I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
4
,
Augus
t
202
5
,
pp.
3214
~
3227
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
4
.
pp
32
14
-
3227
3214
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
C
om
p
a
r
at
iv
e
e
val
u
at
io
n
of
l
e
f
t
ve
n
t
r
ic
l
e
se
gm
e
n
t
at
io
n
u
si
n
g
imp
r
ov
e
d
p
yr
am
i
d
sc
e
n
e
p
a
r
si
n
g n
e
t
w
or
k
i
n
e
c
h
oc
a
r
d
io
gr
a
p
h
y
Jin
Wan
g
1
,
2
,
S
h
ar
if
ah
Alim
an
1
,
S
h
a
f
af
I
b
r
ah
im
1
1
F
a
c
ul
ty
of
C
omput
e
r
a
nd M
a
th
e
ma
ti
c
a
l
S
c
ie
nc
e
s
, U
ni
ve
r
s
it
i
T
e
knol
ogi
M
A
R
A
, S
ha
h A
la
m,
M
a
la
ys
i
a
2
De
pa
r
tm
e
nt
of
E
le
c
tr
ic
a
l
E
ngi
ne
e
r
in
g, T
a
iy
ua
n I
ns
ti
tu
te
of
T
e
c
hnol
ogy, T
a
iy
ua
n, C
hi
na
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
J
ul
25,
2024
R
e
vis
e
d
Apr
8,
2025
Ac
c
e
pted
J
un
8,
2025
A
u
t
o
ma
t
i
c
s
e
g
men
t
at
i
o
n
o
f
t
h
e
l
ef
t
v
en
t
ri
c
l
e
i
s
a
ch
al
l
e
n
g
i
n
g
t
as
k
d
u
e
t
o
t
h
e
p
res
e
n
ce
o
f
ar
t
i
fac
t
s
an
d
s
p
eck
l
e
n
o
i
s
e
i
n
ec
h
o
car
d
i
o
g
ra
p
h
y
.
T
h
i
s
p
a
p
er
s
t
u
d
i
es
t
h
e
ab
i
l
i
t
y
o
f
a
fu
l
l
y
s
u
p
er
v
i
s
ed
n
e
t
w
o
rk
b
a
s
e
d
o
n
p
y
rami
d
s
cen
e
p
ars
i
n
g
n
e
t
w
o
rk
(PSPN
et
)
t
o
i
m
p
l
eme
n
t
ech
o
card
i
o
g
ra
p
h
i
c
l
ef
t
v
en
t
ri
c
u
l
ar
s
eg
me
n
t
a
t
i
o
n
.
Fi
rs
t
,
t
h
e
l
i
g
h
t
w
e
i
g
h
t
Mo
b
i
l
eN
e
t
v
2
w
as
s
el
ect
e
d
t
o
rep
l
ace
Res
N
e
t
t
o
ad
j
u
s
t
t
h
e
co
d
i
n
g
s
t
r
u
ct
u
re
o
f
t
h
e
n
e
u
ral
n
et
w
o
r
k
,
red
u
c
e
t
h
e
co
mp
u
t
a
t
i
o
n
a
l
co
m
p
l
e
x
i
t
y
,
a
n
d
i
n
t
eg
ra
t
e
t
h
e
p
y
ram
i
d
s
c
en
e
an
a
l
y
s
i
s
mo
d
u
l
e
t
o
co
n
s
t
r
u
ct
t
h
e
PSPN
et
;
s
eco
n
d
l
y
,
i
n
t
ro
d
u
ce
d
i
l
at
e
d
co
n
v
o
l
u
t
i
o
n
an
d
feat
u
r
e
fu
s
i
o
n
t
o
p
ro
p
o
s
e
an
i
m
p
ro
v
ed
PSPN
et
mo
d
el
,
an
d
s
t
u
d
y
t
h
e
i
mp
ac
t
o
f
p
re
-
t
rai
n
i
n
g
an
d
t
ra
n
s
fer
l
ear
n
i
n
g
o
n
mo
d
el
s
e
g
men
t
at
i
o
n
p
erf
o
rma
n
ce;
fi
n
al
l
y
,
t
h
e
p
u
b
l
i
c
d
a
t
a
s
e
t
ch
al
l
en
g
e
o
n
en
d
o
card
i
al
t
h
ree
-
d
i
men
s
i
o
n
al
u
l
t
ra
s
o
u
n
d
s
eg
me
n
t
a
t
i
o
n
(CE
T
U
S)
w
as
u
s
ed
t
o
t
ra
i
n
an
d
t
es
t
d
i
f
feren
t
b
ack
b
o
n
e
an
d
i
n
i
t
i
al
i
zed
PSPN
e
t
mo
d
el
s
.
T
h
e
re
s
u
l
t
s
d
emo
n
s
t
rat
e
t
h
at
t
h
e
i
m
p
r
o
v
e
d
PSPN
et
mo
d
e
l
h
as
s
t
r
o
n
g
s
eg
me
n
t
a
t
i
o
n
ad
v
an
t
ag
e
s
i
n
t
e
rms
o
f
accu
racy
an
d
ru
n
n
i
n
g
s
p
eed
.
Co
m
p
are
d
w
i
t
h
t
h
e
t
w
o
c
l
as
s
i
c
a
l
g
o
ri
t
h
ms
V
G
G
an
d
U
n
e
t
,
t
h
e
d
i
ce
s
i
mi
l
ari
t
y
co
eff
i
ci
e
n
t
(D
SC
)
i
n
d
ex
i
s
i
n
creas
ed
b
y
a
n
av
erag
e
o
f
7
.
6
%
,
H
au
s
d
o
rff
d
i
s
t
an
ce
(
H
D
)
i
s
r
ed
u
ced
b
y
2
.
9
%
,
an
d
t
h
e
m
ea
n
i
n
t
ers
ec
t
i
o
n
o
v
er
u
n
i
o
n
(
mIo
U
)
i
s
i
m
p
ro
v
ed
b
y
8
.
8
%
.
A
d
d
i
t
i
o
n
a
l
l
y
,
t
h
e
ru
n
n
i
n
g
t
i
me
i
s
g
reat
l
y
s
h
o
rt
e
n
ed
,
i
n
d
i
cat
i
n
g
g
o
o
d
cl
i
n
i
cal
a
p
p
l
i
ca
t
i
o
n
p
o
t
en
t
i
a
l
.
K
e
y
w
o
r
d
s
:
De
e
p
lea
r
ning
E
c
ho
c
a
r
d
io
g
r
a
ph
ic
L
e
f
t
v
e
n
tr
ic
le
s
e
g
me
nt
a
t
io
n
M
obil
e
Ne
tv2
P
S
P
Ne
t
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
:
S
ha
r
if
a
h
Alim
a
n
F
a
c
ult
y
of
C
omput
e
r
a
nd
M
a
thema
ti
c
a
l
S
c
ienc
e
s
,
Unive
r
s
it
i
T
e
knologi
M
AR
A
S
ha
h
Ala
m,
S
e
langor
,
M
a
lays
ia
E
mail:
s
ha
r
if
a
ha
li
@uit
m
.
e
du.
my
1.
I
NT
RODU
C
T
I
ON
Due
to
the
por
tabi
li
ty
a
nd
low
c
os
t
of
ult
r
a
s
ound
i
mage
s
,
a
c
c
ur
a
te
s
e
gmenta
ti
on
of
the
lef
t
ve
ntr
icle
of
e
c
hoc
a
r
diogr
a
phy
c
a
n
he
lp
doc
tor
s
quickly
a
n
d
e
f
f
e
c
ti
ve
ly
a
na
lyze
c
a
r
d
iac
f
unc
ti
on
,
whic
h
is
of
gr
e
a
t
s
igni
f
ica
nc
e
f
or
c
li
nica
l
r
e
a
l
-
ti
me
moni
tor
ing
a
n
d
diagnos
is
[
1]
.
How
e
ve
r
,
e
c
hoc
a
r
diogr
a
phy
it
s
e
lf
ha
s
the
c
ha
r
a
c
ter
is
ti
c
s
of
low
im
a
ge
e
dge
c
ont
r
a
s
t
a
nd
hi
gh
s
pe
c
kle
nois
e
[
2]
.
I
n
a
ddit
ion,
indi
vidual
phy
s
iol
ogica
l
dif
f
e
r
e
nc
e
s
br
ing
a
bout
s
tr
uc
tur
a
l
di
f
f
e
r
e
nc
e
s
.
I
r
r
e
gular
c
ha
nge
s
make
it
a
we
ll
-
known
c
ha
ll
e
nge
to
a
c
hieve
f
ull
y
a
utom
a
ted
r
e
a
l
-
ti
me
s
e
gmenta
ti
on
of
the
lef
t
ve
ntr
icle
in
e
c
hoc
a
r
diogr
a
phy
.
W
it
h
the
de
ve
lopm
e
nt
o
f
de
e
p
lea
r
ning
ne
twor
k
models
,
a
dva
nc
e
d
tec
hnologi
e
s
ba
s
e
d
on
de
e
p
lea
r
ning
a
r
e
good
a
t
dis
c
ove
r
ing
c
ompl
e
x
f
e
a
tur
e
s
in
a
n
e
nd
-
to
-
e
nd
manne
r
by
lea
r
ning
c
onti
nuous
ly
dir
e
c
tl
y
f
r
om
the
da
ta,
inj
e
c
ti
ng
ne
w
vit
a
li
ty
int
o
the
a
utom
a
ti
c
s
e
gmenta
ti
on
of
the
lef
t
ve
nt
r
icle
in
e
c
hoc
a
r
diogr
a
phy.
T
he
de
e
p
lea
r
ning
a
r
c
hit
e
c
tur
e
s
of
two
dim
e
ns
ional
(
2D
)
s
ingl
e
-
f
r
a
me
e
c
hoc
a
r
diogr
a
m
s
e
gmenta
ti
on,
including
f
u
ll
c
onvolut
ional
ne
ur
a
l
ne
twor
ks
(
F
C
N)
[
3]
–
[
5]
a
nd
Une
t
a
r
c
hit
e
c
tur
e
s
[
6]
–
[
8]
,
ha
ve
r
e
c
e
ived
e
xtens
ive
a
tt
e
nti
on
a
nd
ha
ve
pr
ove
n
their
e
f
f
e
c
t
ivene
s
s
.
