
,
a
nd
out
put
c
ha
nne
ls
, t
he
n...W
e
c
ov
e
r
a
ll
t
he
s
e
ke
yw
or
ds
a
nd t
he
ir
l
in
ka
ge
s
i
n M
obi
le
N
e
t
m
ode
ls
. T
o de
c
oupl
e
t
he
ke
r
ne
l
s
iz
e
f
r
om
th
e
num
be
r
of
out
put
c
ha
nne
ls
,
w
e
b
e
g
in
w
it
h
de
pt
h
-
w
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
ns
.
T
o
c
r
e
a
te
a
ne
w
r
e
pr
e
s
e
nt
a
ti
on,
th
e
c
onve
nt
io
na
l
c
onvolut
io
n
m
e
th
od
us
e
s
c
onvo
lu
ti
ona
l
ke
r
ne
ls
to
f
il
te
r
a
nd
c
om
bi
ne
da
ta
.
T
he
us
e
of
de
pt
h
-
w
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
n
s
a
ll
ow
s
f
or
a
s
ig
ni
f
ic
a
nt
r
e
duc
ti
on
in
c
om
put
a
ti
on
c
os
ts
by
s
pl
it
ti
ng
th
e
f
il
te
r
in
g
a
nd
c
om
bi
na
ti
on
s
ta
ge
s
in
to
two
pa
r
ts
.
T
w
o
la
ye
r
s
m
a
ke
up
de
pt
h
-
w
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
ns
:
T
he
two
m
a
in
ty
pe
s
of
c
onvolut
io
ns
a
r
e
poi
nt
w
i
s
e
a
nd
d
e
pt
h
-
w
is
e
.
A
ppl
yi
ng
de
pt
h
-
w
is
e
c
onvolut
io
ns
(
in
put
de
pt
h)
w
it
h
a
s
in
gl
e
f
il
te
r
pe
r
in
put
c
ha
nne
l.
A
ppl
yi
ng
a
ba
s
ic
11
c
onvolut
io
n
f
ol
lo
w
s
th
e
li
ne
a
r
c
om
bi
na
ti
o
n
of
th
e
de
pt
h
-
w
is
e
la
ye
r
'
s
out
put
us
in
g
poi
nt
w
is
e
c
onvolut
io
n
.
F
or
bot
h
la
ye
r
s
,
M
obi
le
N
e
t
e
m
pl
oys
ba
tc
h
nor
m
a
li
z
a
ti
on
a
nd
R
e
L
U
a
c
ti
va
ti
ons
.
H
e
r
e
is
a
n
e
x
a
m
pl
e
of
how
t
o
e
xpr
e
s
s
nor
m
a
l
c
onvolut
io
n
us
in
g
one
f
il
te
r
f
or
e
a
c
h i
nput
c
ha
nne
l
(
in
put
de
pt
h)
,
,
,
=
∑
+
−
1
,
+
−
1
,
∙
,
,
,
,
,
I
f
w
e
ta
ke
th
e
D
K
D
K
M
de
pt
h
-
w
is
e
c
onvolut
io
na
l
ke
r
ne
l
K
a
nd
a
ppl
y
it
to
th
e
m
th
c
ha
nne
l
in
F
,
w
e
ge
t
th
e
m
th
c
ha
nne
l
of
t
he
f
il
te
r
e
d output
f
e
a
tu
r
e
m
a
p G
.
2.2.2.
N
e
t
w
or
k
s
t
r
u
c
t
u
r
e
an
d
t
r
ai
n
in
g
I
n
or
de
r
to
di
s
c
ove
r
a
good
n
e
twor
k
to
pol
ogy,
w
e
m
a
y
qui
c
kl
y
a
s
s
e
s
s
th
e
m
by
gi
vi
ng
a
c
le
a
r
de
s
c
r
ip
ti
on
of
th
e
ne
twor
k.
F
or
nonl
in
e
a
r
c
la
s
s
if
ic
a
ti
on,
a
ll
la
ye
r
s
e
xc
e
pt
th
e
la
s
t
c
om
pl
e
te
ly
li
nke
d
one
f
e
e
d
in
to
a
S
of
tM
a
x
la
ye
r
.
A
ba
tc
h
nor
m
a
nd
R
e
L
U
non
li
ne
a
r
it
y
f
o
ll
ow
e
a
c
h
la
ye
r
.
W
he
n
c
om
pa
r
in
g
two
la
ye
r
s
,
one
us
e
s
c
onve
nt
io
na
l
R
e
L
U
nonl
in
e
a
r
it
y,
ba
tc
h
-
nor
m
,
a
nd c
on
vol
ut
io
ns
,
w
hi
le
th
e
ot
he
r
i
s
a
f
a
c
to
r
iz
e
d
la
ye
r
th
a
t
us
e
s
de
pt
h
-
w
is
e
c
onvolut
io
n, 1×
1 point
w
i
s
e
c
onvolut
io
n, b
a
tc
h
-
nor
m
, a
nd R
e
L
U
a
f
te
r
e
a
c
h c
onv
ol
ut
io
na
l
la
ye
r
[
22]
.
I
s
th
e
f
ir
s
t
la
ye
r
a
nd
th
e
de
pt
h
-
w
is
e
c
onvolut
io
ns
us
e
s
tr
id
e
c
onvolut
io
n
to
h
a
ndl
e
dow
n
-
s
a
m
pl
in
g.
U
s
in
g
a
la
s
t
a
ve
r
a
ge
pool
in
g
pr
oc
e
dur
e
r
e
duc
e
s
th
e
s
pa
ti
a
l
r
e
s
o
lu
ti
on
to
1
be
f
or
e
th
e
f
ul
ly
li
nke
d
la
ye
r
.
T
he
r
e
a
r
e
a
to
ta
l
of
28
l
a
ye
r
s
in
M
obi
le
N
e
t
w
he
n
you
s
e
pa
r
a
te
th
e
de
pt
h
-
w
is
e
a
nd
poi
nt
w
is
e
c
onvolut
io
ns
.
I
t
is
in
a
de
qua
te
to
de
s
c
r
ib
e
ne
twor
ks
us
in
g
ju
s
t
a
s
m
a
ll
num
be
r
of
m
ul
ti
pl
y
-
a
dds
.
E
f
f
ic
ie
nt
ly
im
pl
e
m
e
nt
in
g
th
e
s
e
a
c
ti
vi
ti
e
s
is
of
th
e
ut
m
os
t
im
por
ta
nc
e
.
A
s
a
n
e
xa
m
pl
e
,
unl
e
s
s
th
e
s
pa
r
s
it
y
le
v
e
l
is
r
e
a
ll
y
hi
gh,
de
n
s
e
m
a
tr
ix
ope
r
a
ti
ons
w
il
l
of
te
n
out
pe
r
f
or
m
uns
tr
uc
tu
r
e
d
s
pa
r
s
e
one
s
.
