Scholarly article on topic 'Automatic Characterization of Benign and Malignant Masses in Mammography'

Automatic Characterization of Benign and Malignant Masses in Mammography Academic research paper on "Computer and information sciences"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{"Benign and malignant mass classification" / "Fuzzy C-Means clustering" / "Sparse representation classifier" / "Mathews correlation coefficient."}

Abstract of research paper on Computer and information sciences, author of scientific article — K. Vaidehi, T.S. Subashini

Abstract The paper aims to develop an automated breast mass characterization system for assisting the radiologist to analyze the digital mammograms. Mammographic Image Analysis Society (MIAS) database images are used in this study. Fuzzy C-means technique is used to segment the mass region from the input image. GLCM texture features namely contrast, correlation, energy and homogeneity are obtained from the region of interest. The texture features extracted from gray level co-occurrence matrix (GLCM) are computed at distance d=1 and θ=0o, 45o, 90o, 135o. These with three classifiers namely adaboost, back propagation neural network and sparse representation classifiers are used for characterizing the region containing either benign mass or malignant mass. The experimental results show the SRC classifier is more effective with an accuracy of 93.75% and with the Mathew's correlation coefficient (MCC) of 87.35%.

Academic research paper on topic "Automatic Characterization of Benign and Malignant Masses in Mammography"

Available online at www.sciencedirect.com -

ScienceDirect PfOCGCl \

Computer Science

Procedia Computer Science 46 (2015) 1762 - 1769

International Conference on Information and Communication Technologies (ICICT 2014)

Automatic Characterization of Benign and Malignant masses in

Mammography

K.Vaidehia*, T.S.Subashinib

aResearch Scholar, bAssociate Professor, Department of Computer Science and Engineering,Faculty of Engineering and Technology, Annamalai University, India

Abstract

The paper aims to develop an automated breast mass characterization system for assisting the radiologist to analyze the digital mammograms. Mammographic Image Analysis Society (MIAS) database images are used in this study. Fuzzy C-means technique is used to segment the mass region from the input image. GLCM texture features namely contrast, correlation, energy and homogeneity are obtained from the region of interest. The texture features extracted from gray level co-occurrence matrix (GLCM) are computed at distance d=1 and 0=0°, 45°, 90°, 135°. The se with three classifiers namely adaboost, back propagation neural network and sparse representation classifiers are used for characterizing the region containing either benign mass or malignant mass. The experimental results show the SRC classifier is more effective with an accuracy of 93.75% and with the Mathew's correlation coefficient (MCC) of 87.35%.

© 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-reviewunderresponsibility of organizing committee of the International Conference on Information and Communication Technologies (ICICT 2014)

Keywords:Benign and malignant mass classification; Fuzzy C-Means clustering; Sparse representation classifier; Mathews correlation coefficient.

CrossMark

Corresponding Author. Tel.: +919600323875; E-mail address: vainakrishna@gmail.com

1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of organizing committee of the International Conference on Information and Communication Technologies (ICICT 2014) doi:10.1016/j.procs.2015.02.128

1. Introduction

Globally, breast cancer statistics is alarming and it is the second leading fatal cancer in women next to cervical cancer. In India, breast cancer cases are expected to double by 2025. The Indian Cancer Society has declared 2013 as breast cancer awareness year and is taking various initiatives to create awareness in people. Achieving benign detection and adequate treatment will lead to better long term survival as well as a better quality of life.

Film Mammography is the highly used imaging modality for breast cancer detection and diagnosis. The availability of digital mammograms today facilitates the computer aided detection and diagnosis of breast cancer as the breast image could be easily stored in digital format directly into the computer memory.Digital mammography process, guidelines and advantages are vividly explained in1. Micro-calcifications and masses are two important early signs of breast cancer. Masses are often obscured in the surrounding parenchymal tissue, so it is a challenging process to distinguish between the mass region and normal breast region. Abnormal growth of cells are considered as mass which is seen as high intensity regions in an mammogram. Masses are considered to benign or malignant where benign masses are cancerous tumors and malignant are non-cancerous tumors.

Masses are found in several shapes namely circumscribed, speculated, ill-defined or lobulated. Masses with more irregular shapes are malignant and masses with regular and smooth boundaries are in benign stage. For these reasons, texture measures have been proposed for distinguishing between benign and malignant masses. In this paper GLCM and statistical texture measures are incorporated to extract the features. A novel sparse representation classification method is proposed to classify the benign and malignant tumors in mammograms.

