Robust Classification of Primary Brain Tumor in MRI Images based on Multi Model Textures Features and Kernel Based SVM

.


INTRODUCTION
Medical image segmentation has core importance to implement high level operations such as tissues recognition and classification.Segmentation of this type of imaging data is still a problem that demands to be probed 1 .Segmentation of medical structures is compulsory in computer assisted surgery or diagnosis system.Medical imaging methods take part in focusing organs and tissues during surgery process 2 .The visible and palpable tumor mass seen by different medical imaging modalities like CT and MRI is usually complicated to identify automatically with existing image processing methods, but still same is achieved manually by clinicians using computer aided drawing software.Segmentation of basic brain MR regions supports in visualization to identify various diseases, morphological and volume estimation, tissue classification and etc. Various methods of segmentation are edge based, thresholding, watershed and region growing etc., which are distinguished on the basis of their application and modality using which image is acquired 3 .Magnetic Resonance Imaging (MRI) is a multi-sequence medical imaging technique with which stacks of images are acquired with different tissue contrasts.Each sequence, namely, T1weighted,T2-weighted, Proton Density (PD), Fluid-Attenuated Inversion Recovery (FLAIR), etc., highlights specific properties of tissues and pathologies, but none of them can provide completely decisive and reliable information.In MR images, lesions usually appear quite different in texture from normal tissues.Texture features provide an important in the perception and discrimination of a tumor.Image classification is one of the typical computer applications widely used in the medical field, especially for abnormality detection in Magnetic Resonance (MR) brain images 4 .The automated image classification systems used for such applications must be significantly efficient in terms of accuracy since false detection may lead to fatal results.
The process of extracting the original image based on the certain features which undergo a transformation of pixel regions.Feature extraction can be based on two approaches; they are human centered and machine centered approaches.Human centered approach measures with perception based features such as texture, and the perfect mathematical representation is selected for it.In machine centered approach, a unified computing scheme is selected for extracting certain ad-hoc features.There are many techniques for feature extraction, e.g.texture Features 5 , Gabor features, feature based on wavelet transform, principal component analysis, minimum noise fraction transforms, discriminant analysis, decision boundary feature extraction, non-parametric weighted feature extraction and spectral mixture analysis.The texture elements act as attributes based on the pixel region that digitizes the individual object into binary forms of information through computer imaging based on the silhouette of image information [6][7] .

Machine
learning classification techniques, including supervised and unsupervised (clustering or fuzzy clustering) [8][9] , are also introduced into brain tumor segmentation.Trained classifiers estimate the probability for each voxel in the testing volume, judging whether the voxel belongs to the target or the background.The threshold of the probability map is calculated to obtain the segmentation result or provide for postprocessing.These techniques make it possible for high-dimensional features to be utilized in order to achieve a better discriminatory power for tumors compared with sole dependence on intensity information 10 .Moreover, the approaches applied in the field of pattern analysis can be transplanted into medical image segmentation, such as a distance metric learning algorithm, to make the intra-class samples closer while keeping extraclass samples as far away from each other as possible.Unfortunately, these classifications based segmentation approaches consider the voxels in the image to be independent of each other, with no spatial correlation both in the training and testing phases 11 .

