TELKOM
NIKA
, Vol. 11, No. 10, Octobe
r 2013, pp. 5
600 ~ 5
608
ISSN: 2302-4
046
5600
Re
cei
v
ed Ma
rch 2
5
, 2013;
Re
vised June
24, 2013; Accepte
d
Jul
y
1
1
, 2013
Ultrasound Image Segmentation based on the Mean-
shift and Graph Cuts Theory
Yun Ting*
1
, Gao Mingxin
g
1
, Wang
y
a
nming
1
1
Departme
n
t of Information & Scienc
e, Nanj
i
ng F
o
restr
y
Un
iversit
y
,
Jian
gsu Provi
n
ce, Nanj
ing
21
002
7, Chi
na.
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: nj
yunti
ng@
q
q
.com*
1
, d806
4
916
39@
qq.co
m
2
, 718515
02
8
@
qq.com
3
A
b
st
ra
ct
T
h
is p
aper
ad
d
r
essed
the
issu
e of v
a
scul
a
r u
l
tr
asoun
d i
m
ag
e se
g
m
entati
o
n a
nd
pro
pose
d
a
nov
e
l
ultraso
n
ic vasc
ular l
o
catio
n
a
nd detecti
on
method.
W
e
con
t
ributed i
n
sev
e
ral as
pects: F
i
rstly usin
g me
an-
sh
i
ft se
gm
en
tati
o
n
al
g
o
r
i
t
hm
to
o
b
t
ai
n
the
i
n
i
t
i
a
l
segm
en
ta
tio
n
re
su
l
t
s o
f
va
scu
la
r im
ag
es; Se
co
n
d
l
y
new
data ite
m
a
nd
smo
o
th ite
m
of
t
he grap
h cut
energy fu
nctio
n
w
a
s constru
c
ted bas
ed o
n
the MRF
mo
d
e
,
then w
e
p
u
t forw
ard sw
ap
and
ex
p
an
si
on
a
i
d
e
a
s
to o
p
t
im
i
z
e segm
en
ta
ti
on
re
su
l
t
s, co
n
s
e
q
uen
tly
accurate
ly loc
a
ted the vess
el
w
a
ll an
d lu
men
in vascu
lar i
m
ages. F
i
n
a
lly c
o
mparis
on w
i
th
experts
ma
nu
all
y
taggi
ng r
e
su
lts, and
ap
ply
i
ng
ed
ge c
o
rrel
a
t
i
on
coeffi
ci
ent
s an
d var
i
anc
e to v
e
rify th
e va
lid
ity of o
u
r
alg
o
rith
m, exp
e
ri
ment
al res
u
l
t
s show
that o
u
r al
gorit
h
m
c
a
n efficie
n
tly co
mb
in
es the
ad
vantag
es of
mean-
shift and gra
p
h
-
cut algor
ith
m
and ac
hi
eve be
tter segmentati
on resu
lts.
Ke
y
w
ords
:
Ult
rasound im
age; Mean-shift; Graph-c
u
t algorithm
;
Gauss
m
i
x
t
ure model.
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Medical ultra
s
ou
nd imag
e
s
is an impo
rtant type of medical image
s and is widely
used in
medical di
ag
nosi
s
,
Comp
ared
with
oth
e
r m
edical im
aging
metho
d
s
, ultra
s
o
und
imaging
ha
s t
h
e
advantag
es
of non-tra
u
m
a
tic to huma
n
body, real-t
ime displ
a
y, low co
st, ease to use. As an
ideal n
on-i
n
vasive di
agno
sis metho
d
it h
a
s b
r
illiant d
e
v
elopment a
n
d
bro
ad p
r
o
s
pect
s
. Ho
wev
e
r,
becau
se of
the imagin
g
mech
ani
sm
lead to be
insufficie
n
t grayscal
e
displ
a
y ran
g
e
or
unreasonable gray distri
bution, so the ultras
ound images auxiliary diagnosi
s
effect
is
con
s
trai
ned,
esp
e
ci
ally in some l
o
cal d
e
tails, if t
he g
r
ay scale diff
eren
ce i
s
not
obviou
s
, that will
bring
a l
o
t of difficult to di
agno
si
s. In o
r
de
r to im
pro
v
e ultra
s
ou
nd
image
s
quali
t
y and en
han
ce
the reada
bility of ultra
s
o
u
nd ima
g
e
s
l
o
cal
detail
s
,
ma
ke im
ag
es
suita
b
le f
o
r
huma
n
e
y
es
observation
or ma
chine
analysi
s
, the
r
efore in
re
cent years
automat
ic se
gmentation
f
o
r
patholo
g
ical
area in ult
r
a
s
ound ima
g
e
s
become the rese
arch focu
s.
Some
schola
r
s dealt
with
the ultrasoun
d
image
segm
entation i
n
th
e fre
quen
cy
domain,
su
ch a
s
litera
t
ure [1] used
wavelet de
co
mpositio
n to
achi
eve wave
let coeffici
ent
s then
com
b
i
ned
with neu
ral n
e
twork meth
o
d
to pro
c
e
ss
segm
ent
ation
probl
em. Sheng Y etc [2]
con
s
tru
c
ted
an
accurate ultraso
und im
ag
e seg
m
entati
on algo
rith
m
in the wavele
t domain with
the Cha
n
-Ve
s
e
model, Ali Kerma
n
i etc [3
] combin
ed the local histo
g
ram a
nd
wa
velet transfo
rm to locate t
he
positio
n of
breast l
e
si
on
s.