D
e
e
p
1
×
1
c
onvolut
io
ns
in
c
lu
de
a
ll
c
a
lc
ul
a
ti
ons
in
s
id
e
th
e
s
y
s
te
m
s
tr
uc
tu
r
e
.
V
e
r
y
e
f
f
ic
ie
nt
ge
ne
r
ic
m
a
tr
ix
m
ul
ti
pl
y
(
G
E
M
M
)
r
out
in
e
s
m
a
ke
th
is
pos
s
ib
le
.
T
he
in
it
ia
l
m
e
m
or
y
r
e
or
de
r
in
g
c
a
ll
e
d
im
2c
ol
is
ne
c
e
s
s
a
r
y
to
a
li
gn
it
w
it
h
a
G
E
M
M
,
w
hi
c
h
i
s
ty
pi
c
a
ll
y
u
s
e
d
to
c
on
duc
t
c
onvolut
io
ns
[
23]
.
T
he
popula
r
c
a
f
f
e
pa
c
ka
ge
,
f
or
in
s
ta
nc
e
,
m
a
ke
s
u
s
e
of
th
is
te
c
hni
que
.
O
ne
of
th
e
be
s
t
num
e
r
ic
a
l
li
ne
a
r
a
lg
e
br
a
m
e
th
ods
,
G
E
M
M
,
a
ll
ow
s
you
t
o
e
a
s
il
y
im
pl
e
m
e
nt
1×
1
c
onvolut
io
ns
w
it
hout
r
e
or
de
r
in
g m
e
m
or
y.
a.
T
hr
e
e
-
di
m
e
ns
io
na
l
r
e
c
ons
tr
uc
ti
on modul
e
T
hr
e
e
-
di
m
e
ns
io
na
l
r
e
c
on
s
tr
uc
ti
on
c
r
e
a
te
s
a
3D
m
ode
l
of
a
s
c
e
ne
f
r
om
a
s
e
t
of
two
-
di
m
e
ns
io
na
l
im
a
ge
s
or
a
n
obj
e
c
t
or
vi
de
o
f
r
a
m
e
.
T
hi
s
te
c
hnol
ogy
is
w
id
e
ly
u
s
e
d
in
e
nt
e
r
ta
in
m
e
nt
,
m
e
di
c
a
l
im
a
gi
ng,
a
nd
c
om
put
e
r
vi
s
io
n. A
3D
r
e
c
ons
tr
uc
ti
on modul
e
t
ypi
c
a
ll
y c
on
s
i
s
ts
of
t
he
f
ol
lo
w
in
g s
te
ps
:
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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4630
-
4641
4636
−
T
he
in
it
ia
l
s
ta
ge
is
to
c
ol
le
c
t
a
c
ol
le
c
ti
on
of
two
-
di
m
e
ns
io
na
l
lu
ng
im
a
ge
s
.
T
he
s
e
im
a
ge
s
a
r
e
c
r
e
a
te
d
by
di
vi
di
ng 3D image
s
i
nt
o s
li
c
e
s
. E
a
c
h
s
li
c
e
i
s
a
2D
i
m
a
ge
obt
a
in
e
d f
r
om
m
e
di
c
a
l
im
a
gi
ng e
qui
pm
e
nt
.
−
I
n
th
e
ne
xt
s
te
p,
im
a
ge
s
a
r
e
pr
oc
e
s
s
e
d
to
e
xt
r
a
c
t
f
e
a
tu
r
e
s
th
a
t
m
a
y
be
us
e
d
to
de
te
c
t
c
om
m
on
poi
nt
s
be
twe
e
n
th
e
m
.
T
he
s
e
e
xt
r
a
c
te
d
f
e
a
tu
r
e
s
a
r
e
c
r
uc
ia
l
in
de
te
c
ti
ng
c
om
m
on
poi
nt
s
be
twe
e
n
th
e
im
a
ge
s
.
T
h
e
y
c
ont
r
ib
ut
e
to
t
he
a
lg
or
it
hm
'
s
a
bi
li
ty
t
o i
de
nt
if
y a
nd
m
a
tc
h ke
y point
s
e
f
f
e
c
ti
ve
ly
.
−
T
h
e
r
e
tr
ie
ve
d
f
e
a
tu
r
e
s
c
a
l
c
u
la
te
th
e
c
or
r
e
s
p
ond
in
g
p
oi
nt
s
b
e
tw
e
e
n
t
he
im
a
g
e
s
i
n
th
i
s
s
t
a
g
e
.
T
hi
s
pr
oc
e
s
s
in
vo
lv
e
s
c
om
pa
r
in
g
t
h
e
f
e
a
tu
r
e
s
of
e
a
c
h
i
m
a
g
e
a
n
d
s
e
le
c
ti
ng
th
e
be
s
t
m
a
tc
h t
o
e
s
ta
bl
i
s
h c
or
r
e
s
po
nd
e
n
c
e
.
T
h
e
m
a
t
c
h
in
g
of
f
e
a
t
ur
e
s
e
nh
a
n
c
e
s
t
h
e
pr
e
c
i
s
io
n
of
i
d
e
n
ti
f
y
in
g
c
or
r
e
s
pon
di
n
g p
oi
nt
s
a
c
r
o
s
s
t
he
i
m
a
g
e
s
[
24]
.
−
I
de
nt
if
yi
ng t
he
i
n
tr
in
s
ic
a
nd e
xt
r
in
s
ic
pr
ope
r
ti
e
s
of
t
he
c
a
m
e
r
a
s
us
e
d t
o c
ol
le
c
t
im
a
ge
s
i
s
known a
s
c
a
m
e
r
a
c
a
li
br
a
ti
on.
T
he
s
e
pa
r
a
m
e
t
e
r
s
de
te
r
m
in
e
th
e
c
a
m
e
r
a
'
s
pos
it
io
n a
nd
or
ie
nt
a
ti
on
in
th
r
e
e
-
di
m
e
ns
io
na
l
s
pa
c
e
.
E
s
ta
bl
is
hi
ng
th
e
s
e
pr
ope
r
ti
e
s
e
nha
nc
e
s
th
e
a
c
c
ur
a
c
y
of
s
p
a
ti
a
l
r
e
la
ti
ons
hi
ps
a
nd
m
e
a
s
ur
e
m
e
nt
s
in
th
e
c
a
pt
ur
e
d i
m
a
ge
r
y
[
25]
.
−
A
3D
m
ode
l
of
th
e
obj
e
c
t
or
s
c
e
ne
is
r
e
c
ons
tr
uc
te
d
u
s
in
g
th
e
e
s
ti
m
a
te
d
c
or
r
e
s
ponde
nc
e
s
a
nd
c
a
m
e
r
a
pa
r
a
m
e
te
r
s
.
