TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4563 ~ 4
5
7
1
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.539
3
4563
Re
cei
v
ed
De
cem
ber 2
8
, 2013; Re
vi
sed
F
ebruary 25,
2014; Accept
ed March 1
2
, 2014
Locating Liver Lesion with Local C-V Level S
e
t and
Image Registration
Zhaohui Luo
*
，
Xi Zaifang
，
Wang J
unn
ian
S
c
hoo
l
o
f
In
fo
rma
t
i
o
n
an
d
Electrical
Engi
nee
ri
ng
, Hu
nan
Unive
r
si
t
y
o
f
Scie
nce
an
d
T
e
chno
logy
Xi
a
ngt
a
n
, C
h
i
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: L
u
o
z
h
a
o
h
u
i
0
1
@
s
i
n
a
.
co
m
A
b
st
r
a
ct
In this p
a
p
e
r, w
e
prop
osed
c
o
mpre
hens
ive
meth
ods
to l
o
c
a
te the
liv
er le
sion
in
multi-p
hase
C
T
imag
es. It
first construct a local liver l
e
sio
n
ima
ge fr
o
m
the imag
e in w
h
ich liver les
i
on
differs from liv
e
r
tissue most ma
rkedly,
th
en pr
e-seg
m
ent
the
lesio
n
w
i
th
OT
SU
meth
od to
get the
in
itial
c
ontour,
an
d ev
olve
the active c
ont
our w
i
th loc
a
l
C-V lev
e
l set
meth
od to
get
t
he fin
a
l co
ntou
r of lesi
on i
n
th
e CT
i
m
a
ge. F
i
nall
y
locate th
e l
i
ve
r lesi
on i
n
CT
imag
es of ot
her p
has
es by
imag
e reg
i
stration. Exp
e
ri
ments show
e
d
thi
s
meth
od ca
n ex
tract liver tumo
r efficiently. A w
e
ll-pre
pare
d
abstract en
abl
es the rea
der to ide
n
tify the b
a
si
c
content of a d
o
cu
me
nt quickl
y and accur
a
tely, to dete
rmi
ne its relev
anc
e to their inter
e
sts, and thus
to
deci
de w
hether
to read the do
cument in its e
n
tirety.
Ke
y
w
ords
: loc
a
l lev
e
l set, OTSU meth
od, i
m
age re
gister
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Hep
a
tocellula
r carcino
m
a
(HCC) is one
of the mo
st
common
malig
nan
cie
s
in th
e world,
with ap
proxi
m
ately 1,000,
000
ca
se
s re
ported
every
year. X-ray computed
tom
ogra
phy (CT
)
is
one of sen
s
it
ive imaging
modalitie
s for the liver tu
mor analy
s
is [
1
]. Spiral CT, with the recent
introdu
ction
o
f
multi detect
o
r-
ro
w sca
n
techn
o
l
ogy (MDCT),
curre
n
tly plays a f
undam
ental
role
in the diagn
o
s
is a
nd sta
g
in
g of HCC.
Multi-ph
ase h
e
lical
CT i
s
the mo
st suit
able te
chniq
u
e
. It usually inclu
d
e
s
the
plain CT
image
s
and
enh
an
ced
CT im
age
s,
whi
c
h
can
incre
a
se th
e dete
c
tion
and i
m
prove
the
cha
r
a
c
teri
zati
on of focal liver lesi
on
s by usin
g co
ntra
st agents.
a quad
ru
ple-pha
se p
r
oto
c
ol that inclu
d
e
s u
nenh
an
ced, hepati
c
a
r
terial, p
o
rtal
venou
s,
and
delaye
d
pha
se im
age
s. It provide
s
dynamic info
rm
ation of
the
bloo
d
sup
p
ly of liver le
sio
n
s.
The differe
nt blood
sup
p
ly to the lesio
n
, in fact,
is the
most impo
rta
n
t CT feature
that may help
differentiate among small
hepatocellul
a
r
lesi
on
s
that have eme
r
ged in a ci
rrhotic liver t
h
e
multipha
se e
x
amination [2
-4].
Ho
wever, it a
l
so cost
s the
m
more b
u
rd
ens b
e
cau
s
e
of the incre
a
sin
g
amo
unt
of data
they need to interpret [1]. Rece
ntly, computer-a
id
ed
diagno
sis (CAD), defined
as a diag
no
sis
introdu
ce
d by a radiolo
g
ist
who u
s
e
s
the
output
from a comp
uteri
z
ed analy
s
is of
medical im
a
ges
in dete
c
ting l
e
sio
n
s, a
s
se
ssing
ex
tent of dise
ase, and
makin
g
di
a
g
nosti
c de
ci
sio
n
s i
s
bei
ng u
s
ed
to reduc
e
the
burdens
[5].
