International Conference on Microwave, Optical and Communcation Engineering, December 18-20, 2015, IIT Bhubaneswar, India

Novel Technique for the Detection of Abnormalities

In Mammograms using Texture and Geometric
Features
Spandana Paramkusham* ,K.M.M.Rao, B.V.V.S.N.Prabhakar Rao
EEE department,Bits-Pilani, Hyderabad Campus,Hyderabad,India
[email protected]

Abstract— The paper investigates on recognition of breast malignancy in the breast mammograms. According to the
abnormalities. A novel feature frame work was proposed on Breast Imaging Reporting and Data Systems (BIRADS),
mammographic patches based on both texture and geometric masses are space occupying lesions seen in at least two
features for classification of breast tissues into normal, different projections. They are characterized by their shape
malignant and benign. The methodology comprises of five
(round, oval, lobular, irregular) and margins (circumscribed,
stages. First step is preprocessing, texture feature extraction
using Local quinary pattern for classifying breast tissues into microlobulated, obscured, indistinct, spiculated).
normal and abnormal, Automatic segmentation of mass using This paper explores use of both texture and geometric
k means algorithm, a new geometric feature descriptors features for classifying the ROI into normal, benign,
extraction to classify them into benign and malignant and two malignant. From past few years many algorithms were
stage classification. Our feature extraction method attained proposed for three class classification methods based on
99.27 for normal and abnormal, 79.41% for benign and texture features. The novelty of this paper lies on choosing
malignant and over all accuracy for three class classification is texture features for classification of normal and abnormal
89.05%. and new geometric features based on boundary for benign
Keywords—Breast cancer, Feature extraction, classification,
and malignant. An ensemble of local ternary pattern and
benign, maligant. local phase quantization texture descriptors were used to
classify benign and malignant breast tissues in [3]. In [4] K-
means algorithm for image segmentation and haralick
I. INTRODUCTION features are used to classify mass and non mass. Multi
resolution analysis texture descriptors were extracted [5]
Breast cancer is the most common cancer in women all based on different wavelets and classified into mass and non
over India and accounts for 25% to 31% of all cancers in mass using polynomial classifier. Variogram function
women in Indian cities [1]. According to WHO for the year texture descriptors were used to detect mass or non-mass
2012, an estimated 70,218 women died in India due to [6].Two stage classification is done using Spherical wavelet
breast cancer, more than any other country in the world. transform [7] to distinguish mass and non mass in first
Computer aided detection is very important to help stage ,ROI basic geometric features with spherical wavelet
radiologist in detecting abnormalities. It aids in accurate coefficients are used to distinguish between benign and
diagnosis to prevent fatality. Digital mammography is an malignant in second stage. Zernike moments are used to
important tool to detect breast cancer in early stages. It has classify shape and margin descriptors of benign and
X-ray device which gives mammogram. Mammograms malignant masses [8]. Joberth de et al [9] proposed
show abnormal lesions in the initial stages. These help to qualitative thresholding for segmenting mass region and
prevent incorrect decisions, unwanted investigations, and used correlogram function, haralick features, geometric
inadequate surgeries, which directly affects the outcome and features to distinguish between normal and abnormal
longevity of the patient. It also helps radiologists to detect regions. Two new features (correlation coefficient and
breast cancer easily in less time and forbid unnecessary mean absolute error) were extracted from 1D boundary of
biopsies. Mammograms show small difference between lesion using polynomial regression [10].
normal and abnormal tissues due to X-ray permeability. In this paper, we proposed two stage classification, where
Image processing algorithms help to assess the abnormality first stage of classification distinguishes normal and
of tumour in mammogram.Masses are the most common abnormal patches using texture features and second stage
asymmetric signs of cancer and appear brighter than the classification distinguishes benign and malignant by
surrounding tissue [2]. Most benign masses possess well- analyzing the boundaries. It also involves automatic mass
defined sharp borders, while malignant tumors often have segmentation using k-means algorithm for extraction of
ill-defined, microlobulated, or spiculated borders [10]. boundaries. The flow chart of the proposed method is shown
Boundaries of these masses are the prominent signatures of in Fig.1.