J.Xie [4] p
r
op
ose
d
a
ne
w
method
whi
c
h combin
ed
texture
and
shap
e
as th
e p
r
ior i
n
formatio
n, then
ene
rgy f
unctio
n
wa
s
con
s
tru
c
ted
a
nd texture
of
patholo
g
ical
area
wa
s cla
s
sifie
d
by the shape p
a
ram
e
ter an
d ga
bor filter
co
efficients. Ot
her
re
sea
r
ch
ers
pro
c
e
s
sed
ultrasoun
d ima
g
e
in th
e time
domain,
Literature [5]
prop
ose
d
segm
en
tation alg
o
rith
m
for vascula
r
i
m
age b
a
sed
on gray pro
b
ability density
function
and
fast matching
idea
s, Literat
u
re
[6] con
s
tru
c
t
ed an i
m
ag
e se
gmentati
on meth
od
based o
n
g
r
aph the
o
ry,
whi
c
h h
a
s t
h
e
advantag
es
of robu
st to
noise, se
nsiti
v
e to the
blu
rre
d ed
ge, lo
w re
sid
ual e
r
ror
rate a
nd
fast
cal
c
ulatio
n speed. After
remove
spe
ckle noi
se,
literature [7,
8] a
dopted
active
cont
our mo
del
combi
n
ing wi
th
prio
r
info
rmation such as sha
pe
tex
t
ure colo
r
to
com
p
lete
p
a
thology
regi
on
division. Ch
ri
stodo
u [9] used ten differe
nt texture
feature incl
ude fi
rst-ord
e
r
statistics, gray le
vel
co-occu
r
ren
c
e-matrix, gra
y
differ
ential
statistics,
n
e
i
ghbo
rho
od gray
differe
nce
matrix, statist
i
cal
feature mat
r
ix, texture energy spe
c
trum,
cha
r
a
c
teri
stic of fractal dim
ensi
on, po
we
r sp
ectrum an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : ISSN: 2302
-40
4
6
5601
sha
pe p
a
ra
m
e
ters etc to
extract
carotid atheros
cle
r
otic pla
que
s,
then extra
c
t
i
ng sepa
ratio
n
results was treated by the K-neig
hbo
rin
g
method [10,
11].
Becau
s
e
the
imagi
ng m
e
cha
n
ism
of
u
l
traso
und
im
age
s i
s
com
p
lex, ch
aract
e
risti
c
s of
patholo
g
ical
area
are di
sturbe
d by noi
se a
nd u
n
-
p
a
t
hologi
cal reg
i
on, so th
e visual fe
ature
s
of
critical re
gion
s are not di
stin
ct, the swap thoug
ht o
f
graph
cut
algorith
m
ca
n only get lo
cal
optimum
sol
u
tion, so for the excessive inte
rfere
n
ce
ultra
s
o
u
nd ima
ge
can not
achi
eve
satisfactory result
s, if increas
e the iteration steps t
hat will
s
pend more time, thus vascul
ar
ultrasoun
d im
age segme
n
tation method
base
d
on m
ean-shift and
graph
cut theory is p
r
op
o
s
ed
in this arti
cl
e, firstly mean-shift method is
a
dop
ted for initia
l vascul
a
r ul
traso
und im
age
segm
entation
,
in image
small adja
c
e
n
t are
a
s
with
similar g
r
ay at
tribute a
r
e
cl
assified into
one
cla
ss,
se
cond
ly a new d
a
ta
and smooth i
t
em of
the en
ergy fun
c
tion
is define
d
an
d the graph
cut
method i
s
u
s
ed to obtain
the global
op
timal solu
tio
n
,
thereby opt
imization
and
corre
c
tion o
f
segm
entation
re
sult
s a
r
e
reali
z
ed
an
d
the vessel
wall ed
ge
and
vessel
lume
n is a
c
curate
ly
locate
d. Final
ly compa
r
ing
with the man
ual taggin
g
re
sults the valid
ity of our algorithm is p
r
ove
d
.
2. Mean-shift Ultras
ound Image Segmenta
tion Me
thod
Mean
-shift co
mputational
model is a
n
e
ffective
tool for analy
s
is in
f
eature sp
ace, it was
widely
used I
n
many
co
m
puter vi
sion
appli
c
ation
s
.
Mean
-shift is a p
r
ob
ability den
sity grad
ient
function
with
no pa
ram
e
ters e
s
timation
algorith
m
, It along the
gra
d
i
ent risi
ng di
re
ction to find t
he
peak of the
probability di
st
ribution. In
the kernel
density esti
mati
on, kernel fu
nction generally
meets the
condition
s:
()
2
,
kd
K
pc
p
æö
÷
ç
=
÷
ç
÷
ç
èø
,where
(
)
(
)
0
Kp
p
³
calle
d
ke
rnel
fun
c
tion an
d
p
rep
r
e
s
ent
s
a sin
g
le pi
xel in the
image
doma
i
n The
normalizatio
n p
o
sitive con
s
tant
ψψ
δ
σ
m
es
r
s
r
sr
L
3
Ts
i
n
2L
L
=
is to
en
su
re
that
(
)
Kp
integ
r
al
is 1. In
me
an-shift meth
od g
a
u
ssi
an
kernel
fun
c
tio
n
an
d uniform
kern
el fun
c
tion is two
kind of com
m
only use
d
kernel fun
c
tio
n
, le
t
(
)
(
)
g
pK
p
=-
, k
e
rnel func
tion
(
)
Gp
define
()
2
,
gd
Gp
c
g
p
æö
÷
ç
=
÷
ç
÷
ç
èø
, and me
an
shift vector i
s
d
e
fined
as
follows
,
i
p
is the sample p
o
int of the current pa
rzen windo
w.
()
()
()
2
1
,
2
1
n
ii
i
hk
n
i
i
pg
p
h
p
h
mp
x
gp
ph
=
=
æö
÷
ç
-
÷
ç
÷
ç
èø
=-
æö
÷
ç
-
÷
ç
÷
ç
èø
å
å
(1)
Mean
-shift so
lving step
s in
clud
e two pa
rts:
1) Cal
c
ul
ation
of mean shift
vector
()
,
hk
mp
;
2) According
to the
()
,
hk
mp
value to transfo
rm n
u
cle
a
r lo
ca
tion, this
proc
ess
ens
u
res
to
conve
r
ge to
the poi
nts tha
t
neighb
orh
o
ods
gra
d
ient
are
ze
ro. Let
{}
1
,
2
,
.