In a liver
ca
n
c
er CA
D
syst
em, the first
step i
s
to locate the liver les
i
ons
,
i.e. extrac
t the
regio
n
of interest (ROI
) for further an
alysis
. A su
cce
s
sful CA
D system depen
ds on the co
rre
ct
segm
entation
of liver lesio
n
s. In abdo
m
i
nal CT im
a
g
es, there are
many orga
n
s
su
ch a
s
h
e
art,
stoma
c
h, an
d sple
en b
e
sides live
r
, wh
ich ma
ke
the
image
s more com
p
licate
d
. On the ot
her
hand, e
a
ch p
a
tient is
uniq
ue, so
the
sh
ape a
nd fe
at
ure
of the liv
er i
s
dive
rsifi
ed. Wh
at’ mo
re,
most of liver lesio
n
s
diverse in
different
pha
se
CT im
age
s [3]. The
s
e a
c
cou
n
t for the h
a
rd
ne
ss of
segm
ent the liver lesio
n
an
d the
intere
sts of many re
searche
r
s.
There a
r
e m
any app
ro
aches fo
r
seg
m
entation
in
m
edical ima
g
e
s
, such a
s
th
reshold,
conto
u
r
ba
se
d techniqu
es,
regi
on b
a
sed
tech
nique
s,
clu
s
terin
g
, an
d template
m
a
tchin
g
. Each
of
these
app
ro
a
c
he
s h
a
s its
advantag
es
and di
sa
dvan
t
ages in te
rm
s of a
ppli
c
ab
ility, suitability,
perfo
rman
ce,
and
comp
utational
co
st [5, 6]. Acti
ve conto
u
rs hav
e bee
n exten
s
ively studi
ed
and
widely u
s
e
d
i
n
medi
cal im
age
seg
m
ent
ation, pa
rticul
arly to lo
cate
boun
da
ries,
whe
r
e
an initi
a
l
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4563 – 4
571
4564
conto
u
r i
s
de
formed to
wa
rds the b
oun
d
a
ry of t
he obj
ect to dete
c
ted by the mi
nimizin
g
ene
rg
y
function .But
the pa
ram
e
tric
deform
able m
odel
s have difficulties
with
segm
entation
of
topologi
cally
compl
e
x stru
cture
s
, To
o
v
erco
me the
s
e p
r
obl
em,
the level set approa
ch
was
introdu
ce
d by S.Osher a
n
d
J.A.Sethian [7]. They m
odel the prop
ag
ating cu
rve a
s
a sp
ecifi
c
le
vel
set of a highe
r dimen
s
io
nal
surfa
c
e [8].
By studying the abdomi
nal CT ima
ges,
we find that although the ima
ges a
r
e
compli
cate
d, insid
e
the regi
on of liver, th
e image
s a
r
e
relatively sim
p
le. In one of
the qua
dru
p
l
e
-
pha
se CT image
s, the lesio
n
ha
s di
fferent CT v
a
lue from
no
rmal liver tissue a
nd
can
be
disting
u
ished
by eyesig
ht. In this p
ape
r,
we p
r
op
ose a
local l
e
vel se
t algorithm
co
mbining i
m
ag
e
regi
ster meth
od to locate li
ver lesio
n
. Th
e step
s of are
as follows:
a) Sel
e
ct th
e
CT
imag
e i
n
whi
c
h
the
le
sion’
s
den
se
differs m
o
st
greatly from t
he liver
tissu
e a
nd
ca
n be
di
stingui
she
d
by
eyesi
ght in
th
e q
u
a
d
rupl
e-pha
se
CT im
age
s
a
s
the
refere
nce
image for the
segme
n
tatio
n
of liver lesio
n
;
b) In the reference image choo
se seve
ra
l point
s insi
de
the liver area
manually an
d form
a polygon
whi
c
h covers the
whole le
sion
to con
s
tru
c
t a
local le
sion i
m
age;
c) Pre-segm
e
n
t the lesion
with multileve
l OT
SU methods
to form the initial c
ontour;
d) Evolve th
e a
c
tive cont
our with
mult
ipha
se
C-V
l
e
vel set met
hod
s to
get t
he final
conto
u
r of lesion;
e) Use th
e i
m
age
regi
ste
r
metho
d
s
a
nd map th
e
area
of lesi
o
n
into othe
r
pha
se
C
T
image.
The remain
d
e
r of this
pap
er is
structu
r
ed a
s
follows. In Sec
t
ion
2, C-V level
s
e
t [7] is
introdu
ce
d bri
e
fly, in sectio
n 3, we de
scribes
the lo
cal
C-V level set
algorithm
co
mbining O
T
SU
method. Sect
ion 4 detail
s
the experim
e
n
tal pro
c
ed
ure of segm
ent
ation, data set. Result
s a
r
e
also exami
n
e
d
and di
scussed.
In sectio
n 5
we de
scrib
e
the image
reg
i
stration m
e
th
ods a
nd deta
il the corre
s
p
ondin
g
experim
ent proce
dure in se
ction 6.
The co
ncl
u
si
on is drawn in Section 7.
Finally
, possi
bilities for future work are outlined.
2. The Des
c
r
i
ption of C-V
Method
The C-V met
hod [10], inte
grating th
e le
vel set and
M
u
mford
-
Sha
h
model, do
es
not use
the gradie
n
t i
n
formatio
n. It minimi
ze
s th
e en
ergy
fu
n
c
tion
app
roa
c
h to evolutio
n
the
curve.
