Machine learning techniques for classification of breast tissue

Available online at www.sciencedirect.com

ScienceDirect Available online at www.sciencedirect.com Procedia Computer Science00 (2017) 000–000

ScienceDirect

www.elsevier.com/locate/procedia

Procedia Computer Science 120 (2017) 402–410

9th 9thInternational InternationalConference Conferenceon onTheory Theoryand andApplication Applicationof ofSoft SoftComputing, Computing,Computing Computingwith with Words and Perception, 22-23 August 2017, Budapest, Hungary Words and Perception, ICSCCW 2017, 24-25 August 2017, Budapest, Hungary

Machine learning techniques for classification of breast tissue Abdulkader Helwana*, John Bush Idokob, Rahib H.Abiyevb a

Biomedical Engineering Department Near East University, POBOX:99138, Nicosia, North Cyprus, Mersin 10, Turkey Engineering Department Near East University, POBOX:99138, Nicosia, North Cyprus, Mersin 10, Turkey

bComputer

Abstract This paper presents an automated classification of breast tissue using two machine learning techniques: Feedforward neural network using the backpropagation learning algorithm (BPNN) and radial basis function network (RBFN). The two neural network models are implored basically to identify the best model for breast tissue classification after an intense comparison of experimental results. An electrical impedance spectroscopy method was used for data acquisition while BPNN and RBFN were the models implored for the execution of the classification task. The approach implored in this paper is made out of the following steps; feature extraction, feature selection and classification steps. The features are obtained using the electrical impedance spectroscopy (EIS) at the feature extraction stage. These extracted features are impedance at zero frequency (I0), the high frequency slope of phase angle, the phase angle at 500KHz, the area under spectrum, the maximum of spectrum, the normalized area, the impedance distance between spectral ends, the distance between the impedivity at I0 and the real part of the maximum frequency point and the length of the spectral curve. Information theoretic criterion is the strategy used in the proposed algorithm for feature selection and classification phase that was executed using the BPNN and RBFN. The performance measure of the two algorithms is the accuracy of the BPNN and RBFN models. The RBFN outperforms the BPNN in terms of accuracy in classifying breast tissues, minimum square error reached, and time to learn as demonstrated in the experimental results. © 2018 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 9th International Conference on Theory and application of Soft Computing, Computing with Words and Perception. Keywords: Breast tissue; electrical impedance spectroscopy; neural networks; radial basis function network classifier.

* Corresponding author. . Tel.: +905488789719 ; fax: +903922236622 E-mail address: [email protected] 1877-0509 © 2018 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of the 9th International Conference on Theory and application of Soft Computing, Computing with Words and Perception.

1877-0509 © 2018 The Authors. Published by Elsevier B.V. Peer-review under responsibility of the scientific committee of the 9th International Conference on Theory and application of Soft Computing, Computing with Words and Perception. 10.1016/j.procs.2017.11.256

2

Abdulkader Helwan et al. / Procedia Computer Science 120 (2017) 402–410 Abdulkader Helwan et al./ Procedia Computer Science00 (2017) 000–000