..,
j
y
jn
=
is location
seq
uen
ce d
u
ring mean
-shif
t
proce
s
s, by the formula (1) ca
n be got:
()
()
2
1
,1
2
1
,1
,
2
,
.
.
.
n
ii
i
ij
n
i
i
pg
p
p
h
yj
gp
p
h
=
+
=
æö
÷
ç
-
÷
ç
÷
ç
èø
==
æö
÷
ç
-
÷
ç
÷
ç
èø
å
å
(2)
formula
(2
)
calcul
ate
weig
ht mean
valu
e with
kern
el
functio
n
G
i
n
,
ij
y
positio
n,
whe
r
e
,
ij
y
is the initial position of kernel.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Ultra
s
ou
nd Im
age Segm
entation ba
sed
on the Mean
-shift and
Gra
ph Cut
s
Theo
ry
(Y
un Ting
)
5602
Ultra
s
ou
nd i
m
age
se
gm
entation m
e
thod b
a
sed
on me
an-shi
ft algorithm
can
be
descri
bed
as: Firstly, the ultrasoni
c im
age can b
e
rep
r
e
s
ente
d
as a two-di
m
ensi
onal g
r
id
in
three
-
dime
nsi
onal spa
c
e,
grid
spa
c
e i
s
called
sp
atial domai
n, gray value sp
ace al
so
call
ed
rang
e dom
ai
n. Throu
gh
searchin
g for t
he next sh
ift point in the i
m
age
spa
c
e,
and set the stop
conditions, until the di
splacement is l
e
ss than
a give
n value, the
n
tran
slation i
s
stopp
ed. L
e
t
i
x
and
i
z
(i
=
1, 2,
…, n) are
re
spectively the input
vector a
nd the filter output re
sults.
For the all d
a
ta points
i
x
i
=
1
,
2, …, n, mean-shift vector
(
)
,
hk
mx
of each
point are
cal
c
ulate
d
, accordi
ng to the
()
,
hk
mx
value tha
t
moving win
dow
cente
r
to the next point. This
pro
c
e
ss i
s
repeate
d
until
conve
r
gen
ce to t
he de
nsity pea
k o
f
the data space, whe
n
the
estimated
de
nsity gra
d
ient
is ze
ro, then
no need to
move, and a
s
sign the
pixel value
x
P
to
i
Z
,
that is
ix
Z
P
=
.
i
Z
is the filtered pi
xel. The output im
age consi
s
ts of m
u
ltiple indep
ende
nt
regio
n
s.
B
a
si
c st
ep
s a
r
e a
s
f
o
llow
s
:
1) Initialize
1
j
=
and
1
ii
y
x
=
2) A
c
cordi
ng
to equ
ation
(2), calculate
,1
ij
y
+
until converg
ence, an
d re
cord
conve
r
g
ence
value is
,
ic
y
.
3) As
s
i
gn
()
,
,
r
ii
s
i
c
zx
y
=
.
Thro
ugh the
above three steps we get final re
su
lts
of mean
-shift filter, wh
ere
sup
e
rscript
s
and sub
s
cri
p
t
r
respe
c
tively rep
r
e
s
ent
spatial d
o
m
a
in and val
u
e ran
ge. Usi
ng mea
n
-shift
filtering meth
od, need to set band
width
vector
(
)
,
hh
s
h
r
=
.Band
width can be
rega
rded a
s
the
resolution of
segm
entation
,
bandwi
d
th is larger,
mo
re details of t
he image
will
be igno
red,
how
to choo
se the
appro
p
ri
ate band
width, is the key of
su
ccessfully usi
ng ke
rnel d
e
n
s
ity function. In
this pap
er rad
i
al gau
ss
kernel functio
n
is adopted for
ultrasoun
d im
age segme
n
tation.
3. Ultras
oun
d Image Segmenta
tion O
p
timi
zation
Bas
e
d On th
e Graph
Cut
Theor
y
Grap
h cut al
gorithm
is
a
global
optimi
z
ation
alg
o
rit
h
m, by u
s
in
g
the
cla
s
s la
bel a
n
d
con
s
tru
c
t ene
rgy function, i
t
convert ima
ge se
gmentat
ion pro
b
lem i
n
to the probl
e
m
of minimizi
ng
the ene
rgy functio
n
. Un
d
e
r the g
u
ida
n
ce
of
the
grap
h cut theory, network is i
ngeni
o
u
sly
con
s
tru
c
ted,
and en
ergy i
s
linked to the netwo
rk
ca
pacity, lastly netwo
rk flo
w
prin
ciple a
b
o
u
t
grap
h the
o
ry
is ad
opted
to
find the g
r
ap
h minimu
m cut, the cut i
s
optimal
soluti
on of the
ene
rgy
function
mini
mization
p
r
o
b
lem, be
sid
e
s
the
imag
e
se
gmentatio
n is complet
ed. Graph
cut
algorith
m
are
two thoug
hts,
includ
e swap
and
ex
p
an
si
on
a
.
A. Ultra
s
oun
d Image Segmenta
tion M
odel Bas
e
d
On MRF
Image se
gm
entation app
roach based
on the MRF
model can b
e
seen a
s
op
timization
probl
em
s of acq
u
isitio
n the label field
f
which ma
ke energy function
(
)
Ef
minimization.