T
h
e
image
()
()
,,
ux
y
x
y
Î
W
is fo
rmed by t
w
o
regio
n
s:
obje
c
t (ui
)
a
nd
b
a
ckgroun
d (u
o), which i
s
sep
a
rate
d by
the evolving
cu
rve
C in
Ω
. Th
e
con
s
tants, c1 a
n
d
c2
de
pendi
n
g
on
C,
are
the
averag
es of i
m
age I insid
e
C and re
spe
c
tively out
sid
e
C. Chan a
n
d
Vese intro
d
u
ce the en
erg
y
function
al
()
E
ccC
12
,,
de
fined by:
()
()
()
()
()
12
22
11
2
2
,,
,,
io
i
uu
Ec
c
C
L
e
n
g
t
h
C
A
r
e
a
u
u
x
y
c
dx
dy
u
x
y
c
dx
dy
mu
ll
=+
+-
+
-
òò
(
1
)
Her
e
()
Le
ng
t
h
C
is the
l
ength
of the
cure
C, a
nd
()
i
A
re
a
u
is th
e a
r
e
a
of the
regi
on
inside C,
μ
,
ν
≥
0,
12
,0
ll
>
are fixe
d paramete
r
s. Therefo
r
e th
e ene
rgy function is minimi
zed if
the cu
rve is on the bo
unda
ry of the obje
c
t. O
p
timization
(1), it can
g
e
t the ultimate
segm
entation
line C, as we
ll as the location of the unknown
cc
12
,
.
Usi
ng
th
e He
aviside
fun
c
tion H(z), and
the
on
e
dim
e
nsio
nal Dirac
mea
s
u
r
e
δ
(z),
and
defined, re
sp
ectively, by:
()
()
()
1,
i
f
z
0
,
0,
i
f
z
<
0
d
H
zz
H
z
dz
d
ì
ï
³
ï
ï
==
í
ï
ï
ï
î
(
2
)
Partial differe
ntial equatio
n
s
, gotten by
Ch
a
n
and Ve
se u
s
ing Eul
e
r-Lag
ran
ge
method,
are as
follows:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Locating Li
ve
r Lesi
on with
Local C-V L
e
v
el
Set and I
m
age Regi
stration (Zh
aoh
ui Luo)
4565
()
()
(
)
()
(
)
()
22
11
2
2
0
di
v
0
,,
0
,
0
uc
uc
t
xy
xy
n
ff
df
m
u
l
l
f
ff
df
f
f
ì
éù
æö
ï
ï
÷
ç
¶Ñ
êú
÷
ï
ç
÷
=-
-
-
+
-
=
ï
êú
ç
÷
ï
ç
÷
êú
¶
ïÑ
ç
÷
ç
èø
ï
êú
ëû
ï
ï
ï
=
í
ï
ï
ï
¶
ï
ï
×=
ï
ï
Ñ
¶
ï
ï
ï
î
u
r
(3)
Whe
r
e the
c
1
，
c
2
can get from the followi
ng e
quation
s
.
()
()
()
()
()
()
1
,,
,
u
x
y
H
x
y
dxdy
c
Hx
y
d
x
d
y
f
f
f
W
W
=
ò
ò
(4)
()
()
()
()
()
()
()
()
2
,1
,
1,
u
x
y
H
x
y
dxdy
c
Hx
y
d
x
d
y
f
f
f
W
W
-
=
-
ò
ò
(
5
)
In the nume
r
i
c
al
cal
c
ulatio
ns, the regul
ariz
i
ng fun
c
ti
on (6
) is
use
d
to repl
ace
H(z),
δ
(z
)
r
e
spec
tively.
22
12
()
1
a
r
c
t
a
n
(
)
,
2
1
()
.
z
Hz
z
z
(
6
)
So that the
gradi
ent flow
Equation
(3
) rol
e
s in all
of the level
set, an
d we ca
n
automatically monitor the
empty goal with the
internal regi
on, a
nd ma
ke the
overall ene
rgy
function to th
e minimum.
Let’s di
spe
r
se the equatio
n in
f
, use a finite differen
c
es impli
c
it scheme. Re
call
first the
u
s
ua
l no
ta
tion
s
:
le
t h
b
e
th
e
s
p
ac
e
s
t
ep
,
∆
t
be th
e ti
me
step, a
n
d
(,
)
(
,
)
ij
x
yi
h
j
h
, be
the
gri
d
points. Here
1,
,
ij
M
. The
app
roxi
mation of
(,
,
)
tx
y
a
r
e set by
,
(,
,
)
n
ij
i
j
nt
x
y
. From
(3), we ca
n get
n
.The
12
()
,
(
)
nn
cc
can
be got
re
spe
c
tively by (4
), (5).
Cha
n
and Ve
se
cal
c
ulate
1
n
through (7).
