The Combination of Mammography and MRI for Diagnosing Breast Cancer Using Fuzzy NN and SVM

R cancer is one of the common illnesses [2]. According to calculations of the national cancer institute of America, from each 8 women one is suffering of breast cancer. Breast cancer is the most common cancer among women and includes 32 percent of all women – related cancer and 19 percent of death reason of women with cancer, in other hand. It is the second reason of death with cancer [3]. If this cancer be diagnosed in time, it can be cured easily. Cancers are the third reason of death in our country and among 101 to 105 persons die due to it every day [3]. Gynecologist’s emphasis that if women with breast cancer take action early, possibility of their treatment is more and their costs will be less [1]. In accordance with survey of Breast cancer research Centre (SID). The majority of patients (about 57 percent) are in second step of illness when they are in initial step [4]. The incidence of this cancer in Iran is z6 cases per 100/000 population of women over 30 years and in western countries it is 130 cases per 100/000, however, age of Iranian women who are infected with cancer is at least a decade less than women in developed countries [1]. Statistics show that incidence of cancer is increasing percent in the world annually. Despite of increments of it death amount caused by it has been decreased which indicates the development of notification methods thereby early detection of it becomes possible. In past articles different methods of cancer detection are used. Technique using a new hybrid approach based on cuckoo algorithm and support vector machine classifier advantages of which are increasing computation performance and accuracy of SVM parameters that for improving it we have applied combination of fuzzy neural network and SVM. In another article namely the combination, decision tree and support vector machine have been used for breast cancer detection advantages of which are high accuracy and sensitivity and disadvantage of which is that its accuracy is relevance to the size of the training set [4]. However in this work we have attempted to train input data with evolutionary algorithm and then give its output to support machine that can be categorized easily. Fig. 1 indicates complete framework of this study.

R east cancer is one of the common illnesses [2]. According to calculations of the national cancer institute of America, from each 8 women one is suffering of breast cancer. Breast cancer is the most common cancer among women and includes 32 percent of all women -related cancer and 19 percent of death reason of women with cancer, in other hand. It is the second reason of death with cancer [3]. If this cancer be diagnosed in time, it can be cured easily. Cancers are the third reason of death in our country and among 101 to 105 persons die due to it every day [3]. Gynecologist's emphasis that if women with breast cancer take action early, possibility of their treatment is more and their costs will be less [1]. In accordance with survey of Breast cancer research Centre (SID). The majority of patients (about 57 percent) are in second step of illness when they are in initial step [4]. The incidence of this cancer in Iran is z6 cases per 100/000 population of women over 30 years and in western countries it is 130 cases per 100/000, however, age of Iranian women who are infected with cancer is at least a decade less than women in developed countries [1]. Statistics show that incidence of cancer is increasing percent in the world annually. Despite of increments of it death amount caused by it has been decreased which indicates the development of notification methods thereby early detection of it becomes possible.
In past articles different methods of cancer detection are used. Technique using a new hybrid approach based on cuckoo algorithm and support vector machine classifier advantages of which are increasing computation performance and accuracy of SVM parameters that for improving it we have applied combination of fuzzy neural network and SVM. In another article namely the combination, decision tree and support vector machine have been used for breast cancer detection advantages of which are high accuracy and sensitivity and disadvantage of which is that its accuracy is relevance to the size of the training set [4]. However in this work we have attempted to train input data with evolutionary algorithm and then give its output to support machine that can be categorized easily. Fig. 1 indicates complete framework of this study.

II. affIne tRansfoRMatIon and fencIng and thReshold IMages
Affine transformation leaps either linearly (all points on a line will stay on a line after transformation) or the distances. In affine transformation straight lines stay straight and parallel lines will stay parallel [10].
Contractions, expansions, congruence, reflections, rotations and shear of similarity transformations, transfer and their combinations are all affine transformation keeps ratios on the line however it doesn't need to keep angles and length too. In Fig. 2 a sample of MRI and mammography pictures are shown.   The main objective of fencing medical images is dividing it to different anatomy structures such as blood useless and tumors from background. For fencing first we use threshold for making dork points to light points in order to increase accuracy. Using threshold method in to transform images with gray scale to black and white images, these methods are accompanied with errors. In cases with threshold that pixels of background will became pixels of image will cause that object becomes thicker or thinner. For solving this problem we will use fuzzy logic. FCM classification method is based on fuzzy theory [2]. In the Fig. 5 threshold method is used for improving injured points.