403

1. Introduction Electrical impedance spectroscopy remains the key screening tool for breast tissue classification as well as the detection of abnormalities, because it allows identification of tumor before being palpable. Vacek et al. (2002) demonstrated that the proportion of breast tumors that were detected in vermont by screening mammography increased from 2% between 1974−1984 to 36% between 1995−1999. However, of all lesions previously diagnosed as suspicious and sent for biopsy, approximately 25% were confirmed malignant lesions, and approximately 75% were diagnosed benign. This high false-positive rate is related to the difficulty of obtaining accurate diagnosis as depicted by Basset and Gold (1987). For this reason, computerized image analysis plays an essential role in improving issues with diagnosis. Computer-Aided Diagnosis (CAD) systems are composed of a set of tools to help radiologists diagnose and detect new cases. Ho and Lam (2003) have shown that the sensitivity of these systems has significantly decreased as the density of the breast increased while the specificity of the systems remained relatively constant. The dataset used in this paper was deduced from the operations of electrical impedance spectroscopy (EIS) and could be found at the UCI repository. The electrical impedance procedures have for quite some time been utilized in classifying tissue as well as impedocardiography applications by Kubicek et al. (1970). These strategies have additionally empowered impedance mapping by Tachibana et al. (1970) and Henderson and Webster (1994) and recently, dynamic imaging by Brown et al. (1994). The AC equivalent of resistivity for DC current is known as impedivity/particular impedance. The electric and dielectric properties dictates impedivity of a tissue and this depend, in addition to other things such as; membrane capacitance, cell concentration, intracellular medium and the electric conductivity in interstitial space as mentioned by Gabriel et al. (1996). Some of the good features of impedance techniques includes; minimum invasiveness, easiness and low cost. Initially in the 80s, estimations of electric and dielectric has been performed in breast tissue under a scope of test conditions incorporating into vivo/ex-vivo estimations as well as utilizing different methods of measurement given in Surowiec et al. (1988), Campbell and Land (1992), Heinitz and Minet (1995). In the 488Hz-1MHz range, Jossinet (1998) found critical contrasts in phase angle and impedivity modulus from among the six groups of breast tissue. EIS is conceivably applicable in breast cancer detection and breast tissue separation as proposed in the above discoveries. Using EIS, this paper demonstrate a strategy for the classification of breast tissues. Feature set utilized in this paper is the same as those features defined by Jossinet and Lavandier (1998) and also extra features chose for their separation capacity. Twelve-point and seven-point spectra were used to choose the statistical hierarchical approach. A non-invasive strategy used to measure the impedance of cells in a scope of frequencies from a surface of tissue is termed an electrical impedance spectroscopy. Changes that occur in the nature of tissues are as a result of changes in impedance. Along these lines, fitness level of the fundamental tissue can be demonstrated by the variation of impedances. The above ideology makes electric impedance spectroscopy an essential strategy for detecting/diagnosing irregularities, cancer and abnormalities particularly to diagnose women breast cancer as it was mentioned by Kerner et al. (2002) and Zheng et al. (2008). NN and RBFN classifiers are applicable in virtually every situation in which a relationship between the predictor variables (independents, inputs) and predicted variables (dependents, outputs) exists, even when that relationship is very complex and not easy to articulate in the usual terms of correlations or differences between groups. The type of problem amenable to solution by a neural network is defined by the way they work and the way they are trained. NN and RBFN work by feeding in some input variables, and producing some output variables. They can therefore be used where there is some known information, and would like to infer some unknown information. The main objective of this research work is to train both NN and RBFN to predict which group of six classes of freshly excised tissue the breast tissue belongs, when it is given other attributes as input. First thing needed to execute this task is to have a dataset. The dataset used in this experiment can be found at UCI repository under classification category. The name of the dataset is breast tissue database. The Dataset contains information about electrical impedance measurements in samples of freshly excised tissue from the breast. Several constraints were placed on the selection of these instances from a larger database. This database includes 106 instances. Each instance belongs to one class. Six classes of freshly excised tissue were studied using electrical impedance measurements: Carcinoma, fibro-adenoma, mastopathy, glandular, connective and adipose. The characteristics (input attributes) that

404

Abdulkader Helwan et al. / Procedia Computer Science 120 (2017) 402–410 Abdulkader Helwan et al./ Procedia Computer Science00 (2017) 000–000

