Scholarly article on topic 'A Study of Different Texture Features Based on Local Operator for Benign-malignant Mass Classification'

A Study of Different Texture Features Based on Local Operator for Benign-malignant Mass Classification Academic research paper on "Materials engineering"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{"Breast Cancer" / Mammography / "Mass Classification" / "Local Binary Pattern(LBP)" / "LBP Variance(LBPV)" / "Completed LBP(CLBP)"}

Abstract of research paper on Materials engineering, author of scientific article — Rinku Rabidas, Abhishek Midya, Jayasree Chakraborty, Wasim Arif

Abstract In this paper, a comparative analysis of different texture features based on local operator has been produced for the determination of mammographic masses as benign or malignant. Local Binary Pattern (LBP), LBP Variance (LBPV), and Completed LBP (CLBP) descriptors are extracted to evaluate their potential for mass classification in a Computer-Aided Diagnosis (CAD) system. An Az value of 0.97 ± 0.02 and an accuracy of 92.25 ± 0.01% have been achieved, while experimenting on 200 mass cases from the DDSM database, by selecting the optimal set of features employing stepwise logistic regression method, followed by classification via Fisher Linear Discriminant Analysis (FLDA) using 10-fold cross validation.

Academic research paper on topic "A Study of Different Texture Features Based on Local Operator for Benign-malignant Mass Classification"

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedía Computer Science 93 (2016) 389 - 395

6th International Conference On Advances In Computing & Communications, ICACC 2016, 6-8

September 2016, Cochin, India

A study of different texture features based on local operator for benign-malignant mass classification

Rinku Rabidasa,H\ Abhishek Midyab, Jayasree Chakrabortyc, Wasim Arifa

aDepartment of Electronics and Communication Engineering,National Institute of Technology Silchar, Assam, 788010, India b Department of Electronics and Instrumentation Engineering,National Institute of Technology Silchar, Assam, 788010, India cDepartment of Surgery, Memorial Sloan Kettering Cancer Center, New York, NY, 10065, USA

Abstract

In this paper, a comparative analysis of different texture features based on local operator has been produced for the determination of mammographic masses as benign or malignant. Local Binary Pattern (LBP), LBP Variance (LBPV), and Completed LBP (CLBP) descriptors are extracted to evaluate their potential for mass classification in a Computer-Aided Diagnosis (CAD) system. An Az value of 0.97 ± 0.02 and an accuracy of 92.25 ± 0.01% have been achieved, while experimenting on 200 mass cases from the DDSM database, by selecting the optimal set of features employing stepwise logistic regression method, followed by classification via Fisher Linear Discriminant Analysis (FLDA) using 10-fold cross validation.

© 2016 The Authors. Published by ElsevierB.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 the Organizing Committee of ICACC 2016

Keywords: Breast Cancer;Mammography;Mass Classification;Local Binary Pattern(LBP);LBP Variance(LBPV);Completed LBP(CLBP)

1. Introduction

Among different types of cancers, breast cancer, though rare in male, frequent in women especially aged above 40 years has been ranked top in estimated number of new cases and deaths in 2015 The only way to reduce this mortality rate is to detect and diagnose breast cancer at its earlier stages. Mammography, an X-ray imaging technique, has been considered as the best effective technique for the detection of abnormalities present in the breast. Anomalies like calcification, architectural distortion, bilateral asymmetry, and masses are the common peculiarities visible in mammogram, but among these, detection of masses are the most difficult due to their subtle nature i.e. variations in shape, size, and margin. The growing number of new cases and the difficulty involves in the detection of masses, have made examination of mammograms a challenging task for the radiologists. Hence, in order to assist them, Computer-Aided Diagnosis (CAD) system has been developed as a second evaluator.

Generally the categorization of a mammographic mass as benign and malignant is done on the basis of shape, margin, and density. A mass having round/oval shape, well defined margin, and low density can be categorized

* Corresponding author. Tel.: +91-9085601375 E-mail address: rabidas.rinku@gmail.com