978-1-4673-6981-7/15/$31.00 ©2015 IEEE 150 ICMOCE-2015

 International Conference on Microwave, Optical and Communcation Engineering, December 18-20, 2015, IIT Bhubaneswar, India

operator labels sign gray levvel differences for all the pixels
by thresholding all the valuees with centre pixel value in a
given neighbourhood. Wherre P corresponds to number of
II. METHODS AND MATERIALLS neighbours, R correspondss to spatial resolution of the
The mammogram patches that we havee used in the operator.
experiment are taken from IRMA database. IRMA database Let gc represents centre pixel
p and g1, g2,..............gP-1
has normal, benign, and malignant masss patches taken represents neighbouring pixeels as in (1).
from DDSM, MIAS, LLNL, IRMA. We W have used ௣ୀ௉ିଵ
mammographic patches from DDSM. ‫ܲܤܮ‬ሺǡ ሻ ൌ σ௣ୀ଴ ‫ݏ‬
‫ݏ‬ሺ݃‫݌‬ െ ݃ܿሻ2p (1)

A. Preprocessing Where s(x) = 1 for, x>=

=0
At first the image is preprocessed by median
m filter of = 0 for, x<0
size 5X5 to remove noise from mammograam patches. The
A Local quinary pattern is an
a extended version of the Local
operations after this stage are divided into two stages. The
binary pattern, which reppresents gray-level differences
first stage includes extraction of texture features
f directly
between the pixels by a quuinary value instead of a binary.
from grayscale mammographic patch. In thhe second stage The signed gray level differrence s(x) can take 5 values by
mammographic patch (ROI) is segmentedd automatically assuming two thresholds t1 and
a t2 as in (2).
using k-means algorithm and intensity feaatures to extract
boundary and shape of the mass. S(x, t1, t2) = 2 x • t2
=1 t1<x<t2
=0 -t1” x”t11
ROI Enhancement with Median Fiilter = -1 -t2”x<-t1
= -2 x<-t2 (2)
Quinary coded labels are asssigned to each image pixel to
form four binary patterns .T
The histograms computed from
Texture Feature Automatic these four label images aree concatenated to form feature
Extraction using vector.
ntation
Segmen
Local Quinary C. Automatic Segmentation of
o mass from ROI
Pattern
Bounddary This method takes mammogrraphic patch as input and gives
Extracction binary mass and boundary as output without any manual
intervention. Fig.2a, 2b, 2c shhows the output of this method.
SVM Classification The steps that are involved inn this process are
Normal/Abnormal Intensityy
featuree 1) Calculate mean (M) and standard
s deviation (STD) of the
Extractioon image
2) A 3X3 window is movedd on the image pixels to select
1D Sign
nature pixel, whose mean value iss greater than M and standard
deviation is less that STD. As mass in ROI have high
intensity and less standard deeviation.
3) K- means algorithm witth fixed number of clusters is
Geometric Featture applied to the output of stepp 2 which divides the image into
Extraction different regions.
4) Mass cluster is selectedd automatically based on two
intensity properties (Third Moment and entropy). As the
SVM Classification size of mass is a big, smalll region whose area is less that
(Benign/Malignant) 100 pixels are removed fromm mass cluster using connected
component analysis.

Fig.1. Flow chart of the proposed method

= ͳȀܰ σேିଵ
௦ୀ଴ ሺ‫ܭ‬ሺܵሻሻ (3
2
3)
B. Texture Feature Extraction with Local Quuinary Pattern
This method is extension of Local binary pattern
proposed in [11]. A Local Binary Patteern (LBP) is a
powerful local texture descriptor operaator. The LBP
Fig.2.a) Mass b) Binnary mass c) Boundary

151 ICMOCE-2015
 International Conference on Microwave, Optical and Communcation Engineering, December 18-20, 2015, IIT Bhubaneswar, India