Energy fun
c
tion expre
s
sed
as follows:
()
()
()
()
{}
()
,
,
,,
s
mo
o
t
h
d
a
t
a
p
q
p
q
p
p
i
pP
pq
N
Ef
E
f
E
f
V
f
f
D
f
w
Î
Î
=+
=
+
åå
(3)
Whe
r
e
(
)
sm
oo
t
h
Ef
calle
d smo
o
th en
ergy, it is th
e puni
shm
e
n
t
to the un-smoothne
ss
cha
r
a
c
t
e
ri
st
ic
s;
(
)
da
t
a
Ef
kno
w
n
as the
data
item ene
rg
y, it is the
puni
shm
ent
to the
disa
gre
e
ment
betwee
n
cu
rrent cla
ss la
b
e
l
f
and ob
se
rvation data
i
w
class. For a gi
ven image,
p
represent ea
ch pixel,
P
is the set of all p
i
xels, all neig
hbori
ng pixel
pair
{
}
,
p
q
that c
o
ns
titute
the set
N
. Where
()
,
,
p
qp
q
Vf
f
use
s
fou
r
neigh
bo
rh
ood Potts model
[10]:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : ISSN: 2302
-40
4
6
5603
()
()
(
)
,
,
s
mo
o
t
h
p
q
p
q
p
q
Ef
V
f
f
f
f
ld
==
¹
,
l
is the
fact
or to
bal
an
ce
the
data ite
m
an
d
smo
o
thing
item, each pi
xel data item energy
()
p
p
Df
in typ
e
(3)
can b
e
got by type (4):
()
(
)
2
11
ˆ
,e
x
p
2
2
i
pi
pp
i
w
p
i
Df
w
f
y
dm
ps
æö
æö
÷
ç
-
÷
÷
ç
ç
÷
÷
ç
ç
==
-
÷
÷
ç
ç
÷
÷
ç
ç
÷
÷
÷
ç
ç
èø
÷
÷
ç
èø
h
(4)
Whe
r
e
p
d
repre
s
ent attrib
ute
value of pixel
p
(su
c
h a
s
gray value
)
.
()
ˆ
i
wp
fd
r
e
pr
es
e
n
t
the probabilit
y of pixel
p
bel
ong
s to the
category
i
w
. We
use th
e mea
n
-shift metho
d
to obtain
the initial ultraso
und i
m
ag
e se
gme
n
tation, and
get
mean val
u
e
s
and va
rian
ces of
w
cla
s
se
s:
(
)
12
1
2
,,
,
,
,
ww
qm
m
m
s
s
s
=L
L
, with ga
uss
RBF
kernel f
unctio
n
(4)
g
e
t likelih
ood
estimation
of
the
pixel
p
to the correspon
ding
w
cla
s
s
e
s.
Wh
ere
h
cont
rol th
e smo
o
thne
ss ra
nge
of fun
c
tion (4),
whe
n
the va
lue of
h
is la
rge, function
curve i
s
mo
re smo
o
th, b
u
t may lose
more d
e
tail
informatio
n; The value of
h
is smaller,
function curve will
be m
o
re
sharper, that will over-
relian
c
e
on t
he ob
se
rvatio
n data, then
the algo
rith
m
perfo
rma
n
ce
will de
gra
d
e
.
In literature[
5]
the meth
od t
o
e
s
timation
h
is given, su
ch as
()
ex
p
fr
q
n
a
md
g
æö
÷
ç
÷
=-
ç
÷
ç
÷
ç
èø
h
,where
(
)
frq
d
is the
occurre
n
ce freque
ncy of
y
in the training
sampl
e
set.
B.
Gra
ph-Cu
t Algorithm
: S
w
a
p
Algorithm Ideas
Thoug
ht of th
e swap
algo
ri
thm is
rep
eat
edly cal
c
ul
ation two
differe
nt categ
o
ri
es
,
ww
ab
,
and g
r
ap
h gri
d
is
con
s
tru
c
t
ed to a
s
soci
a
t
e with en
erg
y
equation
(
3
)
in ord
e
r to
o
b
tain the o
p
timal
solutio
n
of e
q
uation. Let
P
b
e
the
set of a
ll pixels,
whil
e
P
ab
is a pixel
points
set
wh
ich
cla
ss
belon
g to
,
ww
ab
,
grid
con
s
truction as
sh
o
w
n in fig
u
re
1, figure el
ements
are: vertex set
{}
,,
VS
S
P
ab
=-
-
, edge
set
{}
{}
{}
,
,
,,
pp
pq
pP
pq
N
tt
e
ab
e
Î
Î
ìü
ïï
ïï
=
íý
ïï
ïï
î
þ
UU
, where
,
SS
ab
--
is two graph
ver
t
ex,
{
1
,
2
,
3
,
...
,
1
2
}
pp
R=
=
,
,
p
X
a
in figure
1 represe
n
t pixel
p
label
to the
cla
ss
w
a
throug
h
mean
-shift method. Pixel
p
(satisfy
p
fa
=
or
p
fb
=
) in
ab
R
respe
c
tively conne
ct with two
ver
t
ic
es
,
SS
ab
--
th
at co
nstitute
t-lin
k e
dge
s, de
noted
by
,
p
p
tt
ab
, and
th
e ele
m
ents
among
ab
R
con
s
ti
tute n-link ed
ges, den
oted
by
{}
,
p
q
e
, the weight of t-link and n-lin
k assi
gnment
approa
ch respectively se
e equatio
n(5
)
a
nd equ
ation(6
)
.
