Scholarly article on topic 'Adaptive Gradient Descent Backpropagation for Classification of Breast Tumors in Ultrasound Imaging'

Adaptive Gradient Descent Backpropagation for Classification of Breast Tumors in Ultrasound Imaging Academic research paper on "Medical engineering"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{"Breast tumor classification" / backpropagation / "gradient descent" / "variable learning rate" / "classification accuracy."}

Abstract of research paper on Medical engineering, author of scientific article — Bikesh Kumar Singh, Kesari Verma, A.S. Thoke

Abstract This paper investigates and evaluates the performance of three gradient descent based backpropagation artificial neural network (ANN) algorithms in classifying the tumor as benign and malignant in ultrasound imaging. The ultrasound images were preprocessed by wavelet filters for reducing speckle noise. Fifty seven texture and shape attributes were extracted from filtered breast ultrasound images to classify breast tumors. Area under receiving operating curve (AUC), sensitivity, specificity, classification accuracy and CPU time were used as figure of merit for the classifier. Results show that adaptive gradient descent backpropagation based on variable learning rate outperformed other techniques giving highest classification accuracy of 84.6%.

Academic research paper on topic "Adaptive Gradient Descent Backpropagation for Classification of Breast Tumors in Ultrasound Imaging"

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedía Computer Science 46 (2015) 1601 - 1609

International Conference on Information and Communication Technologies (ICICT 2014)

Adaptive gradient descent backpropagation for classification of breast tumors in ultrasound imaging

Bikesh Kumar Singha*, Kesari Vermab, A. S.Thokec

aDepartment of Biomedical Engineering, N.I.T. Raipur,492010, India bDepartment of Computer Applications, N.I.T. Raipur,492010, India cDepartment of Electrical Engineering, N.I.T. Raipur, 492010, India

Abstract

This paper investigates and evaluates the performance of three gradient descent based backpropagation artificial neural network (ANN) algorithms in classifying the tumor as benign and malignant in ultrasound imaging. The ultrasound images were preprocessed by wavelet filters for reducing speckle noise. Fifty seven texture and shape attributes were extracted from filtered breast ultrasound images to classify breast tumors. Area under receiving operating curve (AUC), sensitivity, specificity, classification accuracy and CPU time were used as figure of merit for the classifier. Results show that adaptive gradient descent backpropagation based on variable learning rate outperformed other techniques giving highest classification accuracy of 84.6%. © 2015TheAuthors.PublishedbyElsevierB.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:Breast tumor classification; backpropagation; gradient descent; variable learning rate; classification accuracy.

1. Introduction

Among various types of cancer in women, the second leading cause for death in women is breast cancer. In past few years it is one of the major health issues as its incidence is increased in recent years. Being the most frequent type of cancer, it is responsible for more than 1.6% of the women deaths across world1.

* Corresponding author. Tel.: +91,9826469522 E-mail address: bsingh.bme@nitrr.ac.in

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.2015.02.091

In India, a recent study reported that, one in twenty eight women are diagnosed with breast cancer throughout her

lifespan. In urban area this statistics is even higher of approximately one in twenty two in a lifespan compared to

rural areas. The women age group of 43-46 years is at higher risk in India while in west the women within age group

53-57 years are more susceptible to breast cancer1' 2. Mammography is one of the best existing screening tools for

detection of breast cancer in early stages. However breast ultrasound has proved to be an important appendage to

mammography for patients with dense breast, palpable masses and normal mammograms1' 2' 3' 4. Computer aided

diagnosis (CAD) systems have come out as second diagnostic tool for classification of breast masses. A neural

network to evaluate breast lesions using texture features is reported in5. ROC analysis for classifying breast masses

was proposed in6. Collected data was normalized (Amplitude scaling) before processing. Ten features including

Patients age and radiologists pre biopsy level of suspicion (LOS), two features were computed from Nakagami

distribution for envelope of the backscattered echo, four features were derived using generalized spectrum model

(GS), and two features were based on PLSN (power law shot noise) model of ultrasound RF echo. ROC (Receiver

operating characteristics) analysis was used for performance evaluation. In7 a CAD system for diagnosing solid

breast nodules using ANN based on multiple sonographic features was proposed. Multilayer perceptron using back

propagation algorithm for classification of solid breast nodules was reported in8 using texture features. Some other

