Tumor extraction and elimination of pectoral [601304]

Tumor extraction and elimination of pectoral
muscle based on hidden Markov and region
growing: applied based MIAS

SOUKAINA EL IDRISSI EL KAITOUNI
LIIAN, dept of Informatics Faculty of Science Dhar -Mahraz
University of Sidi Mohamed Ben Abdellah
P.B 1796 Atlas -Fez, Morocco
[anonimizat]

ABDELGHAFOUR ABBAD
LIIAN, dept of Informatics Faculty of Science Dhar -Mahraz
University of Sidi Mohamed Ben Abdellah
P.B 1796 Atlas -Fez, Morocco
[anonimizat]

HAMID TAIRI
LIIAN, dept of Informatics Faculty of Science Dhar -Mahraz
University of Sidi Mohamed Ben Abdellah
P.B 1796 Atlas -Fez, Morocco
[anonimizat]

Abstract— In this article, there we propose an automatic
method for the detection and extraction of the tumor on
mammogram images. Most methods of detection of a tumor
require the extraction of a large number of texture features from
multiple calculations. The study first examines a technique of
pre-processing images to obtain the Otsu thresholding method to
eliminate items that do not belong in. Aft er performing the
thresholding, we estimate the number of base classes of technical
LBP (Local Binary Pattern). To automate the initialization task,
the classification proposed by applying dynamic k -means and
improve the classes obtained by the method of Markov. Then we
calculate the correlation between these classes and the original
image, we deduce the class that contains the tumor and muscle
pectoral. Finally, it uses the method of growing the region to
eliminate pectoral muscle. The result obtained by this approach
shows the quality and accuracy of extracting parts of the tumor
compared to existing approaches in the literature.

Keywords —Classification; tumors; mammogram image; Otsu
thresholding ; LBP (Local Binary Pattern); k -means; Markov.

I. INTRODU CTION
Breast cancer is the most common cancer in women in the
world. It is the leading cause of death among women due to
the lack of early diagnosis. Mammography screening is the
most adopted technique to performing early detection of breast
cancer. In ma mmography images, suspected breast cancer
appears as white spots. Breast density, the presence of tags,
artifacts or even pectoral muscle decrease the sensitivity of
mammography.
In the literature, several stud ies have been developed for the
detection of r egions of interest (ROI) in mammograms. Among these studies, In [1], we find the study of Roula Alayli
using the thresholding algorithm for the detection of breast
cancer. This technique poses the problem of defining the
threshold. Gumaei Abdo et al. [2] proposed a method based on
K-means with a mixture of gamma distributions; Nalini Singh
et al. [3] have used K -means and Fuzzy C -means for the
detection of mass center in mammography; Siddheswar. R and
RH Turi [4] proposed a method based on the determinati on of
the number of class colored images from the K -means method.
Approaches based on K -means algorithms and Fuzzy C -means
have the disadvantage of initializing cluster number and
centers. Elmoufidi [5] chose to combine the LBP (Local
Binary Pattern) and d ynamic k -means algorithm. Nagi et al.
[6] used as a preprocessing morphological and growing seeded
area to detect the pectoral muscle. Liu et all [7] using GVF
snake algorithm for Extraction of extrapolated breast object.
Mustra and Grgic [8] proposed an A daptive histogram
equalization and polynomial curvature estimation. Agrawal,
Vatsa, and Singh [9] proposed Saliency maps for ROI
segmentation and the ROIs classification using entropy
features. These techniques can be cited along other ones
[10, 11, 12, 1 3].
All the works that have been cited previously do not address
efficiently the following issues:
 Unwanted areas.
 Automatic detection of tumor class.
 Tumor and pectoral muscle extraction.
In this paper, we propose new techniques which address the
problems cited in the previous paragraph. Our idea is outlined
as follows: we start with an Otsu’s thresholding method. Next,
an image classification by estimating the number of classes
based on LBP (Local Binary Pattern) Technique. To automate
the initialization task, we have proposed to apply the

classification by k -means dynamic improved by Markov
method. The tumors image is the result of the maximum
correlation.
The rest of this article is organized as follows: In section 2, we
present used methods. In Section 3, we describe proposed
approach. The results along with discussions are presented in
Section 4 and the last section is dedicated to the conclusion .
II. DATA BASE
To test our approach, we have used the mini -MIAS database
[14].This database contains 322 digital mammograms images
of the size 1024 * 1024 pixels and of the PGM type, these
images are in grayscale with a pixel intensity of the interval
[0,255], acquired mammogram images are classified into three
major cases: normal, benign and malignant. The Fig 1 sh ows
the various components of an image of the base used.