T
hi
s
r
e
c
ons
tr
uc
ti
on
pr
oc
e
s
s
ty
pi
c
a
ll
y
e
m
pl
oys
tr
ia
ng
ul
a
ti
on,
s
te
r
e
o
r
e
c
ons
tr
uc
ti
on,
or
s
tr
uc
tu
r
e
d
li
ght
s
c
a
nni
ng
te
c
hni
qu
e
s
.
T
he
s
e
m
e
th
od
s
c
ont
r
ib
ut
e
to
a
c
c
ur
a
te
ly
c
a
pt
ur
in
g
th
e
th
r
e
e
-
di
m
e
ns
io
n
a
l
r
e
pr
e
s
e
nt
a
ti
on of
t
he
obs
e
r
ve
d obje
c
t
or
s
c
e
ne
[
26]
.
−
O
nc
e
th
e
3D
m
ode
l
is
r
e
c
ons
tr
uc
te
d,
te
xt
ur
e
s
c
a
n
be
a
ppl
ie
d
t
o
th
e
m
ode
l'
s
s
ur
f
a
c
e
to
gi
ve
it
a
r
e
a
li
s
ti
c
a
ppe
a
r
a
nc
e
. T
hi
s
pr
oc
e
s
s
in
vol
ve
s
pr
oj
e
c
ti
ng
te
xt
ur
e
s
f
r
om
2D
im
a
ge
s
ont
o
th
e
c
or
r
e
s
ponding
s
ur
f
a
c
e
s
of
th
e
m
ode
l.
T
he
3D
m
ode
l
ga
in
s
a
li
f
e
li
ke
a
ppe
a
r
a
nc
e
by
s
e
a
m
le
s
s
ly
in
te
gr
a
ti
ng
th
e
s
e
te
xt
ur
e
s
,
e
nr
ic
hi
ng
vi
s
ua
l
f
id
e
li
ty
.
T
he
m
a
r
c
hi
ng
c
ube
s
a
lg
or
it
hm
is
w
id
e
ly
us
e
d
f
or
e
xt
r
a
c
ti
ng
a
n
is
o
-
s
ur
f
a
c
e
(
a
s
ur
f
a
c
e
r
e
pr
e
s
e
nt
in
g
a
T
he
m
a
r
c
hi
ng
c
ube
s
te
c
hni
que
,
in
tr
oduc
e
d
by
L
or
e
ns
e
n
a
nd
C
li
ne
[
27]
in
1987,
is
a
w
id
e
ly
us
e
d
m
e
th
od
f
or
e
xt
r
a
c
ti
ng
pol
ygona
l
m
e
s
h
r
e
pr
e
s
e
nt
a
ti
ons
of
is
o
-
s
ur
f
a
c
e
s
f
r
om
3D
s
c
a
la
r
f
ie
ld
s
or
voxe
l
da
ta
,
w
hi
c
h
i
s
of
te
n
us
e
d
f
or
m
e
di
c
a
l
im
a
gi
ng
da
ta
s
uc
h a
s
m
a
gn
e
ti
c
r
e
s
ona
nc
e
im
a
gi
ng
(
M
R
I
)
s
c
a
ns
[
27]
,
[
28]
.
T
he
f
unda
m
e
nt
a
l
c
onc
e
pt
is
to
di
vi
de
3D
s
pa
c
e
in
to
c
ube
s
,
w
it
h
th
e
ve
r
ti
c
e
s
of
e
a
c
h
c
ube
c
or
r
e
s
ponding
to
2D
ne
ig
hbor
in
g
s
li
c
e
s
.
T
he
a
lg
or
it
hm
m
ove
s
ove
r
th
e
s
c
a
la
r
f
ie
ld
,
c
ons
id
e
r
in
g
e
ig
ht
ne
i
ghbor
in
g
poi
nt
s
s
im
ul
ta
ne
ous
ly
,
ge
ne
r
a
ti
ng
a
n
im
a
gi
na
r
y
c
ube
,
a
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
9.
T
hi
s
a
ppr
oa
c
h
f
in
ds
th
e
pol
ygons
ne
e
de
d
to
d
e
pi
c
t
th
e
is
o
-
s
ur
f
a
c
e
pa
s
s
in
g
th
r
ough
th
is
c
ube
. T
he
in
di
vi
dua
l
pol
ygons
a
r
e
m
e
r
ge
d
to
ge
ne
r
a
te
th
e
r
e
qui
r
e
d
s
ur
f
a
c
e
.
T
h
e
m
e
r
gi
ng
m
e
th
od
is
a
id
e
d
us
in
g
a
n
in
de
x
of
a
pr
e
c
om
put
e
d
a
r
r
a
y
of
256
pot
e
nt
ia
l
pol
ygon
c
onf
ig
ur
a
ti
ons
.
T
he
num
be
r
of
di
f
f
e
r
e
nt
in
s
ta
nc
e
s
ha
s
be
e
n
r
e
du
c
e
d
to
15
due
to
s
ym
m
e
tr
y
a
nd
r
ot
a
ti
on
pr
oc
e
dur
e
s
,
a
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
10.
F
ig
ur
e
9. C
ube
de
li
m
it
e
d by 2D a
dj
a
c
e
nt
s
li
c
e
s
F
ig
ur
e
10
. 15
uni
que
c
ube
c
onf
ig
ur
a
ti
ons
ge
ne
r
a
te
d by ma
r
c
hi
ng
E
a
c
h c
ube
'
s
e
ig
ht
s
c
a
la
r
va
lu
e
s
c
or
r
e
s
pond to a
bi
t
in
a
n 8
-
bi
t
in
te
ge
r
. I
f
t
he
va
lu
e
of
a
s
c
a
la
r
e
xc
e
e
ds
th
e
is
o
-
va
lu
e
(
s
how
in
g
th
a
t
it
is
in
s
id
e
th
e
s
ur
f
a
c
e
)
,
th
e
a
s
s
o
c
ia
t
e
d
bi
t
is
s
e
t
to
one
;
ot
h
e
r
w
is
e
,
it
is
s
e
t
to
z
e
r
o.
T
he
f
in
a
l
8
-
bi
t
in
te
ge
r
,
r
e
s
ul
t
in
g
f
r
om
c
he
c
ki
ng
a
ll
e
ig
ht
s
c
a
la
r
s
,
is
us
e
d
a
s
a
n
in
de
x
to
a
c
c
e
s
s
th
e
pol
ygon
in
di
c
e
s
a
r
r
a
y.
