TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 8, August 201
4, pp. 6211 ~ 6216
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.614
8
6211
Re
cei
v
ed Ap
ril 27, 2013; Revi
sed
Ju
n
e
1, 2014; Acce
pted Ju
ne 15,
2014
A Novel Approa
ch for Tumor Detection in
Mammo
graphy Images
Elahe Chag
h
a
ri
1
, Abbas
Karimi*
2
Departme
n
t of Computer En
g
i
ne
erin
g, F
a
cul
t
y
of Engi
ne
eri
ng, Arak Brach
,
Islamic Azad Univers
i
t
y
, Ara
k
381
81-4
6
7
75, Iran
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: elah
e.cha
gha
ri@gma
il.com
1
, akarimi@
iau-
a
r
ak.ac.ir
2
A
b
st
r
a
ct
Breast cancer is one of the ma
jor causes
of death among w
o
m
en in recent decades.
Screening
mammography
is currently the best av
ailable radiological technique for ear
ly detection of breast cancer. In
recent years, several methods
have been used
for automated tumor
detection in mammography
images.
In some methods, due
to a variety
of processing
and
multiple operations on
images, there are
many
computational complexities and much time overhead.
In other methods the recognition accuracy
is
relatively
low
.
In this paper, a
new
method to
detect cancerous
lesions
in mammography images is
presented using cellular
learning aut
omata algorithm.
Cellular learning
automata algorithm
is w
e
ll
suited
for image processing, because
it is
cellular
and belongs pixels like
an image.
Distributed performance
and
parallel
processing properties of this met
hod has
optimal results in image processing.
Experimental results
show
the effectiveness of
the proposed method
.
Ke
y
w
ords
:
medical image processing, mass
detec
tion,
cellular learning automata
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
Brea
st cance
r
is one
of th
e mo
st comm
on
can
c
e
r
s
a
m
ong
wom
e
n
all over the
worl
d. In
the We
st, 10 percent of wome
n su
rvi
v
e with brea
st can
c
e
r
an
d it forms ab
out 19 perce
nt of
death
rate
s
resulte
d
fro
m
can
c
e
r
s. A
b
o
u
t one
out
of
every eig
h
t women i
n
Ame
r
ica
is dia
gno
sed
with brea
st cancer in h
e
r l
i
fetime [1]. in Iran,
bre
a
st
can
c
e
r
ra
nks fifth in cance
r
death
s
am
o
ng
women [2].
Since the ca
use of this cancer is un
known,
it has long bee
n co
nsid
ere
d
imp
o
rtant by
clini
c
ian
s
a
n
d
re
sea
r
che
r
s,
and th
at's why in ma
ny case
s, b
r
ea
st
can
c
e
r
i
s
n
o
t diagn
osed
u
n
til
the adva
n
ce
d sta
g
e
s
. T
h
is l
ead
s to
use the i
m
aging te
ch
ni
que
s be
sid
e
s
the
scree
n
ing
prog
ram
s
. M
any metho
d
s have
bee
n
used fo
r
b
r
ea
st ima
g
in
g in
cludi
ng:
mammog
r
a
p
h
y,
son
ography, and MRI [3]. Among the
s
e
method
s, Mammog
r
ap
hy has be
en u
s
ed more, due
to
some propert
i
es such as
the possibilit
y of lesi
on detection before doctor's detection in the
clini
c
al exami
nation an
d be
fore bein
g
visible on sono
g
r
aphy ima
g
e
s
[1].
In the re
ce
nt years, the tech
nolo
g
y used
in
the me
dical devices
has bee
n
d
e
velope
d
enormou
s
ly and
the
s
e de
vices provid
e
com
p
lete
inf
o
rmatio
n on
the physi
cal
con
d
ition of t
he
patient. Beca
use of thi
s
, reco
nstructio
n
and imag
e p
r
ocessin
g
techniqu
es a
r
e
highly re
gard
ed
by rese
arch
ers.