CAD systems based on cytological features9, morphometric features10, texture, fractal and histogram based

features11, 12 were also reported. A classifier based on unsupervised learning approach was proposed in13 for

recognition and categorization of breast masses in breast ultrasound images using texture and morphological

features. Classification based on combination of multiple sonographic features and texture features was reported

in14. Computer assisted lesion diagnosis in three dimensional breast ultrasound using coronal speculation was

discussed in15. Behavior of co-occurrence texture statistics for classifying breast ultrasound was investigated in16.

Area under ROC, accuracy, specificity, sensitivity, positive predictive value and negative predictive value were used

for evaluating the classifier. Some comparative studies of ANN with other classification techniques were worked out in17, 18.

The convergence time of the classification process highly depends on the learning algorithm and learning rate. Algorithms like Levenberg Marquardt may result in high accuracy and small convergence time but at the same time present 'out of memory' issues for large feature set. Thus appropriate choice of learning rate, time complexity, network model and architecture for a given task is still an important issue in CAD systems. This paper investigates and evaluates the performance of three gradient descent based backpropagation artificial neural network (ANN) algorithms to classify breast masses in to benign and malignant. The experiments were conducted to the new patient data. The remaining part of the paper is organized as follows. In Section two material and methods used in carrying out the experiments is discussed. Section three presents results and discussion followed by conclusion and future directions.

2. Material and methods

2.1. Dataset

The image database was collected consecutively during routine breast diagnostic procedures at Pt. J.N.M

Government Medical College Raipur (C.G), India. All of them were histopathologically proven cases. The dataset

consists of 89 sonograms including 44 benign and 45 benign tumors.

2.2. Preprocessing

Breast ultrasound images suffer from intrinsic artifact called speckle resulting in low resolution, poor contrast and

blurry edges. A lot of speckle reducing methods have been developed by researchers. We employ a wavelet based

despeckle filter due to its effectiveness in providing smoothening while preserving edges, boundaries and other sharp

details. Wavelet decomposition of image results in four bands namely LL, HH, LH and HL. Wavelet filtering

involves eliminating any of the above band or their combinations and then reconstructing the original image19. Since

speckle noise is contained in the higher frequencies, experiments were conducted using Haar wavelet by eliminating

HH band after first level decomposition. To preserve important textures, edges and sharp features of the image, only first level decomposition was performed to avoid over smoothening. A rectangular region of interest (ROI) containing tumor and its neighboring area was then manually cropped under guidance of radiologist. This is illustrated in Fig. 1.

Fig. 1. (a) Original noisy image (b) Wavelet filtered image (c) Cropped ROI

2.3. Feature extraction

Table 1. Texture and shape attributes used in classification of breast tumor

Category of feature Number of features Name of features

First order statistics 5 Mean, variance, median, skewness, kurtosis21

Fractal based texture features 4 Hurst coefficient H(k) for k=1- 423 24, 25 26

Neighbourhood gray tone difference features 5 Coarseness, contrast, busyness, complexity, strength22, 23, 24

Statistical features 4 Coarseness, contrast, periodicity, roughness24, 25, 26, 27

Gray level difference statistics 4 Contrast, angular second moment, entropy, mean23, 24, 25

Haralick texture features 26 Mean and range following features over four angles were calculated: Angular second moment, contrast, correlation, sum of squares, inverse difference moment, sum average, sum variance, sum entropy, entropy, difference variance, difference entropy, information measures of correlation-1 and information measures of correlation-228, 29

Laws texture features 6 LL, EE, SS, LE, ES, LS30

Shape features 3 Area, perimeter and perimeterA2 /area23, 24

Extracting suitable features and selecting relevant ones of them is of paramount importance and vital task in breast cancer representation and classification. Appropriate features and their dimension directly affect the performance of the classifier and execution time. Too many irrelevant features may result in a complicated and reduced accuracy. The features should be selected so that they must result in improved accuracy of the system. Table 1 shows various texture and shape features extracted from cropped ROI. The details of these features can be found in references mentioned therewith. In order to determine the optimal number of features that contributed in improving classification accuracy, initially only first five features were used in training of backpropagation ANN using adaptive gradient descent learning algorithm. New features were added to training dataset and classification accuracy

was recorded after adding each set of features. Fig. 2 shows the variations in accuracy with increase in number of features. It was found that maximum accuracy of 92.1% (combined accuracy of training, testing and validation) was obtained up to 40th feature after which accuracy decreases. Hence only first 40 features were used in training of classification system.