()
()
(,
)
(,
,
,)
q
qN
p
p
q
q
qN
p
p
q
pp
pp
tV
f
p
t
Df
w
w
Df
w
w
Vf
p
a
a
ab
ab
b
ab
a
a
b
b
b
Î
ÏR
Î
ÏR
=+
Î
R
=+
Î
R
å
å
(5)
{,
}
(,
)
{
,
}
,
p
pq
q
eV
p
q
fp
q
f
ab
=Î
N
Î
R
(6)
Whe
r
e
p
N
is a
d
j
ace
n
t area o
f
pixel
p
, litera
t
ure[11] p
r
ov
e the
capa
cit
y
of cut set i
s
:
CE
K
=-
(7
). Where
E
is the
en
ergy
value of
equa
tion(3
)
,
,
ww
ab
is a
con
s
tant, the
optimal
solutio
n
of e
nergy fu
nctio
n
(3
) is the
minima
l cut set
of
g
r
ap
h(1a). swap
al
gorithm ra
nd
omly
sele
ct
t
w
o
cla
s
s
,
ww
ab
from
w
cla
ss, an
d
throug
h th
e ene
rgy fu
nction to
so
lve th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Ultra
s
ou
nd Im
age Segm
entation ba
sed
on the Mean
-shift and
Gra
ph Cut
s
Theo
ry
(Y
un Ting
)
5604
corre
s
p
ondin
g
cut
set, the
n
dete
r
mine t
he pixel b
e
lo
ngs to
which
cla
ss
set, thu
s
compl
e
te the
segm
entation
optimization,
spe
c
if
ically shown in Figu
re 1a.
C.
Gra
ph-Cu
t Algorithm
:
ex
p
an
si
on
a
Algorithm Ideas
The swap
al
gorithm ca
n only
exch
ang
e
w
a
w
b
cla
s
s to
cal
c
ulate
of the minimu
m
energy. If we limit
w
b
and make
w
a
exchan
ge with all other cl
asse
s, then will get a broa
de
r
transfo
rmatio
n app
ro
ach, it is
ex
p
an
si
on
a
algorith
m
. Literatu
re[
14] proved
when fo
rmatio
ns
of
smooth item
satisfy ‘metri
cs’
con
s
train,
then
ex
p
an
si
on
a
algo
rithm idea
s ca
n be ad
opte
d
to
achi
eve more broa
der t
r
an
sform
a
tion
, as sh
o
w
n
in figure 1b, a set of vertices
V,
{}
{}
,
,
,s
i
n
,
,
pq
pq
pq
N
ff
Vs
o
u
r
c
e
k
P
Z
aa
Î
¹
ìü
ïï
ïï
ïï
=-
-
íý
ïï
ïï
ïï
î
þ
U
is the
com
positio
n ele
m
ents
of the graph,
wh
ere
s
ou
r
c
e
a
-
,
sin
k
a
-
re
spe
c
tively the
highe
st
and l
o
west v
e
rtice
s
.
p
is th
e any
one
pix
e
l in
P
,
q
is adjacent pixels of
,
pq
ÎN
. When the two pixel class la
bel
p
f
and
q
f
are unequal, the
n
auxiliary nodes
{}
,
p
q
Z
a
r
e
adde
d, Let
{}
,
,,
p
q
ab
c
Z
Î
are
th
e auxilia
ry n
ode, a
ddin
g
t
he a
u
xiliary
node
()
1,
3
2
,
1
,
X
aX
betwee
n
1,
3
X
and
2,
1
X
, as sh
own in fig 1b,
similarly, there are a
u
xiliary nodes
()
4,
3
5
,
1
,
X
bX
and
()
7,
3
8
,
1
,
X
cX
. Where
edge
set i
s
{}
{}
{}
()
(
)
{}
{}
()
(
)
,,
,,
,,
,
pp
pq
pq
pP
pq
N
p
q
N
di
s
p
p
d
i
s
p
q
di
s
p
p
d
i
s
p
q
tt
e
aa
ex
Î
ÎÎ
=¹
ìü
ïï
ïï
ïï
ïï
ïï
=
íý
ïï
ïï
ïï
ïï
ïï
îþ
UU
U
,
{}
,
p
q
x
are
the ed
ges th
at between
a
u
xiliary no
de
{}
,
p
q
Z
with n
ode
s
p
,
q
,
a
. For exampl
e
figure
1b
sh
ows
that total thre
e ed
ge
s of
c,
the ed
ge
{}
73
,
X
c
e
betwee
n
c
and
73
X
, the e
dge
{}
81
,
cX
e
bet
wee
n
c
and
8,
1
X
; the edge
c
t
a
betwee
n
c
and
a
, similarly, auxiliary node
a
and
b
constru
c
t edge
s with
their adja
c
e
n
t node
s. the pixels of
P
resp
ectively con
n
e
cted
with
a
and
a
to c
o
ns
truc
t t-link
edge
s, the pixels amon
g P conne
ct wi
th each othe
r to constitute
n-link ed
ge
s, edge weig
h
t
s
assignm
ent method is ex
pre
s
sed by formul
a (8
)
a
nd (9
), if existing auxiliary node
c, then
the
edge
weight
betwe
en c a
n
d
adja
c
ent po
ints ca
n refe
r to the followin
g
formula (10
)
.
()
()
,
,
,
pp
i
p
p
pp
i
p
Df
w
w
tp
w
Df
w
tp
p
tp
a
a
a
a
a
aa
=¥
Î
R
=Ï
Î
=Î
R
(8)
{,
}
()
,{
,
}
,
pq
pp
q
eV
p
f
wf
f
q
a
=Î
N
=
(9)
{,
}
{,
}
()
{
,
}
,
()
{
,
}
,
(,
)
{
,
}
,
,
,
pc
cq
p
c
p
q
qp
q
pq
p
q
fw
f
f
wf
eV
p
q
ef
f
ff
f
f
Vp
q
tV
p
q
a
a
a
=Î
N
¹
=Î
N
¹
=Î
N
¹
(10
)
Whe
r
e th
e d
e
finition of d
a
ta item D
see form
ula
(4), sm
ooth it
em
V
as sho
w
n in
formula (3). Literature [14] proved that the c
u
t-s
e
t capac
i
ty is
CE
=
(11
)
whe
r
e E is obtaine
d fro
m
equation (3), then usi
n
g the metho
d
of
ex
p
an
si
on
a
to
optimally solv
e the equatio
n (11). The
steps of
ex
p
an
si
on
a
algori
t
hm is: con
s
truct the netwo
rk
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : ISSN: 2302
-40
4
6
5605
sho
w
n in Fig
u
re 1b, pixel
set
R
, Label is
the same d
e
finition as swa
p
algorithm, f
o
r the
R
cla
ss,
if the pixel
la
bels a
r
e
a
a
nd
a
, then
set t-li
nk
and
n
-
lin
k edg
es to fin
d
the
minimu
m graph
cut
set C, eq
uiva
lent to see
k
t
he minimu
m energy
E
¢
whi
c
h co
rre
sp
ondi
ng to the cl
ass label
L
¢
, if
condition
EE
¢
<
is satisfied, then
modify the cl
ass la
b
e
l by the same mo
de of swap
al
gorithm,
and up
date the cla
s
s set
of every pixel to
L
¢
.if
EE
¢
>
,then excha
nge
a
to other cl
ass
and
cal
c
ulate the
energy.