In mammo
graphy, due to
the low
co
ntrast of im
a
g
e
s
and al
so
be
cau
s
e
of the
fact that
these im
age
s are often n
o
i
s
y, image det
ection i
s
faci
ng so
me chal
lenge
s. Studi
es have
sh
o
w
n
that the sensitivity of
human eye to interp
ret a larg
e volume of image
s de
creases
with the
increa
sing
n
u
mbe
r
of ca
se
s an
alyze
d
espe
cially
that, in the mos
t
majority of the c
a
s
e
s
,
mammog
r
a
p
h
y image
s
do
not p
r
e
s
ent
obviou
s
an
d
visible
sign
s
whi
c
h atte
st t
he exi
s
ten
c
e
of a
tumor. Con
s
i
derin
g the
s
e
asp
e
ct
s, the developm
ent
of su
ch comp
uter aid
ed di
agno
si
s syste
m
s
(CA
D
system
s) to
hel
p e
a
rl
y detectio
n
of
bre
a
st
ca
nce
r
is very im
po
rtant an
d h
e
lp
ful; they wo
uld
analyze mam
m
ograms a
n
d
highlight
s onl
y the abnorm
a
l area
s [4].
According to the research, the number
of false negative diagnoses in
mammography
reports in Iran
is also more
than twice
the a
cceptable level
[2]. Hence, providing
a method
to
help radiologists and other physicians in det
ection of this disease appears necessary.
The
purpose of this study is to provide
a
method for early detection of breast cancer,
independently of the radiologist, and also to dec
rease the rate of false negative detection.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 621
1 –
6216
6212
The first ste
p
in imag
e pro
c
e
ssi
ng i
s
the
pre
p
ro
ce
ssin
g stag
e. At this sta
ge, the i
m
age
s
with poo
r qu
ality are enh
anced. Shi
et al
[5], used Bayesian m
e
thods a
nd wa
velet techniq
ues
for recon
s
tru
c
ting im
age
s
and
enha
nci
n
g their qu
ality. Cha
ngi
zi
et al
[6],
p
r
op
ose
d
a me
th
od
in
whi
c
h firstly t
h
re
shol
d al
g
o
rithm
and
seco
ndly regio
n
g
r
owth
alg
o
rithm
are
u
s
ed
to
sep
a
rate
segm
ents of the ima
ge. Sa
heb Ba
sh
a
et
al
[7], use
d
morp
holo
g
ica
l
ope
rators
(in
c
ludi
ng e
r
o
s
i
on,
dilation, open
ing, and clo
s
i
ng) for imag
e
segme
n
tatio
n
and fuzzy C-mea
n
s cl
ust
e
ring meth
od
for
pattern recog
n
ition and
cl
assificatio
n
o
f
masses. M
encattini
et al
[8], used two-dime
nsio
nal
wavelet to de
tect can
c
e
r
ou
s tumors and
also ap
plied t
he gra
d
ient a
nd Lapla
c
ia
n filters to redu
ce
noise. They
combi
ned
two-dim
e
n
s
iona
l wavel
e
t wit
h
erosi
on
an
d dilation
of
morphol
ogical
operators to
improve th
e
s
egm
entation
method
s. Cheng
et al
[9
], used fu
zzy logic to
det
ect
micro-cal
c
ification type
o
f
can
c
e
r
o
u
s lesi
on
s in
mammog
r
a
p
h
y imag
es.
Wan
g
et
al
[10],
prop
osed a
method to d
e
t
ect breast
cancer u
s
in
g
sup
port ve
cto
r
ma
chin
e (S
VM). Zhen
g
et al
[11], pre
s
ent
ed a
ne
w al
gorithm
co
m
b
ining
artifici
al intellige
n
ce techniqu
es and th
e
Discret
e
Wav
e
let Tr
an
sform
(D
WT
).
In this pa
per, a cellul
a
r l
earni
ng a
u
to
mata
algo
rith
m is u
s
ed to
extract feat
ure
s
an
d
detect can
c
erous le
sio
n
s.
2.
The Propo
sed Me
thod
The propo
se
d method is
shown
in the followin
g
step
s:
Figure 1. Pro
posed Frame
w
ork
for M
a
m
m
ogram Seg
m
entation
2.1. A Da
tab
ase of M
a
mmograms
The UK research group
has generated a
MIAS (Mammographic Image
Analysis
Society)
database of digital mammograms. T
he database contains left and right breast images
of
161 patients. Its quantity consists of 322 im
ages, which belongs to three types such as
Normal,
benign and
malignant. The database
has been reduced
to 200 micron
pixel edge, so
that all images are 1024 x 1024.