()
()
()
()
nn
n
n
ij
i
j
i
j
ij
nn
n
n
i
j
ij
ij
ij
nn
ij
ij
nn
n
n
i
j
ij
ij
ij
x
y
i
j
ij
ij
i
j
th
vu
C
u
C
h
1
,,
1
,
,
1,
,
,
1
,
,1
,
1,
,
,
1
,
22
22
1
1
,,
1
,
,
2
2
,
,
2
,,
22
()
(
)
ff
f
f
m
ff
f
f
ff
m
ll
ff
f
f
+
+
++
+
++
-
-
é
æö
÷
ç
ê
÷
--
ç
÷
ê
ç
÷
ç
=D
÷
ê
ç
÷
ç
D
÷
ê
÷
ç
-+
-
÷
ç
ê
èø
ë
æö
÷
ç
÷
-
ç
÷
ç
÷
ç
+D
-
-
-
-
-
÷
ç
÷
ç
÷
÷
ç
-+
-
÷
ç
èø
n
ij
h
,
()
df
ù
ú
ú
ú
ú
ú
û
(7)
From the Eq
uation (3), we can
see, t
he def
inition
of partial di
fferential eq
u
a
tions
involving ima
ge fun
c
tion
(,
)
Ix
y
is do
main
-wi
de ma
p dat
a
,
and the
de
finition of oth
e
r two
unkno
wn
cc
12
,
is also im
age
d
e
finition of the regi
on, wi
t
h
the overall
cha
r
a
c
teri
sti
cs.
Hen
c
e,
updatin
g leve
l set function i
s
in the entire
defined
re
gio
n
, the comput
ation is very large [11].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4563 – 4
571
4566
3. Local C-V
Lev
e
l Set for the Segme
n
ta
tion of Li
v
e
r
Lesions
The
C-V m
o
d
e
l ca
n o
n
ly p
a
rtition a
n
im
age into
two
region
s, i.e. o
b
ject a
nd
ba
ckgroun
d.
While in a
n
a
bdomin
al CT
image
s, there
are
many o
r
gan
s su
ch a
s
heart, stom
a
c
h, and
sple
e
n
besi
d
e
s
liver,
whi
c
h ma
ke
the image
s m
o
re
com
p
licated an
d con
s
i
s
t of more th
an two
re
gion
s,
so the C-V le
vel set can n
o
t be applie
d for the se
gme
n
t of liver lesions di
re
ctly.
As we can se
e, the liver region is rel
a
tive
ly simple co
mpared with the whol
e abd
omina
l
CT imag
e, an
d in the local
area, the le
si
on ca
n
be ea
sily disting
u
ished from n
o
rmal liver tissu
e
,
so
we fo
cu
s our attentio
n
on the lo
cal
area.
We
sel
e
ct several
p
o
ints to form
a polygon
wh
ich
covers the
whole le
sion a
nd co
nst
r
u
c
t a local le
si
on image
of the smalle
st
size. Th
en
the
evolution of t
he
contou
r i
s
pro
c
e
s
sed
o
n
ly in t
he l
o
cal imag
es. T
h
is m
e
thod
can get
rid
of
the
interferen
ce
o
f
other organ
s. In a
ddition,
be
cau
s
e
the
size of th
e lo
cal im
age
is
much
fewer than
that of the whole abd
omina
l
CT image, the
evolution
of contou
r ca
n be sp
eed
ed
up.
As in
(4
), (5), c1
and
c2
depe
nding
o
n
C,
are the
averag
es of i
m
age I
in
sid
e
C an
d
outsid
e
C
re
spectively. A l
o
cal
imag
e in
clud
es
the m
a
rgin
al regi
on
outsi
de
th
e polygon
be
si
des
lesio
n
and liver organ. To
apply the C-V
level set to
the se
gmentat
ion of liv
er lesion, we re
defi
n
e
c2 a
s
the ave
r
age g
r
ay lev
e
l of liver regi
on, as sho
w
n
in (8).
()
()
()
()
()
()
ma
s
k
ma
s
k
ux
y
H
x
y
H
x
y
d
x
d
y
c
Hx
y
H
x
y
d
x
d
y
2
,[
1
,
]
(
,
)
[1
,
]
(
,
)
f
f
f
W
W
-
=
-
ò
ò
(
1
)
Becau
s
e of the local
cha
r
acter of this
method,
the initialization of
the level set function
s
plays an im
portant role i
n
segm
entati
on of an
image. If the initial conto
u
r is clo
s
e to the
boun
dary
of lesio
n
, the ite
r
ation
of evol
ution
can
be
lessen
ed. We apply th
e
multilevel OT
SU
method p
r
op
ose
d
by Otsu
to pre-segm
ent the
local i
m
age to get ideal initial co
ntour.
In the O
T
SU method,
onl
y the g
r
ay-le
v
el
histo
g
ra
m suffice
s
without othe
r
a pri
o
ri
kno
w
le
dge,
and the fea
s
ibility of evaluating
the
“goo
dne
ss”
of threshold
is done through
exhau
stive search to mini
mize the
with
in-cl
a
ss
vari
a
n
ce b
e
twe
en
dark an
d brig
ht regio
n
s of
the
image. Altho
ugh ma
ny works o
n
thre
shol
d metho
d
s h
a
ve bee
n pro
p
o
s
ed i
n
a num
ber of
literatures, th
e OTS
U
met
hod [9],
whi
c
h is a m
e
thod
that minimi
zes th
e
within-cla
ss varia
n
ce, is
a popul
ar no
n
-
pa
ramet
r
ic
method for its simplicity an
d efficien
cy.