I
n
or
de
r
to
im
p
r
ove
pe
r
f
o
r
manc
e
,
s
ome
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
ompar
ati
v
e
e
v
aluat
ion
of
left
v
e
ntr
icle
s
e
gme
ntat
ion
us
ing
impr
ov
e
d
py
r
amid
s
c
e
ne
…
(
J
in
W
ang
)
3215
s
tudi
e
s
ha
ve
e
nha
nc
e
d
the
s
e
gmenta
ti
on
c
a
pa
bil
it
ies
of
the
F
C
N
ne
twor
k
s
tr
uc
tur
e
by
e
xpa
nding
the
tr
a
ini
ng
da
tas
e
t
[
9]
,
r
e
f
ini
ng
the
los
s
f
unc
ti
on
[
10
]
,
[
11
]
,
a
nd
uti
l
izing
p
r
e
-
tr
a
ini
ng
to
ini
ti
a
li
z
e
the
mod
e
l
[
12
]
,
a
c
hieving
good
s
e
gmenta
ti
on
r
e
s
ult
s
.
On
the
oth
e
r
ha
nd,
Une
t,
tun
ing
Une
t
a
r
c
hit
e
c
tur
e
,
wide
U
ne
t,
a
nd
Une
t++
s
how
obvious
a
dva
nt
a
ge
s
in
a
c
c
ur
a
c
y
a
nd
r
e
a
l
-
ti
me
pe
r
f
or
manc
e
f
or
e
c
hoc
a
r
diogr
a
phy
s
e
gmenta
ti
on
[
13]
,
[
14
]
.
T
o
f
ur
the
r
im
p
r
ove
the
pe
r
f
o
r
manc
e
,
the
r
e
s
e
a
r
c
he
r
s
im
pr
ove
d
the
Une
t
ne
twor
k
by
e
xtr
a
c
ti
ng
global
a
nd
loca
l
f
e
a
tur
e
s
of
the
im
a
ge
a
nd
pr
opo
s
e
d
a
powe
r
f
ul
e
nd
-
to
-
e
nd
s
olut
ion
,
s
uc
h
a
s
the
t
he
ba
tch
nor
malize
d
U
n
e
t
(
B
NU
-
N
et
)
model,
whic
h
e
mpl
o
ys
e
xpone
nti
a
l
li
ne
a
r
unit
(
E
L
U)
a
s
a
c
ti
va
ti
on
f
un
c
ti
ons
a
t
s
uc
c
e
s
s
ive
laye
r
s
in
the
c
oding
pa
th
a
nd
ba
tch
nor
maliza
ti
on
a
f
ter
th
e
c
onvolut
ional
f
il
ter
s
[
15
]
.
C
o
mbi
ning
the
a
dva
ntage
s
of
R
e
s
Ne
t
a
nd
Une
t,
the
R
e
s
_U
model
e
xpa
nds
the
or
igi
na
l
da
ta
f
r
om
the
pr
e
viou
s
laye
r
to
e
a
c
h
block
of
the
c
ur
r
e
nt
laye
r
,
ther
e
by
pr
opa
g
a
ti
ng
a
nd
e
nha
nc
ing
f
e
a
tur
e
s
th
r
oughout
the
mo
de
l
[
16]
.
R
e
s
idual
dil
a
ted
U
n
e
t
(
R
e
s
DN
Une
t
)
model
that
uti
li
z
e
s
Une
t,
c
a
s
c
a
de
dil
a
ted
c
onvolut
ion,
a
nd
r
e
s
idual
blocks
r
ich
in
s
que
e
z
e
-
a
nd
-
e
x
c
it
a
ti
on
ope
r
a
ti
ons
to
e
xtr
a
c
t
global
a
nd
mul
ti
-
s
c
a
le
f
e
a
tur
e
s
[
17]
.
T
he
pyr
a
mi
d
ne
twor
k
a
nd
Une
t
we
r
e
c
ombi
ne
d
to
c
ons
tr
u
c
t
m
ult
i
-
f
e
a
tur
e
pyr
a
mi
d
Une
t
(
M
F
P
-
Une
t)
[
18
]
.
An
a
tt
e
nti
on
mec
ha
nis
m
wa
s
int
r
oduc
e
d
int
o
the
Une
t
model
t
o
a
void
e
xt
r
a
c
ti
ng
many
s
im
il
a
r
f
e
a
tur
e
s
dur
ing
p
a
r
a
mete
r
c
a
lcula
ti
on
[
19]
.
T
he
r
e
s
idual
r
e
s
idual
o
f
r
e
s
idual
-
Une
t
(
R
OR
-
Une
t
)
wa
s
pr
opos
e
d
to
s
olve
the
v
a
nis
hing
gr
a
dient
pr
oblem
a
nd
im
pr
ove
s
e
gmenta
ti
on
pe
r
f
or
manc
e
[
20]
.
De
ns
e
-
Une
t
thr
ough
da
ta
a
ugm
e
ntation
s
tr
a
tegy
[
21]
.
T
he
r
e
s
ult
s
s
how
that
thes
e
models
outper
f
or
m
Une
t
in
da
ta
de
nois
ing
a
nd
pr
ovide
r
e
l
iable
a
nd
s
table
s
e
gmenta
ti
on
r
e
s
ult
s
.
I
n
a
ddit
ion
to
F
C
N
a
nd
Une
t
de
e
p
lea
r
ning
n
e
tw
or
k
s
tr
uc
tur
e
s
,
s
ome
s
tudi
e
s
c
ombi
ne
tr
a
dit
ional
s
e
gmenta
ti
on
tec
hniques
with
mor
phologi
c
a
l
met
hods
[
22]
,
s
na
ke
models
[
23]
,
a
c
ti
ve
s
ha
pe
models
(
ASM
)
[
24]
,
[
25]
a
nd
c
onvolut
ional
ne
ur
a
l
ne
two
r
k
(
C
N
N)
model
a
r
e
c
ombi
ne
d
to
im
pr
ove
the
pe
r
f
o
r
man
c
e
of
lef
t
ve
ntr
icula
r
s
e
gmenta
ti
on
in
e
c
hoc
a
r
diogr
a
phy.
S
o
me
s
tudi
e
s
ha
ve
s
igni
f
ica
ntl
y
c
ont
r
ibut
e
d
to
the
e
xpa
ns
ion
of
de
e
p
lea
r
ning
ne
twor
k
a
r
c
hit
e
c
tur
e
s
by
m
ixi
ng
model
s
tr
uc
tur
e
s
.
F
or
ins
tanc
e
,
S
e
gNe
t
is
bu
il
t
us
ing
17
s
tac
ke
d
c
onvolut
ional
laye
r
s
[
26]
,
c
a
s
c
a
de
d
s
e
gmenta
ti
on
a
nd
r
e
gr
e
s
s
ion
ne
twor
k
(
C
S
R
Ne
t)
c
ombi
ne
s
s
e
gmenta
ti
on
C
NN
models
with
qu
a
nti
z
e
d
r
e
gr
e
s
s
ion
ne
twor
ks
[
27]
,
a
nd
the
VG
GN
e
t
model,
b
a
s
e
d
on
tr
a
ns
f
e
r
lea
r
ning
[
28]
,
de
mons
tr
a
tes
obvious
a
dva
ntage
s
in
s
e
gmenta
ti
on
e
f
f
icie
nc
y
whe
n
us
ing
a
c
ombi
ne
d
ne
twor
k
of
a
n
e
nc
ode
r
a
nd
long
s
hor
t
-
ter
m
memo
r
y
(
L
S
T
M
)
[
29
]
,
[
30]
.
B
a
s
e
d
on
a
n
im
pr
ove
d
a
nd
va
r
iant
model
of
the
Une
t
s
tr
uc
tu
r
e
,
f
e
a
tur
e
f
us
ion
is
pe
r
f
o
r
med
thr
ough
s
kip
c
onne
c
ti
ons
to
maximi
z
e
the
ut
il
iza
ti
on
of
de
e
p
s
e
mantic
inf
or
mation
a
nd
s
ha
ll
ow
de
tail
i
nf
or
mation.
T
his
a
ppr
oa
c
h
s
hows
s
igni
f
ic
a
nt
a
dva
ntage
s
in
ter
ms
of
c
omput
a
ti
ona
l
a
c
c
ur
a
c
y,
s
e
ns
it
ivi
ty,
a
nd
e
f
f
icie
nc
y.
How
e
ve
r
,
it
a
ls
o
f
a
c
e
s
s
ome
li
mi
tations
.
F
or
e
xa
mpl
e
,
a
ll
s
e
mantic
a
dva
ntage
s
of
f
e
a
tur
e
s
a
t
dif
f
e
r
e
nt
s
c
a
les
dur
ing
s
e
gmenta
ti
on
a
r
e
ignor
e
d,
whic
h
hinder
s
the
lea
r
ning
c
a
pa
b
il
it
ies
of
de
e
p
lea
r
nin
g
ne
twor
ks
.
C
ons
e
que
ntl
y,
thes
e
ne
twor
ks
r
e
ly
on
lar
ge
a
mount
s
of
a
nnotate
d
da
ta
a
nd
powe
r
f
ul
s
tor
a
ge
a
nd
c
omput
ing
unit
s
.
Additi
ona
ll
y
,
the
lar
ge
nu
mber
o
f
model
pa
r
a
mete
r
s
r
e
s
ult
s
in
e
xtende
d
p
r
e
diction
ti
mes
.
I
n
view
o
f
the
inabil
it
y
of
F
C
N
a
nd
Une
t
to
ha
ndle
tar
ge
ts
of
dif
f
e
r
e
nt
s
ize
s
a
nd
mul
t
i
-
s
c
a
le
inf
or
mation
we
ll
,
r
e
s
e
a
r
c
he
r
s
buil
t
a
pyr
a
mi
d
s
c
e
ne
pa
r
s
ing
ne
twor
k
(
P
S
P
Ne
t
)
ne
twor
k
to
us
e
dif
f
e
r
e
nt
a
r
e
a
s
to
a
ggr
e
ga
te
global
c
ontext
inf
or
mation
a
nd
lea
r
n
global
im
a
ge
-
leve
l
f
e
a
tur
e
s
a
nd
loca
l
mul
ti
-
s
c
a
le
f
e
a
tur
e
s
a
t
the
s
a
me
ti
me
.
M
or
e
a
nd
mor
e
r
e
s
e
a
r
c
he
r
s
a
r
e
a
p
plyi
ng
thi
s
kind
of
ne
twor
k
to
the
f
ield
o
f
medic
a
l
im
a
ge
s
e
gmenta
ti
on,
s
howing
c
e
r
tain
a
dva
ntage
s
in
pixel
-
leve
l
s
e
gmenta
ti
on
a
nd
a
c
hieving
e
xc
e
ll
e
nt
s
e
g
menta
ti
on
pe
r
f
or
ma
nc
e
on
va
r
ious
da
ta
s
e
ts
[
31]
–
[
33
]
.