1877-0509 © 2016 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 the Organizing Committee of ICACC 2016 doi:10.1016/j.procs.2016.07.225

into benign case otherwise malignant case. As a result, several feature extraction methods based on shape, margin, and texture have been proposed for the classification of mammographic masses in the state-of-the-art. In general, textural information has been widely exploited against shape-2,3,4 and margin5,6,7-based features because the later approaches require accurate segmentation. In case of texture analysis, Haralick's features, extracted from Gray Level Co-occurrence Matrix (GLCM)8, Angle Co-occurrence Matrices (ACM)9, Optical Density Co-occurrence Matrix (ODCM)10 have been extensively utilized. Sahiner et al.11 has measured Haralick's features using GLCM matrix obtained from the Rubber-band straightening transform(RBST) images and achieved an Az value of 0.94 using 160 mammograms. Similar analysis has been produced by Chakraborty et al.12 where Haralick's features are extracted from ACM matrices and an Az value of 0.77 has been reported using 433 Region of interests (ROIs). Multi-resolution analysis of the oriented patterns has been evaluated by Midya and Chakraborty using Haralick's features and 0.86 has been observed as the highest Az value using 433 ROIs13. Tai et al.10 has also computed Haralick's descriptors from ODCM matrix and obtained an Az value of 0.98 for 398 mammograms. Frequency domain analysis has also been performed where coefficients of Wavelet transform have been used as feature vector14 to achieve an accuracy of 98.6% using 140 cases. Similar analysis has been reported by Eltoukhy et al.15 where an accuracy of 97.3% is achieved with curvelet transform. A few works have advocated the efficiency of combinational methods to get improved performance. Laroussi et al. has clubbed Zernike moments and Local Binary Pattern (LBP) features to achieve an Az value of 0.96 using 160 mammograms16.

In this correspondence, the authors have analyzed the discriminative capability of different texture feature extraction methods, based on the local operators, in a CAD system. Local binary pattern (LBP), LBP Variance (LBPV) and Completed LBP (CLBP) features with their variants have been extensively examined with an aim to achieve more precision in the classification of mammographic masses as benign or malignant.

The reminder of the paper is as follows: An overview of the DDSM database, utilized to carry out several experiments, has been provided in section 2. Section 3 presents a brief introduction of a CAD system and feature extraction methods investigated in the present work. The performance analysis of different features for benign-malignant mass classification has been discussed in section 4 and finally, the paper is concluded in section 5.

2. Image Database

To observe the performance of LBP, LBPV, and CLBP features for the classification of masses, several experiments have been conducted on the DDSM database which is an open source database provided by the University of South Florida17. The mammographic images of the DDSM database are of resolution of 42jum/pixel, 43.5jum/pixel, and 50jum/pixel and the boundary points inscribing the abnormalities present in the breast are provided with the database. In this present study, 200 randomly selected mass cases, 100 cases each of benign masses and malignant masses have been considered and the sample mass cases are shown in Fig. 1.

Fig. 1: Selected ROIs containing (a) Benign mass and (b) Malignant mass from the original images of the DDSM database.

3. Methodology

In this paper, considering significant difference in benign and malignant mass textures, textural informations are measured utilizing LBP, LBPV, and CLBP to bring out a comparative analysis of their proficiency in the classification of mammographic masses as benign or malignant. The schematic block diagram of a CAD system for benign-malignant mass classification is shown in Fig. 2.

Fig. 2: Schematic block diagram of a CAD system.

3.1. Region of Interest (ROI) selection

With each and every mammographic images, the boundary points of the suspected mass region have been provided with the DDSM database. Considering 10 pixels extra to the width and height of the provided mass region, a rectangular region inscribing the suspected mass has been cropped and selected as ROI and example of such ROIs are provided in Fig. 1.

3.2. Feature extraction

In a CAD system, feature extraction holds very important place as its overall efficiency is immensely dependent upon discriminative features. Hence, extraction of salient features is a challenging task. Considering the difference in spatial arrangement of pixels in benign and malignant masses, three local features: LBP, LBPV, and CLBP have been measured for benign-malignant mass classification. A brief introduction of LBP, LBPV, and CLBP features have been produced below:

• Local Binary Pattern (LBP): LBP, a successful texture classification technique, was introduced by Ojala et al.18 for extracting features by moving a local operator on each pixel location of the gray scale image. This operator generates patterns of binary digits i.e. LBP codes at any location (x, y) to obtain the LBP histogram of an image having size M x N.