In
Otsu'
s
met
hod we exha
ustively
se
arch
for
th
e thre
shol
d that mi
nimize
s th
e in
tra-cla
ss
varian
ce
(the
varian
ce
wit
h
in the
cla
ss), defin
e
d
a
s
a weighte
d
sum
of varia
n
ce
s
of the t
w
o
cla
s
s
e
s:
22
01
1
2
arg
m
i
n
{
(
)
(
)
(
)
(
)}
h
Tt
t
t
t
ws
w
s
=+
(9)
Her
e
i
w
are the
probabilities of the two
classes
separat
ed by a threshold
h
T
and
2
()
i
t
s
are varia
n
ce
s of these
cla
s
ses.
Ot
su
sho
w
s t
hat
minimi
zin
g
t
he int
r
a
-
cl
as
s
varia
n
ce
is the
same
as maximi
zi
ng inter-
cla
ss v
a
ri
an
c
e
.
To get
rid
of the inte
rferen
ce
of the m
a
rginal
a
r
e
a
of
the lo
cal im
a
ge, we only
calcul
ate
the region
in
side
the
poly
gon. T
he
opt
imal threshol
d
h
T
can be ob
tained by
mi
nimizin
g
the
within-cla
ss
varian
ce i
n
(9). We the
n
pre
-
segme
n
t
the local ima
ge in
side
the
polygon,
an
d
partition it
int
o
two
regio
n
s
, i.e. le
sion
and live
r
ti
ssue. Th
e g
r
ay
level of th
e
margi
nal
are
a
is
zero, and may be mi
staken for l
e
si
on. T
o
eliminate the impact, we
fill the marginal area
with t
he
averag
e g
r
ay
level of liver
tissu
e.
When
the g
r
ay lev
e
l of liver l
e
si
on is hig
her than live
r
tissue,
the initial level set function
is co
nst
r
u
c
ted
by (10), othe
rwi
s
e, it is co
nstru
c
ted by (11).
(
)
,,
0
(
,
)
h
x
yu
x
y
T
f
=-
(10)
(
)
,,
0
(
,
)
h
x
yT
u
x
y
f
=-
(11)
Then the leve
l set functio
n
is upd
ated a
c
cording to
(7). After severa
l iteration
s
, we get
the final co
ntour a
nd extra
c
t the lesi
on.
By mendi
ng
C-V level
set, we can g
e
t rid of the imp
a
ct
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Locating Li
ve
r Lesi
on with
Local C-V L
e
v
el
Set and I
m
age Regi
stration (Zh
aoh
ui Luo)
4567
of other o
r
ga
ns. In additio
n
, becau
se th
e si
ze of
the l
o
cal im
age i
s
much fe
we
r
than that of the
whol
e abdo
m
i
nal CT ima
g
e
, the evolution of conto
u
r can be
spe
e
ded up g
r
eatl
y
. We call the
mende
d meth
od local level set.
4. Experiments of Segm
enta
tion
We teste
d
the prop
osed method on t
w
o imag
es. T
he first one i
s
sh
own in Fi
gure 1
(
a
)
.
There i
s
a tu
mor
of hi
gh
d
ensity,
i.e
hig
her gray leve
l in th
e b
o
tto
m of the
liver. We
sele
ct four
points a
nd
co
nstru
c
t a poly
gon to be
sie
g
e
the tumor,
as sho
w
n in
Figure 1(a).
The lo
cal ima
g
e
is sh
own in Figure 2(a
)
. By using OT
SU me
thod
pre
-
segme
n
t the image first, as sh
own
in
Figure 2(b
)
. To de
cre
a
se the gradi
ent of the bound
ary of the pol
ygon, we fille
d the empty area
outsid
e
the p
o
lygon with t
he mea
n
CT
value of
the liver, as sho
w
n in Figure 2(c). We ca
n see
that the initial contou
r sho
w
n in Fig
u
re
2(d
)
is
n
ear t
he bou
nda
ry of the lesion,
but there a
r
e
some
di
ssoci
a
tive are
a
s i
n
sid
e
o
r
out
side the
tum
o
r.
Then the
contour was
e
v
olved with t
he
local
level
set
method.
The
param
ete
r
s are as
follo
ws:
time-step
0.
3
t
D=
,
850
m
=
,
12
1
ll
==
.
After 50 ite
r
at
ions,
we
got t
he final
conto
u
r, a
s
sh
own
in Figu
re
2(e). The
distri
but
ion of l
e
vel
set
is sho
w
n in Fi
gure 3 an
d the final contou
r in the
source image is sh
own in Figu
re
4, we can se
e
that the conto
u
r is ju
st on t
he bou
nda
ry of the lesion.