The paper is organized as follows. Section 2 briefly reviews some existing technique for segmentation and classification of mammographic mass. Section 3 describes the materials and proposed methodology for automatic characterization of breast mass. Section 4 demonstrates the results and performance and finally conclusions are presented in section 5.

2. Literature Review

Many researchers developed and reviewed automatic mass detection and classification with different CAD approaches2-5. Most of the existing methods differ in the types of features and classifiers that have been used for benign and malignant classification and the way the features have been extracted. Texture is an important characteristic that helps to discriminate and identify the objects. Texture descriptors have been used for detecting normal and abnormal lesion regions in mammograms 6,7

The authors in8 classified the benign and malignant masses using the shape based continuous Zernike orthogonal moment and discrete Krawtchouk orthogonal moment descriptors as features. The extracted feature dimensions are reduced using PCA and K-nn classifier. The accuracy obtained was 81% and 90.2% using Zernike moment and Krawtchouk moment respectively. From the four directions of the gray level co-occurrence matrices, texture features such as Contrast, Energy and Homogeneity features were extracted9. These features classified the mass into either speculated, ill-defined or circumscribed. Decision tree with five criteria wereanalyzed for classification of masses. Morphologic multiple concentric layer analysis is used to identify mammographic masses10. The detection rate of this CAD technique outperformed for identifying malignant masses than benign masses. Gabor filter bank at different scales and orientations is used to extract the texture features for characterizing the mass in mammography. This method achieved 94.92% and 85.53% accuracy with SVM classifier for normal-mass classification and benign-malignant classification respectively11. Segmentation approaches are classified into region based methods12, contour based methods13, clustering and thresholding methods14 and model based methods15. In this work, the clustering based segmentation technique is used.

3. Materials and methods

The block diagram showing the various steps of the proposed method is given in Fig. 1.

3.1. Database description

An organization in UK called Mammography Image Analysis Society (MIAS) has created a mammogram database. This database is used in this study. It contains both the right and left breast images of the same patient with a uniform size of 1024 x 1024 pixels. The mammogram in this database is digitized to a resolution of 50m x 50m, 8 bits represents each pixel. The database contains 322 mammograms of 161 patients, in which 209 are normal mammograms and 113 are abnormal mammograms which includes both mass and microcalcification. The database has the ground truth details such as tissue types, class of abnormality, severity of abnormality and location of abnormality like xy image-coordinates of centre of abnormality, and approximate radius (in pixels) of a circle enclosing the abnormality16.

• Fuzzy C-means clustering

region from

^reas^O^

Feature extraction

GLCM features Zernike moment features

Classification of benign

and malignant

vAdaboostr

• Back propogation neural network (BPNN)

• Sparce representation based classifier (SRC)

Fig. 1. Block diagram showing the various steps of the proposed method

3.2. Mass region Segmentation

For computer aided detection and diagnosis of breast mass, segmenting the mass region accurately is the most vital step. Breast masses are obscured by the normal breast parenchymal tissue. Fuzzy K-means clustering based segmentation approach is used in this work.

Fuzzy C-Means

In the medical domain, FCM is one of the most commonly used unsupervised pattern recognition approach for tumor segmentation. In 1981,As an alternative to C-means clustering algorithm Bezdekformulated the Fuzzy C-means algorithm17. FCM employs fuzzy partitioning, this approach partitions the data points into K clusters and the degree of association of a data point to each cluster is denoted by the membership grades ranging between 0 and 1. In hard C-means algorithm, each group of data points in a dataset completely belongs to one cluster. Whereas, Fuzzy C-means algorithm allows data points nearer to the cluster center have a high degree of belonging and those that are far away from the cluster center have a less degree of association with its cluster centers.

Algorithm for Fuzzy C-Means clustering:

The goal of Fuzzy c-means is to minimize the following objective function of weighted distances of the data to the centers.

Om = ££

i=i j=i

where m>1, the degree of membership of Xj in the cluster Ci is given by Vij ,Ci is the it cluster of the n-dimensional centerand Xj is thei"1 data point of the n-dimensional data.