3

are used in the prediction process include: impedivity (ohm) at zero frequency (I0), phase angle at 500 KHz (PA500), high-frequency slope of phase angle (HFS), impedance distance between spectral ends (DA), area under spectrum (AREA), area normalized by DA (A/DA), maximum of the spectrum (MAX IP), distance between I0 and real part of the maximum frequency point (DR) and the length of the spectral curve (P). The dataset can neither be inserted in NN nor RBFN in its original form for classification unless it is first normalized. 2. Review of related works Many reviews have demonstrated the importance of EIS for the recognition of breast cancer. Since 1926, Fricke and Morse (1926) have been carrying researches on breast tumors electrical properties. Because of shifting outcomes that have been in existence, the agreement has been that breast tumors electrical properties do contrast from healthy breast tissue. Surowiec et al. (1988) demonstrated some vitro tests in order to decide the fluctuation of some properties between samples of healthy tissue, samples with a mix of carcinoma and samples of breast carcinoma including the apparent limit of lesion and samples of healthy tissue as it were. The group reasoned conductivity of cancerous tissues and dielectric constants contrasted between sample groups with frequency measurements from 20KHz to 100MHz, albeit significant variability that existed between data to be measured. Morimoto et al. (1993) exhibited the measurement of electrical impedance of breast tumors in vivo. It was executed by inserting a fine needle electrode into the tumor using three-electrode technique. The group the membrane capacitance, intercellular resistance and extracellular resistance in light of a model circuit and the measured complex impedance. The extracellular resistance with a series combination of the capacitance and the intracellular resistance are within the model circuit. Frequency range of 0 to 200KHz was used to obtain the estimations. The group inferred that there are factually noteworthy contrasts amongst pathology and normal tissue. In any case, the range of values ascertained for every tissue type overlap. Jossinet (1998) and Da Silva et al. (2000) used frequency range of 488Hz-1MHz to study the impedance of six sets of breast tissue. 120 samples of impedance spectra were obtained from 64 patients, with the specimen groups divided into three types of typical breast tissue; carcinoma and two types of benign tissue. Every one of the three articles introduces studies using similar data. In Jossinet’s first research work, he examines how impedance data varies within each group by examining the reduced standard error and the standard deviation . Jossinet’s (1998) research was directed towards computing the Cole-Cole parameters by plotting intricate impedance against frequency. His research included computing parameters that would separate other samples from carcinoma samples. This research recommends that at frequencies more than 125KHz, there is a much distinction in attributes of cancerous tissues. Da Silva et al. (2000) in their research try to characterize a new set of eight parameters by which other tissues can be separated from cancerous tissue. Both infer that several parameters spanning a range of frequencies are appropriate to characterize the tissue. Chauveau et al. (1999) using a range of frequency values calculated bio-impedance parameters. Considering exvivo samples of pathological and normal tissues, estimations were gotten for frequency range from 10KHz to 10MHz. In view of the above measurements and a model that incorporates a constant phase element, the membrane capacitance, the intracellular and the extracellular resistances were computed from the measurement and a model that incorporates a consistent phase element. From these values, three indices were characterized for classifying tissue pathology. The experimental observations differentiated normal tissues and those with fibrocystic changes from cancerous tissues. Zhao et al. (2012) in a bid to enhance the spatial resolution of impedance images built a trans-admittance mammography system. A system with an array of 60*60 electrodes that look like the set-up of an X-ray mammography was developed to accomplish their idea. Kim (2012) used the scanning probe of the electrical impedance scanning to carry out a research on the frequency dependent conduct of induced current. Along these lines, relationship between the difference in the values of conductivity between the tissues with cancer and the encompassing tissues and that of the current measurement was obtained. From the above data, they proposed a formula for the breast tumor size in view of the current measurement. Considering McGivney et al. (2012) recommended electrical impedance spectroscopy as a highly ill-posed and a regularized inverse problem due to prior knowledge from modeling error and mammogram images. They introduced and properly analyzed the computational techniques for solving tissue classification methods and EIS inverse issue

4

Abdulkader Helwan et al. / Procedia Computer Science 120 (2017) 402–410 Abdulkader Helwan et al./ Procedia Computer Science00 (2017) 000–000