D. Geometric Feature Extraction malignant. Second stage classsifier takes abnormal patches as
input from first stage classifiier. We obtain the result of three
Five features are extracted from 1D signnal of boundary class classification by connsidering the outputs of two
and two features from 2D boundary. classifiers. SVM, Naive baayes classifier, Self organizing
1D Geometric features: map neural network is used forf these classifications.
a) Correlation coefficient and mean squarre error: These III. RESULTS
R
features are extracted by fitting polyynomial on 1D 137 mammographic patchhes were taken from IRMA
signature of the boundary as in [10]. A siggnature is a 1D database of size 128X128. Inn those images 68 patches have
functional representation of a boundary. Fiig.3a shows the mass or abnormality where as other 69 images are normal.
boundary signature of above mass shown inn Fig.2c. It is a The following values have been used during the experiment.
plot of the distance from the centroid to thee boundary as a No of clusters in k means is equal to 3. The t1 and t2 values
function of angle [13]. in Local quinary pattern are 3 and 5.
b) Bending energy: The Average Bending Ennergy is defined A. Two class classification sttudy (Normal/Abnormal):
as
SVM classification is carriedd out via 5 fold cross validation
BE = ͳȀܰ σேିଵ
௦ୀ଴ ሺ‫ܭ‬ሺ‫ݏ‬ሻሻ
2
(3) and Least square SVM withh linear kernel, which gave good
result of 99.7%.The inputt to this classifier is texture
descriptors taken from LQP P. Tabel I shows results of SVM
classifier by considering diifferent local rotation invariant
where K(s) as in (3) denotes the curvaturre of the shape texture descriptors (LBP, LTP, LQP). Table II gives
parameterized by arc length . N is the numbeer of points on a confusion matrix for two claass classification of normal and
contour and s is the arc length parameter [12]. abnormal patches using SVM M which shows that detection of
abnormal patches is 100%.
c) Fractal dimension of 1D signature usinng box counting
TA
ABLE I
method.
Accuracy
d) Number of concave points to No of convvex points Ratio
with Triangular area Representation (TAR): k LBP LTP LQP

TAR signature computed directly from arrea of triangles 1 fold 86.13 91.97 99.27
formed by points on shape boundaary [12]. Let
Pn = ( Xn, Yn) denotes boundary points. TheT area formed 2 fold 86.13 94.89 99.27
by any two points on boundary and centroid gives triangular 3 fold 86.13 96.35 99.27
area. As boundary point changes, area also varies and this
function is called triangular area function. Negative TAR 4 fold 86.06 96.35 99.27
represents concave points, Positive TAR reppresents convex 5 fold 84,67 96.35 99.27
points. Fig .3b is TAR of the boundary show
wn Fig.2c.
LBP: Local binary pattern, LTP: Loocal ternary pattern, LQP: Local quinary
pattern
ABLE II
TA
Normal Abnormal
68 0
1 68

Fig.3. a) 1D Signature b) TAR

T
150
N/AB
e) Perimeter: Length of boundary.
100
2D Geometric Features: B/M
f) Compactness 50
g) Eccentricity The eccentricity is the ratio of the distance
between the foci of the ellipse and its major axis
a length. 0
SVM Naïve Bayes
B SOP
E. Classification
The proposed method consists of two classsification stages. Fig.4.Plot of accuracy with
w different classifiers for
First stage of classification distinguishees normal and N(Normal),AB(Abnorm mal),B(Benign),M(Malignant)
abnormal patches, second stage classifiees benign and

152 ICMOCE-2015
 International Conference on Microwave, Optical and Communcation Engineering, December 18-20, 2015, IIT Bhubaneswar, India