LBPpr(x, y) = £ s(In - Ic)2b, (1)

s(z) = I1' Z * 0 , (0, z < 0

where In represents intensity value of any neighboring pixel centered around Ic indicating the pixel value of the center pixel at (x, y); P refers to the number of neighboring pixels and R denotes the radius of neighborhood. Sometimes the generated patterns are further classified into different variants like uniform patterns (LBPp^), rotation invariant uniform patterns (LBPPR2) and have been analyzed to classify texture images18. LBP variance (LBPV): Usually LBP histograms contains local spatial information but to make the features more resistance to rotation variations, additional contast information has been added by defining a joint histogram of LBP and Rotation invariant variance measures (VAR ) LBPprIVARpr 18. However, it has some drawbacks which has been overcome in LBPV proposed by Guo et al.19. In this method, an adaptive weight has been assigned to the LBP codes to obtain the LBPV histogram which can be mathematically expressed as:

LBPVpr(I) = 22 w(LBPpR(x, y), I), I e [0 , L] (2)

x=1 y= 1

rmp r ^ A iVARP'R(x,y) , LBPp^(x,y) = I w(LBPpr(x, y), l) = ^ . ,

10 , otherwise

VARpr = P 2(4 - I-)2 , (3)

Iav = P 2 In' (4)

where L indicates the maximal value of LBP codes.

Completed LBP (CLBP): Unlike LBP, instead of considering the sign difference only, magnitude difference between the center pixel and neighboring pixels in the local operator has been considered. In addition, the center pixel itself is encoded into a binary code using global thresholding. All these three distinct methods have been combined (jointly/concatenated) to obtain CLBP histogram which was proposed by Gou et al.20. In this method, the difference between In and Ic is observed in a local region to obtain a difference vector db which is further splitted into two components

db = Sb * Mb and \ b g ( b , (5)

b b b \Mb = |db|

where Sb indicates the sign of db given by Sb = 1 if db > 0 , otherwise 0 and Mb represents magnitude difference.

CLBP_Sign (CLBP_S) operator is same as traditional LBP where as CLBP_Magnitude (CLBP_M) operator can expressed as

CLBPMpr = 2 t(Mb , T)2b, (6)

t( j'T ) =

1, j > T 0 , j<T'

where T indicates the threshold estimated adaptively. Similarly CLBP_Center (CLBP_C) operator can be defined as

CLBPCpr = t(In,Ai), (7)

where t is as defined in equation (6) and AI represents the mean gray scale value of the entire image.

To extract the above mentioned features, initially the selected ROIs are resized to a size of 256 x 256 where the resized ROIs are divided into sub-images of size [M/WJ x [N/WJ obtained from the resized ROI of size M x N. LBP, LBPV, and CLBP features are measured from each sub-images and concatenated to form the feature vector of that ROI.

3.3. Feature selection and Classification

Since the measured descriptors are high dimensional and all the extracted features do not posses significant discriminating potential, hence, stepwise logistic regression method21 has been employed to select the optimal subset of features with Fisher Linear Discriminant Analysis (FLDA)22 for benign-malignant mass classification.

4. Experimental set up, Results, and Discussions

The experimental results provided, in this paper, have been obtained using MATLAB 2013a software in a computer having Intel(R) Core(TM) i7-3770 CPU processor clocked at 3.40 GHz with 8GB RAM.

The performance characteristics of LBP, LBPV, and CLBP have been evaluated in terms of area under the receiver operating characteristic (ROC) curve (Az) and accuracy (Acc) in percentage. A ten-fold cross validation technique has been incorporated with FLDA and the process is repeated 10 times to observe the mean of area under the ROC curve (Az) and mean accuracy (Acc) with their standard deviations as provided in Table 1, 2, and 3. The quantative analysis of the three LBP based feature extraction methods has been done by varying the values of P, R and W to show their efficacy in categorization of mammographic masses as benign or malignant.

In Table 1, performance analysis of LBP and its variants are provided where an Az value of 0.97 ± 0.01 and an accuracy of 92.25 ± 0.01% have been achieved as the best results. Similar analysis has been mentioned in Tables 2 and 3 for LBPV and CLBP features respectively, where an Az value of 0.95 ± 0.004 and an accuracy of 87.70 ± 0.008% for LBPV and the same for CLBP are 0.96 ± 0.001 and 90.60 ± 0.006% have been observed as their best results. Hence, it is clearly evident from the Table 1, 2, and 3 that LBP outperforms the other methods. The optimum results obtained with the individual methods have been compared with other currently developed schemes in Table 4 and also a comparison of ROC curves for the best results obtained from different methods have been demonstrated in Fig. 3.

Table 1: Performance analysis of different feature sets in terms of mean of area under the ROC curve (Az) and mean accuracy (Acc) with their respective deviations for different values of W. The best results are highlighted in bold face.