405

for the breast. Finally Perlet et al. (2000) explored the dependability of impedance measurement of breast tissue in a healthy state. Ones in every week, they recorded measurements over two successive menstrual cycles to figure out if electrical impedance spectroscopy images rely on hormones. They presumed that impedance is reliant on hormone levels since the images got differed all through the cycles with some consistency. 3. Implored neural network models In this study, we implored BPNN and RBFN for the implementation of the task. The considered neural network model uses backpropagation learning algorithm and it is abbreviated as BPNN. The experimental results of the BPNN and RBFN models demonstrated that RBFN outperformed BPNN with a remarkable performance. 3.1. Neural network model based on backpropagation algorithm The back propagation based neural network (BPNN) uses a feed forward process, a back propagation updating method, and supervised learning topology. This algorithm was the reason of neural networks development in the 80s of the last century. Back propagation is a general purpose learning algorithm. Although it is very efficient, it is costly in terms of processing requirements for learning. A back propagation network with a given hidden layer of elements can simulate any function to any degree of accuracy as mentioned by Helwan and Tantua (2016). These networks have been used by Helwan and Abiyev (2016) for solving of different problems. These are related breast cancer identification Helwan and Abiyev (2016), face recognition Abiyev (2016), iris recognition Abiyev and Altunkaya (2008), Abiyev and Kilic (2009), for control purpose Aliev et al (1994), for channel equalization Mamedov and Abiyev (2001). The back propagation algorithm is still as simple as it was in its first days. That is due to its simple principle and efficient algorithm. The input set of training data is presented at the first layer of the network, the input layer passes this data to the next layer where the processing of data happens. The results after being passed through the activation functions are then passed to the output layers. The result of the whole network is being then compared with a desired output. The error is used to make a one update of the weights preparing for a next iteration. After the adjustment of the weights, the inputs are passed again to the input, hidden, and output layers and a new error is calculated in a second iteration and vice versa. The back propagation is an algorithm that uses the theory of error minimization and gradient descent to find the least squared error. Finding the least squared error imposes the calculation of gradient of the error for each of the iterations. As a result, the error function must be continuous derivable function. These conditions lead to the use of continuous derivable activation functions as they are the precedents of error calculation Helwan et al. (2016). In most of cases, the tangent or logarithmic sigmoid functions are used. The pseudocode algorithm for BPNN is also given by Oyedotun et al. (2015).  Randomly choose the initial weights  While error is too large -For each training pattern (presented in random order)  Apply the inputs to the network  Calculate the output for every neuron from the input layer, through the hidden layer(s), to the output layer  Calculate the error at the outputs  Use the output error to compute error signals for pre-output layers  Use the error signals to compute weight adjustments  Apply the weight adjustments -Periodically evaluate the network performance  Apply the value of each input parameter to each input node  Input nodes compute only the identity function

406

Abdulkader Helwan et al. / Procedia Computer Science 120 (2017) 402–410 Abdulkader Helwan et al./ Procedia Computer Science00 (2017) 000–000

