You are on page 1of 6

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/262158861

Mammography Feature Analysis and Mass Detection in Breast Cancer Images

Conference Paper · January 2014


DOI: 10.1109/ICESC.2014.89

CITATIONS READS

10 135

2 authors:

Dr. Bhagwati Charan Patel Professor G R Sinha


Shri Shankaracharya Group of Institutions Myanmar Institute of Information Technology Mandalay Myanmar
18 PUBLICATIONS   110 CITATIONS    199 PUBLICATIONS   401 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Reliable CAD system for Breast cancer detection View project

Breast Cancer View project

All content following this page was uploaded by Professor G R Sinha on 27 July 2015.

The user has requested enhancement of the downloaded file.


2014 International Conference on Electronic Systems, Signal Processing and Computing Technologies

Mammography Feature Analysis and Mass detection in Breast Cancer Images

Bhagwati Charan Patel G. R. Sinha


Associate Professor, Department of Information Professor and Associate director ,Department of
Technology, Faculty of Engineering and Electronics and Telecommunication, Faculty of
Technology,Shri Shankaracharya Group of Institutions, Engineering and Technology,Shri Shankaracharya
SSTC, Bhilai,(C.G.), India Group of Institutions, SSTC, Bhilai,(C.G.), India
bhagwaticharanpatel@gmail.com drgrsinha@ieee.org

Abstract— This paper introduces a novel approach for


accomplishing mammographic feature analysis through detection II. RELATED WORKS
of tumor, in terms of their size and shape with experimental work Large numbers of segmentation methods have been developed
for early breast tumor detection. The objective is to detect the
abnormal tumor/tissue inside breast tissues using three stages:
in the field of image analysis and many of them have been
Preprocessing, Segmentation and post processing stage. By using used to segment masses in mammograms. Kom et al. (2007)
preprocessing noise are remove and than segmentation is applied presented an approach to segment the suspicious mass regions
to detect the mass, after that post processing is applied to find out by a local adaptive thresholding technique. It did not consider
the benign and malignant tissue with the affected area in the the case where a mass contains the small window, the centre
cancers breast image. Size of tumor is also detected in these steps. region of a suspicious lesion is not detected, and it gives an
The occurrences of cancer nodules are identified clearly. empty area in the segmentation result [6]. Kai Hu et al.(2011)
Compared with an expert observer reading the Mammography, proposed Wavelet transforms method and combination of
our algorithm achieves 96.5% sensitivity, 89% specificity, 95.6% adaptive global thresholding segmentation and adaptive local
accuracy value.
thresholding segmentation is used to segment the
.
Keywords- breast cancer;mammography image;segmentation;
multiresolution sub images of the original mammogram.
tumor;sensitivity;specificity;accuracy. Results for some types of lesions mainly characterized by
texture feature may be improved [7]. Indra et al. (2011)
proposed the identification technique is divided into two
I. INTRODUCTION
distinct parts i.e. formation of homogeneous blocks and colour
Breast cancer is a cancer that starts in the tissues of the breast. quantization after pre-processing. Texture features are
According to Indian Council of Medical Research (ICMR), it extracted using gray level co-occurrence matrices (GLCM).
is reported that one in 22 women in India is likely to suffer By using gray level values eliminate the non-masses area from
from breast cancer during her lifetime, while the figure is the mammogram to identify the presence of abnormality [8].
definitely more in America with one in eight being a victim of Nithy et al. (2011) proposed valuation and comparison of the
this deadly cancer. In Mumbai, 30% , in Delhi and Bangalore performance of feature extraction methods for classification of
it was almost 26.9% while the incidence in Chennai 26.5%. In normal and abnormal patterns in mammogram. They use Grey
Kolkata, it accounted for 27.2% of cancer cases among Level Co-occurrence Matrix to calculate the performance of
women [1], [2]. Mmammography is the procedure of using the each feature extraction method is evaluated on accuracy.
low-dose X-rays to examine the human breast for the purposes [9]. Fatima et al. (2011) proposed muscle segmentation, hard
of identifying breast cancer or other abnormalities. Detection density zone detection and texture analysis of regions of
and diagnosis of breast cancer in its early stage increases the interest. It was found that classification rate to be 77% on
chances for successful treatment and complete recovery of the average. These rates were significantly improved, achieving
patient. Finding an accurate and efficient breast region 95% in average, when applied the classification method on
segmentation technique still remains a challenging problem in pre-segmented mammograms by the maxima thresholding.
digital mammography [3]. Round and oval shaped masses [10]. Minavathi et al. (2012) suggested that ultrasound images
with smooth and circumscribed margins usually indicate are pre-processed using Gaussian smoothing to remove
benign changes. On the other hand, a malignant mass usually additive noise and anisotropic diffusion filters to remove
has a speculated, rough and blurry boundary [4]. Benign multiplicative noise [11]. After segmenting the suspected mass
calcifications are usually larger and coarser with round and region, features of the segmented region should be examined
smooth contours. Malignant calcifications tend to be to verify whether the extracted region contains mass or not. A
numerous, clustered, small, varying in size and shape, angular, key stage of mass detection and classification is feature
irregularly shaped and branching in orientation [5]. analysis and extraction. Lie et al. (2001) implemented AI
techniques that include fractal dimension analysis,
multiresolution markov random field. The classification stage