Methods Block Size (W=2) Block Size (W=3) Block Size (W=4) Block Size (W=5)

Feature setsfi Az Acc (%) Az Acc (%) Az Acc (%) Az Acc (%)

lbpfj 0.95 ± 0.003 90.05 ± 0.008 0.94 ± 0.007 87.50 ± 0.01 0.95 ± 0.01 89.85 ± 0.003 0.97 ± 0.002 92.95 ± 0.01

lbp^2 0.93 ± 0.004 85.20 ± 0.01 0.95 ± 0.002 91.20± 0.007 0.96 ± 0.002 89.60± 0.007 0.97 ± 0.001 91.55±0.008

lbp™2 0.75 ± 0.004 72.35 ± 0.01 0.83 ± 0.005 77.10 ± 0.01 0.84 ± 0.006 74.70 ± 0.01 0.89 ± 0.009 80.00 ± 0.001

lbp™| 0.74 ± 0.003 71.60 ± 0.007 0.76 ± 0.01 69.85 ± 0.01 0.88 ± 0.007 79.75 ± 0.01 0.89 ± 0.008 81.05 ± 0.01

Table 2: Performance analysis of different feature sets in terms of mean of area under the ROC curve (Az) and mean accuracy (Acc) with their respective deviations for different values of W. The best results are highlighted in bold face.

Methods Block Size (W=2) Block Size (W=3) Block Size (W=4) Block Size (W=5)

Feature set^ Az Acc (%) Az Acc (%) Az Acc (%) Az Acc (%)

lbpvg2 lbpv82,2 lbpv8'f lbpv™2 0.88 ± 0.01 0.82 ± 0.01 0.66 ± 0.005 0.69 ± 0.01 81.95 ± 0.02 76.90 ± 0.01 60.45 ± 0.009 64.15 ± 0.02 0.92 ± 0.01 0.89 ± 0.009 0.77 ± 0.009 0.75 ± 0.007 85.20 ± 0.01 80.25 ± 0.01 71.05 ± 0.008 70.25 ± 0.01 0.93 ± 0.006 0.95 ± 0.004 0.79 ± 0.01 0.87 ± 0.009 86.20 ± 0.01 88.56 ± 0.001 72.15 ± 0.01 82.00 ± 0..07 0.95 ± 0.004 0.94 ± 0.006 0.82 ± 0.01 0.86 ± 0.01 87.70 ± 0.008 86.05 ± 0.009 75.50 ± 0.001 78.55 ± 0.01

Table 3: Performance analysis of different feature sets in terms of mean of area under the ROC curve (Az) and mean accuracy (Acc) with their respective deviations for different values of W. The best results are highlighted in bold face.

Methods

Block Size (W=2) Block Size (W=3) Block Size (W=4) Block Size (W=5)

Feature set,

Acc (%) Az Acc (%) Az Acc(%) Az Acc (%)

0.006 78.80 ± 0.01 0.89 ± 0.009 81.75 ± 0.008 0.90 ± 0.006 81.00 ± 0.01 0.93 ± 0.007 83.15 ± 0.01

0.009 80.95±0.001 0.94 ± 0.003 86.15 ± 0.01 0.91 ± 0.007 82.85 ± 0.009 0.95 ± 0.002 87.50 ± 0.01

0.006 65.50 ± 0.01 0.75 ± 0.01 68.60 ± 0.01 0.78 ± 0.005 71.90 ± 0.006 0.75 ± 0.005 70.05 ± 0.01

0.006 71.25 ± 0.009 0.77 ± 0.005 72.30±0.005 0.74 ± 0.009 69.30±0.008 0.74 ± 0.007 69.30±0.007

0.004 85.10 ± 0.01 0.93 ± 0.005 85.75 ± 0.01 0.92 ± 0.006 82.95 ± 0.01 0.93 ± 0.004 85.80 ± 0.01

0.008 81.20 ± 0.01 0.92 ± 0.004 84.05 ± 0.01 0.94 ± 0.003 85.20 ± 0.01 0.94 ± 0.004 85.45 ± 0.007

0.002 71.15 ± 0.001 0.77 ± 0.009 70.35 ± 0.01 0.80 ± 0.005 70.85 ± 0.01 0.75 ± 0.01 68.75 ± 0.01

0.004 69.90±0.007 0.87 ± 0.006 78.35 ± 0.01 0.84 ± 0.007 77.30 ± 0.01 0.83 ± 0.008 76.55 ± 0.01

0.004 89.00 ± 0.01 0.94 ± 0.004 87.10 ± 0.01 0.93 ± 0.007 84.45 ± 0.01 0.94 ± 0.002 88.65 ± 0.008