5

3.2. Radial basis function neural networks (RBFN) Radial basis function networks are very similar to the back-propagation networks in architecture; the fundamental difference being in the analogy behind weight computation, the activation function used at the neurons’ outputs and that they basically have one hidden layer. In the context of a neural network, the hidden units provide a set of “functions” that constitute arbitrarily the “basis” for input patterns when they are expanded into hidden space; these functions are called radial-basis functions. The motivation behind RBFN and some other neural classifiers is based on the knowledge that pattern transformed to a higher-dimensional space which is nonlinear is more probable to be linearly separable than in the low-dimensional vector representations of the same patterns (cover’s separability theorem on patterns). The output of neuron units are calculated using k-means clustering similar algorithms, after which Gaussian function is applied to provide the unit final output. During training, the hidden layer neurons are centered usually randomly in space on subsets or all of the training patterns space (dimensionality is of the training pattern) as mentioned by Strumiłło and Kamiński (2003) after which the Euclidean distance between each neuron and training pattern vectors are calculated, then the radial basis function (also or referred to as a kernel) applied to calculated distances. The radial basis function is so named because the radius distance is the argument to the function as shown Ng et al. (2007). (1) Weight  RBFN (dis tan ce) It is to be noted that while other functions such as logistic and thin-plate spline can be used in RBFN networks, the Gaussian functions is the most common. During training, the radius of Gaussian function is usually chosen; and this affects the extent to which neurons have influence considering distance. The best predicted value for the new point is found by summing the output values of the RBF functions multiplied by weights computed for each neuron. The equation relating Gaussian function output to the distance r2 from data points (r>0) to neurons center is given below. 2 2 (2) ( r )  e Where, σ is used to control the smoothness of the interpolating function, r is the Euclidean distance from a neuron center to a training data point. 3.3. BPNN training The breast tissue classification system was trained and 70 cases (almost 70% of data) and 36 cases (almost 30% of data), respectively. The developed BPNN network for this classification task comprises of 9 input neurons in its input layer to accommodate the 9 features. The output layer has 6 output neurons to accommodate the 6 classes, breast tissues, (Carcinoma, Fibro-adenoma, Mastopathy, Glandular, Connective, and Adipose tissue); the appropriate number of hidden layer neurons is found experimentally while training the network. Fig. 1 below shows the BPNN network topology. In figure X1, X2,…., X9 represents the parameters or features used for each different tissue. The developed back propagation based neural network (BPNN) was trained on 70 cases of the 6 breast tissues as discussed above. For optimizing the network; different values of learning rate, momentum rate and number of hidden neurons are considered. Thus, three networks were compared, each with different values of learning rate, momentum rate and number of hidden neurons. Table 1 shows the training parameters of the three networks. It can be seen in Table 2 that BPNN3 with 80 hidden neurons, learning rate of 0.3, and momentum rate of 0.85 achieved the lowest mean square error (MSE) (0.03352s). This MSE was reached in 40 seconds; although this training time is slightly greater than one BPNN1 needed to reach 0.03558s. The learning curve that shows the convergence of the network toward the mean square error versus the increase of iteration number is shown in fig. 2.

6

Abdulkader Helwan et al. / Procedia Computer Science 120 (2017) 402–410

407

Abdulkader Helwan et al./ Procedia Computer Science00 (2017) 000–000

Fig. 1. BPNN topology Table 1. BPNNs training parameters BPNN1 Network parameter # of training samples 70 20 # hidden neurons Learning rate (η) 0.03 Momentum rate (α) 0.55 Maximum epochs 3000 Training time (secs) 149 Mean Square Error 0.03558

BPNN2 70 50 0.13 0.65 3000 152 0.03452

BPNN3 70 80 0. 3 0.85 3000 120 0.03352

Fig. 2. BPNN error plots Table 2. Testing of the BPNNs Network parameter Number of training samples Correctly classified training samples Recognition rate on training Number of test samples Correctly classified test samples Recognition rate on testing Overall recognition rate

BPNN1 70 65 91.7% 36 30 83.33% 89.62%

BPNN2 70 66 92.6% 36 30 83.33% 90.56%

BPNN3 70 66 93.7% 36 33 91.67% 93.39%

3.4. RBFN training Similarly, three radial basis function networks (RBFNs) with different values of hidden neurons and spread constant are trained on the same data used for training the BPNNs. i.e. 70 for training and 36 for testing. This aims

Abdulkader Helwan et al. / Procedia Computer Science 120 (2017) 402–410 Abdulkader Helwan et al./ Procedia Computer Science00 (2017) 000–000

408

7

to observe the networks’ performances when trained with different values of spread constant. Table 3 describes the training parameters of three RBFN networks used in simulation. Table 3. RBFNs training parameters Network parameter RBFN1 # of training samples # hidden neurons Spread constant Maximum epochs Training time (secs) Mean Square Error

RBFN2

RBFN3

70 50 0.5 50 7 0.0309

70 75 1.0 75 9 0.0319

70 30 0.14 10 10 0.0320

Fig. 3. RBFN learning curve

Table 4 depicts the result of simulations using different RBFN structures. It can be observed from Table 4 that RBFN2 with 50 hidden neurons and spread constant of 0.5 achieved the lowest mean square error (MSE) (0.0309s). This network was capable of reaching that low MSE in 7 seconds. The learning curve for RBFN2 is shown in Fig. 3. Table 4. Testing of the RBFNs Network parameter Number of training samples Correctly classified training samples Recognition rate on training Number of test samples Correctly classified test samples Recognition rate on testing Overall recognition rate