Fig. 1 :Example of an image of the mini -MIAS base
III. USED METHODS
A. Otsu method
The principle of Otsu method is to find an optimal threshold
that maximizes the difference between two classes [15]. It is
performed based on the variance. The optimal threshold
optimalS
is one that maximizes the following functions :

2
2B
Wttt

2
2B
Ttt

2
2T
Wttt

If
t is chosen, then
[min,max] optimal tS argmax t  ò
(1)
Where
2 2 2,,T B W   are successively the total variance of
the image, the inter -class variance (between -class variance)
and intra -class variance (within -class variance) .
2 2 2( ) ( ) ( ) B T Wt t t  
(2)
max2 2
minTTt i m (3)
max
min*Ti
im i p

: The total average of all the image points
2 2 2**B font font objet objett P t t P t t  
(4)
ip
: The probability of occurrence of the gray level i in the
image.
of pixels whose gr y level ( )
*inumber a i h ipnumbre of pixels in image M N
(5)
 ,font objetP t P t
: The sum of the probabilities of
occurrence of gray levels of pixels of the background and that
of the object by taking the threshold t.
  max
min 1, 1 t
objet i font i objet
i i tP t p P t p P t
     
(6)
,font objetmm
: The average of the pixels belonging to the
background and that of the pixels of the object.
 max
1 min* *
,t
i
it ii
objet font
objet fontip ip
m t m tPP  
(7)
22,font objettt
: The variance of the class background
and the variance of the class object.

2
2 min
max
2
2 1( ) *
,
( ) *t
objet i
i
objet
objet
font i
it
font
fonti m P
tP
i m P
tP






(8)
[min, max] is the dynamic range of the image.

B. LBP (Label Binary Pattern)
The descriptor LBP (Local Binary Pattern) was proposed by
Ojala et al [16,17] in 1996 for the texture classification.
We consider an image
I(x, y) and
cg representing the gray
level of the central pixel
x, y . Moreover,
pg the gray value
of its neighbors and P represents the total number of neighbors
concerned and R is the radius of the neighborhood:
( ) , 0,……, 1 p p pg I x y p P   
(9)

2cos( ) ppx x Rp
(10)

2sin( ) ppy y Rp
(8)

LBP operator is defined as follows:
1
,
0( )2 p
p
P R p c
pLBP S g g

(12)

The thresholding function
Sx is defined by:
1, 0 0, 0xSxx
(13)

C. K-means
k-means is the simplest unsupervised learning algorithm that
solve the problem of classification.
k-means is to minimize the sum of squared distances between
all the points and the class center [18].
2
11 ick
ji
ijJ V x v

(14)
Where:

 K : Is the number of cluster centers;

ic : Is the number of data points in
thi cluster;

jixv : Is the Euclidean distance between
jx and
iv
;

iv : Is the mean of
thi in
ic during each iteration; it is
as Follows :
1 ic
i
j
j
i
ix
vc
(15) Let
 1 2 3, , ,……,n X x x x x be the set of data points and
 1 2 3, , ,……,c V v v v v
be the set of centers.
The main steps of the method "K -means" can be summarized
as follows:
 Randomly selecting K objects.
 Assigning each object to the nearest class, each of
these classes is characterized by a center.
 Calculate the new representatives for classes.
 Repeat 2 and 3 until the centers cease moving.

The intra -cluster distance : is the sum of squared distance
from all points to their cluster centers (see equation 16).
2
ji
111intra cluster xN ic k
ijv
  
(16)
Where : N is the number of pixels in the image, k is the
number of clusters, and
iv is the cluster centre of cluster
ic .
The inter -cluster distance : is the distance between cluster
centers (see equation 17).