2.2. Image Denoising an
d Qualit
y
En
hanceme
n
t
of Image
Due
to low
contrast of mammogram
images, it
is
difficult to detect
signs such as
masses, so before doing the main operation of
image processing, the noise must be removed
and image must be enhanced. For this, a combinat
ion of morphological operators was used.
2.3. Backg
ro
und and Pec
t
oral Mus
c
le Remov
a
l
As
a preprocessing step, the breast area is
separated from the background image. This
saves the processing time and also the memory space.
The margin
of some
database images
c
ontains labels
and frills
and the
degree
of
brightness of lateral muscles
and the masses is
close together. Because of
these, in the
next
phase
of research, frills and lateral muscles of
the breast are detected and removed from the
image.
Because of
the large
size of
mammography
images,
a high
volume of
calculations
required to find
the damage
and lesion
areas, t
hus by
reducing the
area under
study, image
processing will be done more rapidly. so image sizes were reduced to 256* 256.
Then, using the threshold method and
“fi
nd” and “bwlabel” MATLAB functions,
labels
and frills of image are eliminated [12].
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Novel Ap
proach for Tum
o
r Dete
ction i
n
Mam
m
ography Im
ages
(Elahe Ch
agh
ari)
6213
2.4. Mass De
tec
t
ion
w
i
th
Cellular Lear
ning Automa
ta Algori
t
hm
The main idea for using CLA to segment r
egions is to use the adjacency relation
among
regions for better
segmentation. To
do so,
it
is assumed
that each
image pixel
is mapped
to a
cell in automata and each
LA related to its
pixel
in input image.
Then, a
dynamic structure
LA
with L
R
-
ε
P
learning algorithm [13] is allocated
to each cell of CLA and Moore neighborhood
is
considered
for the cells.
Each learning automaton ta
kes
eight actions. Each
action is related to
eight
neighbors of the
central LA in a
3×3 adjac
ency which has
a selection probability. Selection
probability of each action shows the similarity of
the central pixel to its neighbors. Selection
of
an
action by
central LA means
t
hat the central
pixel in input
image and the
selected neighbor lie
in the same region.
For receiving an
award from the
environm
ent, each selected
action increases
its
selection possibility for
the next steps
and reduces
the
other action selections
and for
receiving
penalty from the environment, it reduces its sele
ction possibility for the next steps and
increases
the selection possibility of the other actions.
A law which will be used to calculate the reward
or penalty is that at first the
brightness
distance of pixel in a cell to all its neighbors is calculated [14].
D
i
(x,y) =
|
I(x,y) – I(x
i
,y
i
)| i=
1..8
The m
ean
of
these
inte
rvals i
s
cal
c
ulate
d
an
d di
spl
a
yed
with
D
M
(x, y). The
la
w
appli
e
d
to this cell is:
C× D
i
(x
,y
)
≤
D
M
(
x
,
y
)
R
e
w
a
r
d
(
1
)
C× D
i
(x
,y
) > D
M
(x,y)
Penalty
(2)
After determi
ning the li
st
of neigh
bori
n
g cell
s, the d
y
e relea
s
e
o
peratio
n o
c
cu
rs th
ose
pixels lo
cate
d on
a
seg
m
ent are
clo
s
e
r
togeth
e
r in
term
s of
col
o
r. To
do
this, the foll
owi
n
g
formula i
s
used:
′
,
∑
,
|
,
|
∑
|
,
|
(
3
)
Where I is
the input
image,
′
is the
output soft segmented
image,
L(x,y)
is the
list of
chain
elements
corresponding to pixel at location
(x,y
), |.| is the number of
chain elements, L
i
is the ith
chain
element and w
i
is the
weight related to the it
h
chain element which determines the
effectiveness ratio of the ith chain pixel.
Weights wi are designed to have an descending
behavior.
This is true and significant that the in
itial chain elements is more important and must
have higher weights and
the final chain
elem
ents are less important
and must have
lower
weights.
Figure 2. Gen
e
ric P
s
eu
do-code for the CLA Algorithm
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 621
1 –
6216
6214
This cycle
continues until
a stop
condition. Fi
nally, using
a threshold,
the final
step of
integration phase was started and the mass
is separated from the segmented image.