(a)
(b)
Figure 1. Hig
h
Den
s
ity Lesion
(a)
(b)
(c
)
(d)
(e)
Figure 2 .The
Proce
s
s and
Re
sult of the First Experi
m
ent
Figure 3. The
Distrib
u
tion o
f
Level Set of
the
First Experi
m
ent
Figure 4.The
Final Co
ntou
r Line of High
Den
s
ity Lesi
o
n
0
10
20
30
40
0
10
20
30
40
-100
-50
0
50
100
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4563 – 4
571
4568
(a)
(b)
Figure 5. Low Den
s
ity Lesi
o
n
(a)
(b)
Figure 6. The
Pre-segm
ent
ation
(a)
(b)
Figure 7. The
Contou
r Line
of Lesion
Figure 8. Level Set Distrib
u
tion after 85
Iterations
The
se
co
nd i
s
sho
w
n
in
Fi
gure
5
(
a).
Th
ere
is a
hypo
den
se tu
mor,
i.e lo
we
r g
r
a
y
level in
the centre of
the liver.
We
sele
ct five p
o
i
nts a
nd
co
nstruct
a p
o
lyg
on to
be
sieg
e the tu
mor,
as
sho
w
n i
n
Fig
u
re
5(b
)
. By usin
g OTS
U
method
pre
-
segment th
e i
m
age first, a
s
sho
w
n
in Fig
u
re
6(a
)
. To de
crease the gra
d
ient of the b
ound
ary
of the polygon, we filled the empty area ou
tside
the polygon
with the me
a
n
CT valu
e of
the liver,
as
sho
w
n in
Fig
u
re 6
(
b
)
. We
can
se
e that the
initial co
ntour sho
w
n i
n
Fi
gure
7(a) i
s
near t
he b
o
u
ndary
of the
lesio
n
, but t
here
are so
me
0
20
40
60
80
10
0
0
20
40
60
80
-1
50
-1
00
-5
0
0
50
10
0
15
0
x
y
L
e
l
v
el
s
e
t
D
i
dt
r
i
but
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Locating Li
ve
r Lesi
on with
Local C-V L
e
v
el
Set and I
m
age Regi
stration (Zh
aoh
ui Luo)
4569
disso
c
iative area
s in
side
or outsi
de the tumor.
Then the conto
u
r wa
s evolved with the loca
l
level set
met
hod. Th
e pa
ramete
rs are
as foll
ows:
time-ste
p
0.
3
t
D=
,
850
m
=
,
12
1
ll
==
.
After 85 ite
r
at
ions,
we
got t
he final
conto
u
r, a
s
sh
own
in Figu
re
7(b). The
distri
but
ion of l
e
vel
set
is sh
own in Figure 8.
W ca
n see that the
c
ontou
r is ju
st on the bou
ndary of the lesio
n
.
5. Image Re
gistra
tion fo
r Locating Li
v
e
r
Lesion in other Pha
s
e
Images
Most HCCs a
r
e hypo
den
se
or iso
den
se
whe
n
visuali
z
ed on plai
n CT image
s, as
sho
w
n
in Figure 9(b). Du
e to their predo
m
i
nant arte
rial
supply, HCC are se
en
as tran
sie
n
t
ly
hyperd
e
n
s
e
masse
s
in th
e arte
rial p
h
a
s
e of h
epati
c
enha
ncement
as
sho
w
n i
n
Figure 9(a). T
hey
become isod
ense with h
epatic p
a
re
n
c
hyma o
r
hypode
nse in the portal ve
nou
s pha
se
of
enha
ncement
. On d
e
lay
ed ima
g
e
s
, the
cap
s
u
l
e an
d
sep
t
a demo
n
st
rate p
r
olon
g
ed
enha
ncement
, wherea
s contra
st
wa
sh
-out from th
e tumor m
a
ke
s the le
si
on ag
ain ap
pea
r
hypode
nse. So the HCC
may be seen
by eyesig
ht whe
n
its d
e
n
s
e
differs fro
m
liver greatl
y
and
undete
c
ted
when its
den
se is
simila
r t
o
liver [2]. If
the liver le
sio
n
’s d
e
n
s
ity is simila
r to liv
er
tissu
e, it is dif
f
icult to lo
cat
e
the l
e
sio
n
b
y
segm
entati
on. The
only
clue
come
s from the i
m
age
of
other ph
ases. Because of influenc
e of breath, the locatio
n
of liver lesio
n
m
a
y be differe
nt
slightly. We u
s
e the imag
e regi
stratio
n
to align image
s of different phases.
Image regi
stration is the p
r
ocess of ove
r
layi
ng two o
r
more imag
e
s
of the sam
e
scene
taken
at different time
s, from differe
nt viewpoi
nt
s, a
nd/or
by different
sen
s
o
r
s.
It geometri
cally
align
s
two
im
age
s—the ref
e
ren
c
e and sensed
ima
g
e
s
. The p
r
e
s
e
n
t
differences
betwe
en ima
ges
are introdu
ce
d due to different imaging
condition
s.
Regi
stratio
n
method
s ca
n
be catego
ri
zed with
re
spect to vario
u
s criteria. T
he one
s
usu
a
lly use
d
are the a
ppli
c
ation a
r
ea,
dimen
s
io
n
a
lity of data, typ
e
and compl
e
xity of assu
med
image d
e
form
ations, comp
utational cost
, and the ess
ential idea
s o
f
the regist
rati
on algo
rithm.