E
a
c
h
ve
r
te
x
of
th
e
c
r
e
a
te
d
pol
ygon
s
is
pl
a
c
e
d
a
t
th
e
c
or
r
e
c
t
lo
c
a
ti
on
a
lo
ng
th
e
c
ube
'
s
e
dge
by
li
ne
a
r
ly
in
te
r
pol
a
ti
ng
be
twe
e
n
th
e
two
s
c
a
la
r
va
lu
e
s
c
onne
c
te
d
by
t
ha
t
e
dge
to
c
ons
tr
uc
t
th
e
f
in
a
l
m
e
s
h.
T
he
m
a
r
c
hi
ng_c
ube
s
_l
e
w
in
e
r
f
unc
ti
on
a
c
c
e
pt
s
a
th
r
e
e
-
di
m
e
ns
io
n
a
l
b
in
a
r
y
im
a
ge
p
a
s
in
put
a
nd
out
put
s
a
tr
ia
ngul
a
r
m
e
s
h
r
e
pr
e
s
e
nt
a
ti
on
of
th
e
bi
na
r
y
im
a
ge
'
s
s
ur
f
a
c
e
.
T
he
in
di
c
e
s
of
th
e
th
r
e
e
ve
r
ti
c
e
s
th
a
t
m
a
ke
up
a
f
a
c
e
a
r
e
s
to
r
e
d i
n e
a
c
h a
r
r
a
y r
ow
. W
e
c
a
n
s
e
e
t
h
e
r
e
s
ul
t
in
F
ig
ur
e
11
[
29]
.
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
3D
v
is
ual
iz
at
io
n di
agnos
ti
c
s
f
o
r
l
ung c
anc
e
r
de
te
c
ti
on
(
R
ana M
. M
ahm
oud
)
4637
F
ig
ur
e
11. L
ung thr
e
e
-
di
m
e
ns
io
na
l
w
it
h t
um
or
3.
R
E
S
U
L
T
A
N
D
D
I
S
C
U
S
S
I
O
N
3.1
.
E
xp
e
r
im
e
n
t
r
e
s
u
lt
an
d
an
al
ys
is
W
e
us
e
th
e
s
iz
e
m
e
a
s
ur
e
m
e
nt
of
nodule
s
,
a
n
e
s
s
e
nt
ia
l
f
a
c
to
r
f
or
r
a
di
ol
ogi
s
ts
to
c
la
s
s
if
y
be
ni
gn
a
nd
m
a
li
gna
nt
nodule
s
.
A
ddi
ti
ona
ll
y,
w
e
u
s
e
2
D
te
xt
ur
e
f
e
a
tu
r
e
s
,
s
ta
ti
s
ti
c
a
l
a
nd
g
e
om
e
tr
ic
,
w
hi
c
h
gi
ve
u
s
im
pr
ove
d
r
e
s
ul
ts
.
T
o
a
na
ly
z
e
th
e
pe
r
f
or
m
a
nc
e
of
th
e
pr
opos
e
d
a
lg
or
it
hm
qua
li
ta
ti
ve
ly
a
nd
qua
nt
it
a
ti
ve
ly
,
w
e
s
e
le
c
t
a
M
obi
le
N
e
t
f
r
a
m
e
w
or
k
to
i
m
pl
e
m
e
nt
th
e
c
or
r
e
s
ponding
ne
twor
k
m
ode
l.
T
hi
s
m
ode
l
is
be
ne
f
ic
ia
l
f
or
ta
s
ks
in
w
hi
c
h
c
om
pl
ic
a
te
d
e
va
lu
a
ti
ve
s
pe
e
d
is
c
r
it
ic
a
l.
T
he
e
xpe
r
im
e
nt
a
l
ha
r
dw
a
r
e
pl
a
tf
or
m
is
th
e
N
vi
di
a
T
e
s
la
K
80
G
P
U
(
12
G
B
)
a
nd N
vi
di
a
T
e
s
la
K
80(
12G
B
)
. W
e
now
pr
e
s
e
nt
t
hi
s
s
e
c
ti
on i
n de
ta
il
.
3.1.1. Dat
a an
d
an
al
ys
is
T
he
L
U
N
A
16
da
ta
s
e
t
c
ha
ll
e
nge
is
pa
r
t
of
I
S
B
I
2016.
I
t
is
a
publ
i
c
ly
a
va
il
a
bl
e
m
e
di
c
a
l
im
a
gi
ng
da
ta
s
e
t
th
a
t
c
ont
a
in
s
C
T
s
c
a
ns
of
th
e
c
he
s
t
f
r
om
1,018
pa
ti
e
nt
s
.
I
t
e
xc
lu
de
d
s
c
a
n
s
w
it
h
s
li
c
e
th
ic
kn
e
s
s
gr
e
a
te
r
th
a
n
2.50
m
m
(
a
bout
0.1 in)
f
r
om
lu
ng i
m
a
ge
da
ta
ba
s
e
c
on
s
or
ti
um
a
nd i
m
a
ge
da
ta
ba
s
e
r
e
s
our
c
e
i
ni
ti
a
ti
ve
(
L
I
D
C
-
I
D
R
I
)
.
T
he
da
ta
s
e
t
pr
ovi
de
s
a
va
lu
a
bl
e
r
e
s
our
c
e
f
or
r
e
s
e
a
r
c
he
r
s
w
or
ki
ng
on
lu
ng
nodule
de
te
c
ti
on
a
nd
c
l
a
s
s
if
ic
a
ti
on,
w
hi
c
h
a
r
e
e
a
r
ly
s
ig
n
s
of
lu
ng
c
a
nc
e
r
.
T
he
im
a
g
e
s
a
r
e
in
R
A
W
f
or
m
a
t
a
nd
la
be
le
d
w
it
h
nodule
a
nnot
a
ti
ons
f
r
om
m
ul
ti
pl
e
r
a
di
ol
ogi
s
ts
.
T
he
da
ta
s
e
t
a
ls
o
in
c
lu
de
s
a
C
S
V
f
il
e
w
it
h
a
ddi
ti
ona
l
pa
ti
e
nt
a
nd
nodule
m
e
ta
da
ta
.
T
he
L
U
N
A
16
da
ta
s
e
t
is
w
id
e
ly
u
s
e
d
in
m
a
c
hi
ne
le
a
r
ni
ng
r
e
s
e
a
r
c
h
to
de
ve
lo
p
nodule
m
e
th
ods
f
or
de
te
c
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
0
[
30]
.
3.1.2. E
xp
e
r
im
e
n
t
al
d
at
a s
e
t
s
an
d
e
val
u
at
io
n
c
r
it
e
r
ia
T
hi
s
pa
pe
r
ut
il
iz
e
s
th
e
L
U
N
A
16
da
ta
s
e
t
f
or
bot
h
te
s
ti
ng
a
nd
tr
a
in
in
g
of
th
e
ne
twor
k.