RBFN1 70 69 97.6% 36 30 83.33% 93.39%

RBFN2 70 67 95.5% 36 33 91.66% 94.33%

RBFN3 70 68 97.28% 36 31 86.11% 93.39%

4. Experimental results In this paper, an automated classification of breast tissues using EIS features is presented. Neural networks have recently shown good and fast classification systems that are capable of distinguishing new cases, i.e. breast tissues, with reasonable accuracies. Both networks were tested on the same data; 30% of the available data. The neural network models based on back propagation shows a good generalization capability as shown in table below. Although, overall recognition rates of the three implemented networks are found to be slightly different. It can be seen that the BPNN3 that uses 80 neurons in the hidden layer reached the highest recognition rate of 93.39%. This means that this network has motivating generalization capabilities when unseen data of breast tissues are applied. Moreover, this network “BPNN3” outperformed BPNN1 and BPNN3 in terms of minimum error reached in the shortest time. Similarly the Radial Basis Function networks were also tested using same images. Only 30 % of the data was used for testing purposes. As shown in table 4 the RBFN2 achieved the highest recognition rate (97.16%) among the three networks when tested on 30% of the data. Note that this network didn’t outperform the other

8

Abdulkader Helwan et al. / Procedia Computer Science 120 (2017) 402–410 Abdulkader Helwan et al./ Procedia Computer Science00 (2017) 000–000

409

networks in the training phase, as it achieved 95.5% recognition rate while the RBFN1 and RBFN3 reached 97.6% and 97.28% respectively. 4.1. Results discussion The developed classification system based neural networks (BPNN and RBFN) are shown to be capable of classifying breast tissues into 6 different tissues, Carcinoma, Fibro-adenoma, Mastopathy, Glandular, Connective, and Adipose tissue. The used networks in this work showed promising performances when tested on unseen cases; breast tissues cases other than those were used in the training phase. Best overall classification accuracies obtained from the BPNNs and RBFNs are 93.39% for BPNN3 and 94.33% for RBFN2. It is remarked that the RBFNs network achieved higher recognition rates on the test data. i.e. better generalization capability as compared to BPNNs. Also, it is noted that the RBFN2 network’s performance was the highest in the testing phase, but it wasn’t in the training phase. This proofs that a neural network can learn fast and accurately in the training phase, however it can be weak in generalizing; that is recognizing unseeing data. This may be due to training parameters values that allow the networks to get stuck in local optima. Moreover, the mean square error reached for the RBFNs after convergence was less than that of BPNNs. Note that the difference of the error was not that high between both types of networks. However, the time taken for the networks to reach that error is what matters; since it was very high for BPNNs as compared to that taken for the RBFNs to reach its minimum error. Also it is important to spot on the training times for the BPNNs which are roughly 6 times those of the RBFNs. 5. Conclusion This paper proposed an intelligent system for classifying the breast tissues from the features derived from the EIS. Two types of neural networks were used for performing this task. Feed forward neural networks based back propagation and radial basis function network were selected for this classification task. A comparison between the same types of these networks was made based on different parameters set during the training phase. Also, a comparison between the two types of networks was conducted to evaluate the performance of each and to discover the network that performs better in this classification task. It was discovered that a BPNN with more hidden layer performs better when trained and tested on unseen data. In addition, this network reached the least minimum square error in a shorter time than the other back propagation networks. For these networks, the one that achieved the highest training recognition rate is the one that achieved the highest testing recognition rate. In contrast, for the RBFNs; the highest network’s performance in the training and testing phase were for different networks. In other words, the network that reached the highest training recognition rate wasn’t capable of achieving the highest recognition rate in the testing phase. This means that a network can be weak in generalization even if it performed well in the training phase. This is the reason why different networks of the same type were used. As a result, it should be noted that the Radial basis function network outperformed the back propagation network for classifying six different breast tissues. This outperformance is in terms of accuracy, minimum error, maximum epochs and training time. References Vacek P, Geller B, Weaver D, and Foster R, 2002. Increased mammography use and its impact on earlier breast cancer detection in vermont. Cancer; 94, 2160–2168. Basset L, and Gold R., 1987. Breast Cancer Detection: Mammograms and Other Methods in Breast Imaging. Grune & Stratton, New York Ho W, and Lam P., 2003. Clinical performance of computer-assisted detection (cad) system in detecting carcinoma in breasts of different densities. Clin Radiol, 58, 133–136. Kubicek W.G, Patterson R.E and Witsoe D.A, 1970. Impedance cardiography as a non- invasive method of monitoring cardiac function and other parameters of the cardiovascular system. Ann. N.Y. Acad. Sci, 170, 24-732. Tachibana S, Aguilar JA, and Birzis L., 1970. Scanning the interior of living brain by impedography. J of App. Physiol 1970; 28:534-539. Henderson R.R, and Webster J.G., 1978. An impedance camera for spatially specific measurements of the thorax, IEEE Trans. Biomed. Eng, 27, 250-254. Brown B.H., Barber D.C., Morice A.H. and Leathard A.D., 1994. Cardiac and respiratory related electrical impedance changes in the human thorax. IEEE Trans. Biomed. Eng, 41, 729-734.