95-1-1 i-1 i

50 1-1-1-1--1-

0 10 20 30 40 50 60

Number of features

Fig. 2. Graph showing classification accuracy vs. number of features

2.4. Classification using ANN

Of various neural network models, the most widely accepted one is backpropogation neural network [BPNN] classifier. It provides the needed weight adjustments in the backward sweep13. Gradient descent backpropogation artificial neural network has been successively used for classification of breast tumors by considerable number of researchers. Its performance can be improved by using variable learning rate during the training process i.e instead of using constant learning rate; adaptive learning rate can be employed. The affect of using variable learning rate will be larger step size and stable learning. In this study adaptive gradient descent (AGD) algorithm based on variable learning rate is used for classification of breast tumors. The results were compared with that of basic gradient descent (GD) and gradient descent with momentum (GDM) algorithm. The weight update rule of GD, GDM and AGD algorithms are summarized in table 2.

In adaptive gradient descent, the first step is to determine the output of initial network and corresponding error. Then at each iteration new weights and biases are determined using the present learning rate. Again new outputs and errors are determined. Then if the new error is above the previous one by more than a predefined value (1.04 in our case), the new weights and biases are rejected. Further, the learning rate is reduced (typically by scaling with constant say k1). Otherwise, the new weights are accepted. If the present error is smaller than the old error, the learning rate is augmented (typically by scaling with constant say k2)31. The network architecture consisted of one input layer, two hidden layers and one output layer. Number of neurons in hidden layers was decided after experimenting multiple times so that maximum classification accuracy can be obtained. As a consequence number of neurons in first hidden layer was fixed at 20 and that of second layer at 1. Transfer function used were hyperbolic tangent segment at the output of first layer and linear at the output of second and third layer respectively. Learning rate, momentum (for GDM) and performance goal were fixed at 0.01, 0.9 and 0 respectively. The value of constants

k1 and k2 for AGD were fixed at 0.7 and 1.05 respectively. The network architecture is illustrated in Fig. 3.

Table 2. Weight update rule in GD, GDM and AGD algorithms

Name of algorithm Weight update equation

Remarks

Wk+i = Wk - a gk Wk+i = Wk - a gk + ^ Wk-i Wk+i = Wk - ak+i gk

a is learning rate fixed at 0.01.

^ is momentum fixed at 0.9

ak+i = Y ak and y = ki (if new error > 1.04 (old error) and k2 (if new error < 1.04 (old error), k1 = 0.7 and k2 = 1.05.

Fig. 3. Architecture of ANN

2.5. Performance metrics

In this work, AUC (area under receiver operating characteristic curve), sensitivity (true positive rate), specificity (true negative rate), accuracy (on test samples) and elapsed time were used to assess the performance of classifiers. AUC, sensitivity, specificity and accuracy are determined using following equations-

AUC = -

21 TP + FN TN + FP TP

Sensitivity (%) = Specificity (%) = Accuracy (%) =

FN + TP TN

TN + FP

TP + TN TP + FN + TN + FP

(i) (2)

where TP, TN, FP, FN are true positive rate, true negative rate, false positive rate and false negative rate respectively.

3. Results and discussions

Out of 89 samples containing 44 benign and 45 malignant cases, 70% were used for training of classification models i.e. GD, GDM and AGD 15% each for testing and validation. Fig. 4 shows receiver operating characteristic obtained for these classifiers. It is evident from Fig. 4 that area under ROC is highest for AGD and lowest for GDM. Table 3 shows the summary of results for various classifier models in terms of performance metrics.

o-■-1-1-1-1-1-'-'-1-1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 False Positive Rate

Fig. 4. ROC curve for (a) GD; (b) GDM; (c) AGD

Table 3. Summary of results obtained for backpropagation classifiers.