0.004 82.60 ± 0.01 0.96 ± 0.003 89.85±0.009 0.95 ± 0.003 87.60 ± 0.01 0.94 ± 0.004 87.10 ± 0.01

0.007 75.10 ± 0.009 0.87 ± 0.01 80.90 ± 0.01 0.86 ± 0.007 77.20±0.007 0.82 ± 0.008 75.85 ± 0.01

0.01 78.25 ± 0.01 0.85 ± 0.009 76.65 ± 0.01 0.86 ± 0.01 77.55 ± 0.01 0.86 ± 0.01 77.85 ± 0.01

0.003 87.10 ± 0.009 0.94 ± 0.006 86.60 ± 0.01 0.96 ± 0.002 90.30±0.008 0.96 ± 0.001 90.60 ± 0.006

0.008 80.65 ± 0.01 0.88 ± 0.006 80.40 ± 0.01 0.91 ± 0.008 82.10 ± 0.01 0.87 ± 0.01 78.35 ± 0.01

0.007 81.85 ± 0.001 0.92 ± 0.002 85.05 ± 0.01 0.87 ± 0.004 78.80 ± 0.01 0.92 ± 0.004 83.95 ± 0.01

0.003 87.60±0.001 0.95 ± 0.005 89.15 ± 0.01 0.97 ± 0.002 91.30 ± 0.009 0.96 ± 0.004 89.60±0.009

0.004 82.15 ± 0.009 0.93 ± 0.005 85.00 ± 0.01 0.89 ± 0.009 81.85 ± 0.01 0.88 ± 0.007 81.90±0.009

0.002 88.40±0.006 0.89 ± 0.007 81.55 ± 0.01 0.95 ± 0.003 87.80±0.009 0.93 ± 0.005 87.60 ± 0.01

clbp_mfj

clbp_m^2

clbp_m8''f

clbp_m™2

clbpJm/c8 clbp^/q6 2 clbp^/c™2

0.87 d 0.90 d 0.71 d 0.76 d 0.91 d 0.91 0.77

-16.2 ^ clbp-s-m/c^ 0.95 d

clbp_s_mc8|20.9i d clbp_s_m/c8,6i20.81 -clb^sm/c'j'6'20.85 d clbp-s/m^/ ' 0.95 d

clbp_s/m8'/2 0.90 d clbp_s/m™2 0.89 d clbp_s/m/c8;/ 0.95 d clbp-s/m/cs'f 0.90d

Table 4: Comparison of the best results of different methods with other recently developed competing schemes in terms of area under the ROC curve (Az).

Methods/References Database Az

LBP^2, LBPVÙj CLBP-S/M^2 DDSM DDSM DDSM 0.97 ± 0.002 0.95 ± 0.004 0.96 ± 0.001

Chakraborty et al.12 Midya and Chakraborty13 Nascimento et al.23 Laroussi et al.16 Tahmasbi et al.24

DDSM DDSM DDSM DDSM mini-MIAS

0.77 0.86

0.96 ± 0.04

1 M 'V."2

'I'll --, .■■ -, ■, u2

False Positive Rate

Fig. 3: Comparison of ROC curves for the best results obtained with different feature extraction methods.

5. Conclusion

This paper presents a comparative study of different texture features for the categorization of mammographic masses as benign or malignant. LBP, LBPV, and CLBP features are rigorously evaluated and the obtained results

provide satisfactory performance when compared with other competing schemes. Since CLBP, a combination of basic LBP, CLBP_M, and CLBP_C, has large feature size which performs lower than LBP, hence, other feature selection technique can be examined in future to have more promising CAD system for benign-malignant mass classification.

References

1. National Cancer Institute(NCI), Cancer stat fact sheets: cancer of the breast, http://seer.cancer.gov/statfacts/html/breast.html. accessed in 2015.

2. P. Grgel, A. Sertbas, O. N. Ucan, Mammographical mass detection and classification using local seed region growing-spherical wavelet transform (LSRG-SWT) hybrid scheme, Computers in Biology and Medicine 43 (2013) 765 - 774.

3. A. Vadivel, B. Surendiran, A fuzzy rule-based approach for characterization of mammogram masses into BI-RADS shape categories, Computers in Biology and Medicine 43 (2013) 259 - 267.

4. A. Serifovic-Trbalic, A. Trbalic, D. Demirovic, N. Prljaca, P. Cattin, Classification of benign and malignant masses in breast mammograms, in: 37th International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO), 2014, pp. 228-233.