In
this pa
pe
r we used A
r
ea
-based m
e
tho
d
s. Area-
ba
sed meth
od
s, sometim
e
s
called
co
rrel
a
tion-
like metho
d
s
or template
matchin
g
[12] merge t
he fe
ature dete
c
tio
n
step with th
e matchin
g
p
a
rt.
These metho
d
s de
al with
the image
s without atte
mp
ting to detect
salient o
b
je
cts. Wind
ows
of
pred
efined
si
ze or eve
n
en
tire image
s are use
d
for the
corre
s
po
nde
nce e
s
timatio
n
.
The
cla
ssi
cal
rep
r
e
s
entati
v
e of the are
a
-ba
s
e
d
met
hod
s is th
e n
o
rmali
z
e
d
Co
rrel
a
tion
ij
r
in (12) an
d its modification
s [13].
()
()
()
()
1
2
2
2
(,
)
(
,
)
(,
)
(
,
)
(,
)
(
,
)
(
,
)
(
,
)
ij
ij
xy
ij
ij
ij
xy
x
y
f
x
yf
x
y
f
x
yf
x
y
f
x
yf
x
y
f
x
yf
x
y
r
éù
éù
--
êú
êú
ëû
ëû
=
ìü
ïï
ïï
éù
éù
--
íý
êú
êú
ëû
ëû
ïï
ïï
îþ
åå
åå
åå
(
1
2
)
(a)
( b)
Figur
e 9. The
SHCC Ima
g
e
s
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4563 – 4
571
4570
Whe
r
e
(,
)
f
xy
is th
e
refe
ren
c
e i
m
age,
(,
)
f
xy
is the
avera
ge
gra
y
level of
(,
)
f
xy
an
d
(,
)
ij
f
xy
is the
re
giste
r
ed ima
ge
whi
c
h
ha
s be
en t
r
an
sform
ed ,
(,
)
ij
f
xy
is the
average
gray l
e
vel of
the regi
stered
image.
This m
e
asure of similarity is computed f
o
r wi
ndo
w p
a
i
rs from the
sensed a
nd re
feren
c
e
image
s and it
s maximum i
s
sea
r
ched. Th
e wind
ow
p
a
irs for which the maximum i
s
achieved a
r
e
set as the
co
rrespon
ding o
nes.
(a)
( b)
(c
)
Figure 10. Th
e Process of Image Regi
stration
6. Experiment of Image
Regis
t
er
We te
sted th
e method
s af
ter se
gme
n
ting the liver l
e
sio
n
with lo
cal level
-
set. The final
conto
u
r
of the SHCC i
s
shown in
Fi
gu
re 1
0
(a
). Th
e plain
CT i
m
age, a
s
sh
own i
n
Fig
u
re, the
gray level of
the lesi
on is very clo
s
e t
o
the
liver ti
ssue, ne
arly
undete
c
tabl
e. Before ima
ge
regi
stratio
n
,
we
mapp
ed
the
regio
n
of
the le
sion
int
o
the
plain
CT imag
e, a
s
sho
w
n
in
Fig
u
re
10(b
)
, we
ca
n se
e that th
e co
ntou
r wa
s mig
r
ated
a
little. We trie
d to alig
n the
plain
CT im
age
with the refe
ren
c
e ima
ge.
We tra
n
sl
ated the un
re
g
i
stere
d
imag
e to regi
ster
it with the b
a
se
image, an
d
Correl
ation
betwe
en the
unre
g
iste
re
d image
an
d the refe
re
nce im
age
wa
s
comp
uted. When its maxi
mum is
sea
r
ched out, we
g
o
t the regi
ste
r
ed ima
ge. T
he co
rrespon
ding
conto
u
r
of lesion in th
e plai
n CT
imag
e
wa
s sho
w
n
i
n
figure 10
(c). We can
see t
hat co
ntou
r is on
les
i
on’s
true loc
a
tion.
7. Conclusio
n
In this pape
r, we pro
p
o
s
ed co
mpreh
ensive meth
ods to lo
cat
e
liver lesio
n
s in CT
multipha
se i
m
age
s. It first segme
n
t liver lesio
n
by combing local
C-V level set algorithm, OTSU
method in th
e CT im
age
s in whi
c
h th
e
lesio
n
differs
from liver ti
ssue mo
st ma
rkedly, then
m
a
p
the regi
on of
lesio
n
into CT image
s of
other p
h
a
s
e
s
by image regi
strati
on. Be
cause the cont
our
evolves in the
local area, the method can
not only get
rid of the interferen
ce
of oth
e
r organ
s, bu
t
also
red
u
ce
the co
mput
ational
comp
lexity. T
he result
s of ex
perim
ents
h
a
ve testified
the
feasibility of t
he m
e
thod
s.
We
will
apply
othe
r m
e
th
o
d
s to
segm
en
t the oth
e
r le
sion
in
ou
r fut
u
re
wor
k
.
Ackn
o
w
l
e
dg
ements
This
work was supp
orted
by scientific res
earch fu
nd of Hun
a
n
provinci
al e
ducation
depa
rtment (No. 11
c05
3
8
)
and T
e
chn
o
logy plan
fo
undatio
n of Hun
an p
r
ovi
n
cial, China
(No.