The membership matrix V=[vjj] have elements with the values ranging between 0 and 1. The objective function (Om) is optimized with an iterative function to achieve fuzzy partitioning and the elements of the membership matrix and the cluster center is updated and is given by :

f xj - ci 2/(m-1) Cj

V xj - ck J

In medical clustering, the pixel intensity variation is very useful to segment the tumor region and normal region. When the high membership values are assigned to the pixel intensities which are close to a cluster center and low membership values are assigned to the pixel intensities which are far away from the cluster center, then the objective function is minimized. In this work Fuzzy K-means algorithm is applied to segment the mass region.

3.3. Feature Extraction

Texture characteristics give more significant information in pattern recognition area and in this work and for mass classification GLCM features are derived from the mass region. GLCM is one of the widely used techniques for texture analysis. Four texture descriptors namely contrast, correlation, energy and homogeneityused for classification of benign and malignant masses are computed at different orientations with the distance of 1 pixel. Intensitycontrast between a current pixel and its neighbour in an image is given by the contrast descriptor. Correlation tells how current pixel and its neighbourhood pixel is related to one another. Energy is the angular second moment. Homogeneity gives the idea about the closeness of the distribution of GLCM elements to its diagonal

3.4. Classification Ada Boost

In 2003, Yoav Freund and Robert Schaphire formulated a machine learning meta-algorithm called Adaptive Boosting (AdaBoost). It is very simple to implement and good for generalization. It improves the classification accuracy and not prone to over fitting. It is an iterative algorithm and during an each iteration of the training phase, a new weak learner is added to create a strong learner that is only slightly correlated to the classifier. The weighting vector is adjusted every time the weak learner is added to the ensemble to focus on examples that were misclassified in the earlier iteration. Hence it is called adaptive and finally results with a classifier with better accuracy.

Back Propogation Neural Network (BPNN)

A BPNN is a supervised machine learning technique and uses a feed-forward architecture. The BPNN is based on the gradient descent technique for solving an optimization problem, which involves the minimization of the network cumulative error. Error is the difference between the target output and the actual output. This BPNN is designed in such a way as to update the weights in the direction of the gradient descent of the cumulative error. This is done in an iterative way. The weights are adjusted during each iteration by propagating the errors backwards. This is continued until the mean square error is minimized to an acceptable level18.

The Sparse Representation based classifier (SRC)

In this work for mass classification SRC is used. In SRC, the given test sample can be represented as a linear

combination of the training sample and does not require any formal training process19, 20. Let A be the training sample matrix of p classes.

A = W A2 ,.....Ap }= kl, ^i12,........vini }

A test mammaogram image y can be well approximated as a linear combination of the training sample taken from Aiand is given by

yav (4)

wherenrepresents the total samples in the ith class. Now eqn. (4) can be rewritten as

y=Ax0 (5)

Since A is the dictionary containing all the training samples.

x0 =^0,....0,aj1,aj 2,...aj ^ ,0,...0jr is thecoefficient vector where only a few coefficients are non- zero

whereas others are zero. The nonzero coefficients are alone related to class i. Thus sparse solution of the coefficient vector which is same as solving the following optimization problem (l0 minimization) could be evaluated as given in equation (6).

X0 = arg min||x||0 subject to Ax=y (6)

Because the lack of l0norm's mathematical representation, l0 minimization is regarded as an NP-hard problem, as it is too complex and almost impossible to solve. In many cases, l0 minimization problem is relaxed to be higher order norm problem such as l1 minimization and l2 minimization. The SRC approximated the l0 norm coefficients by l1 minimization problem21.

If the test sample y belongs to a certain class, the coefficients in the estimated x0 not within this class should all be zeros. But given the noise and the modelling error, the noisy model is modified as:

y=Ax0+e (7)

wheree is the noise level. Then, the equation (6) is converted into:

xx0 = arg min||x||0 subject to ||Ax - y||2 < s

We consider using the residual error to classify y. After estimating x1 , the given test mass image y is approximated as:

yi = (xi) (9)

where Si (xci) is a new vector. The nonzero entries in x1 are alone related to class i. The residual error ri (y)is:

r iy )=y - A$i (xi) (10)

The test sample y is classified with the class having the minimal residual error. The SRC algorithm has good

generalization ability. So it is more suitable for medical applications. 3.5. Performance measures

Table 1 presents a confusion matrix for binary classification, where Arepresents true positive, Brepresents false positive, Crepresents false negative, and Drepresents true negative counts. Accuracy, sensitivity, specificity, PPV, NPV, MCC are some of the measures used for evaluating the performance of the classifiers used in this work namely adaboost, SRC 22.