a
swa
p
k
si
n
s
ou
r
c
e
t
X
7
t
X
4
t
X
1
t
X
3
t
X
6
t
X
9
t
X
5
t
X
8
t
c
t
b
t
a
t
X
2
b
ex
p
an
sion
Figure 1. Gra
ph cut alg
o
rit
h
m prin
cipl
e diagram
4. Experimental Re
sults
The expe
rime
nt of three ultrasoun
d imag
es is
sho
w
n i
n
Figure 2
Then
varia
n
ce
an
d co
rrel
ation coeffici
ents ar
e u
s
e
d
for comp
arison
of te
st
results,
varian
ce rep
r
esents
devi
a
tion deg
ree
of the two bound
ary p
o
ints, whil
e the co
rrel
a
tion
coeffici
ents i
ndicate simil
a
rity of two boun
dar
y sh
ape. Firstly starting p
o
sit
i
ons in the t
w
o
boun
dari
e
s
a
r
e cho
s
e, the
n
two L
su
b-seque
nce (
1
R
an
d
2
R
) with fixed l
ength in t
w
o
curve
s
a
r
e
taken,
The
si
milarity of
1
R
an
d
2
R
notes for
()
12
,
L
s
im
il
R
R
, it depi
cts
by
12
RR
-
varianc
e
, that
is
(
)
(
)
12
1
2
,
LL
s
im
i
l
R
R
S
E
R
R
=-
, whe
r
e variance
()
(
)
2
12
1
2
1
L
Li
i
i
SE
R
R
R
R
m
e
a
n
L
=
éù
-=
-
-
êú
ëû
å
,
()
12
1
L
ii
i
me
a
n
R
R
L
=
=-
å
,
L
SE
is
smaller, t
he
1
R
and
2
R
ar
e mo
re
s
i
mila
r, w
h
er
e
1
i
R
rep
r
e
s
ent
s the di
stan
ce
betwe
en
sampling
poin
t
s on a
cu
rve to the o
r
igin. Correla
t
ion
coeffici
ents
L
r
are ado
pted for
ed
ge co
mpari
s
o
n
.
(
)
12
12
,
L
Co
v
R
R
r
dd
=
,
(
)
12
,
Co
v
R
R
rep
r
e
s
ent
s
covari
an
ce
b
e
twee
n
cu
rve
1
R
and
curv
e
2
R
,
1
d
and
2
d
re
spe
c
tively de
note the
sta
ndard
deviation of
1
R
and
2
R
,
()
()
()
12
1
1
2
2
1
1
,
n
ii
i
Co
v
R
R
R
R
R
R
n
=
=-
-
å
(12
)
()
2
11
1
1
1
n
i
i
RR
n
d
=
=-
å
,
()
2
22
2
1
1
n
i
i
RR
n
d
=
=-
å
,
1
R
2
R
rep
r
e
s
e
n
t the m
ean
distan
ce
of
all points o
n
curves to the o
r
igin.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Ultra
s
ou
nd Im
age Segm
entation ba
sed
on the Mean
-shift and
Gra
ph Cut
s
Theo
ry
(Y
un Ting
)
5606
a
b
c
d
a
b
c
d
a
b
c
d
Image 1 re
sul
t
Image 2 re
sul
t
Image 3 re
sul
t
Figure 2. Segmentation of ultras
oun
d im
age
s, First co
lumns
(a): the
original ult
r
a
s
ou
nd imag
e
s
;
S
e
con
d
col
u
mns (
b
):
t
h
re
e sna
p
s
hot
f
o
r only
mean
-shift segmenta
t
ion results; T
h
ird colum
n
s
(c): only the graph
cut algo
rithm used to
get t
he segm
entation re
sul
t
s. Fourth col
u
mns
(d):
Segmentatio
n of our algo
ri
thm, using g
r
aph cut
re-se
g
mentation af
ter mean
-shift segme
n
tatio
n
results, wh
ere label 1 re
prese
n
ts the ve
ssel
wall a
nd
label 2 re
presents lume
n of blood vessel.
The last row i
n
Figure 2 is
comp
ari
s
o
n
chart
abo
ut ab
ove three ultraso
und ima
g
e
segm
entati
on
results, blue li
nes rep
r
e
s
ent
s the medi
cal
experts ha
nd
-label
ed resul
t
s of vessel
cell and lume
n
edge, the re
d
line is the lab
e
ling re
sult
s of our algo
rith
m.