2013F
J3
058
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Locating Li
ve
r Lesi
on with
Local C-V L
e
v
el
Set and I
m
age Regi
stration (Zh
aoh
ui Luo)
4571
Referen
ces
[1]
Yuanz
ho
ng L
i
,
Shoj
i Har
a
, Ka
zuo Sh
imura.
A Machi
ne
Lea
rnin
g Ap
proac
h for Loc
atin
g
Boun
dari
e
s o
f
Liver Tumors in CT Im
ages.
Procee
din
g
s of
the18th
Intern
ation
a
l C
onfer
ence
on P
a
ttern Rec
ogn
itio
n
(ICPR). Hong
Kong. 20
06; 1: 400 – 4
03
[2]
E Rummeny
, R Baro.
Imag
i
ng of Liver Di
seases. Dis
ea
ses
of the Abdo
me
n and P
e
lvis
. Davos
:
Sprin
ger. 20
06
.
[3]
Christia
ne K
u
li
nna M
D
, W
o
lfg
ang Sc
hima M
D
.
Imag
in
g F
e
atures of H
e
p
a
t
ic Metastases:
CT
and M
R
Diseas
es of T
he Abdo
men a
n
d
Pelvis
. Dav
o
s: Springer. 20
06.
[4]
M Lang
er J, T
W
i
nterer, E Kotter, N Ghan
em. De
tection
and ch
aracter
i
zation of b
eni
gn focal l
i
ver
lesio
n
s
w
i
th m
u
ltislic
e CT
. Eu
rope
an Ra
di
olo
g
y
. 2006; 1
6
(1
1): 2427-
24
43.
[5]
Seon
g-Ja
e Li
m, Yong-Ye
o
n
Jeo
ng, Yo-
S
ung
Ho
. Au
tomatic Liv
e
r Segme
n
tatio
n
for Volum
e
Measur
ement
in CT
Images.
Jour
nal
of Vi
sual
Co
mmuni
cation
an
d I
m
age
Re
pres
ent
ation
. 20
06
;
17(4): 86
0-8
7
5
.
[6]
S Mukhop
ad
h
y
a
y
, B Cha
n
d
a
. Multiscal
e
morp
h
o
lo
gica
l segme
n
tatio
n
of gra
y
-sc
a
l
e
images.
IE
EE
T
r
ans. Imag
e Process
. 200
3;
12(5): 533
–5
4
9
.
[7]
SJ Osher, JA Sethia
n. F
r
o
n
ts prop
ag
atin
g
w
i
th c
u
rvat
ure d
e
p
end
en
t speed:Al
gor
ithms bas
e
d
onH
amilto
n
-Ja
c
obi formu
latio
n
s.
Comput Ph
ys.
1998; 79(
1)
: 12-49.
[8]
XXu Ji
ng, C
h
e
n
Ke
n, Ya
ng
Xi
ang
do
ng, W
u
Dan.
A
daptiv
e
Leve
l
Set M
e
th
od for
Seg
m
en
tation
of Liv
e
r
Tumors
in
Min
i
mally
Invas
i
ve
Surg
ery Us
in
g Ultras
o
u
nd I
m
a
ges.
Pr
oceedings
of the
IEEE the 1s
t
Internatio
na
l
C
onfere
n
ce on Volum
e
Bio
i
nf
orma
tics a
nd
Biome
d
ica
l
En
gin
eeri
ng (ICB
BE). W
uhan
.
200
7; 109
1-10
94.
[9]
OT
SU N. A thresho
l
d s
e
lecti
o
n metho
d
from
gra
y
-l
evel
hist
ograms.
IEEE
Trans. Syst. Man Cy
bern
.
197
9; 9(1): 62-
66.
[10]
CHan
T
F
, Vese LA. Activ
e
c
ontours
w
i
th
ou
t edg
es.
IEEE
Transactions on Image Pr
ocessing
. 2
0
01;
10(2): 26
6-2
7
7
.
[11]
JJinsh
eng
Xia
o
, Li
ngl
ing
Xu,
Bens
hun
Yi.
T
he Improv
ement of C-V
Le
vel Set
metho
d
for I
m
a
g
e
Se
gm
en
ta
ti
on
. Procee
din
g
s of
the 20
08 Inter
natio
nal
Co
nfer
ence
on C
o
mp
uter Scie
nce a
nd Soft
w
a
r
e
Engi
neer
in
g (C
SSE). Wuhan. 200
8; 2: 1106-
110
9.
[12]
LMG F
onseca,
BS Manj
un
ath. Reg
i
stratio
n
tec
hni
qu
es fo
r multise
n
sor
remotel
y
s
ens
ed im
ager
y.
Photogr
a
m
met
r
ic Engi
neer
in
g
and Re
mote S
ensi
ng.
19
96; 562(
9): 104
9–1
056.
[13]
BBarbar
a Z
i
tov
á
, Jan F
l
uss
e
r.
Image registr
a
tion m
e
thods:
a surve
y
.
Ima
ge a
nd Vis
i
on
Co
mp
uting
.
200
3; 21(1
1
): 977–
10
00
Evaluation Warning : The document was created with Spire.PDF for Python.