W
he
n
a
pplyi
ng
the
P
S
P
Ne
t
de
e
p
le
a
r
ning
ne
twor
k
to
the
pr
e
diction
of
tum
or
mar
ke
r
s
,
a
dice
s
im
il
a
r
it
y
c
oe
f
f
icie
nt
(
Dic
e
)
index
o
f
91.
3
%
wa
s
a
c
hieve
d,
a
lon
g
with
a
f
a
s
ter
pr
oc
e
s
s
ing
s
pe
e
d
[
34]
,
S
im
i
lar
ly,
whe
n
a
p
plyi
ng
the
P
S
P
Ne
t
model
to
p
r
os
tate
magne
ti
c
r
e
s
ona
nc
e
im
a
ging
(
M
R
I
)
s
e
gmenta
ti
on,
a
lea
ding
s
e
gmenta
ti
on
a
c
c
ur
a
c
y
of
98.
65%
wa
s
a
c
hieve
d
[
35]
.
T
he
P
S
P
Ne
t
ne
twor
k
b
a
s
e
d
on
De
ns
e
N
e
t
wa
s
us
e
d
f
or
br
e
a
s
t
c
a
nc
e
r
im
a
ge
s
e
gmenta
ti
on,
a
c
hieving
a
s
e
g
menta
ti
on
a
c
c
ur
a
c
y
of
94.
68%
higher
than
the
e
xis
ti
ng
method
[
36]
.
T
he
s
e
mantic
s
e
gmenta
ti
on
of
na
tur
a
l
im
a
ge
s
wa
s
a
c
hieve
d
ba
s
e
d
on
P
S
P
Ne
t,
a
nd
the
s
e
gmenta
ti
o
n
pe
r
f
o
r
manc
e
wa
s
we
ll
ve
r
i
f
ied
on
pub
li
c
da
ta
s
e
ts
[
37]
.
S
ome
s
tudi
e
s
ha
ve
us
e
d
P
S
P
Ne
t
f
or
ne
twor
k
f
us
ion
a
nd
a
c
hieve
d
good
r
e
s
ult
s
f
or
medic
a
l
im
a
ge
s
e
gmenta
ti
on
tas
ks
[
38]
–
[
40
]
.
How
e
ve
r
,
f
e
a
s
ibi
li
ty
s
tudi
e
s
of
lef
t
ve
nt
r
icula
r
s
e
gmenta
ti
on
in
e
c
hoc
a
r
diogr
a
phy
a
r
e
s
ti
ll
s
c
a
r
c
e
.
At
thi
s
s
tage
,
de
e
p
lea
r
ning
-
ba
s
e
d
methods
a
r
e
f
a
c
e
d
with
the
dua
l
c
ha
ll
e
nge
s
of
im
pr
ov
ing
pr
oc
e
s
s
ing
a
c
c
ur
a
c
y
a
nd
a
c
c
e
ler
a
ti
ng
pr
oc
e
s
s
ing
s
pe
e
d
whe
n
a
ppli
e
d
to
2D
e
c
hoc
a
r
diogr
a
phic
lef
t
ve
ntr
icle
s
e
gmenta
ti
on.
T
his
s
tudy
a
im
s
to
ove
r
c
ome
the
l
i
mi
tations
of
c
ur
r
e
nt
s
e
gmenta
ti
on
models
s
uc
h
a
s
F
C
N
a
nd
Un
e
t
in
de
a
li
ng
with
s
e
gmenta
ti
on
tas
ks
.
I
t
a
ls
o
s
e
e
ks
to
e
xplor
e
a
n
a
lgor
it
hm
that
ba
lanc
e
s
ne
twor
k
lea
r
ning
de
pth
a
nd
pe
r
f
or
manc
e
,
a
nd
it
pr
ov
ides
a
ne
w
s
ol
uti
on
f
o
r
r
e
a
l
-
ti
me
s
e
gmenta
ti
on
of
the
le
f
t
ve
nt
r
ic
le
in
2D
e
c
hoc
a
r
diogr
a
phy.
T
his
model
c
a
n
e
f
f
e
c
t
ively
e
xt
r
a
c
t
mul
ti
-
s
c
a
le
global
f
e
a
tur
e
s
f
r
om
im
a
ge
s
,
f
ull
y
uti
li
z
e
the
pos
it
ion
a
nd
s
ha
pe
pr
io
r
s
of
the
im
a
ge
,
f
us
e
global
a
nd
loca
l
inf
o
r
mation,
a
nd
im
pr
ove
s
e
gmenta
ti
on
a
c
c
ur
a
c
y
a
nd
s
pe
e
d.
T
his
pa
pe
r
mainly
f
oc
us
e
s
on
the
f
oll
owing
thr
e
e
a
s
pe
c
t
s
of
r
e
s
e
a
r
c
h:
i
)
t
o
e
xplor
e
the
us
e
of
P
S
P
Ne
t
ne
twor
k
to
a
c
hieve
r
e
a
l
-
ti
me
s
e
gmenta
ti
on
of
the
lef
t
ve
ntr
icle
,
a
djus
t
the
ba
c
kbone
f
e
a
tur
e
e
xtr
a
c
ti
o
n
s
tr
uc
tu
r
e
t
o
im
pr
o
ve
M
obi
leN
e
t
v2
ins
te
a
d
of
R
e
s
Ne
t
f
o
r
be
tt
e
r
pe
r
f
o
r
ma
nc
e
,
a
nd
c
o
mp
r
e
h
e
ns
ively
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
202
5
:
321
4
-
3227
3216
e
va
luate
the
s
e
gm
e
ntat
ion
a
c
c
u
r
a
c
y
a
nd
e
f
f
icie
nc
y
;
ii
)
t
he
e
f
f
e
c
ts
of
tw
o
d
if
f
e
r
e
nt
mode
l
i
nit
ial
iza
ti
on
metho
ds
,
pr
e
-
tr
a
ini
ng
a
nd
tr
a
ns
f
e
r
lea
r
n
ing
,
o
n
s
e
gme
ntat
io
n
pe
r
f
o
r
ma
nc
e
,
s
uc
h
a
s
a
lg
or
i
thm
ic
a
c
c
ur
a
c
y
a
nd
lea
r
nin
g
e
f
f
icie
nc
y,
of
the
ne
two
r
k
m
ode
l
in
th
e
le
f
t
ve
ntr
ic
ular
s
e
g
menta
ti
o
n
tas
k
we
r
e
s
tu
died
;
a
nd
iii
)
a
n
e
x
ha
us
ti
ve
c
ompa
r
is
on
o
f
s
e
gmen
tati
on
pe
r
f
o
r
ma
nc
e
wa
s
c
onduc
te
d
be
twe
e
n
the
im
pr
o
ve
d
P
S
P
Ne
t
a
nd
the
op
ti
ma
l
models
of
c
las
s
ica
l
s
e
gme
nta
ti
on
ne
t
wor
ks
,
inc
lu
ding
VG
G
[
41
]
,
U
ne
t
[
42
]
,
a
n
d
R
e
s
_U
[
1
6]
.
E
x
p
e
r
im
e
ntal
r
e
s
ult
s
s
h
ow
t
ha
t
the
i
mp
r
ove
d
P
S
P
Ne
t
s
ign
if
ic
a
ntl
y
out
pe
r
f
o
r
ms
t
he
R
e
s
_U
ne
two
r
k
(
ba
s
e
d
o
n
Une
t
)
in
f
e
a
tur
e
e
xtr
a
c
ti
on.
S
pe
c
if
ica
ll
y,
it
e
nha
nc
e
s
the
Dic
e
index
by
2.
3%
,
r
e
duc
e
s
model
pa
r
a
mete
r
s
,
s
hor
tens
pr
oc
e
s
s
ing
ti
me
by
33.
4
%
,
a
nd
thus
boos
ts
r
e
a
l
-
ti
m
e
pe
r
f
or
manc
e
a
nd
a
c
c
ur
a
c
y.
2.
M
E
T
HO
D
B
a
s
e
d
on
the
e
xis
ti
ng
r
e
s
e
a
r
c
h
on
e
c
hoc
a
r
diogr
a
phy
s
e
gmenta
ti
on
methods
,
thi
s
pa
pe
r
pr
opos
e
s
a
P
S
P
Ne
t
ne
twor
k
model
f
o
r
s
e
gmenting
the
lef
t
ve
ntr
icle
in
2D
e
c
hoc
a
r
diogr
a
phy.
T
he
ove
r
a
ll
f
r
a
mew
or
k
of
the
a
lgor
it
hm
is
il
lus
tr
a
ted
in
F
igu
r
e
1:
i
)
in
pr
oc
e
s
s
ing
the
publi
c
thr
e
e
-
dim
e
ns
ional
(
3D)
c
ha
ll
e
nge
on
e
ndoc
a
r
dial
3D
ult
r
a
s
ound
s
e
gmenta
ti
on
(
C
E
T
US)
da
tas
e
t,
2D
s
li
c
e
s
a
r
e
obtaine
d
by
s
a
mpl
ing
a
long
the
s
hor
t
a
xis
,
a
nd
the
im
a
ge
s
unde
r
go
p
r
e
pr
oc
e
s
s
ing
to
e
nha
nc
e
de
tails
a
nd
r
e
duc
e
nois
e
,
without
a
lt
e
r
ing
the
s
ha
pe
of
the
he
a
r
t
[
43]
.
ii
)
the
pr
oc
e
s
s
e
d
2D
im
a
ge
is
pa
s
s
e
d
to
the
P
S
P
Ne
t
s
e
gmenta
ti
on
model,
w
hich
is
a
pyr
a
mi
d
a
nd
de
e
p
c
onvolut
ional
ne
twor
k,
f
or
a
ut
omatic
f
e
a
tur
e
e
xtr
a
c
ti
on
a
nd
to
pr
e
dict
the
l
e
f
t
v
e
ntr
icula
r
s
e
gmenta
ti
on
r
e
s
ult
s
.
iii
)
e
va
luate
the
s
e
gmenta
ti
on
r
e
s
ult
s
.