Name of algorithm AUC Sensitivity Specificity Accuracy CPU time(seconds)

GD 0.73 60% 87.5% 76.9% 1.92

GDM 0.68 36% 100% 46.2% 0.65

AGD 0.90 100% 81.8% 84.6% 2.32

GDM Classifier model

Fig. 5. Performance of the classifier based on (a) AUC ;(b) Sensitivity ;(c) Specificity ;(d) Classification accuracy ;(e) CPU time.

Fig. 5 shows the performance of classifier based on AUC, sensitivity, specificity, accuracy and CPU time. In terms of AUC and accuracy AGD algorithm outperforms other classifiers giving 84.6% classification accuracy. This indicates that AGD algorithm has highest capability in classifying the dataset even with less number of samples and features. However the CPU time for AGD classifier is quite high compared to other classifiers. It is found that GDM classifier converges faster as compared to GD and AGD but its classification accuracy is extremely poor. All simulations were carried out on MATLAB® platform.

4. Conclusions

Accurate diagnosis and classification of breast tumors is an important issue to reduce inadequate surgeries and unnecessary number of biopsies. Gradient descent backpropogation artificial neural network has been successively used for classification of breast tumors by considerable number of researchers. However appropriate choice of learning rate, time complexity, network model and architecture for a given classification task is still an important issue. This work presented a comparative analysis of gradient descent based backpropagation neural network for classifying breast tumors using 40 texture and shape features. Experiments were conducted on new database of 89 historically confirmed real breast ultrasound images containing 44 benign and 45 malignant breast cases using MATLAB® software platform. Three algorithms namely gradient descent (GD), gradient descent with momentum (GDM) and adaptive gradient descent (AGD) were evaluated in terms of AUC, sensitivity, specificity, accuracy and CPU time. It was shown that gradient descent back propagation neural network can be successfully used for classification of breast tumors. It was found that AGD based backpropagation outperformed other algorithms due to its highest capability in classifying benign and malignant tumors while achieving classification accuracy up to 84.6%. However AGD algorithm suffers from time complexity, hence in future we think to evaluate some other classification algorithms which can converge faster and at the same time can give high classification accuracy. Further the performance of classifiers on enlarged dataset with more number of descriptive features will be evaluated.

References

1. About breast cancer, online available: http://tmc.gov.in/cancerinfo/breast.

2. National centre for disease informatics and research, Indian council of medical research, online available: http://www.pbcrindia.org.

3. American Cancer Society, Inc., Surveillance Research, 2012.

4. Cancer Prevention and Early Detection: Facts and Figures, 2012 (American Cancer Society Cancer Action Network, 2012).

5. Zheng Y, Greenleaf James F, Gisvold John J, Reduction of breast biopsies with a modified self organizing map, IEEE Transactions On Neural Networks, 1996; 8: 1386-1396.

6. Gefen S, Tretiak Oleh J, Catherine WP, Kevin DD, Athina PP, Mohana SP, Vishruta AD, Huang L, Kutay MA, Genis V, Forsberg F, Reid JM and Goldberg BB, ROC analysis of ultrasound tissue characterization classifiers for breast cancer diagnosis. IEEE Transactions on Medical Imaging, 2003; 22: 170-177.

7. Joo S, Yang YS, Moon WK and Kim C, Computer Aided Diagnosis of solid breast nodules: Use of an artificial neural network based on multiple sonographic features. IEEE Transactions on Medical Imaging, 2004; 23:1292-1300.

8. Alvarenga AV , Pereira WCA, Infantosi AFC, Azevedo CM, Classifying Breast tumors on ultrasound images using a hybrid classifier and texture features. Proceedings of IEEE International Symposium on Intelligent Signal Processing (WISP 2007), 2007, p.1-6.

9. Subhashini TS, Ramalingam V and Palanivel S, Breast mass classification based on cytological patterns using RBFNN and SVM. Expert Systems with Applications, 2009; 36:5284-5290.

10. Pereira WCA, Alvarenga AV and Infantosi AFC, A non-linear morphometric feature selection approach for breast tumor contour from ultrasonic images. Computers in Biology and Medicine, 2010; 40:912-918.