Related work
Lots of research has been performed for the segmentation of normal and abnormal tissues in MRI brain images.Some of the recent related works regarding the classification of brain tissues are reviewed in this section.
Kekre et al 8 have proposed a vector quantization segmentation technique to identify a cancerous mass from MRI images.In order to improve the radiologists' diagnostic performance, computer-aided diagnosis (CAD) scheme has been introduced to enhance the recognition of primary signatures of this disease: masses and micro calcifications.As well, to tackle the class distinguishability as well as feature space sparseness and solution space intricacy problems in multivariate image segmentation, a Markov random field (MRF) based multivariate segmentation algorithm called "multivariate iterative region growing using semantics" (MIRGS) has been proposed by Shan Shen 13 .In MIRGS, the impact of intra-class variation and computational cost has been minimized by means of the MRF spatial context model integrated with adaptive edge penalty and applied to regions.To restrain the initialization sensitivity, a region-level means (RKM) based initialization technique has been utilized, which always provides exact initial conditions at low computational cost.Experiments have demonstrated the pre-eminence of RKM relative to two frequently used initialization techniques.
As well, a template-based framework for multi-object segmentation of deep brain structures (caudate nucleus, putaen and thalaus) in medical brain images has been presented by Jue Wu and Albert C.S. Chung 14 .This framework combines the information of edge features, region statistics, and inter-structure constraints for identifying and locating all target brain structures.The multi-object template has been structured in the form of a hierarchical Markov Dependence Tree (MDT), and manifold objects have been successfully matched to a target image via a top-to-down optimization approach.The final segmentation has been achieved through refinement by a B-spline based non-rigid registration between the exemplar image and target image.The approach necessitates only one example as training data.The technique has been validated using a publicly available T1weighted MRI database with expert-segmented brain structures and obtained satisfactory results as 0.80 Dice score for the caudate nuclei, 0.81 Dice score for the putaina, and 0.84 Dice score for the thalai on average.
In order to enhance the performance of automated image segmentation, especially in the field of brain tissue segmentation from 3D MRI towards classical image deterioration including the noise and bias field artifacts that arise in the MRI acquisition process, caldairou et al 15 have proposed to integrate into the FCM segmentation methodology concepts stimulated by the Non-Local (NL) framework.The major algorithmic contributions of this paper were the definition of an NL data term and an NL regularization term to effectively handle the intensity inhomogeneity and noise in the data.Then, the resulting energy formulation was built into an NL/FCM brain tissue segmentation algorithm.Experiments carried out on both the synthetic and real MRI data, leading to the classification of brain tissues into gray-matter, white matter, and cerebro-spinal fluid, have shown a substantial enhancement in performance in the case of higher noise levels, when compared to a range of standard algorithms.

Feature extraction
The purpose of feature extraction is to reduce the original data set by measuring certain properties, or features, that distinguish one input pattern from another pattern 16 .The extracted feature is expected to provide the characteristics of the input type to the classiûer by considering the description of the relevant properties of the image into a feature space.The proposed method,feature extraction method consists of Three steps are used for calculation.Texture features can be described using this co-occurrence matrix [17].In our proposed method, Feature Vector F(V1) (Five features such as, ASM, entropy, IDM, contrast and Maximum probability) is extracted from the co-occurrence matrix.

Computation of Feature Vector F (V2)
According to the neuropsychological findings, different types of incentive are processed disjoint, yet concurrently, by dissimilar neural mechanism previously to the stimulus is intentionally perceived as a whole.In the proposed method, feature extraction process is done with the help of enhanced Texton Co-occurrence Matrix (ETCM).In this method, both Histogram and cooccurrence matrix are used for feature extraction process.The relationship between the values of neighboring pixels is characterized by TCM.Histogram based techniques are simple to compute, but highest indexing performance.The cooccurrence matrix directly uses a feature representation of the image.If the dimension of the image is high, then the performance is decreased.The spatial information is lost when the histogram is used only for feature representation of the image.Hence combine both histogram and co-occurrence matrix for feature extraction and representation.In the ETCM method, four special types of textons are used for detecting the texton in the original image.It is shown in Figure 1.
Consider a 2 × 2 matrix in the image with four pixels P1, P2, P3 and P4.If two pixels have the same, then these pixels form a texton.The possible textons formed with different combination of pixels with same intensity values are denoted by T1, T2, T3 and T4 which are shown in Figure 1.The texton image is generated using these four texton templates with two pixel length as shown in Figure 2.
In Figure 2, the 2 × 2 matrices are shown in Figure 2 (a), the experimental image data are shown in Figure 2 (b), the four texton templates that slide over the entire experimental image from left to right and top to bottom with two pixel length to detect four textons is shown in Figure 2(c).The four different type textons are given in Figure 2(d).The four texton component images that are composed to form a final texton image is shown in Figure 2(e).
After the formation of final texton image,the feature vector F(V2) (Five features such as, ASM, entropy, IDM, contrast and Maximum probability) is extracted from the final texton image.

Concatenated of the two vectors
Hence, total Feature vector uses dimensional vector as the concluding image features in the classification.