2
ji inter cluster min v v    (17)
where: i = 1,2,….,k -1 and j = i+1,….,k .
intra clusterRatiointer cluster
(18)

D. Hidden Markov
The hidden Markov models (Hidden Markov Models or
HMM) model random phenomena that are assumed to
comprise a first level of a random process of transition
between unobservable states (hidden states) and on second
level, other random process in each state generates observable
values . Assume that Z is a 2D gray-level matrix (M*N). The
T
iZ
denotes the intensity measurement at pixel
i . Given an
image
 12, ,…,N y y y y . Each
iy associated with pixel
i
is an unknown class label
ixL where
L is regarded as the
set of all possible labels. The Gaussian Hidden Markov
Random Field (HMRF) can be specified as :

ii i N i i N
lLp y | x ; θ g y;θ q(l|x )

(19)

Where
1 2 N x (x , x , , x ) ,
ilg y ;  is a Gaussian
probability density function with parameter
l l l2 , µ
and
i q l | x N is a conditional probability mass function for
the class label l.
We use the MAP and EM algorithm to estimate the parameter
set x and θ.
 MAP algorithm

We seek a labeling of an image, which is an estimate of the
true labeling , according to the MAP criter ion:
 
xχ x χarg max max p y|x; θ f(x) X

(20)
It is assumed that
iy and
ix are pair -wise independent so
N
ii
i1p y|x;θ p y |x

(21)
and the probability density function for x is the so -called
Gibbs distribution (proposed by Geman and Geman [19]) is
given by:
U(x)1f x eZ
(22)
where Z is a normalizing constant called the partition
function, and
Ux is an energy function given by the form:
c
cCU x V (x)
ò
(23)
where
cV (x) is the clique potential and C is the set of all
possible cliques (see more details in [20]). In this paper, it is
assumed that each pixel has at most 4 neighbors in the image
domain. Then, on pairs of neighboring pixels, the clique
potentials is calculated by :
ij c i j x ,x1V x , x (1 I )2
(24)
The MAP estimation is equivalent to minimizing the posterior
energy function
 xχ x χarg min min U y|x U(x)X

(25)
where
i
i
i2
ix 2
x 2
i xyμ 1U y|x log log σ2σ2
 For
solving the MAP problem we can use the same approach
proposed in [22] .
 EM algorithm
We use the EM algorithm to estimate the parameters θ. Below,
it is briefly explained:
– At the kth iteration, we have
k , and We compute the
EM functional :
  kkQ(θ|θ ) E loglogp(y,x;θ)|y,θ  (26)
– For obtaining the next estimate we maximize the EM
functional.
  k 1 k
θθθ argmaxmaxQ(θ|θ )
(27)
More details can be found in [21, 22].
E. Cross -Correlation
The cross -correlation measurement normalized centered,
noted ZNCC (Zero mean Normalized Cross -Correlation) is
given by:
 .
( , )
.gd gd
gd
gd gdf f f f
ZNCC f f
f f f f


(28)
( , )gd ZNCC f f
Values belong to the
1,1 interval. This
measure corresponds to the coefficient of linear correlation
classic statistics. This measurement is one of the most used,
particularly in [23]. It has the advantage of exhibiting gain and
bias type of invariance.
IV. PROPOSED APPROACH
The problem we want to solve in this articl e is how to extract
and detect the tumor region in mammograms images; Fig 2
shows the proposed approach
Our approach is based on the following steps:
Step1: The preprocessing phase: (This phase is applied on all
the base images) Applying a pretreatment on each image of
the database MIAS using Otsu's method for the binarization
and the removal of unwanted areas. Obtained images are
stored in a new base (treated MIAS).
Step2: The number of recovery phase of average classes: In
this step, we want to recover t he average number of classes
from the MIAS treated base to utilize it in the next phase. To
this end, we apply the LBP method.
Step3: The recovery phase of the number of optimal classes
To extract the optimal number of classes, we use the algorithm
propose d in [6], this algorithm take as parameters input an
image and the number of average classes recovered in the
previous phase and as output, the number of optimal class of
input image.
Step4: The extraction phase of the classes After the recovery
of the opt imal number classes, this number is used to initialize
the k -means algorithm, as a result of the application of this
algorithm; we obtain a set of images, each representing a
class.
Step5: The adjustment classes phase To get a good
classification, we adjus ted the classes obtained in the previous
phase using the method of Markov. To make this adjustment,
this method that uses the original image as a reference image
to correct the classes obtained in the extraction phase of the
classes.
Step6: The selection p hase of the tumor class In this step, we
want to choose the class that contains the tumor in an
automatic manner, to this end; we compute the correlation
class obtained in step 5 for each image of the database