Table 1. Confusion matrix for mass classification_

Class/Recognised Benign Malignant

Benign True Positive (A) False Negative (C)

Malignant False Positive (B) True Negative (D)

Accuracy = -

Sensitivit y = Specificity =

A + B + C + D

D B + D

PPV = NPV = MCC =

D D + C

A x D - B x C

A + B)( A + C )( D + B)( D + C )

Mathews Correlation Coefficient (MCC) is another accuracy evaluation measure which could give a better picture of the performance of the classifier. MCC is used as a metric rather than accuracy when the number of samples in the two classes is unbalanced.

4. Experimental Results

The MIAS database contains ground truth of the image which includes the center of the mass and approximate radius of the circle enclosing the mass in terms of number of pixels. Totally 48 abnormal mammograms containing mass is considered for this study. Out of which 16 are malignant mammograms and 32 are benign mammograms. A square region of area 174x174 pixels is taken as the ROI for further processing. The value 174 is chosen in consultation with the radiologist because it is the radius of the largest mass present in the database and moreover mass will not be more than the size of 174x174. Fig.2. shows the segmented input breast region.

Fig. 2. 174 x 174 size segmented input breast region

From the input ROI thus obtained which is shown in Fig. 3a., the mass region is segmented using fuzzy K-means clustering algorithm and Fig. 3b. shows the segmented mass region. To obtain the contour of the mass image, morphological erosion applied image is subtracted from the original image which is nothing but the gradient of the original image. The contour thus obtained is shown in Fig. 3c. Then morphological operation "areaopen" is applied, to retain only those regions which contain more than 600 pixels and this removes smaller regions. Fig.3d. shows the mass region alone without smaller regions. Then the original image is compared with Fig.3d and all the pixels which

is white is replaced with the original intensity values to obtain the segmented mass and it is shown in Fig. 3e.

□ ■□Q

Fig. 3. a) input ROI b) FCM segmented mass image c) contour of mass d) after applying morphological operation e) orginal pixel values of the segmented mass.

Four features derived from the GLCM (contrast, correlation, energy and homogeneity) are calculated at 0 = 0°, 45°, 90°, 135° and d=1, Since the reliable information cannot be given by a single direction.Therefore, 4directions from the co-occurrence matrix are used for extracting the second order texture information from the mammograms. The derived features are fed into the classifiers namely Adaboost, SRC and BPNN.

Table 2. Performance measures of different classifiers with local binary pattern texture features Type Accuracy Sensitivity Specificity PPV (%) NPV (%) MCC

(%) (%) (%) (%)

Adaboost 81.25 78.12 87.5 92.59 66.66 62.36

BPNN 77.08 75.0 81.25 88.88 61.90 53.45

SRC 93.75 90.62 99.99 99.99 84.21 87.35

The performance measures used for classification are accuracy, sensitivity, specificity, PPV, NPV and MCC. Leave one out procedure has been adopted in testing the performance of the various classifiers. Table 2 shows the performance of the three classifiers with GLCM features. It could be seen that the SRC outperformed Adaboost and BPNN in classifying benign and malignant masses. The accuracy obtained was 93.75%. Mathews Correlation Coefficient (MCC) is calculated to get a better picture of the performance of the classifier, since the number of samples in the two classes is unbalanced. When compared to accuracy MCC is used in cases where the number of samples in each of the classes differs considerably. SRC obtained the highest MCC of 87.35% whereas the other classifiers reported relatively poor MCC.

5. Conclusions

In this paper, it is proposed an automatic system which detects and categorizes benign mass and malignant mass regions from the breast ROI. Mini-mias database images are used for this study. This automatic detection of masses is beneficial to the radiologist for finding the early stage of (benign) mass and cancerous stage (malignant) without confusion. Even though the mass regions are obscured in the dense regions, the study reveals the usefulness of fuzzy C-means algorithm for segmenting mass regions from the ROI. The experimental results show that the extracted GLCM descriptors along with SRC classifier could be effectively used in classifying breast masses in digital mammograms. The SRC classifier obtained highest accuracy is 93.75%. The proposed work was carried out using MATLAB 2012a.

Acknowledgement

The financial support received from UGC to carry out this work under UGC major research project is highly acknowledged.