11. Shi X, Cheng HD, Hu L, Ju W, and Tian J, Detection and classification of masses in breast ultrasound images. Digital signal Processing, 2010; 20: 824-836.

12. Liu B, Cheng HD, Huang J, Tian J, Tang X and Liu J, Fully automatic and segmentation robust classification of breast tumors based on local texture analysis of ultrasound images. Pattern Recognition, 2010; 43: 280-298.

13. Su Y, Wang Y, Jiao J and Guo Y, Automatic Detection and Classification of Breast Tumors in Ultrasonic Images Using Texture and Morphological Features. The Open Medical Informatics Journal, 2011; 5: 26-37.

14.Liao R, Wan T and Qin Z, Classification of benign and malignant breast tumors in ultrasound images based on multiple sonographic features and textural features. Proceedings of third international conference on intelligent human machine systems and cybernetics, 2011, p. 71-74.

15. Tan T, Platel B, Huisman H, Sanchez CI, Mus R and Karssemeijer N, Computer Aided Lesion diagnosis in automated 3-D Breast Ultrasound using coronal speculation. IEEE Transactions on Medical Imaging, 2012; 31:1034-1042.

16. Gomez W, Pereira WCA, and Infantosi AFC, Analysis of Co-occurrence Texture Statistics as a Function of Gray-Level Quantization for Classifying Breast Ultrasound. IEEE Transactions on Medical Imaging, 2012; 31:1889 -1899.

17. Ren J, ANN vs. SVM: Which one performs better in classification of MCCs in mammogram imaging. Knowledge based systems, 2012; 20: 144-153.

18. Ubaidillah SHSA, Sallehuddin R and Ali NA, Cancer detection using artificial neural network and support vector machine: A comparative study. Jurnal Teknologi, 2013; 65:73-81.

19. Mateo JL and Caballero AF, Finding out general tendencies in speckle noise reduction in ultrasound images. Expert systems with applications, 2009; 36:7786-7797.

20. Wu JCM., and Chen YC., Statistical Feature Matrix for Texture Analysis. Computer Vision, Graphics, and Image Processing; Graphical Models and Image Processing, 1992; 54: 407-419.

21. Gonzalez RC and Woods R E, Digital Image Processing Using MATLAB. 2nd edition, 2010.

22. Moses A and King R, Textural Features Corresponding to Textural Properties. IEEE Trans. On Systems, Man and Cybernetics, 1989; 19:1264-1274.

23. Loizou CP, Pattichis CS, Pantziaris M, Tyllis T and Nicolaides A, Quality evaluation of ultrasound imaging in the carotid artery based on normalization and speckle reduction filtering. Med. Biol. Eng. Comput., 2006; 44: 414-426.

24. Loizou CP, Pattichis CS, Christodoulou CI, Istepanian RSH, Pantziaris M and Nicolaides A, Comparitive evaluation of despeckle filtering in ultrasound imaging of the carotid artery. IEEE Transactions on Ultrasonics, Ferroelectrics and Frequency Control, 2005; 52: 1653-1669.

25. Chen DR, Chang RF, Chen CJ, Ho MF, Kuo SJ, Chen ST, Hung SJ, and Moon WK, Classification of breast ultrasound images using fractal feature. Journal of Clinical Imaging, 2005; 29: 235-245.

26. Wu CM, Chang CY, Sheng HK, Texture features for classification of ultrasonic liver images. IEEE Trans. Med. Imag., 1992; 11: 141-152.

27. Wu CM and Chen YC, Statistical Feature Matrix for Texture Analysis. CVGIP: Graphical Models and Image Processing, 1992; 54: 407-419.

28. Haralick RM, Statistical and Structural Approaches to Texture. Proc. IEEE, 1979; 67: 786-804.

29. Haralick RM, Shanmugam K, and Dinstein I, Textural Features for Image Classification. IEEE Trans. On Syst., Man & Cybernetics, 1973; SMC-6: 610-621.

30. Laws KI, Rapid texture identification. SPIE, 1980; 238:376-380.

31. Mathworks, Neural network toolbox: User's Guide. Retrieved 31 August 2014, from http:// www. mathworks .in.