T
he
da
ta
s
e
t
c
om
pr
is
e
s
1,018
pa
ti
e
nt
c
a
s
e
im
a
ge
s
,
di
vi
de
d
in
to
two
s
e
c
ti
ons
.
W
e
pa
r
ti
ti
one
d
th
e
L
U
N
A
16
da
ta
s
e
ts
r
a
ndoml
y
a
nd e
qua
ll
y i
nt
o
t
e
n
s
ubs
e
t
s
. N
in
e
of
t
he
s
e
s
ubs
e
ts
, t
ot
a
li
ng
800
C
T
s
c
a
n
s
, w
e
r
e
ut
il
iz
e
d
f
or
t
r
a
in
in
g, w
hi
le
one
s
ubs
e
t
c
ont
a
in
in
g
88
C
T
s
c
a
ns
w
a
s
r
e
s
e
r
ve
d
f
or
te
s
ti
ng.
T
e
nf
o
ld
c
r
os
s
-
va
li
da
ti
on
w
a
s
pe
r
f
or
m
e
d
on
th
e
te
s
t
m
ode
l.
A
m
ong
th
e
888
C
T
im
a
g
e
s
,
a
to
ta
l
of
36,378
nodul
e
s
w
e
r
e
id
e
nt
if
ie
d,
w
it
h
1186
nodule
s
s
e
le
c
te
d
f
or
a
na
ly
s
is
.
P
r
e
c
is
io
n a
nd
s
e
ns
it
iv
it
y
m
e
tr
ic
s
w
e
r
e
e
m
pl
oye
d
in
th
is
s
tu
dy
to
a
s
s
e
s
s
th
e
lo
c
a
li
z
a
ti
on
pe
r
f
or
m
a
nc
e
of
th
e
pr
opos
e
d
s
ys
te
m
.
W
e
ut
il
iz
e
d
a
c
c
ur
a
c
y
to
a
s
s
e
s
s
how
m
a
ny
e
xpe
c
te
d
po
s
it
iv
e
s
w
e
r
e
s
uc
c
e
s
s
f
ul
ly
r
e
c
ogni
z
e
d,
s
e
ns
it
iv
it
y
to
e
s
ta
bl
is
h
how
m
a
ny
e
xpe
c
te
d
pos
it
iv
e
s
w
e
r
e
r
e
c
ogni
z
e
d, a
nd
s
e
ns
it
iv
it
y
to
e
s
ta
bl
is
h
how
m
a
ny
ge
nui
ne
po
s
it
iv
e
s
w
e
r
e
a
c
c
ur
a
t
e
ly
de
te
c
t
e
d.
T
h
e
F
1
s
c
or
e
w
a
s
a
ls
o
us
e
d
to
c
om
put
e
th
e
ha
r
m
oni
c
m
e
a
n
of
s
e
ns
it
iv
it
y
a
nd
a
c
c
ur
a
c
y,
a
s
w
e
ll
a
s
th
e
ha
r
m
oni
c
m
e
a
n
of
s
e
ns
it
iv
it
y
a
nd
a
c
c
ur
a
c
y.
B
e
c
a
us
e
th
e
F
1
s
c
or
e
m
e
a
s
ur
e
s
a
c
c
ur
a
c
y
a
nd
s
e
ns
it
iv
it
y,
it
a
lwa
y
s
a
s
s
ig
ns
e
qua
l
w
e
ig
ht
to
bot
h
m
e
a
s
ur
e
s
.
T
hi
s
is
be
c
a
u
s
e
th
e
r
e
is
a
lwa
ys
a
tr
a
de
-
of
f
be
twe
e
n
th
e
two
s
in
c
e
th
e
m
e
a
n
va
lu
e
is
a
lwa
ys
s
tr
ongl
y
in
f
lu
e
nc
e
d
a
t
th
e
e
xpe
ns
e
of
e
it
he
r
va
lu
e
.
3.
1.
3. T
r
ai
n
in
g, t
e
s
t
in
g, an
d
i
m
p
le
m
e
n
t
at
io
n
d
e
t
ai
ls
T
he
im
a
ge
s
iz
e
of
th
e
ne
twor
k
in
put
is
224×
224
×
3.
T
he
M
obi
le
N
e
t
a
r
c
hi
te
c
tu
r
e
us
e
s
de
pt
hw
i
s
e
s
e
pa
r
a
bl
e
c
onvolut
io
ns
,
c
om
bi
ni
ng
a
poi
nt
w
is
e
(
w
hi
c
h
a
ppl
ie
s
a
1x1
f
il
te
r
to
th
e
out
put
of
th
e
de
pt
hw
is
e
c
onvolut
io
n)
a
nd a
de
pt
hw
is
e
c
onvolut
io
n (
w
hi
c
h a
ppl
ie
s
a
s
in
g
le
f
il
te
r
t
o e
a
c
h i
nput
c
ha
nne
l)
. T
he
de
pt
hw
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
n
r
e
duc
e
s
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
a
nd
th
e
a
m
ount
of
c
om
put
a
ti
on
ne
e
de
d
c
om
pa
r
e
d
to
a
s
ta
nda
r
d
c
onvolut
io
na
l
la
ye
r
.
T
hi
s
a
ls
o
in
c
lu
de
s
a
c
us
to
m
i
m
pl
e
m
e
nt
a
ti
on
of
th
e
de
pt
hw
is
e
c
onvolut
i
on,
w
hi
c
h
a
ppl
ie
s
a
s
e
pa
r
a
te
2D
c
onvolut
io
n
to
e
a
c
h
in
put
c
h
a
nne
l
us
in
g
a
s
ha
r
e
d
ke
r
ne
l.
T
hi
s
im
pl
e
m
e
nt
a
ti
on
a
ll
ow
s
f
or
m
or
e
e
f
f
ic
ie
nt
c
om
put
a
ti
on
th
a
n
a
s
ta
nda
r
d
d
e
pt
hw
i
s
e
c
onvolut
io
n,
w
hi
c
h
a
ppl
ie
s
a
s
in
gl
e
f
il
te
r
to
e
a
c
h
in
put
c
ha
nne
l.
T
he
M
obi
le
N
e
t
a
r
c
hi
te
c
tu
r
e
c
ons
i
s
ts
of
a
s
e
r
i
e
s
of
de
pt
hw
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
na
l
la
ye
r
s
a
nd
a
n
in
it
ia
l
c
onvolut
io
na
l
la
ye
r
.
T
he
out
put
o
f
th
e
la
s
t
c
onvolut
io
na
l
la
ye
r
is
f
e
d
in
to
a
gl
oba
l
a
ve
r
a
ge
pool
in
g
la
ye
r
, w
hi
c
h c
om
put
e
s
t
he
a
ve
r
a
ge
va
lu
e
of
e
a
c
h
f
e
a
tu
r
e
m
a
p,
a
nd
a
f
ul
ly
c
onne
c
te
d output
l
a
y
e
r
. T
h
e
n t
he
out
s
how
w
it
h
voxe
l
s
iz
e
is
32×
32×
32.