3D
L
e
f
t
V
e
nt
r
i
c
ul
a
r
E
c
hoca
r
di
ogr
a
phy
3D
l
a
bel
s
I
m
a
ge
P
r
e
pr
oc
e
s
s
i
ng
2D
s
l
i
c
e
s
2D
l
a
bel
s
I
m
pr
oved
P
S
P
N
e
t
2D
S
e
gm
e
nt
a
t
i
on
S
e
gm
e
nt
a
t
i
on
R
e
s
ul
t
s
L
a
bel
s
R
e
s
ul
t
s
a
nd
A
na
l
ys
i
s
I
nput
S
e
gm
e
nt
a
t
i
on
M
od
e
l
P
e
r
f
or
m
a
nce
E
val
uat
i
on
F
igur
e
1.
Ove
r
a
ll
a
lgor
it
hm
f
r
a
mew
or
k
2
.
1.
I
n
p
u
t
an
d
im
age
p
r
e
p
r
oc
e
s
s
in
g
T
he
da
tas
e
t
f
or
thi
s
e
xpe
r
im
e
nt
uti
l
ize
s
the
publi
c
C
E
T
US
da
tas
e
t,
whic
h
c
ompr
is
e
s
45
3D
e
c
hoc
a
r
diogr
a
phy
s
e
que
nc
e
s
.
T
he
s
e
s
e
que
n
c
e
s
f
o
r
m
a
n
e
c
hoc
a
r
diogr
a
phy
da
tas
e
t
that
is
e
ve
nly
di
s
tr
ibut
e
d
a
mong
thr
e
e
dif
f
e
r
e
nt
s
ubgr
oups
:
he
a
lt
hy
s
ubjec
ts
,
pa
ti
e
nts
wit
h
pr
e
vious
mus
c
le
da
mage
,
a
nd
pa
ti
e
nts
with
dil
a
ted
c
a
r
diom
yopa
thy.
T
his
da
tas
e
t
ha
s
be
e
n
e
xt
e
ns
ively
va
li
da
ted
in
numer
ous
c
las
s
ic
a
nd
s
tate
-
of
-
the
-
a
r
t
a
lgor
it
hms
[
44
]
.
I
n
t
his
e
xpe
r
im
e
nt
,
the
45
3D
vol
ume
da
ta
we
r
e
s
li
c
e
d
int
o
2D
im
a
ge
s
a
long
the
s
hor
t
a
xis
.
S
li
c
e
s
that
did
not
c
ontain
a
ny
c
a
r
diac
inf
o
r
matio
n
we
r
e
f
il
ter
e
d
out
a
nd
de
lete
d.
Due
to
the
a
xial
s
ymm
e
tr
y
of
the
2D
s
li
c
e
s
,
only
ha
lf
of
them
we
r
e
s
e
lec
ted.
Ulti
mate
l
y,
3616
va
li
d
2D
s
li
c
e
im
a
ge
s
we
r
e
obtai
ne
d.
2
.
2.
S
e
gm
e
n
t
at
ion
m
od
e
ls
2
.
2.
1.
P
S
P
Ne
t
m
od
e
l
T
he
e
c
hoc
a
r
diogr
a
phy
input
,
main
s
tr
uc
tur
e
,
a
n
d
s
e
gmenta
ti
on
pr
oc
e
s
s
of
the
P
S
P
Ne
t
ne
twor
k
model
a
r
e
il
lus
tr
a
ted
in
F
igu
r
e
2,
whic
h
c
ompr
i
s
e
s
f
our
main
pa
r
ts
:
f
e
a
tur
e
e
xtr
a
c
ti
on
,
pyr
a
mi
d
pooli
ng,
f
e
a
tur
e
f
us
ion,
a
nd
de
e
p
s
upe
r
vis
ion.
F
i
r
s
t,
the
f
e
a
tur
e
e
xtr
a
c
ti
on
modul
e
o
f
the
C
NN
is
us
e
d
to
o
btain
the
f
e
a
tur
e
map
of
the
input
im
a
ge
.
T
h
e
n
,
thi
s
f
e
a
tur
e
map
is
input
int
o
the
pyr
a
mi
d
pooli
ng
modu
le
(
P
P
M
)
to
obtain
a
4
-
laye
r
pooled
f
e
a
tu
r
e
map.
F
inally
,
th
e
pooled
f
e
a
tur
e
map
is
c
onc
a
tena
ted
with
the
ba
c
kbone
f
e
a
tur
e
map
a
nd
e
nter
e
d
in
to
the
F
C
N
modul
e
to
obtain
the
pr
e
dicte
d
s
e
gmenta
ti
on
r
e
s
ult
s
.
T
he
P
P
M
is
the
c
or
e
of
the
P
S
P
Ne
t
ne
twor
k
.
T
his
modul
e
a
ggr
e
g
a
tes
f
our
f
e
a
tur
e
laye
r
s
of
di
f
f
e
r
e
nt
dim
e
ns
ions
,
u
s
e
s
1×
1
c
onvolut
ion
to
r
e
duc
e
the
dim
e
ns
ionalit
y,
a
nd
th
e
n
pe
r
f
or
ms
a
n
ups
a
mpl
ing
ope
r
a
ti
on
to
s
upe
r
im
pos
e
the
r
e
s
tor
e
d
f
e
a
tur
e
s
with
th
e
ini
ti
a
l
f
e
a
tur
e
s
,
ther
e
by
f
o
r
mi
ng
r
iche
r
global
inf
o
r
mation
a
nd
c
ha
r
a
c
ter
is
ti
c
r
e
pr
e
s
e
ntations
of
s
ub
-
r
e
gion
inf
or
mation
.
I
n
F
igu
r
e
2
,
the
obtaine
d
f
e
a
tur
e
laye
r
is
divi
de
d
int
o
f
our
s
ub
-
r
e
gions
of
d
if
f
e
r
e
nt
dim
e
ns
ions
of
1×
1,
2×
2
,
3×
3,
a
nd
6×
6,
a
nd
then
a
ve
r
a
ge
pooli
ng
withi
n
the
s
ub
-
r
e
gions
is
pe
r
f
or
med
.
C
ompar
e
d
wi
th
di
r
e
c
t
global
pooli
ng,
whic
h
wi
ll
c
a
us
e
the
los
s
of
pa
r
t
o
f
the
loca
ti
on
in
f
or
mation
,
py
r
a
mi
d
pooli
ng
is
us
e
d
to
take
int
o
a
c
c
ount
the
global
in
f
or
mation
a
nd
the
r
e
latio
ns
hip
be
twe
e
n
e
a
c
h
s
ub
-
r
e
gion
a
nd
r
e
a
li
z
e
the
a
gg
r
e
ga
ti
on
of
c
ontext
inf
or
mation
in
di
f
f
e
r
e
nt
r
e
gions
.
T
he
P
P
M
modul
e
e
na
bles
the
P
S
P
Ne
t
ne
twor
k
to
f
ull
y
obtain
s
e
mantic
inf
or
mation
a
t
a
ll
leve
ls
a
nd
s
c
a
les
in
e
c
h
oc
a
r
diogr
a
phy
a
nd
ha
s
s
tr
ong
a
ppli
c
a
ti
on
potential
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
ompar
ati
v
e
e
v
aluat
ion
of
left
v
e
ntr
icle
s
e
gme
ntat
ion
us
ing
impr
ov
e
d
py
r
amid
s
c
e
ne
…
(
J
in
W
ang
)
3217
CONV
CONV
CONV
CONV
POOL
CONV
UPSAMPLE
Improv
ed Enco
der
CONCAT
Orig
inal
ima
ge
Pre
dic
tive im
ag
e
Feat
ure
ext
rac
tion
Netwo
rk
PPM
mod
ule
F
igur
e
2.
P
S
P
Ne
t
s
e
gmenta
ti
on
model
a
lgo
r
it
hm
s
t
r
uc
tur
e
diagr
a
m
2
.
2.
2.
I
m
p
r
ove
d
P
S
P
Ne
t
m
o
d
e
l
T
h
e
i
mp
r
ov
e
d
P
S
P
N
e
t
ne
tw
or
k
is
ma
in
ly
f
oc
us
e
d
on
t
he
f
e
a
tu
r
e
e
x
t
r
a
c
ti
on
pa
r
t
,
a
s
s
h
ow
n
i
n
F
ig
u
r
e
2
.
T
h
e
l
i
gh
twe
i
gh
t
M
o
bi
le
Ne
t
is
us
e
d
to
r
e
pl
a
c
e
t
he
R
e
s
Ne
t
.
A
t
r
o
us
c
on
vo
lu
t
io
n
a
n
d
f
e
a
t
ur
e
f
us
io
n
a
r
e
f
u
r
t
he
r
i
n
t
r
o
duc
e
d
to
a
c
h
ie
ve
th
e
g
oa
l
o
f
e
f
f
e
c
t
ive
l
y
e
x
t
r
a
c
t
in
g
f
e
a
t
ur
e
s
a
nd
s
ho
r
ten
i
ng
the
m
ode
l
r
u
nn
in
g
ti
me
.
2
.
2.
3
.
F
e
a
t
u
r
e
e
xt
r
ac
t
ion
n
e
t
wor
k
T
he
t
r
a
dit
ional
P
S
P
Ne
t
us
e
s
a
R
e
s
Ne
t
-
ba
s
e
d
C
N
N
f
or
ba
c
kbone
f
e
a
tur
e
e
xtr
a
c
ti
on,
whic
h
f
e
a
tu
r
e
s
many
laye
r
s
a
nd
a
lar
ge
r
e
c
e
pti
ve
f
ield
,
but
it
ha
s
l
im
it
a
ti
ons
in
c
a
ptur
ing
global
inf
or
mat
ion
[
45
]
.
I
n
c
ontr
a
s
t,
M
obil
e
Ne
t
e
mpl
oys
de
pthwis
e
s
e
pa
r
a
ble
c
onvolut
ions
,
r
e
duc
ing
c
omput
a
ti
ona
l
c
ompl
e
xit
y
a
nd
e
ns
ur
ing
it
is
li
ghtwe
ight
,
s
wif
t
,
a
nd
pr
e
c
is
e
[
46
]
.