410

Abdulkader Helwan et al. / Procedia Computer Science 120 (2017) 402–410 Abdulkader Helwan et al./ Procedia Computer Science00 (2017) 000–000

9

Gabriel S, Lau R.W, and Gabriel C., 1996. The dielectric properties of biological tissues: III. Parametric models for the dielectric spectrum of tissues. Phys Med Biol, 11, 2271. Surowiec A.J, Stuchly S.S, Barr J.R, and Swarur A., 1988. Dielectric properties of breast carcinoma and the surrounding tissue. IEEE Trans. Biomed. Eng 35, 257-263. Campbell A.M, and Land D.V., 1992. Dielectric properties of female breast tissue measured in vitro at 3.2 GHz. Phys. Med. Biol, 37, 193-210. Heinitz, J., and Minet, O., 1995. Dielectric properties of female breast tumors. In Ninth International Conference on Electrical Bio-Impedance, Heidelberg, 55, 356-359. Jossinet J., 1998. The impedivity of freshly excised human breast tissue. Physiol Meas, 19, 61-75. Jossinet J, and Lavandier B., 1998. The discrimination of excised cancerous breast tissue samples using impedance spectroscopy. Bioelectrochem Bioenerg, 45, 161-167. Kerner T.E, Paulsen K.D, Hartov A, Soho S.K, Poplack S.P., 2002. Electrical impedance spectroscopy of the breast: clinical imaging results in 26 subjects. IEEE Trans Med Imag. 6, 638–645. Zheng B, Zuley ML, Sumkin JH, Catullo VJ, Abrams GS, Rathfon GY, Chough DM, Gruss MZ, and Gur D., 2008. Detection of breast abnormalities using a prototype resonance electrical impedance spectroscopy system: a preliminary study. Med Phys, 35, 3041–3048. Fricke H, and Morse S., 1926. The electrical properties of tumors of the breast. J Cancer Res, 16, 310-376. Surowiec A, Stanislaw S.S, Barr J.R, and Swarup A., 1988. Dielectric properties of breast carcinoma and the surrounding tissues. IEEE Trans Biomed Eng, 35, 257-263. Morimoto T, Kimura S, Konishi Y, Komaki K, Uyama T, Monden Y, Kinouchi Y, and Iritani T., 1993. A study of the electrical bio-impedance of tumors. J Invest Surg, 6:25-32. Jossinet J., 1998. The impedivity of freshly excised human breast tissue. Physiol Meas 1998;19:61-75. Da Silva J.E, De Sá J.M, Jossinet J.,2000. Classification of breast tissue by electrical impedance spectroscopy. Med Biol Eng Comput 38,26-30 Chauveau N, Hamzaoui L, Rochaix P, Rigaud B, Voigt JJ, and Morucci JP., 1999. Ex vivo discrimination between normal and pathological tissues in human breast surgical biopsies using bioimpedance spectroscopy. Ann NY Acad Sci, 873, 42-50. Zhao M, Wi H, Mostofa Kamal A.H, McEwan A.L, Woo E.J, and Oh T.I., 2012. High density trans-admittance mammography development and preliminary phantom tests. Biomed Eng Online, 11, 75. Kim S., 2012. Assessment of breast tumor size in electrical impedance scanning Inverse. Probl, 28, 45-51. McGivney D, Calvetti D, and Somersalo E., 2012. Quantitative imaging with electrical impedance spectroscopy. Phys Med Biol, 57, 7289–7302. Perlet C, Kessler M, Lenington S, Sittek H, and Reiser M., 2000. Electrical impedance measurement of the breast: effect of hormonal changes