References

1. Bowes MP. Digital Mammography: Process, Guidelines, and Potential Advantages.

2. Oliver A, Freixenet J, Marti J, Pérez E, Pont J, Denton ER, et al. A review of automatic mass detection and segmentation in mammographic images. Medical Image Analysis. Elsevier; 2010;14(2):87-110.

3. Rojas Dominguez A, Nandi AK. Toward breast cancer diagnosis based on automated segmentation of masses in mammograms. Pattern Recognition. Elsevier; 2009;42(6):1138-1148.

4. Tang J, Rangayyan RM, Xu J, El Naqa I, Yang Y. Computer-aided detection and diagnosis of breast cancer with mammography: recent advances. Information Technology in Biomedicine, IEEE Transactions on. IEEE; 2009;13(2):236-251.

5. Elter M, Horsch A. CADx of mammographic masses and clustered microcalcifications a review. Medical physics. American Association of Physicists in Medicine; 2009;36(6):2052-2068.

6. Székely N, Toth N, Pataki B. A hybrid system for detecting masses in mammographic images. Instrumentation and Measurement, IEEE Transactions on. IEEE; 2006;55(3):944-952.

7. Nunes AP, Silva AC, Paiva ACD. Detection of masses in mammographie images using geometry, Simpson's Diversity Index a nd SVM. International Journal of Signal and Imaging Systems Engineering. Inderscience; 2010;3(1):40—51.

8. Narvaez F, Romero E. Breast mass classification using orthogonal moments. Breast Imaging. Springer; 2012. p. 64-71.

9. Khuzi AM, Besar R, Zaki WW, Ahmad N. Identification of masses in digital mammogram using gray level co-occurrence matrices. Biomedical Imaging and Intervention Journal. Department of Biomedical Imaging, University of Malaya; 2009;5(3):e17.

10. Eltonsy NH, Tourassi GD, Elmaghraby AS. A concentric morphology model for the detection of masses in mammography. Medical Imaging, IEEE Transactions on. IEEE; 2007;26(6):880-889.

11. Hussain M, Salabat Khan GM, Ahmad I, Bebis G. Effective Extraction of Gabor Features for False Positive Reduction and Mass Classification in Mammography. Appl. Math. 2014;8(1L):397-412.

12. Wei J, Chan H-P, Sahiner B, Hadjiiski LM, Helvie MA, Roubidoux MA, et al. Dual system approach to computer-aided detection of breast masses on mammograms. Medical physics. American Association of Physicists in Medicine; 2006;33(11):4157-4168.

13. Sahiner B, Chan H-P, Petrick N, Helvie MA, Hadjiiski LM. Improvement of mammographic mass characterization using spiculation measures and morphological features. Medical Physics. American Association of Physicists in Medicine; 2001;28(7):1455-1465.

14. Bellotti R, De Carlo F, Tangaro S, Gargano G, Maggipinto G, Castellano M, et al. A completely automated CAD system for mass detection in a large mammographic database. Medical physics. American Association of Physicists in Medicine; 2006;33(8):3066-3075.

15. Cheng H, Cui M. Mass lesion detection with a fuzzy neural network. Pattern recognition. Elsevier; 2004;37(6):1189-1200.

16. Suckling J, Parker J, Dance D, Astley S, Hutt I, Boggis C, et al. The mammographic image analysis society digital mammogram database. Excerta Medica; 1994;

17. Bezdek JC. Pattern recognition with fuzzy objective function algorithms. Kluwer Academic Publishers; 1981.

18. Yegnanarayana B. Artificial neural networks. PHI Learning Pvt. Ltd.; 2009.

19. Wright J, Yang AY, Ganesh A, Sastry SS, Ma Y. Robust face recognition via sparse representation. Pattern Analysis and Machine Intelligence, IEEE Transactions on. IEEE; 2009;31(2):210-227.

20. Zhang L, Zhou W-D, Chang P-C, Liu J, Yan Z, Wang T, et al. Kernel sparse representation-based classifier. Signal Processing, IEEE Transactions on. IEEE; 2012;60(4):1684-1695.

21. Sokolova M, Japkowicz N, Szpakowicz S. Beyond accuracy, F-score and ROC: a family of discriminant measures for performance evaluation. AI 2006: Advances in Artificial Intelligence. Springer; 2006. p. 1015-1021.

22. Donoho DL. For most large underdetermined systems of linear equations the minimal ??1-norm solution is also the sparsest solution. Communications on pure and applied mathematics. Wiley Online Library; 2006;59(6):797-829.