III. Feature Extraction with Gabor Filter Banks
While working on texture of image, features such as local spatial frequency and in formations related to pixel directions will be addressed. A certain approach for determining these features is calculating the power spectrum of image however Fourier transformation of an image calculates just a total estimation of image frequency in formation. A Gabor filter can be considered as a sinusoidal part with certain directions and frequencies that have been modeled by a Gaussian cover.
Gabor filter can be formulated as below (1)

IV. Improved Neural Network Bundle
The classification of improved neural network Fig. 7 indicates a total structure of RFNN model where N is input. And the classifier algorithm of support vector machine. M is tern for each variable and p is output variable. Considering to Fig. 7 it is obvious that RFNN method has 6 layers. Each date function has a return relations ship with its structure. Kith rule with M inputs and N outputs in RFNN method is defined as below

A. RFNN Network Training Based on Descending Gradient
RFNN network training is based on descending Gradient the cost function used in RFNN method training is de fined as below in formula T(k) is target vector of K th sample in RFNN method rules of linear combination consists of inputs in addition a constant. The derivative of w st function considering to parameters of date function is so important. When normalization comes in every output of 4 th layer entails a function of all rules in RFNN method that this issue will complicate calculations.

B. Network Training Based on Learning by Combinatorial Optimization Algorithm
RFNN network training is based on learning by combinatorial optimization algorithms. In this part RFNN network training based on learning by combinatorial optimization algorithms will be addressed.
Since cost function with parameters of first part is complex, we use 4 optimization algorithms for optimizing these parameters.

C. Discussion
Technique using a new hybrid approach based on cuckoo algorithm and support vector machine classifier advantages of which are increasing computation performance and accuracy of SVM parameters that for improving it we have applied combination of fuzzy neural network and SVM.
The main advantage of it is to avoid over training and if it occurs, model will be able to predict similar data, which are used in learning step, be applied, system will have improper operation. Each data function in first part from RFNN model has 3 parameters for training. The centers of Fuzzy data functions are shown with m and standard deviation is determined by Q. O weight provides mobility for Fuzzy data functions should be trained. Parameters of teleport are trained with training algorithm of descending gradient. Every components of each part, includes fuzzy return data function, m, q, o, in this algorithm and in every stage of combinatorial optimization algorithm parameters are completed in the first interaction and mean square error is calculated. Semi code [1] indicates steps of training of RFNN network based on learning by combinatorial optimization algorithms.
Assume that we have data point sets and we want to segregate them pair wise. Every vector is consists of real numbers that are variables of software behavior. Linear classification methods try to segregate data with super surface [5]. Support vector machine classifier method that is the best linear classification, will find this super surface [4].  Goal of these classifications understand if a gland is benign or malignant. A common criterion for rating used is information gain 9 criterions defined as below.

(4)
Where P(lc) is fractions of documents that has class LC. P(tj=1) and P(tj=0) means number of documents that have term tj or don't have it. IG(ti=0) is calculated for all terms and terms with vary law IG will be skipped.

v. evaluatIon cRIteRIa and PRoPosed Method evaluatIon
Compare to digital and normal radiographies, digital radiographies are more a accurate [10]. Level of accuracy or measurement is determined by accurate. Goal of this study is to verify if this proposed method will increase accuracy in breast cancer detection or not. (5) In this section we compare our proposed method with other methods. In table we can conclude that results from Fuzzy C-Mean are better than network neural so this algorithm will be chosen as basic algorithm in other steps. Also for input and output variables training we will use particle swarm optimization Algorithm (PSO), Genetic Algorithm (GA), Imperialist competitive Algorithm (ICA) and simulated annealing Algorithm (SA). In accordance with results of performance, simulation and evaluation and also comparing classification errors we can conclude that PSD method is better.   You can see that after applying SVM final result has been improved and is 98.14.

vI. conclusIon and futuRe woRks
Breast cancer is a common cancer among women. In this study in order to early breast cancer detection we have proposed a method based on processing mammography and MRI images presented here. This method uses classification methods and different so that detection and determine the exact place of lump. This method effectively decease human errors. Proposed algorithms applied to 216 cases of images and all results verified efficiency of it. According to results and comparing them, ability of Gabor bank fitter in producing feature vector has been proved. Tests and performance evaluations are done on images such as MRI and mammography. A sample image which had cancer lumps gotten from processing images and can be seen as segregated parts. Then inputs entered into improved neural network and results calculate precisely. In order to comparing proposed method evaluation with other methods, results shown in table V. This study is a unique work in breast cancer detection realm. So results of which can be expanded. For this purpose suggestions are told. Multi -objective evolutionary algorithms and classification technology so that achieving better results are proposed.