Next, true po
sitive ratio (T
rue po
sitive ra
tio, TP
), fals
e-pos
i
tive ratio
(Fal
se po
siti
ve ratio,
FP) an
d total
simila
rity de
gree
(Simila
ri
ty, SI)
are a
d
opted, the
s
e
three i
ndi
cato
rs
evaluate
the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : ISSN: 2302
-40
4
6
5607
tumor regio
n
differen
c
e
s
b
e
twee
n the e
x
pert man
ual
labeled
re
sul
t
s and
our
al
gorithm
calib
rate
segm
entation
result
s [15]:
M
S
M
A
A
TP
A
Ç
=
,
MS
M
M
A
AA
FP
A
È-
=
,
M
S
M
S
A
A
SI
A
A
Ç
=
È
(13
)
whe
r
e
S
A
rep
r
e
s
ent
s tumo
r region
of ou
r
algorith
m
seg
m
entation,
M
A
r
e
p
r
es
en
ts
the
docto
rs man
ually label tu
mor a
r
e
a
,
TP
is hig
her, th
a
t
our
segm
e
n
tation result
s cover th
e
highe
r deg
re
e of expert
manual
calib
ration area; F
P
index is lo
wer, the
n
the
covered wro
n
g
area
is less;
SI index is
hi
gher, th
at ou
r se
gment
atio
n re
sult i
s
clo
s
er to the
ma
nual la
bel
are
a
.
In this pape
r we defin
e:
(
)
2
wa
l
l
l
u
m
e
n
TP
TP
T
P
=+
,
(
)
2
w
a
ll
lu
m
e
n
FP
FP
F
P
=+
,
(
)
2
w
a
ll
lu
m
e
n
SI
SI
SI
=+
Table 1 sho
w
s the sp
ecifi
c
experim
ental para
m
eters a
nd experi
m
en
tal result
s, includi
ng
the
sel
e
ctio
n threshold, an
d
the comp
arison
between
our
algo
rith
m and
man
u
a
l se
gme
n
tation
results, by calcul
ating the
correlation
coeffici
ent
s an
d varian
ce of
the curve
s
to comp
are t
h
e
simila
rity betwee
n
them.
From th
e ex
perim
ental
re
sults
ca
n b
e
see
n
, ou
r de
sign
ed al
gorit
hm
can
com
p
lete
ly extract vascula
r
lume
n and a
c
curat
e
ly locate the
positio
n of the vessel wall
, it
has ex
celle
nt perfo
rman
ce i
n
vascular ult
r
asoun
d imag
e segm
entati
on.
Table 1. Sele
cted
coeffici
e
n
ts and Expe
rimental resul
t
s
The
h
of
equal(1)
The
of
equal(4)
vessel w
a
ll segmentation
comparison w
i
th
hand-
labeling
vessel lumen
segmentation comparison
w
i
th ha
nd-labelin
g
TP
(%
)
FP
(%
)
SI
(%)
va
r
i
a
n
c
e
correlati
on
coef
ficien
t
va
r
i
a
n
c
e
correlati
on
coef
ficien
t
w
a
ll+
lumen
w
a
ll+
lumen
w
a
ll+
lumen
ultras
oun
d
image1
h =17
60 25.7
0.7789
13.2
0.8524
92.1
22.07
77.92
ultras
oun
d
image2
h =13
90
12.6
0.9121
8.4
0.9456
95.3
8.90
86.80
ultras
oun
d
image3
h =13
90
21.2
0.7833
11.8
0.9219
89.6
17.21
80.31
5. Conclusio
n
and Futu
r
e
Work
This p
ape
r p
r
esents
a no
vel vascular
ultrasoun
d i
m
age
seg
m
e
n
tation meth
od which
combi
n
ing m
ean-shift an
d
grap
h cut al
gorithm, fi
rstly mean-shift algorith
m
is
adopte
d
for the
initial segm
entation, then initial classification
re
su
lt
s whi
c
h in
clud
e sp
ecif
i
c
ch
ara
c
t
e
ris
t
ic
para
m
eters o
f
each cl
ass
are obtai
ned
according to
the histog
ra
m classification, sub
s
e
que
ntly
prob
abili
stic
model is u
s
e
d
to const
r
u
c
t
data and smooth items
of energy fu
nction, and t
h
e
n
usin
g g
r
aph
cut meth
od t
o
get the
opti
m
al solution
of the en
ergy
function, th
e
r
eby lo
cating
th
e
positio
n of ve
ssel wall an
d
vessel lum
e
n, com
pari
s
o
n
with expert manual
segm
entation re
sul
t
s,
our al
gorith
m
achi
eves va
scular l
o
catio
n
functi
o
n
. O
u
r future work mainly in
cl
ude two aspe
cts,
firstly how t
o
accu
rately
and effe
ctively locate th
e ultra
s
ou
nd
vascular im
age e
dge
when
acq
u
isitio
n effect of ultra
s
o
und imag
e are not we
ll, se
con
d
ly whe
n
the amount o
f
medical ima
g
e
data is too la
rge, how to im
prove
calcula
t
ion
spe
ed of maximum flo
w
metho
d
s a
nd enh
an
ce the
grap
h-cut rea
l
-time pe
rformance for be
tter adapt fo
r
the division o
f
ultraso
und i
m
age
s, these
will
be the focu
s
of our future
work.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Ultra
s
ou
nd Im
age Segm
entation ba
sed
on the Mean
-shift and
Gra
ph Cut
s
Theo
ry
(Y
un Ting
)
5608
Referen
ces
[1]
İş
can Z
,
Kurn
az MN. Ultra
soun
d Imag
e
Segme
n
tatio
n
b
y
Us
in
g
W
a
velet T
r
ansform and
Se
lf
Organiz
i
ng Ne
ural
N
e
t
w
ork.
Neur
al Infor
m
a
t
ion Proc
essin
g
-Letters a
nd
Review
s
. 20
06
, 10(8): 18
3-
190.
[2]
Yan S, Y
uan
J
P
, Hou
CH. Se
gmentati
on
of
medic
a
l u
l
traso
und
ima
ges
ba
sed o
n
l
e
ve
l se
t method
w
i
t
h
edg
e repr
ese
n
ting m
a
sk.
3
rd International Conference on Adv
anc
ed Com
puter
T
heor
y
and
Engi
neer
in
g (ICACT
E). Chengdu,
Ch
in
a. 20
10; 2: 85-8
8
.