T
he
m
ode
l
us
e
s
th
e
lo
s
s
f
unc
ti
on
a
nd
th
e
s
to
c
ha
s
ti
c
gr
a
di
e
nt
de
s
c
e
nt
(
S
G
D
)
.
I
t
us
e
s
t
he
S
G
D
opt
im
iz
e
r
, w
it
h t
he
l
e
a
r
ni
ng r
a
te
s
e
t
to
0. de
c
a
y
, m
om
e
nt
um
t
o 0.9, a
nd w
e
ig
ht
de
c
a
y.
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
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4630
-
4641
4638
3.
1.
4. P
e
r
f
or
m
an
c
e
m
e
t
r
ic
s
P
e
r
f
or
m
a
nc
e
m
e
tr
ic
s
a
r
e
e
s
s
e
nt
ia
l
to
ol
s
th
a
t
h
e
lp
or
ga
ni
z
a
ti
ons
m
e
a
s
ur
e
a
nd
e
va
lu
a
te
how
w
e
ll
th
e
y
a
r
e
a
c
hi
e
vi
ng
th
e
ir
goa
ls
.
T
he
y
pr
ovi
de
a
qua
nt
i
ta
ti
ve
ba
s
is
f
or
a
s
s
e
s
s
in
g
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
va
r
io
us
pr
oc
e
s
s
e
s
,
s
ys
te
m
s
,
or
te
a
m
s
.
I
n
th
is
di
s
c
us
s
io
n,
w
e
w
il
l
f
oc
us
on
s
pe
c
if
ic
pe
r
f
or
m
a
nc
e
m
e
tr
ic
s
,
s
uc
h
a
s
s
e
ns
it
iv
it
y,
s
pe
c
if
ic
it
y, a
c
c
ur
a
c
y,
a
nd me
a
n s
qu
a
r
e
d e
r
r
or
(
M
S
E
)
.
‒
S
e
ns
it
iv
it
y
S
e
ns
it
iv
it
y
m
e
a
s
ur
e
s
th
e
a
bi
li
ty
to
de
te
c
t
th
e
tr
ue
pos
it
iv
e
s
f
r
om
th
e
te
s
t.
I
t
is
c
a
lc
ul
a
te
d
by
c
ons
id
e
r
in
g
th
e
de
te
c
te
d
tr
ue
pos
it
iv
e
s
in
r
e
la
ti
on
to
th
e
to
ta
l
num
be
r
of
a
c
tu
a
l
pos
it
iv
e
s
.
I
n
(
3)
pr
ovi
de
s
th
e
s
pe
c
if
ic
m
a
th
e
m
a
ti
c
a
l
r
e
pr
e
s
e
nt
a
ti
on f
or
s
e
ns
it
iv
it
y i
n t
hi
s
c
ont
e
xt
[
31]
.
=
+
(
3)
‒
S
pe
c
if
ic
it
y
S
pe
c
if
ic
it
y
m
e
a
s
ur
e
s
th
e
a
bi
li
ty
to
de
te
r
m
in
e
be
ni
gn
c
a
s
e
s
c
or
r
e
c
tl
y.
I
t
is
qua
nt
if
ie
d
by
(
4)
,
w
hi
c
h
de
f
in
e
s
t
he
m
a
th
e
m
a
ti
c
a
l
r
e
pr
e
s
e
nt
a
ti
on f
or
s
pe
c
if
ic
it
y. T
hi
s
e
qua
ti
on c
a
pt
ur
e
s
t
he
a
bi
li
ty
of
a
t
e
s
t
to
c
or
r
e
c
tl
y
c
la
s
s
if
y t
r
ue
ne
ga
ti
ve
s
a
m
ong the
t
ot
a
l
num
be
r
of
a
c
tu
a
l
ne
ga
ti
v
e
s
[
31]
.
S
p
e
c
if
ic
it
y
=
TN
TN
+
FP
(
4)
‒
A
c
c
ur
a
c
y
A
c
c
ur
a
c
y
m
e
a
s
ur
e
s
th
e
a
bi
li
ty
to
di
f
f
e
r
e
nt
ia
te
th
e
m
a
li
gna
nt
a
nd
be
ni
gn
c
a
s
e
s
c
or
r
e
c
tl
y.
I
n
(
5)
pr
ovi
de
s
a
m
a
th
e
m
a
ti
c
a
l
r
e
pr
e
s
e
nt
a
ti
on
f
or
a
c
c
ur
a
c
y,
c
a
pt
ur
in
g
th
e
r
a
ti
o
of
c
or
r
e
c
tl
y
id
e
nt
if
ie
d
c
a
s
e
s
to
th
e
to
ta
l
num
be
r
of
c
a
s
e
s
.
T
hi
s
e
qua
ti
on
s
e
r
ve
s
a
s
a
qua
nt
it
a
ti
ve
m
e
a
s
ur
e
of
th
e
ove
r
a
ll
pe
r
f
or
m
a
nc
e
of
th
e
c
la
s
s
if
ic
a
ti
on
or
di
a
gnos
ti
c
pr
oc
e
s
s
[
31]
.
=
+
+
+
+
(
5)
‒
M
e
a
n
s
qu
a
r
e
d e
r
r
or
M
S
E
is
a
s
t
a
ti
s
ti
c
a
l
m
e
a
s
ur
e
u
s
e
d
to
qu
a
nt
if
y
th
e
a
ve
r
a
ge
s
qua
r
e
d
de
vi
a
ti
on
be
twe
e
n
pr
e
di
c
te
d
a
nd
obs
e
r
ve
d
va
lu
e
s
in
r
e
gr
e
s
s
io
n
t
a
s
ks
.
I
t
qua
nt
if
ie
s
th
e
a
ve
r
a
ge
s
qua
r
e
d
e
r
r
or
or
va
r
ia
nc
e
of
th
e
m
ode
l'
s
pr
e
di
c
ti
ons
.
I
n
(
6)
r
e
pr
e
s
e
nt
s
M
S
E
[
32]
.
=
∑
(
−
)
2
(
6)
I
n
th
is
c
ont
e
xt
,
T
P
r
e
pr
e
s
e
nt
s
th
e
c
ount
of
tr
ue
pos
it
iv
e
s
,
F
N
r
e
pr
e
s
e
nt
s
th
e
c
ount
of
f
a
ls
e
ne
ga
ti
ve
s
,
a
nd F
P
r
e
pr
e
s
e
nt
s
t
he
c
ount
of
f
a
ls
e
pos
it
iv
e
s
.