I
n
thi
s
pa
p
e
r
,
R
e
s
Ne
t50
a
nd
M
obil
e
Ne
tv2
a
r
e
uti
li
z
e
d
a
s
b
a
c
kbone
f
e
a
tur
e
e
xtr
a
c
ti
on
ne
twor
ks
f
or
a
ddr
e
s
s
ing
s
e
gmenta
ti
on
de
tails
a
nd
f
a
c
il
it
a
ti
ng
pe
r
f
o
r
manc
e
c
ompar
is
ons
.
a)
R
e
s
Ne
t50
a
s
e
nc
ode
r
R
e
s
Ne
t
,
a
s
a
n
e
nc
ode
r
,
he
lps
im
p
r
ove
the
a
c
c
ur
a
c
y
of
s
e
gmenta
ti
on
ne
twor
ks
by
r
e
taining
s
pa
ti
a
l
inf
or
mation
thr
ough
inc
r
e
a
s
e
d
pa
r
a
mete
r
s
.
I
t
f
e
a
t
ur
e
s
va
r
ious
laye
r
s
with
di
f
f
e
r
ing
c
onvolut
ional
a
nd
ba
tch
nor
maliza
ti
on
c
ounts
,
R
e
s
Ne
t34
a
nd
R
e
s
Ne
t50
be
ing
two
wide
ly
us
e
d
typi
c
a
l
a
r
c
hit
e
c
tur
e
s
that
a
c
hieve
f
e
a
tur
e
map
dim
e
ns
ionalit
y
r
e
duc
ti
on
th
r
ough
s
p
a
ti
a
l
c
onvolut
ion,
a
s
s
hown
in
F
igur
e
3
[
47]
.
R
e
s
Ne
t50
de
mons
tr
a
tes
s
igni
f
ica
nt
a
dva
ntage
s
in
s
e
gmenta
ti
on
a
c
c
ur
a
c
y
a
nd
mi
ti
ga
tes
the
is
s
ue
of
va
nis
hing
gr
a
dients
in
de
e
p
ne
twor
ks
.
I
n
thi
s
pa
pe
r
,
R
e
s
Ne
t50
is
s
e
lec
ted
a
s
the
ba
c
kbone
ne
twor
k
f
or
c
ompar
a
ti
ve
e
xpe
r
i
ments
.
7
×
7
c
o
n
v
,
6
4
/
2
P
o
o
l
I
n
p
u
t
F
c
1
0
0
0
P
o
o
l
S
o
f
t
m
a
x
3
3
,
64
3
3
3
,
64
3
3
,
128
4
3
3
,
128
3
3
,
256
6
3
3
,
256
3
3
,
512
3
3
3
,
512
7
×
7
c
o
n
v
,
6
4
/
2
P
o
o
l
I
n
p
u
t
F
c
1
0
0
0
P
o
o
l
S
o
f
t
m
a
x
1
1
,
64
3
3
,
64
3
1
1
,
25
6
1
1
,
128
3
3
,
128
4
1
1
,
512
1
1
,
256
3
3
,
256
6
1
1
,
1024
1
1
,
512
3
3
,
512
3
1
1
,
204
8
F
igur
e
3.
R
e
s
Ne
t34
a
nd
R
e
s
Ne
t50
ne
twor
k
s
tr
uc
tur
e
diagr
a
m
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
202
5
:
321
4
-
3227
3218
b)
I
mpr
ove
d
M
obil
e
Ne
tv2
a
s
e
nc
ode
r
T
he
ba
c
kbone
f
or
f
e
a
tur
e
e
xtr
a
c
ti
on
e
mpl
oys
th
e
li
ghtwe
ight
ne
ur
a
l
ne
twor
k
M
obil
e
Ne
t,
whic
h
mi
nim
ize
s
the
r
e
quir
e
ment
f
o
r
ne
twor
k
pa
r
a
me
ter
s
a
nd
e
nha
nc
e
s
the
ne
twor
k's
r
e
a
l
-
ti
me
pe
r
f
or
manc
e
.
M
obil
e
Ne
t
pr
im
a
r
il
y
e
xis
ts
in
th
r
e
e
ve
r
s
ions
:
V1,
V2,
a
nd
V3.
T
o
ba
lanc
e
a
lgor
it
hm
a
c
c
ur
a
c
y
a
nd
s
pe
e
d,
thi
s
pa
pe
r
a
dopts
M
obil
e
Ne
tV2
a
s
the
ba
c
kbone
f
e
a
tu
r
e
e
xtr
a
c
ti
on
ne
twor
k
f
or
the
e
nc
ode
r
.
T
o
thi
s
e
n
d,
a
t
r
ous
c
onvolut
ions
a
r
e
int
r
oduc
e
d
int
o
the
C
onv6
f
e
a
tur
e
laye
r
,
while
f
e
a
tur
e
s
f
r
o
m
d
if
f
e
r
e
nt
laye
r
s
of
C
o
nv5
a
nd
C
onv6
a
r
e
blende
d
a
nd
s
upe
r
im
pos
e
d.
T
his
a
ppr
oa
c
h
s
tr
e
ngthens
int
e
r
-
laye
r
f
e
a
tur
e
f
us
ion
without
in
c
r
e
a
s
ing
the
c
omput
a
ti
ona
l
load
,
e
xpa
nds
the
r
e
c
e
pti
ve
f
iel
d,
a
nd
r
e
duc
e
s
the
los
s
of
a
mbi
guous
inf
o
r
mation
a
t
e
dge
s
.
F
igur
e
4
il
lus
tr
a
tes
the
s
tr
uc
tur
e
of
the
im
pr
ove
d
e
nc
ode
r
.
I
npu
t
(
473
×
473
×
3)
C
onv
2d
(
237×
237×
32)
I
nver
t
e
d_R
e
s
_C
onv
, S
t
r
i
de
s
=
1
(
237
×
237
×
16)
I
nve
r
t
e
d_R
e
s
_C
onv ×
2, S
t
r
i
de
s
=
2
(
119
×
119
×
24)
I
nver
t
e
d_R
e
s
_C
onv
×
3, S
t
r
i
de
s
=
2
(
60×
60×
32)
I
nver
t
e
d_R
e
s
_C
onv
×
4, S
t
r
i
de
s
=
2
(
30×
30×
64)
I
nve
r
t
e
d_R
e
s
_C
onv ×
3, S
t
r
i
de
s
=
1
(
30×
30×
96)
A
t
r
ous
_c
onv
×
3
(
30×
30×
160)
M
obi
l
e
N
e
t
v2
C
onv1
C
onv
2
C
onv
3
C
onv4
C
onv
5
C
onv
6
C
onv2d
(
30×
30×
160
)
C
onca
t
(
30×
30×
320)
O
ut
pu
t
(
30×
30×
320)
F
igur
e
4.
I
mp
r
ove
d
M
obil
e
Ne
tv2
e
nc
ode
r
s
tr
uc
tur
e
T
o
pr
e
ve
nt
the
g
r
a
dient
f
r
om
va
nis
hing,
the
M
o
bil
e
Ne
tv2
ne
twor
k
is
de
s
igned
with
a
n
inver
ted
r
e
s
idual
s
tr
uc
tur
e
.
F
igu
r
e
5
s
hows
the
s
c
he
matic
d
iagr
a
m
of
the
inver
ted
r
e
s
idual
c
onvolut
ion
s
tr
uc
t
ur
e
whe
n
the
s
tr
ide
is
1
.
T
he
pr
oc
e
s
s
be
gins
with
1×
1
c
on
volut
ion
to
inc
r
e
a
s
e
the
dim
e
ns
ion
of
the
input
f
e
a
tur
e
s
,
f
oll
owe
d
by
3×
3
s
pa
ti
a
l
c
onvolut
ion
to
ob
tain
m
or
e
f
e
a
tur
e
inf
or
mati
on,
a
nd
f
inally
c
onc
ludes
wi
th
a
1×
1
point
-
wis
e
c
onvolut
ion
to
r
e
duc
e
the
dim
e
ns
ion
a
nd
pe
r
f
or
m
f
e
a
tur
e
c
omp
r
e
s
s
ion.
T
his
method
e
f
f
e
c
ti
ve
ly
r
e
duc
e
s
the
number
of
pa
r
a
mete
r
s
by
f
ir
s
t
inc
r
e
a
s
ing
a
nd
then
de
c
r
e
a
s
ing
the
di
mens
ion.
I
n
a
ddit
io
n,
a
t
the
e
nd
of
the
inver
ted
r
e
s
idual
s
tr
uc
tur
e
,
a
li
ne
a
r
a
c
t
ivation
f
unc
ti
on
is
us
e
d
ins
tea
d
of
the
tr
a
dit
ional
r
e
c
ti
f
ied
li
ne
a
r
unit
6
(
R
e
L
U6
)
a
c
ti
va
ti
on
laye
r
,
whic
h
a
voi
ds
the
los
s
o
f
inf
or
mation
c
a
us
e
d
by
the
a
c
ti
va
ti
on
f
unc
ti
on
in
lowe
r
dim
e
ns
ions
a
nd
im
pr
ove
s
the
p
e
r
f
or
ma
nc
e
of
the
ne
twor
k.
F
u
r
ther
mor
e
,
the
c
ombi
na
ti
o
n
of
thi
s
inver
ted
r
e
s
idual
s
tr
uc
tur
e
with
de
pthwis
e
s
e
pa
r
a
ble
c
onvolut
ions
f
ur
ther
e
nha
nc
e
s
c
omput
a
ti
ona
l
e
f
f
icie
nc
y
a
nd
r
e
duc
e
s
model
s
ize
,
making
M
obil
e
Ne
tv2
high
ly
s
uit
a
ble
f
or
e
c
hoc
a
r
diogr
a
phic
s
e
g
menta
ti
on
tas
ks
.
T
his
pa
pe
r
f
oc
us
e
s
on
the
s
hor
t
-
a
xis
2D
s
li
c
e
of
the
lef
t
ve
ntr
icle
in
e
c
hoc
a
r
diogr
a
ms
,
whic
h
ha
s
na
tur
a
l
s
pa
ti
a
l
or
de
r
a
nd
r
ich
low
-
dim
e
ns
ional
f
e
a
tur
e
s
.
T
o
boos
t
M
obil
e
Ne
tv2's
f
e
a
tur
e
e
xt
r
a
c
ti
on,
a
tr
ous
c
onvolut
ions
a
r
e
int
r
oduc
e
d
.
T
he
s
e
e
xpa
nd
the
c
o
nvolut
ion
ke
r
ne
ls
'
c
ove
r
a
ge
wi
thout
inc
r
e
a
s
ing
c
ompl
e
xit
y,
c
a
ptur
ing
mor
e
de
tailed
global
in
f
or
mation
.