[3]
Kerman
i
A, A
y
atoll
ahi A. Med
i
cal Ultras
o
u
n
d
Image Segm
e
n
tation
b
y
M
odi
fied Loc
al H
i
stogram R
ang
e
Image Metho
d
.
Biomed
ical Sc
ienc
e an
d Engi
neer
ing
, 2
010,
3: 1078-
10
84.
[4]
Xi
e J, Jian
g Y,
T
s
ui H
T
. Se
gmentati
on of
Ki
dn
e
y
F
r
om
Ultraso
und Im
ages Bas
ed o
n
T
e
xture an
d
Shap
e Priors.
IEEE Transaction on Medic
a
l Im
aging
. 20
05, 24(1): 53
9-5
5
1
.
[5]
Hélè
ne
Ma, C
a
rdi
nal
R, Me
uni
er J, et
al.
In
travascul
a
r
Ultraso
und
Image
Segm
enta
t
ion: A T
h
ree
-
Dimens
io
nal
F
a
st-Marchi
ng
Method
Bas
e
d
on
Gra
y
Lev
el
Distrib
ution
s
.
IEEE Trans
On Medic
a
l
Im
aging
. 200
6, 25(5): 590-
60
2.
[6]
Grad
y
L, F
u
n
k
a-le
a G. Multi-lab
e
l ima
ge
segm
e
n
tatio
n
for medica
l ap
plicati
ons b
a
s
ed on gr
aph-
theoretic
el
ectrical
pote
n
tia
l
s.
Confer
ence
o
n
Comp
uter V
i
si
on
and
Math
e
m
atical M
e
tho
d
s i
n
Me
dica
l
and Bi
ome
d
ica
l
Image Ana
l
ysi
s
. Pr
ague, America. 20
04; 23
0-24
5.
[7]
Cremers
D, R
ousso
n M, D
e
riche
R. A R
e
vi
e
w
of
Statistic
a
l Ap
pro
a
ches
to Lev
el S
e
t
Segme
n
tatio
n
:
Integratin
g Co
l
o
r,
T
e
xt
ure, Mo
tion an
d Sha
p
e
.
Int.
J. Com
put
e Vision.
200
7; 72(2): 195-
21
5.
[8]
D
y
de
nko I, Ja
mal F
,
Bernard
O, D'
hooge J.
A Leve
l
Set F
r
ame
w
ork W
i
th
a Sha
pe a
nd M
o
tion Pr
ior fo
r
Segme
n
tatio
n
and R
e
g
i
on T
r
ackin
g
in Ech
o
c
ardi
ogra
p
h
y
.
Medic
a
l Imag
e
Analysis
. 2
0
0
5
; 10(2): 16
2-
177.
[9]
Christodoulou CI,
Pattichis CS
. T
e
xture Bas
ed C
l
ass
i
ficatio
n
of Ath
e
roscl
e
r
otic Car
o
tid
Pl
aqu
es.
IEEE
T
r
ansaci
on on
Medic
a
l Imagi
n
g
. 2003; 2
2
(7): 902-
912.
[10]
Li Zhe
ngz
hou,
Liu M
e
i, Wan
g
Huig
ai, Ya
ng
Y
ang,
C
h
e
n
Ji
n, Jin Gan
g
. Gra
yscale E
d
g
e
De
tection a
n
d
Image S
egmentation Algorith
m Based on M
ean Shift.
T
E
L
K
OMNIKA Ind
ones
ian
Jo
urn
a
l
of Electric
a
l
Engi
neer
in
g
. 2013; 11(
3): 141
4-14
12.
[11]
Jun Sun, Ya
n W
ang, Xiao
ho
ng W
u
, Xi
aod
ong Z
h
a
ng, H
ong
ya
n Gao.
A Ne
w
Im
age
Segme
n
tatio
n
Algorit
hm an
d
It’s Applic
ation
in lettuc
e
o
b
j
e
ct segme
n
tati
on.
T
E
LKOMN
I
KA Indon
esia
n Jour
nal
of
Electrical E
ngi
neer
ing
. 2
012;
10(3): 55
7-5
6
3
.
[12]
Fjortoft R, Delignon Y, Piec
z
y
nsk
i
W. Unsupervis
ed Class
i
fica
tion of Radar Images Using Hidden
Markov Ch
ai
ns
and H
i
d
den
Markov Ra
nd
o
m
F
i
elds.
IEEE Transactio
n
on Geosc
i
enc
e
and R
e
mot
e
Sensi
n
g
. 20
03;
41(3): 675-
68
6.
[13]
Szeliski R, Z
a
b
i
n R, Scharst
ei
n D. A Compa
r
ative Stud
y
of Energ
y
M
i
nim
i
zation Meth
od
s for Markov
Ran
dom F
i
e
l
d
s
w
i
t
h
Smoot
h
ness-Bas
ed Pr
iors.
IEEE Transactio
n
on P
a
ttern Ana
l
ysis
and M
a
chi
n
e
Intelli
genc
e
. 20
08; 30(5): 1
068
-108
0.
[14]
Bo
y
k
ov Y, Ve
ksler O, Z
abi
h R. F
a
st ap
pro
x
imat
e en
e
r
g
y
m
i
nim
i
zati
on vi
a gra
p
h
cuts.
IEEE
T
r
ansactio
n
s o
n
Pattern Ana
l
ysis and Mac
h
i
ne Intell
ig
ence
.
2001; 3
3
(3): 1
81-2
00.
[15]
Mada
bhus
hi A
,
Metaxas
DN
. Combin
in
g L
o
w
-
, Hig
h-lev
e
l an
d Empiric
a
l Dom
a
i
n
Kn
o
w
le
dg
e F
o
r
Automated Se
gmentati
on of
Ultraso
nic Bre
a
st Lesi
ons.
IE
EE T
r
ansactio
n
s on M
edic
a
l
Imag
in
g
. 20
03;
22(2): 15
5-1
6
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.