R
e
c
a
ll
,
pr
e
c
is
io
n,
a
c
c
ur
a
c
y
, F
-
s
c
or
e
,
a
nd M
S
E
got
t
he
va
lu
e
s
of
0.5,
1.0,
0.93,
0.67,
a
nd
0.067,
r
e
s
pe
c
ti
ve
ly
.
T
he
r
e
a
r
e
oppor
tu
ni
t
ie
s
f
or
f
ur
th
e
r
im
pr
ove
m
e
nt
of
th
e
s
e
e
va
lu
a
ti
on
c
r
it
e
r
ia
,
w
hi
c
h
w
e
a
r
e
c
ur
r
e
nt
ly
w
or
ki
ng
on.
W
e
c
a
n
s
e
e
th
is
in
th
e
c
onf
us
io
n
m
a
tr
ix
.
T
he
pr
opos
e
d
m
e
th
odol
ogy pe
r
f
or
m
s
be
tt
e
r
t
ha
n t
he
pr
e
vi
ous
a
ppr
oa
c
he
s
, a
s
s
how
n i
n F
ig
ur
e
12
.
F
ig
ur
e
12.
C
onf
us
io
n m
a
tr
ix
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
3D
v
is
ual
iz
at
io
n di
agnos
ti
c
s
f
o
r
l
ung c
anc
e
r
de
te
c
ti
on
(
R
ana M
. M
ahm
oud
)
4639
3.
1.
5. T
h
r
e
e
-
d
im
e
n
s
io
n
al
r
e
c
on
s
t
r
u
c
t
io
n
r
e
s
u
lt
s
E
xa
m
in
e
th
e
m
e
di
c
a
l
r
e
c
or
ds
of
a
s
in
gl
e
p
a
ti
e
nt
.
T
h
e
or
ig
in
a
l
C
T
s
c
a
n
s
of
hi
m
ha
d
a
s
iz
e
of
224×
224×
133, 133
w
it
h t
he
s
i
z
e
224×
224
photos
w
e
r
e
a
c
qui
r
e
d
by pr
e
-
pr
oc
e
s
s
in
g
a
s
i
ndi
c
a
te
d i
n s
e
c
ti
on 3. 1,
th
e
M
obi
le
N
e
t
r
e
tr
ie
ve
d s
om
e
i
m
a
ge
s
w
it
h nodule
s
, a
nd no
dul
e
nor
m
a
li
z
a
ti
on w
a
s
a
c
hi
e
ve
d. N
or
m
a
li
z
in
g t
he
or
ig
in
a
l
im
a
ge
yi
e
ld
s
th
e
nodule
s
e
que
nc
e
.
W
e
ke
pt
th
e
nodule
s
e
que
nc
e
s
iz
e
of
224×
224×
133
to
e
ns
ur
e
th
a
t
nodule
s
a
nd
lu
ngs
a
r
e
in
th
e
s
a
m
e
c
oor
di
na
te
.
L
a
s
tl
y,
w
e
obt
a
in
e
d
th
r
e
e
-
di
m
e
ns
io
na
l
m
ode
ls
of
pul
m
ona
r
y
nodule
s
(
a
s
s
how
n
in
F
ig
ur
e
12)
a
nd
lu
ngs
us
in
g
r
a
y
-
c
a
s
ti
ng
vol
um
e
r
e
nde
r
in
g.
F
ig
ur
e
12
de
pi
c
t
s
th
r
e
e
-
di
m
e
ns
io
na
l
r
e
pr
e
s
e
nt
a
ti
ons
of
l
ung
a
nd pulm
ona
r
y nodule
s
. C
u
r
ve
s
of
c
ol
or
a
nd
opa
c
it
y va
lu
e
s
i
n
r
a
y
-
c
a
s
ti
ng
vol
um
e
r
e
nde
r
in
g
a
r
e
s
how
n
in
th
e
im
a
ge
'
s
lo
w
e
r
r
ig
ht
c
or
ne
r
.
I
n
te
r
m
s
of
ove
r
a
ll
s
ys
te
m
pe
r
f
or
m
a
nc
e
,
th
e
m
a
jo
r
it
y
of
th
e
ti
m
e
is
s
pe
nt
obt
a
in
in
g
nodule
m
a
s
ks
th
r
ough
th
e
M
obi
le
N
e
t,
w
it
h
a
s
pe
e
d
of
1.29
s
e
c
ond
s
r
e
qui
r
e
d
pe
r
s
c
a
n
a
ba
s
ic
P
C
e
qui
ppe
d
w
it
h
a
n
N
vi
di
a
T
e
s
la
K
80
G
P
U
(
12
G
B
)
a
nd
R
A
W
(
13
G
B
)
,
w
hi
c
h
ta
ke
s
142 s
e
c
onds
i
n t
ot
a
l.
T
he
t
hr
e
e
-
di
m
e
ns
io
na
l
r
e
c
ons
tr
uc
ti
on s
te
p
oc
c
upi
e
s
t
he
m
a
jo
r
it
y of
m
e
m
or
y, w
i
th
a
t
ot
a
l
c
a
pa
c
it
y
of
6.081
G
B
f
or
th
r
e
e
-
di
m
e
ns
io
na
l
m
ode
ls
of
th
e
lu
ng
a
nd
th
r
e
e
nodule
s
.
O
ve
r
a
ll
,
a
s
lo
ng
a
s
th
e
nodule
m
a
s
ks
a
r
e
pr
e
c
is
e
e
nough,
w
e
c
a
n
obt
a
in
th
r
e
e
-
di
m
e
ns
io
na
l
m
ode
ls
of
th
e
pr
oj
e
c
t
e
d
nodule
s
a
nd
lu
ng
s
u
s
in
g
our
a
ppr
oa
c
h.
A
ls
o,
w
e
c
a
n
obs
e
r
ve
m
or
e
a
bout
lu
ng
a
nd
nodule
s
ti
s
s
ue
s
by
a
lt
e
r
in
g
th
e
c
ol
or
a
nd
opa
c
it
y
o
f
th
e
r
a
y
-
c
a
s
ti
ng, whic
h i
s
i
nc
r
e
di
bl
y us
e
f
ul
f
or
di
a
gnos
is
a
nd
s
ub
s
e
que
nt
t
he
r
a
py.
3.2.
D
is
c
u
s
s
io
n
T
he
pr
opos
e
d
s
y
s
te
m
i
s
us
e
d
f
or
pul
m
ona
r
y
nodule
s
e
gm
e
nt
a
ti
on,
de
te
c
ti
on,
a
nd
th
r
e
e
-
di
m
e
ns
io
na
l
vi
s
ua
li
z
a
ti
on
s
ys
t
e
m
.
T
he
s
ys
te
m
us
in
g
th
e
L
una
da
t
a
s
e
t
is
a
pub
li
c
a
nd
a
va
il
a
bl
e
da
ta
s
e
t.