C
on
s
ider
ing
the
s
ize
of
the
input
f
e
a
tu
r
e
map
a
nd
th
e
de
s
ign
pr
inciples
of
hyb
r
id
d
il
a
ted
c
onvolut
ion
(
HD
C
)
,
t
he
pa
pe
r
uti
li
z
e
s
a
s
e
r
ies
of
th
r
e
e
a
tr
ous
c
onvolut
i
ons
[
48]
.
T
he
s
tr
uc
tur
e
is
s
hown
in
F
igur
e
6.
T
he
dil
a
ti
on
r
a
te
(
r
)
r
e
pr
e
s
e
nts
the
s
pa
c
ing
be
twe
e
n
e
a
c
h
pixel
whe
n
the
c
onvolut
ion
ke
r
ne
l
pe
r
f
or
ms
c
onvolut
ion
c
a
lcula
ti
ons
,
a
nd
is
s
e
t
to
2,
3,
a
nd
5,
r
e
s
pe
c
ti
ve
ly.
T
his
s
e
tup
a
void
s
the
gr
iddi
ng
e
f
f
e
c
t
while
e
nha
nc
ing
inf
or
m
a
ti
on
uti
li
z
a
ti
on.
I
n
a
ddit
ion,
c
ombi
ning
M
obil
e
Ne
tv2
with
a
tr
ous
c
onvolut
ions
f
a
c
il
it
a
tes
the
c
a
ptur
e
o
f
s
ubtl
e
c
ha
nge
s
in
lef
t
ve
nt
r
icula
r
s
ha
pe
a
nd
f
unc
ti
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
ompar
ati
v
e
e
v
aluat
ion
of
left
v
e
ntr
icle
s
e
gme
ntat
ion
us
ing
impr
ov
e
d
py
r
amid
s
c
e
ne
…
(
J
in
W
ang
)
3219
I
n
p
u
t
1
×
1
C
o
n
v
,
B
N
,
R
e
L
U
6
3
×
3
D
e
p
t
h
w
i
se
C
o
n
v
,
B
N
,
R
e
L
U
6
1
×
1
C
o
n
v
,
B
N
O
u
t
p
u
t
I
n
p
u
t
(
3
0
×
3
0
×
9
6
)
A
t
r
o
u
s
_
C
o
n
v
(
3
0
×
3
0
×
1
6
0
)
r
=2
A
t
r
o
u
s
_
C
o
n
v
(
3
0
×
3
0
×
1
6
0
)
r
=3
A
t
r
o
u
s
_
C
o
n
v
(
3
0
×
3
0
×
1
6
0
)
r
=5
O
u
t
p
u
t
(
3
0
×
3
0
×
1
6
0
)
F
igur
e
5.
I
nve
r
ted
r
e
s
idual
c
onvolut
ion
s
tr
uc
tur
e
F
igur
e
6.
Atr
ous
c
onvolut
ion
s
tr
uc
tur
e
2.
3.
P
e
r
f
or
m
an
c
e
e
valu
at
io
n
T
o
mea
s
ur
e
the
a
c
c
ur
a
c
y
of
e
c
hoc
a
r
diogr
a
phic
lef
t
ve
ntr
icula
r
s
e
gmenta
ti
on,
we
us
e
d
th
r
e
e
dif
f
e
r
e
nt
metr
ics
:
Dic
e
,
Ha
us
dor
f
f
dis
tanc
e
(
HD
)
,
a
nd
int
e
r
s
e
c
ti
on
ove
r
union
(
mI
oU
)
,
a
s
r
e
por
ted
in
[
49
]
–
[
51
]
.
T
he
s
e
metr
ics
we
r
e
us
e
d
to
e
va
luate
the
s
e
gmenta
ti
on
a
c
c
ur
a
c
y.
L
e
t
U=
{u1,
u2,
.
.
.
,
um}
be
the
pr
e
diction
a
r
e
a
.
L
e
t
R
=
{r
1,
r
2,
.
.
.
,
r
m}
be
the
r
e
f
e
r
e
nc
e
a
r
e
a
.
Dic
e
is
a
mea
s
ur
e
of
the
s
im
il
a
r
it
y
be
twe
e
n
two
s
e
ts
.
I
t
e
va
luate
s
the
s
im
il
a
r
it
y
be
twe
e
n
the
ne
two
r
k
pr
e
diction
s
tr
uc
tur
e
a
nd
the
human
a
nnotation
r
e
s
u
lt
.
T
he
s
e
gmenta
ti
on
tas
k
c
las
s
if
ies
the
pixels
in
the
im
a
ge
.
S
e
t
s
im
il
a
r
it
y
e
va
luate
s
the
s
im
il
a
r
it
y
be
twe
e
n
tw
o
c
on
tour
s
,
ge
ne
r
a
ll
y
r
e
qui
r
e
s
the
index
to
be
gr
e
a
ter
than
0.
7,
a
nd
the
s
e
gmenta
ti
on
e
f
f
e
c
t
is
r
e
latively
good.
Dice
=
2
|
∩
|
|
|
+
|
|
(
1)
HD
is
the
maximum
dis
tanc
e
f
r
om
one
s
e
t
to
the
ne
a
r
e
s
t
point
in
a
nother
s
e
t.
Nota
bly,
thi
s
dis
tanc
e
is
dir
e
c
ti
ona
l;
s
pe
c
if
ica
ll
y,
h
(
U,
R
)
is
not
e
qua
l
to
h
(
R
,
U)
.
H
take
s
the
lar
ge
r
of
the
two
dis
tanc
e
s
.
F
or
pa
r
a
mete
r
s
that
a
r
e
s
e
ns
it
ive
to
di
f
f
e
r
e
nc
e
s
i
n
lo
c
a
ti
on
inf
or
mation,
the
s
maller
the
va
lue,
the
hi
ghe
r
the
de
gr
e
e
of
r
e
pe
ti
ti
on
.
T
he
c
a
lcula
ti
on
f
or
mul
a
is
a
s
f
oll
ows
:
HD
=
max[
h
(
U,
R
)
,
h
(
R
,
U)
]
,
s
e
t
U=
{u
1
,u
2
,
.
.
.
,
u
m
}
,
R
=
{r
1
,r
2
,
.
.
.
,
r
m
}
,
whe
r
e
h
u
s
e
d
to
c
a
lcula
te
the
one
-
wa
y
Hough
dis
tanc
e
be
twe
e
n
two
s
ur
f
a
c
e
s
.
ℎ
(
R,
U
)
=
∈
{
∈
‖
u
-
r
‖
}
(
2
)
ℎ
(
U
,
R
)
=
∈
{
∈
‖
r
-
u
‖
}
(
3
)
mI
oU
is
the
a
ve
r
a
ge
o
f
the
int
e
r
s
e
c
ti
on
a
nd
uni
on
r
a
ti
os
a
c
r
os
s
two
c
a
tegor
ies
:
he
a
r
t
a
r
e
a
a
nd
ba
c
kgr
ound
a
r
e
a
.
I
nter
s
e
c
ti
on
ove
r
union
(
I
oU
)
i
s
us
e
d
to
mea
s
ur
e
the
ove
r
lapping
a
r
e
a
of
e
a
c
h
c
a
tegor
y,
Io
U
=
int
e
r
s
e
c
ti
on
a
r
e
a
of
a
c
e
r
tain
c
a
tegor
y/uni
on
a
r
e
a
of
a
c
e
r
tain
c
a
tegor
y
.
mI
oU
is
then
c
omput
e
d
a
s
the
s
um
of
the
I
oUs
of
a
ll
c
a
tegor
ies
divi
de
d
by
the
nu
mber
of
c
a
tegor
ies
.
m
I
o
U
=
1
2
×
(
ff
+
bf
+
bb
+
fb
)
(
4
)
A
mong
them,
n
ff
r
e
p
r
e
s
e
nts
the
number
o
f
c
o
r
r
e
c
tl
y
c
las
s
if
ied
f
or
e
gr
ound
pixels
,
t
f
r
e
pr
e
s
e
nts
the
nu
mber
of
pixels
be
longi
ng
to
the
f
o
r
e
gr
ound,
n
bf
r
e
pr
e
s
e
nts
the
number
o
f
inco
r
r
e
c
tl
y
c
las
s
if
ied
ba
c
kgr
ound
pi
xe
ls
,
n
bb
r
e
pr
e
s
e
nts
the
number
o
f
c
or
r
e
c
tl
y
c
las
s
if
ied
ba
c
k
gr
ound
pixels
,
t
b
r
e
p
r
e
s
e
nts
the
numbe
r
o
f
pixels
b
e
longi
ng
to
the
ba
c
kgr
ound,
a
nd
t
he
number
o
f
n
fb
r
e
pr
e
s
e
nt
s
the
number
o
f
mi
s
c
las
s
if
ied
f
or
e
gr
ound
pixels
.
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
his
s
tudy
r
e
plac
e
d
di
f
f
e
r
e
nt
ba
c
kbone
f
e
a
tur
e
ne
t
wor
ks
ba
s
e
d
on
the
tr
a
dit
ional
P
S
P
Ne
t
a
lgor
it
hm
a
nd
s
e
lec
ted
the
c
onvolut
ion
-
ba
s
e
d
R
e
s
Ne
t50
ba
c
kbone
ne
twor
k
a
nd
the
im
p
r
ove
d
li
ghtwe
ig
ht
ne
ur
a
l
ne
twor
k
M
ob
il
e
Ne
tv2
to
tr
a
in
the
s
e
gmenta
ti
on
model.
T
wo
methods
we
r
e
us
e
d
to
ini
ti
a
li
z
e
the
we
ight
s
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
202
5
:
321
4
-
3227
3220
the
model:
lef
t
ve
ntr
icula
r
e
c
hoc
a
r
diogr
a
phy
pr
e
-
tr
a
ini
ng
a
nd
na
tu
r
a
l
im
a
ge
tr
a
ns
f
e
r
.
T
he
im
pa
c
t
o
f
the
two
ini
ti
a
li
z
a
ti
on
methods
of
pr
e
-
tr
a
ini
ng
a
nd
tr
a
ns
f
e
r
lea
r
ning
on
model
pe
r
f
or
manc
e
indi
c
a
tor
s
wa
s
c
ompar
a
ti
ve
ly
a
na
lyze
d
thr
ough
e
xpe
r
im
e
nts
.
T
he
e
xpe
r
im
e
nt
is
c
onduc
ted
on
Ka
ggle
us
ing
th
e
P
yT
or
c
h
de
e
p
lea
r
n
ing
f
r
a
mew
or
k.
T
he
ini
t
ial
lea
r
ning
r
a
te
is
0.