978-1-4799-2102-7/14 $31.00 © 2014 IEEE 474


DOI 10.1109/ICESC.2014.89
uses criteria to identify regions that are suspicious for tumor. B. Segmentation
The verification results show that the proposed algorithm has a The aim of the segmentation is to extract ROIs containing
sensitivity of 97.3% and the number of false positive per all masses and locate the suspicious mass candidates from the
image is 3.92. [12]. Shekhar et al. (2011) propose a method for ROI. Region growing is one of the most popular methods to
automatic Breast cancer detection, classification, scoring and segment regions in images. Given a starting pixel or region,
grading to assist pathologists. They presented the classification connecting pixels or regions are added if their properties are
of micro cancer object of breast tumor based on feed forward similar to the already segmented region. Assume a region of
back propagation Neural Network. The sensitivity, specificity interest I with intensity values described by f (x, y). A seed
and accuracy were found to be equal 97.10%, 95.70% and point (xs, ys)  I with an intensity value of f (xs, ys) is used as
96.60% respectively [13]. Pagonis et al. (2010) evaluated the the starting point and is the initial segmented area.
effect of using multiple modalities on the accuracy achieved
by a CAD system, designed for the detection of breast cancer.
They proposed Probabilistic Neural Network (PNN) classifier
scored 78.05% in classification accuracy. While it increased at
82.95%, and than increase accuracy 95.12% [14]. Kowal et al.
(2011) presented a decision support system that allows
distinguish malignant from the benign breast tumors.
Diagnostic accuracy obtained varies according to different (a) (b)
classification methods and fluctuates up to 96% for quasi Figure.1 Result of homomorphic filtering. (a) Original image (b) Image
optimal subset of features [15]. We have proposed a method after homomorphic filter
for easy identification of abnormal masses in mammography TABLE I
images that finds lower values of false positive (FP) and false EVALUATION OF HOMOMORPHIC FILTER
negative (FN) detection. Image Parameter
  PSNR(dB) ASNR(dB)
III. MASS DETECTION AND FEATURE ANALYSIS METHODS Original image 5.47 3.58
45.78 32.5
As already defined, a mass is space occupying lesion seen in Filtered image 4.970 3.233
normal and abnormal mass. We divided our mass
segmentation and feature extraction of breast cancer in 3
stages i.e. preprocessing, segmentation, postproessing or CAD
evaluations. Critical issue in algorithm design is the feature
selection where the best set of features is selected for
eliminating false positives (FP) and for classifying lesion types
are performed in the post processing stage.
A. Pre processing (a ) (b)
Figure.2 Result of mass segmentation of breast cancer image. (a) Filtered
Mammographic images are normally affected by various types image (b) Segmented image
of noise. Removal of these noises without destroying the
desired information is often a significant challenge. We use All neighboring pixels with an intensity value larger than a
homomorphic filter to remove the noise. A filter is applied to threshold are added to this segmented region. This procedure
in frequency domain of the image which allows the is repeated until no more pixels are found with values above
enhancement of the high frequency values and decrement of the threshold. A contour is found and the whole procedure is
the low frequency values. Fig. 1 shows the result of the repeated with a lower threshold value, resulting in a larger
enhancement of real mammograms with the homomorphic segmented region. This way, a series of contours is created.
filtering implemented. In general, contrast was slightly Several criteria can be used to select the best contour. Another
decreased when compared with the original image. Table 1 simple option is to use a fixed threshold level to stop the
shows that the background noise level was decreased when growing process. In this work, two different criteria were
compared with the original image. Mean () and standard ( ) implemented: a fixed threshold value that depends on f (xs, ys)
values are calculated as statistical measures. However, the and a probabilistic method [16]. Based on estimations of the
peak and average signal to noise ratio were slightly inferior to probability distributions of intensity values of background
that of the original image. The PSNR (peak signal to noise tissue and pixels inside a tumor, for each contour that is found
ratio) and ASNR (average signal to noise ratio) are ratio of the in the growing process a possibility is computed. A Gaussian
wanted signal and the unwanted background noise. After function, centered at (xs; ys), was used in a multiplication,
performing filtering method, good filtered imges were giving each pixel (x; y) in I, the value:
obtained. h(x; y) = f (x; y) ×N(x; y; xs; ys;2)
The value of  is important, especially when fixed thresholds
are used to segment the region. In Fig.2 the results for the
region growing method are shown. After preprocessing, a