I
t
is
us
e
d
f
or
de
te
c
ti
ng
tu
m
or
s
by
th
e
M
obi
le
N
e
t
m
ode
l
a
nd
th
e
n
vi
s
ua
li
z
e
d
in
th
r
e
e
-
di
m
e
ns
io
na
l
ba
s
e
d
on
th
e
vol
um
e
r
e
nde
r
in
g
a
lg
or
it
hm
of
r
a
y
-
c
a
s
ti
ng.
T
hi
s
s
ys
te
m
c
a
n
r
e
nde
r
th
e
th
r
e
e
-
di
m
e
ns
io
na
l
m
ode
ls
of
th
e
de
te
c
te
d
lu
ngs
a
nd
nodule
s
,
a
s
s
how
n
in
F
ig
ur
e
13.
N
ot
e
th
a
t
th
r
e
e
-
di
m
e
ns
io
n
a
l
r
e
c
ons
tr
uc
ti
on
m
odul
e
s
a
nd
th
e
de
te
c
ti
on
a
r
e
de
s
ig
ne
d
to
be
s
e
pa
r
a
te
d.
T
he
r
e
f
or
e
,
w
e
c
a
n
e
nha
nc
e
th
e
s
e
gm
e
nt
a
ti
on
ne
twor
k
to
ge
t
de
s
ig
ne
d
to
be
s
e
pa
r
a
te
d
a
nd
a
de
di
c
a
te
d
c
ont
our
to
de
te
c
t
ot
he
r
ki
dne
y,
c
a
nc
e
r
,
or
le
s
io
ns
tu
m
or
s
.
A
ls
o,
th
e
vol
um
e
r
e
nde
r
in
g
a
lg
or
it
h
m
c
a
n
be
r
e
pl
a
c
e
d
a
nd
opt
im
iz
e
d
to
m
e
e
t
pr
a
c
ti
c
a
l
ne
e
ds
.
I
n
th
e
de
te
c
ti
on
a
nd
s
e
gm
e
nt
a
ti
on
m
odul
e
,
w
e
us
e
d
M
obi
le
N
e
t.
A
s
s
how
n
in
T
a
bl
e
1,
M
obi
le
N
e
t
out
pe
r
f
or
m
s
pr
e
vi
ous
s
y
s
te
m
s
e
va
lu
a
te
d
und
e
r
s
im
il
a
r
e
nvi
r
onm
e
nt
s
,
w
hi
c
h
in
di
c
a
t
e
s
th
a
t
it
ha
s
pr
om
is
e
f
or
two
-
s
ta
g
e
obj
e
c
t
r
e
c
ogni
ti
o
n
in
m
e
di
c
a
l
im
a
ge
a
na
ly
s
is
.
T
he
M
S
E
f
or
t
hi
s
m
ode
l
is
e
qua
l
to
0.067.
F
ig
ur
e
13. I
nt
e
r
m
e
di
a
te
a
nd r
e
s
ul
ts
of
3D
r
e
c
ons
tr
uc
ti
on
T
a
bl
e
1
.
C
om
pa
r
is
on of
a
c
c
ur
a
c
y a
lg
or
it
hm
s
R
e
f
e
r
e
nc
e
M
ode
l
D
a
t
a
s
e
t
T
r
a
i
ni
ng s
e
t
T
e
s
t
i
ng s
e
t
A
c
c
ur
a
c
y
(%)
[
11]
C
D
P
-
R
e
s
N
e
t
L
I
D
C
447
539
82
[
12]
CA
-
N
e
t
D
S
B
2017
1595
506
83.79
[
15]
A
L
I
A
S
C
T
i
m
a
ge
s
23979
7989
83.8
[
13]
S
A
N
e
t
P
N
9
6707
2091
87.4
[
14]
M
a
s
k R
-
C
N
N
L
U
N
A
800 c
a
s
e
s
88 c
a
s
e
s
88.1
[
16]
3D
-
M
S
V
i
T
L
U
N
A
800 c
a
s
e
s
88 c
a
s
e
s
91.1
[
17]
3D
S
K
-
R
e
s
N
e
t
L
U
N
A
800
c
a
s
e
s
88 c
a
s
e
s
91.75
P
r
opos
e
d a
l
g.
M
obi
l
e
N
e
t
L
U
N
A
800 c
a
s
e
s
88 c
a
s
e
s
93.3
O
ur
ne
twor
k
ut
il
iz
e
d
onl
y
a
s
in
gl
e
two
-
di
m
e
ns
io
na
l
vi
e
w
im
a
g
e
,
w
he
r
e
a
s
S
e
ti
o
e
t
al
.
[
33]
e
m
pl
oye
d
m
ul
ti
-
vi
e
w
2D
i
m
a
ge
s
,
a
nd
F
a
r
ha
t
e
t
a
l.
[
34
]
in
c
or
po
r
a
te
d
th
r
e
e
-
di
m
e
ns
io
na
l
s
pa
ti
a
l
da
ta
.
T
hi
s
s
ugge
s
t
s
th
a
t
in
c
or
por
a
ti
ng
m
or
e
s
pa
ti
a
l
in
f
or
m
a
ti
on
m
a
y
e
nha
nc
e
pe
r
f
or
m
a
n
c
e
.
R
e
ga
r
di
ng
th
r
e
e
-
di
m
e
ns
io
na
l
vi
s
ua
li
z
a
ti
on,
th
e
r
e
'
s
r
oom
f
or
opt
im
iz
in
g
r
e
nde
r
in
g
e
f
f
e
c
ts
.
F
or
in
s
ta
nc
e
,
di
s
ti
nc
t
vol
um
e
r
e
nde
r
i
ng
te
c
hni
que
s
or
pa
r
a
m
e
te
r
s
c
oul
d
be
ta
il
or
e
d
f
or
lu
ngs
a
nd
pul
m
ona
r
y
nodule
s
.
A
ddi
ti
ona
l
ly
,
m
e
m
or
y
opt
im
iz
a
ti
on
pos
e
s
a
c
ha
ll
e
nge
in
pr
a
c
ti
c
a
l
im
pl
e
m
e
nt
a
ti
ons
.
A
s
de
pi
c
te
d
in
F
ig
ur
e
12,
w
e
c
a
n
di
s
c
e
r
n
th
e
r
e
la
ti
ve
pos
it
io
ns
a
nd
s
iz
e
s
of
pul
m
ona
r
y
nodule
s
w
it
hi
n
th
e
lu
ng.
T
he
de
ve
lo
pm
e
nt
of
a
s
e
gm
e
nt
a
ti
on,
de
te
c
ti
on,
a
nd
th
r
e
e
-
di
m
e
ns
io
na
l
Evaluation Warning : The document was created with Spire.PDF for Python.