Final classification
After feature extraction process, In-order to detect the presence of the tumor in the input MRI image, we perform the final classification step.Here we use the Support Vector Machine classifier to classify the image into tumorous or not.Support
and a vector of labels, SVM provides the optimal hyperplane ( ) .For a nonlinear classifier, SVM maps the data points into a higher dimensional space H, where a separating hyperplane with maximal margin is constructed.The following quadratic optimization problem has to be solved

Subject to
Where training data are mapped to the higher dimensional space H by the function A set of slack variables is introduced for each training vector and C is a penalty parameter on the training error.
Under this mapping the solution of an SVM has the form : As can be seen, it is sufficient to compute the scalar products of the form.A kernel function which defines an inner product in H performs the respective mapping leading to the following decision function f(x): The optimal hyperplane is the one with a maximal distance to the closest image ( )

Input data set
For our proposed method, we have collected the various tumor and non tumor MRI images from south Indian area severity analysis which is undergone for processing the images.This image dataset contains 100 brain MRI images.In which, a total of 80 T1-weighted gadolinium enhanced MR images were tumorous .These 3D DICOM real images were obtained from the Government Medical College Hospital, Tirunelveli, Tamilnadu, India, using SIEMENS 1.5 Telsa MR unit.In each case ,only T1-weighted post contrast(Gadolinium) images ,Spin-Echo (SE) sequence (TR=480 ms ,TE=8.7 ms) ,Matrix size is 256 *256 and the slice thickness is 1 mm used for analysis.

Performance evaluation of proposed system
Classifier performance evaluation of this work is conducted with widely used statistical measures, sensitivity, specificity, accuracy and error rate [22].True Positive (TP) is defined as the number of correctly identified positive pixels; True Negative (TN) is defined as correctly identified negative pixels.For example, in a diagnostic test, evaluation focusing on the presence of abnormal tissues, tumor samples is considered in the positive category and normal tissues will be in the negative category.False Error rate = 1 -Accuracy

Comparative analysis
We have compared our proposed tumor detection technique of FVF against the neural network techniques.The neural networks, we have utilized for comparative analysis is Feed Forward Neural Network (FFNN) and Radial Basis Function (RBF) neural network.The performance analysis has been made by plotting the graphs of evaluation metrics such as sensitivity, specificity and the accuracy are shown in Table -1

CONCLUSION
In this paper, we have developed an automated brain MRI diagnostic system with normal and abnormal classes.The medical decision making system was designed with the Texture features and kernel based Support Vector Machine.The proposed approach comprises feature extraction and classification.The benefit of the system is to assist the physician to make the final decision without hesitation.According to the experimental results, the proposed method is efficient for the classification of the human brain into normal and abnormal.For comparative analysis, our proposed approach is compared with other neural networks RBF and FFNN.The accuracy level (94%) for our proposed method proved that the proposed algorithm graph is good at detecting the tumors in the brain MRI images.

Fig. 1 :Fig. 2 :
Fig. 1: Special texton types of ETCM ( a) original 2x2 matrix (b)Texton T1 (c) Texton T2 (d) Texton T3 (e) Texton T4 aims to separate the training patterns.In the case of linearly separable classes this hyperplane maximizes the sum of the distances to the closest positive and negative training patterns.This sum is called margin.To construct the maximum margin or optimal separating hyperplane, we need to classify correctly the vectors Xi of the training set into two different classes Yi using the smallest norm of coefficients W.
data.The dual formulation can be stated as follows:Subject to

Fig. 4 :Fig. 3 :
Fig. 4: Comparison error bar of the proposed Texture features with various classifiers

. 2 .
By analyzing the plotted graph; the performance of the proposed technique has significantly improved the tumor detection compared with Feed Forward Neural Network (FFNN) and Radial Basis Function (RBF) neural network classifier.The evaluation graphs of the sensitivity, specificity and the accuracy graph are shown in Fig.Based on the experimental results our proposed method produces better results compared to other neural network based classifiers.The brain tumor classification error bar is also given in Fig 3.
Histogram based features are local in nature.These features do not consider spatial information into consideration.So for this purpose gray-level spatial co-occurrence matrix h d (i,j) based features are defined which are known as second order histogram based features.These features are based on the joint probability distribution of pairs of pixels.Distance d and angle è within a given neighborhood are used for calculation of joint probability distribution between pixels.Normally d=1,2 and è=0 o ,45 o , 90 o ,135 o