B. Two class classification study (Benign/Malignant): geometric (1D and 2D) features for benign-malignant
classification. This work also includes automatic
The output of first stage is taken as input to this classifier. segmentation method to extract mass from the background.
69 patches were detected as abnormal patches in stage one This method gives over all accuracy of 89.05% for three
of which one is normal. SVM with radial basis polynomial class classification. Our future work focuses mainly on
and sigma=0.9 gave 79.41% accuracy with seven geometric extraction of abnormal breast tissues from whole
features. The normal patch of first stage is detected as mammogram and detect whether that breast tissues are
benign in the second stage. Table III shows 5fold SVM benign or malignant with large set of database.
classification with LQP texture descriptors, Geometric
features and Low order Zernike moments. Table IV shows AKNOWLEDGEMENT
confusion matrix of benign/ malignant classification which
has high accuracy. The output graph of 2fold SVM, Naive We Thank Thomas M. Deserno, Department of Medical
bayes and self organizing map neural network classifiers Informatics, Division of Image and Data Management,
results for normal/Abnormal and Benign/Malignant are Aachen, German for providing images from IRMA
shown in Fig. 4. The outputs of stage1 and stage 2, gives database.
truth table for three class classification shown in Table V. REFERENCES
[1]. http://www.breastcancerindia.net/..
TABLE III [2]. Sheila Timp, Celia Varela, and NicoKarssemeijer “Temporal Change
Analysis for Characterization of Mass Lesions in Mammography”
Accuracy IEEE Int..Trans. in Medical imaging, Vol 26,7, pp.945-953, 2007.
k LQP Geometric Zernike [3]. Loris Nanni ,Sheryl Brahnam, Alessandra Lumini , “A very high
performing system to discriminate tissues in mammograms as benign
1fold 58.82 79.41 61.76 and malignant”,Expert Systems with Applications ,vol.39 , pp.1968–
1971,2012.
2fold 61.76 72.06 63.24 [4]. Leonardo de Oliveira Martins_, Geraldo Braz Junior, Aristofanes
Correa Silva, “Detection of Masses in Digital Mammograms using K-
3fold 58.82 70.59 60.29 means andSupport Vector Machine”, Electronic Letters on Computer
Vision and Image Analysis vol.8(2), pp.39-50, 2009.
4fold 60.29 76.47 57.35
[5]. Marcelo Zanchetta do Nascimento ,Alessandro Santana Martins
5fold 60.29 70.59 61.76 ,Leandro Alves Neves ,Rodrigo Pereira Ramos ,Edna Lúcia Flores
Gilberto Arantes Carrijo , “Classification of masses in a mamographic
Image using wavelet domain features and polynomial classifier,Expert
Systems with Applications”, vol.40, pp.6213-6221, 2013.
TABLE IV [6]. Daniel Rodrigues Ericeira, Aristófanes Correa Silva Anselmo Cardoso
Benign Malignant dePaiva , MarceloGattass, “Detection of masses based on asymmetric
regions of digital bilateral mammograms using spatial description with
18 5 variogram and cross-variogram functions” Computers and Biology
and Medicine,vol.43, pp.987-999, 2013
9 36
[7]. PelinGörgel, AhmetSertbas,OsmanN.Ucan“Mammographical mass
detection and classification using Local Seed RegionGrowing–
Spherical wavelet Transform (LSRG–SWT) hybridscheme”,
TABLE V Computers inBiologyandMedicine, vol.43, pp. 765-774, 2013.
Normal Benign Malignant [8]. Tahmasbin, FatemehSaki, ShahriarB.ShokouhiAmir, “Classification of
benign and malignant masses based on Zernike moments”,
Normal 68 0 0 Computers in Biology and Medicine, vol.41, pp.726- 735,2011.
Benign 1 18 5 [9]. Joberth de Nazare Silva , Antonio Oseas de Carvalho Filho Aristofanes
Correa Silva , Anselmo Cardoso de Paiva ,Marcelo Gattass,
Malignant 0 9 36 “Automatic Detection of Masses in Mammograms Using Quality
Threshold c lustering, Correlogram Function,and SVM”, Journal of
Digital Imaging, June 2015, Volume 28, Issue 3, pp 323-337.
The proposed two stage classification scheme with Local [10]. Spandana Paramkusham,Shivakshit Patri, K.M.M.Rao,B.
quinary pattern texture descriptors and seven efficient V.S.N.Prabhakar Rao , “A new features extraction method based on
polynomial regression for the assessment of breast lesion
geometrical features give over all accuracy of 89.05% for Contours”,Industrial Instrumentation and Control (ICIC), International
three class classification. Conference on,pp.579 – 583, 28-30 May 2015,Pune.
IV. CONCLUSION [11]. Timo Ojala, KimmoValkealahti, ErkkiOja, MattiPietikaKinen,
“Texture discrimination with multidimensional distributions of signed
The proposed method uses two stage classification for gray-level differences”,Pattern Recognition, vol. 34, pp.727-739, 2001.
delineating normal, malignant, benign mammographic [12]. Mingqiang Yang, Kidiyo Kpalma, Joseph Ronsin. A Survey of
Shape Feature Extraction Tech-niques. Peng-Yeng Yin. Pattern
patches using texture features and new geometric features. Recognition, IN-TECH, pp.43-90, 2008. <hal-00446037>.
First stage of classification uses texture descriptors. Authors [13]. Rafael C.Gonzalez,Richard E.Woods,Digital Image
have tested three rotation invariant local texture methods Processing,Third addition.
(LBP, LTP, LQP). LQP (Local quinary pattern) gave best
results with accuracy 99.27%. Second stage of
classification attained 79.41% accuracy with seven

153 ICMOCE-2015