01
,
a
nd
the
mi
nim
um
lea
r
ning
r
a
te
is
0
.
0001.
T
he
opti
mi
z
e
r
e
mpl
oys
s
tocha
s
ti
c
gr
a
dient
de
s
c
e
nt
(
S
GD
)
with
a
mom
e
ntum
pa
r
a
mete
r
of
0.
9.
T
he
ba
tch
s
ize
is
s
e
t
to
8,
a
nd
a
c
os
ine
lea
r
ning
r
a
te
s
tr
a
tegy
is
s
e
lec
ted.
T
he
we
ight
de
c
a
y
is
c
onf
igur
e
d
a
s
0.
0001
.
All
thes
e
pa
r
a
mete
r
s
e
tt
ings
a
r
e
c
hos
e
n
ba
s
e
d
on
tr
a
c
king
th
e
model
t
r
a
ini
ng
pr
oc
e
s
s
to
e
nha
nc
e
s
e
gmenta
ti
on
pe
r
f
or
manc
e
.
M
ode
l
s
e
gmenta
ti
on
pe
r
f
or
manc
e
e
va
luation
is
done
thr
ough
a
na
lys
is
of
the
E
poc
h_los
s
c
ur
ve
a
nd
E
poc
h_M
iou
f
o
r
both
the
tr
a
ini
ng
s
e
t
a
nd
va
li
da
ti
on
s
e
t.
Dur
i
ng
t
he
mo
de
l
tr
a
ini
n
g
p
r
oc
e
s
s
,
the
pr
e
-
t
r
a
in
e
d
model
wa
s
f
ir
s
t
uti
li
z
e
d
to
ini
ti
a
l
ize
t
he
we
i
ghts
o
f
the
ba
c
kbo
ne
f
e
a
tur
e
e
x
tr
a
c
ti
o
n
ne
two
r
k
o
f
the
s
e
gmenta
ti
o
n
mode
l.
T
he
im
a
ge
s
in
the
da
ta
s
e
t
a
r
e
s
tor
e
d
i
n
VO
C
f
o
r
mat
a
nd
a
ll
i
mage
s
a
r
e
r
e
s
iz
e
d
un
if
or
m
ly
t
o
473
×
4
73
.
T
he
da
ta
s
e
t
is
d
ivi
de
d
i
n
to
a
t
r
a
i
nin
g
s
e
t
a
nd
a
va
li
da
ti
o
n
s
e
t
a
c
c
or
d
ing
to
the
r
a
t
io
o
f
9
:1
.
W
he
n
R
e
s
Ne
t50
is
s
e
lec
ted
a
s
the
ba
c
kb
one
n
e
two
r
k
,
t
he
e
poc
h=
30
a
nd
e
poc
h=
50
mode
l
t
r
a
i
ning
r
e
s
u
lt
s
(
E
poc
h_
los
s
a
nd
E
poc
h_M
iou
)
a
r
e
s
how
n
in
F
i
gu
r
e
s
7(
a
)
a
nd
7
(
b
)
.
W
he
n
the
ba
c
kbone
ne
t
wor
k
us
e
s
the
im
pr
ove
d
M
obil
e
Ne
tv2,
the
e
poc
h=
30
a
nd
e
poc
h=
50
model
tr
a
ini
n
g
r
e
s
ult
s
(
E
poc
h_los
s
a
nd
E
poc
h_M
iou)
a
r
e
s
hown
in
F
igu
r
e
s
8(
a
)
a
nd
8
(
b)
.
T
he
r
e
s
ult
a
nt
diagr
a
m
s
hows
th
a
t
whe
n
M
obil
e
Ne
tV2,
whic
h
incor
por
a
tes
dil
a
ted
c
onvol
uti
on
a
nd
f
e
a
tur
e
f
us
io
n,
is
us
e
d
a
s
the
ba
c
kbone
f
e
a
tur
e
e
xtr
a
c
ti
on
ne
twor
k
,
the
s
e
gmenta
ti
on
model
a
c
hie
ve
s
c
onve
r
ge
nc
e
withi
n
30
e
poc
hs
.
T
h
is
is
e
quival
e
nt
to
the
tr
a
ini
ng
e
f
f
e
c
t
of
50
e
poc
hs
whe
n
R
e
s
Ne
t50
is
us
e
d
a
s
the
ba
c
kbone
ne
two
r
k.
Additi
ona
ll
y
,
a
h
igh
e
r
mea
n
M
iou
va
lu
e
is
a
c
hieve
d
withi
n
5
e
poc
hs
.
T
he
s
e
r
e
s
ult
s
de
mons
tr
a
te
s
igni
f
ica
nt
pe
r
f
o
r
manc
e
a
dva
ntage
s
a
nd
f
a
s
ter
c
onve
r
ge
nc
e
c
a
pa
bil
it
ies
.
(
a
)
(
b)
F
igur
e
7.
R
e
s
Ne
t50
ba
c
kbone
ne
twor
k
f
o
r
(
a
)
E
po
c
h=
30
a
nd
(
b)
E
poc
h=
50
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
ompar
ati
v
e
e
v
aluat
ion
of
left
v
e
ntr
icle
s
e
gme
ntat
ion
us
ing
impr
ov
e
d
py
r
amid
s
c
e
ne
…
(
J
in
W
ang
)
3221
(
a
)
(
b)
F
igur
e
8
.
M
obil
e
Ne
tv2
ba
c
kbone
ne
twor
k
f
or
(
a
)
E
poc
h=
30
a
nd
(
b)
E
poc
h=
50
T
o
f
ur
ther
e
xplo
r
e
the
s
pe
c
if
ic
im
pa
c
t
o
f
two
dif
f
e
r
e
nt
we
ight
ini
ti
a
li
z
a
ti
on
s
c
he
mes
on
the
pe
r
f
or
manc
e
of
the
s
e
gmenta
ti
on
model
,
the
in
it
ializa
ti
on
method
of
t
r
a
ns
f
e
r
lea
r
ning
wa
s
the
n
us
e
d.
S
pe
c
if
ica
ll
y,
the
model
we
ight
s
obtaine
d
by
pr
e
-
tr
a
ini
ng
on
a
wide
r
a
nge
o
f
na
tur
a
l
im
a
ge
da
tas
e
ts
we
r
e
us
e
d
to
c
onduc
t
f
ine
-
we
ight
ini
ti
a
li
z
a
ti
on
o
f
the
ba
c
kb
one
f
e
a
tur
e
e
xtr
a
c
ti
on
ne
twor
k
in
the
s
e
gmenta
ti
on
model.
T
he
a
bove
model
tr
a
ini
ng
e
xpe
r
im
e
nts
we
r
e
c
onti
nue
d
to
be
r
e
pe
a
ted
us
ing
both
R
e
s
Ne
t50
a
nd
the
im
pr
ove
d
M
obil
e
Ne
tv2
ba
c
kbone
ne
twor
ks
,
a
nd
whe
n
the
b
a
c
kbone
ne
twor
k
s
e
lec
ts
R
e
s
Ne
t50
with
e
poc
hs
s
e
t
to
30
,
the
r
a
ndom
ini
ti
a
li
z
a
ti
on
a
nd
t
r
a
ns
f
e
r
ini
ti
a
l
iza
ti
on
model
tr
a
ini
ng
r
e
s
ult
s
(
E
poc
h_los
s
a
nd
E
poc
h_M
iou)
a
r
e
s
hown
in
F
igur
e
s
9(
a
)
a
nd
9(
b
)
.
W
he
n
the
ba
c
kbon
e
ne
twor
k
us
e
s
the
im
pr
ove
d
M
obil
e
Ne
tv2
with
e
poc
hs
s
e
t
to
30
,
the
r
a
ndom
ini
ti
a
li
z
a
ti
on
a
nd
t
r
a
ns
f
e
r
ini
ti
a
li
z
a
ti
on
model
t
r
a
ini
ng
r
e
s
ult
s
(
E
poc
h_l
os
s
a
nd
E
poc
h_M
io
u)
a
r
e
s
hown
in
F
igur
e
s
10
(
a
)
a
nd
10(
b
)
.
I
t
c
a
n
be
obs
e
r
ve
d
f
r
om
the
tr
a
ini
ng
r
e
s
ult
s
of
s
e
g
menta
ti
on
models
ini
ti
a
li
z
e
d
with
dif
f
e
r
e
nt
we
ight
s
that
the
P
S
P
Ne
t
s
e
gmenta
ti
on
model
bui
lt
us
in
g
R
e
s
Ne
t50
a
s
the
ba
c
kbone
ne
twor
k
ha
s
s
ign
if
ica
ntl
y
im
pr
ove
d
model
pe
r
f
or
manc
e
,
including
c
onve
r
g
e
nc
e
s
pe
e
d
a
nd
s
e
gmenta
ti
on
a
c
c
ur
a
c
y,
s
uppor
ted
by
two
ini
ti
a
li
z
a
ti
on
s
c
he
mes
:
pr
e
-
tr
a
ini
ng
a
nd
tr
a
ns
f
e
r
le
a
r
ning.
How
e
ve
r
,
whe
n
the
im
p
r
ove
d
M
obil
e
Ne
tv2
is
us
e
d
a
s
the
ba
c
kbone
ne
twor
k,
the
ini
ti
a
li
z
a
ti
on
method
ha
s
li
tt
le
e
f
f
e
c
t
on
the
pe
r
f
or
manc
e
a
nd
c
onve
r
ge
n
c
e
s
pe
e
d
of
the
s
e
gmenta
ti
on
model.
E
ve
n
with
r
a
ndom
i
nit
ializa
ti
on,
it
outper
f
or
ms
the
R
e
s
Ne
t50
-
ba
s
e
d
P
S
P
Ne
t,
a
c
hieving
s
im
il
a
r
pe
r
f
or
manc
e
with
30
e
poc
hs
c
ompar
e
d
to
50
f
or
R
e
s
Ne
t50.
Due
to
it
s
li
ghtwe
ight
de
s
ign,
th
e
im
pr
ove
d
M
obil
e
Ne
tv2
-
ba
s
e
d
P
S
P
Ne
t
r
e
duc
e
s
ove
r
a
ll
r
unning
ti
me
by
40%
c
ompar
e
d
to
the
R
e
s
Ne
t50
-
ba
s
e
d
model.