associated with the menstrual cycle. Eur Radiol, 10, 1550-1554. Helwan A, and Tantua D.P, 2016. IKRAI: Intelligent Knee Rheumatoid Arthritis Identification. Int J Intell Syst Appl, 8, 18-24. Helwan A, and Abiyev R.H., 2016. Shape and Texture Features for Identification of Breast Cancer. In Proceedings of the World Congress on Engineering and Computer Science, 2, 19-21. Helwan A and Abiyev R.H., 2015. An Intelligent System for Identification of Breast Cancer. In Proceedings of the 3-rd IEEE International Conference Advances in Biomedical Engineering, 3, 17-20. Abiyev R.H., 2014. Facial Feature Extraction Techniques for Face Recognition. J Comput Sci, 10(12), 2360-2365. Abiyev R.H, and Altunkaya K., 2009. Neural Network Based Biometric Personal Identification with fast iris segmentation. International Journal of Control, Automation and Systems,7 (1), 17-23. Abiyev R.H, and Kilic K., 2009. Adaptive Iris segmentation. In International Conference on Advances in Information Security and Assurance, 90-99. Abiyev R.H, and Altunkaya K., 2008. Personal Iris Recognition Using Neural Networks. Int J Sec Syst Appl, 2(2), 41-50. Abiyev R.H, and Altunkaya K., 2007. Neural network based biometric personal identification. In Proceedings Frontiers in the Convergence of Bioscience and Information Technologies, 682-687. Abiyev R.H, and Altunkaya K., 2007. Iris Recognition for Biometric Personal Identification Using Neural Networks. In Proceedings International Conference on Artificial Neural networks, 554-563. Lecture Notes in Computer Science, 4669, Springer. Aliev R.A, Aliev R.R, and Abiev R.H., 1994. Synthesis of industrial neural controllers. In Neural Networks, Proceedings of the IEEE World Congress on Computational Intelligence, 1644-1649, Mamedov F., Abiyev R., Alshanableh T., 2004. Recurrent Neural Network for Equalization of Channel Distortion. Proc. Of Electrical, Electronics and Computer Engineering Symposium NEU-CEE2004. Nicaosia, Turkey, 256-260. Helwan A, Khashman A, Olaniyi EO, Oyedotun OK, and Oyedotun OA.,2016. Seminal Quality Evaluation with RBF Neural Network. Bull. Transilv. Univ. Bras¸ov Ser. B (N.S.) 8, 137-145. Oyedotun O.K, Olaniyi E.O, Helwan A, and Khashman A., 2015. Hybrid auto encoder network for iris nevus diagnosis considering potential malignancy. In Proceedings of the 3-rd IEEE International Conference Advances in Biomedical Engineering, 3, 274-277. Strumiłło P, and Kamiński W., 2003. Radial basis function neural networks: theory and applications. In Neural Networks and Soft Computing, Physica-Verlag HD, 19, 107-119. Ng W.W, Dorado A, Yeung D.S, Pedrycz W, and Izquierdo E., 2007. Image classification with the use of radial basis function neural networks and the minimization of the localized generalization error. Pattern Recogn, 40, 19-32. Garg S, Patra K, Pal S.K, and Chakraborty D., 2008. Effect of different basis functions on a radial basis function network in prediction of drill flank wear from motor current signals. Soft Comput, 12, 777-787.