Figure 3. Overall Structu
r
e
of
Propo
sed
Segmentatio
n Method
3. Experimental Re
sults
The
proposed method was
carried out on a
2.
67GHz processors with
4 GB RAM on
Windows 7
platform and
MATLAB R2011b have
been
used. The
proposed algorithm
was tested
on MIAS dataset. The Moore neighborhood with r
=1 has been considered for CLA cells.
In the case
of mammographic
image analysis
, the
results produced
using a
certain
method can be presented in a few ways.
The inte
rpretation being mostly used is the
confusion
matrix
.this matrix c
onsists
of true negative (TN),
false
positive (FP),
false negative (FN) and
true positive (TP).
There
are some often
mentioned terms such
as
accuracy, sensitivity,
precision. In this
research, sensitivity is used
for verify of proposed method.
Sensitivity =
(
4
)
In the followi
ng, results of
the propo
se
d
method im
plementatio
n and compa
r
i
ng it with
other meth
od
with the sam
e
image data
base are p
r
e
s
ented.
Table 1.
Result of Implementing the Propo
sed Meth
od on 25 Ima
ges of MIAS Datab
a
se
The numbe
r of i
m
ages has been
processed w
i
th 2
5
the prop
osed m
e
thod
.
TP
TN
FP
FN
25
20
2
2
1
Sensitivity =
20
21
=
0
.
9
5
(
5
)
Table 2.
Re
sult of Implementing the Propo
sed Meth
od on 25 Ima
ges of MIAS Datab
a
se
Methods
Au
t
h
o
r
Year
Sensiti
v
i
t
y
A Comparison of
Different
Gabo
r
features for M
a
ss Classif
i
cation
in Mammograph
y
Hussain et
al.(15)
2012
92
Proposed Metho
d
2014
95
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Novel Ap
proach for Tum
o
r Dete
ction i
n
Mam
m
ography Im
ages
(Elahe Ch
agh
ari)
6215
Origin
al Imag
e
Den
o
ised Image
Pectoral muscle re
moval i
m
age
Mass Dete
cte
d
Figure 4. Re
sult of the Proposed Metho
d
4. Conclusio
n
In this pa
per, cellula
r lea
r
ning
autom
a
t
a algorith
m
is presented
for mamm
o
g
rap
h
y
image p
r
o
c
e
s
sing
and d
e
te
ction of ma
sses. The
pu
rp
ose
of this re
sea
r
ch is to
p
r
esent a meth
od
for ea
rly dete
c
tion of b
r
e
a
s
t ca
ncer
an
d also in
d
e
p
ende
ntly of the radiolo
g
ist
in dete
c
ting
and
redu
cin
g
the
num
ber of f
a
lse
dete
c
tio
n
s
(e
sp
eciall
y red
u
cin
g
th
e nu
mbe
r
of
false
ne
gati
v
es,
whi
c
h ha
s a h
i
gh co
st due t
o
exclud
e
the
patient from treatme
nt cycl
e).
To evaluate the pro
p
o
s
ed
method, 25 I
m
age
s of MIAS database
were a
nalyzed, work
output sho
w
s the Sen
s
itivity of 95 pe
rce
n
t. Due to
th
e high
Sen
s
itivity and low
numbe
r of fal
s
e
negative
s
in the dete
c
tion, t
he result is a
c
ceptabl
e.
Studies
sho
w
that using i
n
telligent meth
ods fo
r imag
e pre
p
roce
ssi
ng can imp
r
o
v
e the
outcom
e
.
So the
next step in
continuin
g
this
re
sea
r
ch
can
be
prese
n
ting a
n
opti
m
ize
r
al
gorith
m
to enha
nce i
m
age
quality
and
better
removal of th
e pe
ctoral m
u
scled i
n
the
mammo
gra
m
image
s.
Referen
ces
[1]
Sch
w
artz. Sch
w
a
r
tz'
s
princi
pl
es of surger
y. Editor. 201
0; 8.
[2]
M HH. Compa
r
ing mammo
gr
aph
y a
nd so
n
ogra
p
h
y
rep
o
rt
s
w
i
th p
a
tho
l
o
g
y
re
ports of malig
na
nt and
ben
ign breast dise
ase.
Breas
t Diseases.
20
09; 2(2).