MIAS_treated. After observing results, we obse rve that for
each image, a class having a high cross -correlation represents the tumor class table1. So later this criterion is used to select
tumor class.

mdb017 mdb081 Mdb244
Class 1 NaN 0.49818352 NaN
Class 2 0.53845742 NaN 0.67249959
Class 3 0.19264549 0.43255219 0.44819682
Class 4 0.75923741 0.27170504 0.31612903
Class 5
0.12294306 0.16887433
Class 6
0.40209193 0.073541461
Class 7
0.029786986 0.0079965883

Table .1: the values of the correlation classes
Well Noted : NAN meant that the denominator used to calculate the correlation is zero

Step7: The pectoral muscle elimination phase
Most of the tumor classes selected contain several objects,
these objects repre sent parts of tumors and another part
represent (pectoral muscle); so to distinguish the tumor objects
in this phase we want to eliminate the pectoral muscle. In
order to do this, we apply the method applied in Growing
region. It begins with the starting p ixel research of this
method. This pixel is found either in the left or right corner of
the image. After the elimination of the muscle, we clearly see
that the objects that remain in the tumor class represent only
the tumor.

Fig 2: Proposed approach

V. RESULT AND DISCUSSION
A. Result

mdb017 mdb081 mdb244
original
image

Image after
preprocessing

tumors
extracted by
k means
clustering [6]

The classes
tumors
extracts by
Markov

The final
classes
tumors

The final
result

Fig. 3 Examples of the results obtained on some images
B. Discussion
We used the MIAS basis to test our approach, each image
contains a defect, we have information about the center and an
approximation on the circle radius around the anomaly
presented.

In Fig.4, we presented a comparison between the results
obtained by our approach and the information provided by the
MIAS base in order to mount the effectiveness of proposed
solution.
Result
Obtained