5. K. Bojar, M. Nieniewski, New features for classification of cancerous masses in mammograms based on morphological dilation, in: 5th International Conference on Visual Information Engineering, (VIE), 2008, pp. 111-116.

6. L. Sun, L. Li, W. Xu, W. Liu, J. Zhang, G. Shao, A novel classification scheme for breast masses based on multi-view information fusion, in: 4th International Conference on Bioinformatics and Biomedical Engineering (ICBBE), 2015, pp. 1-4.

7. C.-H. Wei, S. Y. Chen, X. Liu, Mammogram retrieval on similar mass lesions, Computer Methods and Programs in Biomedicine, 106 (2012) 234 - 248.

8. R. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification, IEEE Transactions on Systems, Man and Cybernetics, 3 (1973)610-621.

9. J. Chakraborty, R. M. Rangayyan, S. Banik, S. Mukhopadhyay, J. E. Leo Desautels, Statistical measures of orientation of texture for the detection of architectural distortion in prior mammograms of interval-cancer, Journal of Electronic Imaging 21 (2012) 033010-1-03301013.

10. S.-C. Tai, Z.-S. Chen, W.-T. Tsai, An automatic mass detection system in mammograms based on complex texture features, IEEE Journal of Biomedical and Health Informatics, 18 (2014) pp. 618-627.

11. B. S. Sahiner, H.-P. Chan, N. Petrick, M. A. Helvie, M. M. Goodsitt, Computerized characterization of masses on mammograms: The rubber band straightening transform and texture analysis,, Medical Physics 24 (1998) 516-526.

12. J. Chakraborty, A. Midya, S. Mukhopadhyay, A. Sadhu, Automatic characterization of masses in mammograms, in: 6th International Conference on Biomedical Engineering and Informatics (BMEI), 2013, pp. 111-115.

13. A. Midya, J. Chakraborty, Classification of benign and malignant masses in mammograms using multi-resolution analysis of oriented patterns, in: 12th International Symposium on Biomedical Imaging (ISBI), 2015, pp. 411-414.

14. S. Joseph, K. Balakrishnan, Local binary patterns, Haar wavelet features and Haralick texture features for mammogram image classification using artificial neural networks, in: Advances in Computing and Information Technology, Vol. 198, 2011, pp. 107-114.

15. M. M. Eltoukhy, I. Faye, B. B. Samir, A statistical based feature extraction method for breast cancer diagnosis in digital mammogram using multiresolution representation, Computers in Biology and Medicine, 42 (2012) 123 - 128.

16. M. Laroussi, N. Ben Ayed, A. Masmoudi, D. Masmoudi, Diagnosis of masses in mammographic images based on Zernike moments and Local binary attributes, in: World Congress on Computer and Information Technology (WCCIT), 2013, pp. 1-6.

17. K. Bowyer, R. Moore, D. Kopans, M. Heath, W. P. Kegelmeyer, Digital database for screening mammography, in: Proceedings of the Fifth International Workshop on Digital Mammography, Medical Physics Publishing, 2001.

18. T. Ojala, M. Pietikainen, T. Maenpaa, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns, IEEE Transactions on Pattern Analysis and Machine Intelligence, 24 (2002) 971-987.

19. Z. Guo, L. Zhang, D. Zhang, Rotation invariant texture classification using LBP variance (LBPV) with global matching, Pattern Recognition 43 (3) (2010) 706-719.

20. Z. Guo, D. Zhang, D. Zhang, A completed modeling of local binary pattern operator for texture classification, IEEE Transactions on Image Processing 19 (6) (2010) 1657-1663.

21. F. L. Ramsey, D. W. Schafer, The Statistical Sleuth: A Course in Methods of Data Analysis, CA: Duxbury Press, 1997.

22. R. Duda, P. Hart, D. G.Stork, Pattern Classification, 2nd Edition, New York, NY: Wiley-Interscience, 2001.

23. M. Z. do Nascimento, A. S. Martins, L. A. Neves, R. P. Ramos, E. L. Flores, G. A. Carrijo, Classification of masses in mammographic image using wavelet domain features and polynomial classifier, Expert Systems with Applications 40 (2013) 6213 - 6221.

24. A. Tahmasbi, F. Saki, S. B. Shokouhi, Classification of benign and malignant masses based on zernike moments, Computers in Biology and Medicine, 41 (2011) 726 - 735.