475
lower threshold should be used; otherwise very small regions SSGIDB-11 1358 23 83 135 94.2 85.5 93.4
will be the result. Segmentation with =15 mm and a SSGIDB-23 1519 23 56 186 96.5 89 95.6
threshold of 75% gave best performance, which was SSGIDB-37 1392 29 69 166 95.3 85.2 94
considerably better than the performance that was achieved SSGIDB-52 1573 19 87 146 94.8 88.5 94.2
without preprocessing.
The performance of the proposed method is evaluated using
C. Postprocessing
perfect test method which gives the sensitivity and specificity
After the segmentation is performed on breast region, the of the result with graphical representation. SSGIDB-23 have
features can be obtained from it and the diagnosis rule can be highest sensitivity of the proposed method is 96.5% and
designed to exactly detect the cancer nodules in the breasts. specificity is 89% shown in Table 2. Hence the proposed
This diagnosis rules can eliminate the false detection of cancer method is highly desirable in order to assist the radiologist in
nodules resulted in segmentation and provides better diagnosis the detection of malignant region and to improve the
Correct mammography detection of lesions, is essential to diagnostic accuracy. In Fig. 3 binary image is obtained by
discover early breast cancer phases, increasing the treatment thresholding the gray level than converted into pseudo colored
options and survival rate [17]. In the feature extraction and and affected region are labeled with numbered from top to
selection step the features that characterize specific region are bottom and left to right. After finding the region than
calculated and the ones that are important are selected for the boundary is defined and brightest area are displayed.
classification of the mass as benign or malignant. The
efficiency of a CAD system can be classified in four
perspectives [13] i.e
• True Positive (TP), when the suspected abnormality
is in fact malignant;
• True negative (TN), when there is no detection of
abnormality in a healthy person;
• False positive (FP), when occurs detection of
(a) (b)
abnormality in a healthy person;
• False negative (FN), when there is no detection of a
malignant lesion
The terminology which is used to determine the performance
of a CAD System is defined as follows:
• Sensitivity: correctly classified percentage of ROI by
radiologist is calculated as: TP/(TP+FN)
(c) (d)
• Specificity: correctly classified percentage of ROI by Figure.3 Result of segmentation for defined number of area. (a)Binary image
non radiologist defined as: TN/(TN+FP) obtained, (b) Pseudo colored labeled., (c) Outline from boundary
• Accuracy: percentage of correctly classified (d) Brightest area are displayed
pathological and non-pathological cases is calculated
ROC curves illustrate the performance on a binary
by : (TP+TN)/(TP+FP+FN+TN)
classification problem where classification is based on simply
IV. RESULTS AND DISCUSSIONS thresholding a set of scores at varying levels. Low thresholds
give high sensitivity but low specificity and vice versa; the
We used a mammogram database developed by BSR Apolo,
ROC curve plots this trade-off over a range of thresholds.
center for research and diagnosis of the database for Shri
Using the results from Table 2 for the malignant cases, the
Shankaracharya Group of institution (SSGIDB), that are
receiver operating characteristic (ROC) curve of Fig. 4 has
collected from local hospitals mentioning that biopsy has done
been extracted and reflects the system’s performance in terms
on the patients, so we already know the results of benign or
of sensitivity and specificity (ROC analysis).
malignancy. There are 127 mammograms in the database taken
from 89 different patients. Here we present few of them. Using
a small threshold is more likely to detect true lesions, but also
to generate more false positive responses. Using a large
threshold gives fewer false positive responses, but may miss
more true lesions. Table 2 shows the results of quantitative
analysis and from the results we can also prove the
effectiveness of the proposed algorithm.
TABLE II
PERFORMANCE EVALUTION
Database TP FP FN TN SEN SPE ACC
% % Figure.4 ROC curve for SSGIDB-23, depicting the performance CAD system
% with sensitivity of 96.5% and specificity of 89%
SSGIDB-3 1480 19 71 121 95.4 86.4 94.7