Result
Obtained
+ circle

Fig .4 comparison between results obtained and the information provided
with the base
VI. CONCLUSION
In this article, we have proposed a method for
classification and automatic detection of the tumor on
mammogram images. To improve the quality of detection of
the tumor, first, we presented a technique of preprocessing to
remove objects that not belong to the breast through the Otsu
method. After the preprocessing step, we estimated the
number of classes based LBP Technique
(Local Binary Pattern). Then we pe rformed a classification
from k -means and we improved the classes obtained with this
method based on the method of hidden Markov. Finally, we
calculated the correlations between these classes and the
original image to detect automatically the class that co ntains
the tumor and the pectoral muscle and eliminate it, we applied
the region growing method.
References:
[1] M. A. Roula, and A. Y. El -Zaar, An Iterative
Mammographic Image Thresholding Algorithm For
Breast Cancer Detection, ACIT’2013, Oman, 10 -13
Dec., 2 013.
[2] J. Nagi, S.A. Kareem, F. Nagi, S.K. Ahmed, Automated
breast profile segmentation for ROI detection using
digital mammograms, IEEE, EMBS Conference on
Biomedical Engineering & Sciences, 87 –92, 2010.
[3] Gumaei, Abdu, et al. Breast segmentation using k -mean s
algorithm with a mixture of gamma distributions, IEEE
Symposium on Broadband Networks and Fast Internet
(RELABIRA),97 -102 , 2012.
[4] Singh, Nalini, Ambarish G. Mohapatra, and Gurukalyan
Kanungo, Breast cancer mass detection in mammograms
using K -means and f uzzy C -means
clustering, International Journal of Computer
Applications, (0975 –8887) 22.2 (2011).
[5] Bon, Abdul Talib. Developing K -Means Clustering on
Beltline Moulding Contours, Journal of Applied Sciences
Research, (2189 -2193). 5.12 (2009).
[6] Elmoufidi, Abdelali, et al. Detection of regions of
interest's in mammograms by using local binary pattern,
dynamic k -means algorithm and gray level co -occurrence
matrix, IEEE Fifth International Conference on Next Generation Networks and Services (NGNS), 118-123,
2014.
[7] Liu, C. C., Tsai, C. Y., Tsui, T. S., & Yu, S. S. (2012).
An improved GVF snake based breast region
extrapolation scheme for digital mammograms. Expert
Systems with Applications, 39, 4505 –4510.
[8] Mustra, M., & Grgic, M. (2013). Robust automati c breast
and pectoral muscle segmentation from scanned
mammograms. Signal Process, 93, 2817 –2827.
[9] Agrawal, P., Vatsa, M., & Singh, R. (2014). Saliency
based mass detection fromscreening mammograms,
Signal Processing, 99, 29 –47.
[10] Chan, T. F., & Vese, L. A. Active contours without
edges, IEEE transactions on Image processing, 266 -277.
10(2) (2001).
[11] Nagi, Jawad, et al. Automated breast profile
segmentation for ROI detection using digital
mammograms, IEEE EMBS Conference on Biomedica l
Engineering and Sciences (IECBES),87 -92 , 2010.
[12] Gumaei, Abdu, et al. Breast segmentation using k -means
algorithm with a mixture of gamma distributions, IEEE
Symposium on Broadband Networks and Fast Internet
(RELABIRA), 97 -102, 2012.
[13] Mohanty, Aswini Kumar , et al. Detection of Masses
from Mammograms Using Mass shape
Pattern, International Journal of Computer Technology
and Applications ,2.4 (2011).
[14] Suckling, John, et al. The mammographic image analysis
society digital mammogram database, Exerpta Medica,
International Congress Series. Vol. 1069. 1994.
[15] Otsu, Nobuyuki, A threshold selection method from
gray-level histograms, Automatica , 23-27, 11.285 -296
(1975).
[16] Ojala, T., Pietikäinen, M., & Harwood, D , A
comparative study of texture measures with classific ation
based on featured distributions, Pattern
recognition, 29(1), 51 -59. (1996)
[17] Ojala, T., Pietikainen, M., & Maenpaa, T.
,Multiresolution gray -scale and rotation invariant texture
classification with local binary patterns, IEEE
Transactions on Pattern An alysis and Machine
Intelligence, 24(7), 971 -987, (2002)
[18] Ray, Siddheswar, and Rose H. Turi, Determination of
number of clusters in k -means clustering and application
in colour image segmentation,Proceedings of the 4th
international conference on advances in pattern
recognition and digital techniques,137 -143, 1999.
[19] S. Geman and D. Geman, Stochastic relaxation, Gibbs
distributions, and the Bayesian restoration of images,
IEEE Trans. Patt. Anal. Machine Intell., vol. PAMI -6,
no.6, pp. 721 –741, nov. 1984
[20] S. Gem an and D. Geman, Stochastic relaxation, Gibbs
distributions,and the Bayesian restoration of images,
IEEE Trans. Pattern Anal. Machine Intell., no. PAMI –6,
pp. 721 –741, June 1984.
[21] Zhang, Y., Brady, M., & Smith, S. (2001). Segmentation
of brain MR images thr ough a hidden Markov random

field model and the expectation -maximization algorithm
, IEEE Transactions on Medical Imaging, 20(1), 45 -57.
[22] Y. Zhang, M. Brady, and S. Smith, Segmentation of
brain MR images through a hidden Markov Random
Field model and the Ex pectation -Maximization
algorithm, IEEE Trans. on Medical Imaging, vol. 20,
no.1, pp. 45 – 57, 2001.
[23] Aschwanden, P., and W. Guggenbuhl, Experimental
results from a comparative study on correlation -type
registration algorithms, Robust computer vision, 268 –
289, (1992). [24] Jen, Chun -Chu, and Shyr -Shen Yu, Automatic detection
of abnormal mammograms in mammographic images,
Expert Systems with Applications,42.6 (2015): 3048 –
3055