[3]
Joche
l
son M.
Advanc
ed Ima
g
in
g T
e
chniq
u
e
s for the Det
e
ction of Bre
a
st Canc
er.
Ameri
c
an Soc
i
ety of
Clin
ical Onc
o
lo
gy
. 2012.
[4]
I Brodie RAG. Radi
ogr
ap
hic i
n
formatio
n
the
o
r
y
a
nd a
p
p
lica
t
ion to mammo
grap
h
y
. Me
d Ph
y
s
. 1
982; 4.
[5]
Jingl
un S
h
i Z
S
. Image reso
lutio
n
en
ha
nc
ement us
ing
statistical esti
mation
in
w
a
velet d
o
mai
n
.
Bio
m
e
d
ica
l
Signa
l Processi
ng and C
ontr
o
l
. 201
2.
[6]
Cha
ngiz
i
V G
M
, Arab Kh
er
adma
nd A
ppl
i
c
ation
of com
puted
ai
ded
d
e
tection
in
br
east masses
dia
gnos
is.
Indi
an Jour
nal
of Cancer
. 45(
4).
[7]
S Saheb Bas
ha
DKSP. Automa
tic Detection Of Breast
Cancer
M
a
s
s
In Mamm
ograms Using
Morph
o
lo
gica
l
Operators And F
u
zz
y
C-
Means C
l
uste
ring.
Jour
nal
of T
heoretica
l
and Ap
pli
e
d
Information T
e
chno
logy
. 2009.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 621
1 –
6216
6216
[8]
Arian
na M
enc
attini MS. Ma
mmogra
phic I
m
ages E
n
h
a
n
c
ement a
n
d
D
eno
isin
g for B
r
east Ca
nce
r
Detectio
n Us
in
g D
y
a
d
ic Wav
e
let Proc
essi
n
g
.
IEEE trans
actions
on ins
t
rum
e
ntation m
e
asur
em
ent
.
200
8; 57(3).
[9]
Hen
g
-Da C
h
e
ng YML. A Novel A
ppr
oa
ch to Microc
alcificati
on D
e
tection Usi
n
g
F
u
zz
y
L
ogi
c
T
e
chnique.
IE
EE transaction
s on medic
a
l i
m
a
g
i
n
g
. 19
98; 17(3).
[10]
Defen
g
W
a
n
g
L, Phe
ngA
n
n
He
ng. Autom
a
tic
detecti
on
of breast c
anc
ers in m
a
mmo
grams us
in
g
structured support vector machines.
Neuroc
omputi
ng.
20
0
9
; 72.
[11]
Cha
n
LZ
aAK.
An Artificial I
n
telli
ge
nt Alg
o
rithm for T
u
mor Detectio
n in S
c
reen
ing M
a
m
m
ogram.
IEEE
transactio
n
s on
med
i
ca
l imag
i
n
g
. 201
1; 20(7)
.
[12]
R Su
bash
Ch
andr
a Boss
K
T
, D Arul P
o
n
Dan
i
e
l
. Auto
matic Mammo
gram im
ag
e B
r
east R
egi
o
n
Extractio
n
an
d
Remov
a
l of Pe
ctoral Muscl
e. 201
3.
[13]
Me
ybo
d
i MRB
H
. Cellu
lar L
e
a
r
nin
g
Automata
and It'
s Applic
ations. 20
03.
[14]
AA Abin MF, S Kasaei.
Ski
n
Se
gm
en
ta
ti
on b
a
s
ed
o
n
C
e
llu
l
a
r Le
a
r
n
i
ng
Au
tom
a
ta
.
6th
Internatio
na
l
Confer
ence
on
Advances i
n
Mobil
e
Com
put
ing & Multim
ed
ia (MoMM20
0
8
)
. 2008.
[15]
Muhamm
ad H
u
ssai
n
SK, Ghul
am Muh
a
mmad, Georg
e
Bebis.
A Co
mparis
on of Diff
e
rent Gab
o
r
features for Mass Class
ificati
on in Ma
mmo
grap
hy
. Eighth
Internatio
nal
Confer
ence
on
Signa
l Imag
e
T
e
chnolog
y
an
d Internet Base
d S
y
stems. 20
12.
Evaluation Warning : The document was created with Spire.PDF for Python.