476
The area under the ROC curve is a measure of the TABLE III
classification performance. A higher area indicates better SEGMENTATION INTO SMALL NUMBER OF AREA
classification performance because a larger value of TP is
achieved at each value of FP. From Table 2 the general Area #1 Area #2 Area #3 Area #4
accuracy of the CAD system is good referring to their area
under curve value of 95.6%. High values of sensitivity imply Area #5 Area #6 Area #7 Area #8
minimal false negatives detection or higher true positive
detection. The specificity of the test is the fraction of the true Area #9 Area #10 Area #11 Area #12
negatives case over the real negative cases. High values of
specificity imply minimal false positive detection. Feature TABLE IV
selection is thus useful for improving the result of accuracy in PARAMETER FOR THE SEGMENTED AREA FOR SSGI DB-1
our experiment. Therefore, radiologists are particularly (NUMBER OF PIXELS)
# Mean Intensity Area Perimeter Centroid Diameter
interested to observe the size, shape and texture of the organs 1 252.3 153.0 261 21.3 152 18.2
or parts of the tissues. These routine assessments are 2 31 3.0 15 3.7 4.0 2.0
commonly subjective and quantitative, and reports typically 3 278.3 117.0 21.1 16.6 65.6 22.0
refer to lesions as large, small, and prominent. In Table 3 and 4 249.5 182.0 164 28.5 135.0 14.6
5 194.8 13.0 10.8 20.0 48.3 4.1
Table 4 show the various numbers of segmented sub-areas and 6 399.7 351.0 177 36.7 67.2 21.1
their corresponding measurement respectively. The 7 189.4 8.0 8.5 29.5 117.5 3.2
mammograms were digitized at 60m per pixel (12-bit pixels) 8 183.7 19.0 253 35.7 92.0 17.9
with a Lumisys 90 digitizer and averaged down to 0.1 mm per 9 156.5 84.0 84.3 49.7 83.0 10.3
10 249.4 5.0 6.8 40.6 89.4 2.5
pixel. It can easily be seen that the size and diameter of the 11 243.0 2.0 2.0 45.0 91.5 1.6
area which can be tumor are shown in table 4. 12 199.4 5.0 5.7 58.0 64.0 2.5
The classification method proceeds as follows. As objects
are isolated by the user, they may be given an area label. A TABLE V
count is maintained of the number of objects from each class BENIGN/MALIGNANT COUNTS FOR MAMMOGRAPHY IMAGE
that are represented in each Area. Table 5 shows the counts for Image Benign Malignant
SSGIDB-3 4 2
the benign and malignant and also size is a more important
SSGIDB-11 6 1
factor than shape for cell classification. An area number 3, 4, 6 SSGIDB-23 4 3
and 9 gives the more prominent shape of speculated shape for SSGIDB-37 7 2
tumor and their intensity is also higher than other segmented SSGIDB-52 5 3
area. So they may be malignant but area number 5, 7,8,12
TABLE VI
having the shape of circular or oval i.e. they represent the EVALUATION RESULT FOR CAD SYSTEM
benign tumor. As on the area size is not more than 2 cm (3/4 Accuracy Sensitivity Specificity
of an inch) this tumor can be called as primary tumor of T1 Approaches
(%) (%) (%)
category but cancerous cell is present. Our CAD system has Maitra et al. 95.8 96.4 90
been successful in indicating malignant micro calcifications Bick et al 94.7 95 88
Proposed 95.6 96.5 89
seen mammographically and can improve the cancer detection
rate during the screening. Previously, Maitra et al. [8] and V. CONCLUSIONS
Bick et al [18] suggested seeded region growing based
segmentation approaches for mass segmentation and feature The aim of this paper is to improve the detection of suspected
extraction of breast cancer images, but the size and shape areas containing some type of lesion. In the proposed work we
analysis was not incorporated in their work. This is why the have designed a new computer aided detection method to
proposed approaches produced good evaluation result of CAD detect the mass region in the mammogram. Segmented image
system on the basis of segmentation methods, as shown in contains the suspected region which is given for feature
Table 6. extraction process. The extracted features are classified in to
To calculate fastness of CAD system time complexity method normal and abnormal region .The obtained accuracy was
is used that is based on the expression that relates size of the 95.6% whereas the sensitivity and specificity were found to be
input and the run time for the algorithm. A computer is 96.5% and 89% respectively. The proposed system gives fast
assumed to execute a million instructions a second, and then and accurate classification of breast tumors.
the amount of time required to execute f (103) instructions on
CAD system is 17.5 second of the order O (n*log2(n)).