T
o
ve
r
if
y
the
s
e
gmenta
ti
on
model's
s
upe
r
ior
it
y,
T
a
ble
1
s
umm
a
r
ize
s
the
pe
r
f
or
manc
e
of
VG
G
[
41]
,
Une
t
[
42]
,
R
e
s
_U
[
16]
,
a
nd
im
p
r
ove
d
P
S
P
Ne
t
in
l
e
f
t
ve
ntr
icula
r
e
c
hoc
a
r
diogr
a
phy
s
e
gmenta
ti
on,
f
oc
us
ing
on
s
e
gmenta
ti
on
a
c
c
ur
a
c
y
(
a
s
s
e
s
s
e
d
by
Dic
e
,
HD
,
m
I
oU
)
,
pr
oc
e
s
s
ing
ti
me
(
a
r
e
a
l
-
ti
me
pe
r
f
or
manc
e
me
a
s
ur
e
)
,
a
nd
s
e
gmenta
ti
on
e
f
f
e
c
t
(
dis
playe
d
vis
ua
ll
y)
.
Not
a
ti
ons
'T
'
a
nd
'Q'
de
note
p
r
e
-
tr
a
ini
ng
a
nd
tr
a
ns
f
e
r
lea
r
ning
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
202
5
:
321
4
-
3227
3222
ini
ti
a
li
z
a
ti
ons
,
r
e
s
pe
c
ti
ve
ly.
P
S
P
Ne
t_R
,
P
S
P
Ne
t
_M
,
a
nd
P
S
P
Ne
t_M
K
us
e
R
e
s
Ne
t,
M
obil
e
Ne
tv2,
a
nd
im
pr
ove
d
M
obil
e
Ne
tv2
a
s
ba
c
kbone
s
.
(
a
)
(
b)
F
igur
e
9
.
R
e
s
N
e
t50
ba
c
kbone
ne
twor
k
f
o
r
(
a
)
r
a
nd
om
ini
ti
a
li
z
a
ti
on
a
nd
(
b)
tr
a
ns
f
e
r
ini
ti
a
li
z
a
ti
on
T
he
R
e
s
_U
a
lgor
it
hm
c
ombi
ne
s
R
e
s
Ne
t
a
nd
Une
t
to
opti
mi
z
e
the
ne
twor
k
s
tr
uc
tur
e
o
f
the
f
e
a
tur
e
e
xtr
a
c
ti
on
pa
r
t
a
nd
c
a
ptur
e
mo
r
e
e
f
f
e
c
ti
ve
f
e
a
tur
e
s
that
a
r
e
be
ne
f
icia
l
to
s
e
gmenta
ti
on.
C
ompar
e
d
with
the
c
las
s
ic
a
lgor
it
hms
Une
t
a
nd
VG
G,
the
s
e
gmenta
ti
on
e
f
f
e
c
t
is
be
tt
e
r
,
with
Dic
e
r
e
a
c
hing
83%
a
nd
mI
oU
r
e
a
c
hing
84%
.
T
he
pa
pe
r
s
tudi
e
s
the
P
S
P
Ne
t_M
K
a
lgor
it
hm
,
whic
h
opti
m
ize
s
the
e
nc
oding
pa
r
t
of
P
S
P
Ne
t's
M
obil
e
Ne
tv2
by
incor
po
r
a
ti
ng
a
li
ghtwe
ight
ne
tw
or
k
model
.
Addi
ti
ona
ll
y,
it
int
r
oduc
e
s
a
tr
ous
c
onv
olut
ion
to
int
e
gr
a
te
c
ontextua
l
inf
or
mation
,
ther
e
by
a
c
quir
in
g
r
iche
r
global
inf
o
r
mation
a
nd
a
c
hieving
the
hig
he
s
t
Dic
e
a
nd
mI
oU
va
lues
.
R
e
s
_U
of
f
e
r
s
a
s
li
ght
a
c
c
ur
a
c
y
e
dge
bu
t
the
f
a
s
tes
t
pr
e
diction,
ba
lanc
ing
s
e
gmenta
ti
on
a
c
c
ur
a
c
y
a
nd
e
f
f
icie
nc
y.
Nota
bly,
P
r
e
-
tr
a
ini
ng
a
nd
tr
a
ns
f
e
r
lea
r
ning
e
nha
nc
e
P
S
P
Ne
t_R
's
pe
r
f
or
ma
nc
e
,
with
Dic
e
incr
e
a
s
ing
by
3.
7
%
,
HD
de
c
r
e
a
s
ing
by
0
.
6%
,
a
nd
m
I
oU
incr
e
a
s
ing
by
3.
6
%
.
W
he
n
ini
ti
a
li
z
e
d
with
the
s
a
me
we
ight
s
,
M
obil
e
Ne
tv2
r
uns
s
igni
f
ica
ntl
y
f
a
s
ter
but
a
c
hieve
s
lowe
r
Dic
e
(
-
2.
5%
)
a
nd
mI
oU
(
-
1.
2%
)
c
ompar
e
d
to
R
e
s
Ne
t50.
B
y
e
nha
nc
in
g
M
obil
e
Ne
tv2
with
a
tr
ous
c
onvolut
ion,
f
e
a
tur
e
f
us
ion,
a
nd
a
n
i
mpr
ove
d
ba
c
kbone
,
s
e
gmenta
ti
on
pe
r
f
or
manc
e
im
pr
ove
d
ov
e
r
R
e
s
Ne
t
with
Dic
e
+
3.
7%
,
m
I
oU
+
3.
6%
,
a
nd
H
D
-
0.
6%
.
T
his
a
c
hieve
d
opti
mal
s
e
gmenta
ti
on
without
pe
r
f
o
r
manc
e
los
s
a
nd
r
e
duc
e
d
pr
oc
e
s
s
ing
ti
m
e
by
33
.
4%
.
R
e
a
l
-
ti
me
e
c
hoc
a
r
diogr
a
phic
lef
t
ve
nt
r
icula
r
s
e
gmenta
ti
on
is
e
xtr
e
mely
c
ha
ll
e
nging
due
to
a
r
ti
f
a
c
ts
a
nd
s
pe
c
kle
nois
e
in
im
a
ge
s
.
F
igur
e
11
p
r
ovides
a
qua
li
tative
vis
ua
l
c
ompar
is
on
of
di
f
f
e
r
e
nt
im
a
ge
qua
li
ti
e
s
,
their
c
or
r
e
s
ponding
s
e
gmenta
ti
on
m
a
s
ks
,
a
nd
pr
e
d
iction
r
e
s
ult
s
.
T
he
R
e
s
_U
model
a
nd
the
im
p
r
ove
d
P
S
P
Ne
t
model
with
M
obil
e
Ne
tV2
a
s
the
ba
c
kbone
ne
tw
or
k
a
c
hieve
s
a
ti
s
f
a
c
tor
y
s
e
gmenta
ti
on
r
e
s
ult
s
in
the
lef
t
ve
ntr
icula
r
r
e
gion
of
e
c
hoc
a
r
diogr
a
phy
.
I
t
c
los
e
ly
matc
he
s
the
s
e
gment
a
ti
on
mas
k
in
b
ounda
r
y
a
c
c
ur
a
c
y
a
nd
a
r
e
a
ove
r
lap,
e
s
pe
c
ially
outper
f
o
r
mi
ng
c
las
s
ic
VG
G
a
nd
Une
t
models
in
im
a
ge
s
with
a
r
ti
f
a
c
ts
,
n
ois
e
,
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
ompar
ati
v
e
e
v
aluat
ion
of
left
v
e
ntr
icle
s
e
gme
ntat
ion
us
ing
impr
ov
e
d
py
r
amid
s
c
e
ne
…
(
J
in
W
ang
)
3223
blur
r
e
d
a
na
tom
ica
l
bounda
r
ies
.
I
mpr
ove
d
P
S
P
Ne
t
,
s
tudi
e
d
in
thi
s
pa
pe
r
,
r
uns
f
a
s
tes
t,
while
VG
G
is
s
lowe
s
t.
C
ompar
e
d
to
R
e
s
_U,
P
S
P
Ne
t
s
hor
tens
c
a
lcula
ti
on
ti
me
by
33
.
4%
f
o
r
the
s
a
me
r
e
s
ult
s
,
ba
lanc
ing
s
e
gmenta
ti
on
pe
r
f
or
manc
e
a
nd
s
pe
e
d,
with
good
c
l
ini
c
a
l
potential.
(
a
)
(
b)
F
igur
e
10
.
M
obil
e
N
e
tv2
ba
c
kbone
ne
twor
k
f
or
(
a
)
r
a
ndom
ini
ti
a
li
z
a
ti
on
a
nd
(
b
)
t
r
a
ns
f
e
r
ini
ti
a
li
z
a
ti
on
T
a
ble
1.
C
ompar
is
on
of
s
e
gmenta
ti
on
r
e
s
ult
s
unde
r
dif
f
e
r
e
nt
c
onf
igur
a
t
ions
of
P
S
P
Ne
t
s
e
gmenta
ti
on
ne
twor
k
S
e
gme
nt
a
ti
on
a
r
c
hi
te
c
tu
r
e
D
ic
e
HD
mI
oU
P
r
oc
e
s
s
in
g
t
ime
(
s
)
VGG
VGG
0.77
4.65
0.76
34.6
V
GG
_T
0.79
4.61
0.79
33.2
V
GG
_Q
0.79
4.63
0.78
33.7
U
ne
t
U
ne
t
0.78
4.69
0.76
27.4
U
n
et
_T
0.80
4.65
0.80
26.5
U
n
et
_Q
0.79
4.67
0.78
26.8
R
e
s
_U
R
e
s
_U
0.81
4.62
0.82
30.2
R
e
s
_U
_T
0.83
4.59
0.84
28.6
R
e
s
_U
_Q
0.82
4.60
0.83
29.4
P
S
P
N
e
t
P
S
P
N
e
t_
R
0.79
4.61
0.82
27.8
P
S
P
N
e
t_
R
_
T
0.82
4.59
0.85
28.5
P
S
P
N
e
t_
R
_
Q
0.81
4.58
0.84
29.1
P
S
P
N
e
t_
M
0.77
4.55
0.81
19.6
P
S
P
N
e
t_
M
_
T
0.80
4.60
0.82
19.8
P
S
P
N
e
t_
M
_
Q
0.81
4.59
0.82
21.3
P
S
P
N
e
t_
M
K
0.82
4.52
0.85
18.5
P
S
P
N
e
t_
M
K
_
T
0.84
4.51
0.86
19.4
P
S
P
N
e
t_
M
K
_
Q
0.83
4.50
0.85
20.2
Evaluation Warning : The document was created with Spire.PDF for Python.