477
REFERENCES [10] Fatima Eddaoudi and Fakhita Regragui “Masses Detection Using
SVM Classifier Based on Textures Analysis” Applied
[1] Cancer Research in ICMR Achievements in Nineties, Mathematical Sciences, vol.5,pp.8367–8379, 2011.
icmr.nic.in/cancer.pdf. [11] Minavathi, Murali. and M.S. Dinesh “ Classification of Mass in
[2] B.C. Patel and G. R. Sinha. “Early Detection of Breast Cancer Breast Ultrasound Images” International Journal of Computer
using Self Similar Fractal Method”. International Journal of Applications,vol.42,no.10pp.29-36, 2012.
Computer Application, vol.10, no.4.pp.39-43, 2010. [12] Lei Zheng and Andrew K. Chan “ An Artificial Intelligent
[3] Recent Advances in, Breast Imaging, Mammography, and Algorithm for Tumor Detection in Screening Mammogram” IEEE
Computer Diagnosis of Breast Cancer. SPIE, Bellingham, 2006. Transactions On Medical Imaging, vol.20,no.7,pp.235-242,2001.
[4] Cheng, H.D., Shi, X.J., Min, R., Hu, L.M.,“Approaches for [13] Shekhar Singhand and P. R. Gupta “ Breast Cancer detection and
Automated Detection and Classification of Masses in Classification using Neural Network “, International Journal Of
Mammograms”,Pattern Recognition vol.39,pp.646–668, 2006. Advanced Engineering Sciences And Technologies,vol.6,no.1,pp.
[5] Rangayyan, R.M., Ayres, F.J., Desautels, J.E.L. “A Review of 4–9, 2011.
Computer Aided Diagnosis of Breast Cancer: Toward the [14] Pagonis, Dionisis Cavouras, Kostas Sidiropoulos,“ Improving The
Detection of Subtle Signs”, Journal of the Franklin Institute, Classification Accuracy Of Computer Aided Diagnosis Through
vol. 344,no.3,pp.312–348, 2007. Multimodality Breast Imaging”, e-Journal of Science &
[6] G. Kom, A. Tiedeu, and M. Kom, “Automated detection of masses Technology (e-JST),vol. 2, no.5,pp.33-39,2010.
in mammograms by local adaptive thresholding,” Computer [15] Marek Kowal, Pawe Filipczuk, Andrzej Obuchowicz, Józef
Biology Medicine., vol.37,no.1,pp.37–48, 2007. Korbicz “Computer-Aided Diagnosis Of Breast Cancer Using
[7] Kai Hu, Xieping Gao, and Fei Li “ Detection of Suspicious Lesions Gaussian Mixture Cytological Image Segmentation”, Journal Of
by Adaptive Thresholding Based on Multiresolution Analysis in Medical Informatics & Technologies.vol,17,pp.257-262,2011.
mammograms” IEEE Transactions on Instrumentation and [16] M A Kupinski and M L Giger. “Automated seeded lesions
Measurement,vol.60,no2,pp.462-472, 2011. segmentation on digital mammograms”. IEEE transactions on
[8] Indra Kanta Maitra , S. Nag and S. Kumar B “Identification of medical imaging, vol. 17, no.4, pp. 510–517, 2008.
Abnormal Masses in Digital Mammography Images” ,International [17] B.C. Patel and G. R. Sinha. “An Adaptive K-means Clustering
Journal of Computer Graphics, vol.2,no1pp.17-30,2011. Algorithm for Breast Image Segmentation”. International Journal of
[9] R.Nithya, and B.Santhi “Comparative study on Feature Extraction Computer Application vol.10, no.4. pp.35-38., 2010.
Method for Breast Cancer Classification” journal of theoretical [18] Bick, U., Giger, M.L., Schmidt, R.A., Nishikawa, R.M., Wolverton,
and applied information technology, vol.33,no.2,pp.220-226, 2011. D.E., Doi, K.,“Automated segmentation of digitized mammograms
”, Academic Radiology,vol.2,pp.1-9,2009

478

View publication stats

You might also like