Abnormality Classification of Brain Tumor in MRI Images using Multiclass SVM

 

C. Saranya Jothi, V. Usha, S. Alex David, Hijaj Mohammed

Department of Computer Science and Engineering

Vel Tech Rangarajan Dr. Sagunthala R&D Institute of science and Technology, Avadi, Chennai-62

 

ABSTRACT:

Automatic image classification is most important process in medical imaging for analyzing difference between normal patients and those who have the possibility of having abnormalities or tumor. MRI image consists of different gray levels and intensities. Brain tumor is varied in size and shape. Radiologist can diagnose and find location of lesions based on visual diagnosis with help of available software. While performing visual diagnosis, error may be introduced in finding location of tumor. To solve this problem there is need to develop a computer based automatic technique for detecting and classifying tumor from brain MRI images. The proposed system classifies the abnormalities in brain tumor by extracting the feature from brain tumor image and normal image using the GLCM and DWT. From the features classified by the classifier into two classes normal and abnormal along with different type of tumor by using the multiclass SVM which will give a promising result.

 

 

 

1. INTRODUCTION:

Restorative imaging method is most usually used to picture the inner structure and capacity of the body. Cerebrum tumor extraction and its examination are testing errands in restorative picture preparing in light of the fact that mind picture and its structure is convoluted that can be broke down just by master radiologists. X-ray (Magnetic Resonance Imaging)⁶ has turned into an especially valuable therapeutic indicative instrument for conclusion of cerebrum and other medicinal pictures. Tumors can be destructive (threatening) or non harmful (considerate).

 

 

The manual understanding of mind tumor cuts in view of visual examination by a doctor may prompt missing determination and tedious when a substantial number of MRI cerebrum pictures are broke down. To keep away from human based indicative mistake, PC helped analysis framework is required. There are heaps of strategies for programmed and self-loader picture order, however the vast majority of them come up short on account of obscure commotion, poor picture differentiate, in homogeneity and feeble limits that are normal in restorative pictures. Medicinal pictures for the most part contain convoluted structures and their exact grouping is fundamental for clinical finding⁷.

 

The objective of the system is to provide quality software for detection and classification of brain tumor in humans. The main motivation of this is to develop an efficient method for classification of normal and abnormal MRI brain images.

 

2. LITERATURE SURVEY:

The SVM classifiers¹ are use to arrange the pictures into two classifications either ordinary and unusual cerebrum picture. The assurance of ordinary and strange mind picture depends on symmetry which is shown in the pivotal and coronal pictures. The key idea of SVM is the utilization of hyper planes to characterize choice limits isolating between information purposes of various classes as 1 for ordinary where as 0 for anomalous². The thought behind SVM is to delineate unique information indicates from the information space a high dimensional or even interminable dimensional element space with the end goal that the arrangement issue winds up plainly easier in the element space. The mapping is finished by an appropriate decision of a portion work. Robotized MRI³ (Magnetic Resonance Imaging) cerebrum tumor division is a troublesome errand because of the fluctuation and unpredictability of tumors. In this, a measurable structure examination based tumor division conspire is introduced, which concentrates on the auxiliary investigation on both tumorous and typical tissues. The fundamental idea is that neighborhood surfaces in the pictures can uncover the run of the mill regularities of natural structures. In this way textural highlights have been separated utilizing co-event lattice approach. In text features⁴ clarifies the half and half calculation for identification mind tumor in attractive Resonance pictures utilizing measurable highlights and Fuzzy Support Vector Machine (FSVM) classifier. The proposed procedure comprises of four phases to be specific Noise decrease, Feature extraction, Feature diminishment and Classification. In brain stroke detection¹⁰ the primary stage anisotropic channel is connected for commotion decrease and to make the picture appropriate for separating highlights. In classification of MRI brain images⁵ comprises two phases specifically highlight extraction and order. In the principal organize, we have gotten the highlights identified with MRI pictures utilizing Discrete Wavelet Transformation (DWT). Wavelet change based techniques are an outstanding instrument for removing recurrence space data from non-stationary signs.

 

3. SYSTEM OVERVIEW:

The proposed system is used to classification between the normal and abnormal MRI brain images is shown in (Fig.1). The dataset obtained from reputed diagnostics center, which contain 100 images, 70 images is used for training phase and 30 images is used for testing phase. The image should under goes some preprocessing step in order to filter the noise. Then the features are extracted from the images using feature extraction algorithms Gray Level Co- occurrence Matrix (GLCM) and Discrete Wavelet Transform (DWT). After the feature extraction Multi SVM classifier is used to classification between normal and abnormal images along with the types of an existing abnormality. Once the classification is made performance analysis is carried out to measure the accuracy of the result.

 

Fig.1: Architecture Design

 

3.1 Preprocessing:

The Preprocessing of pictures is to upgrade the visual appearance of pictures and to enhance the control of datasets by expelling clamor, normalizing the power of the individual particles pictures, evacuating reflections. This pre-handling step is an endeavor to decrease the nearness of these undesirable highlights in both the preparation and test information. Here middle channel is utilized to expel the clamor exhibit in the pictures.

 

3.2 Feature Extraction:

The Feature extraction is an uncommon type of dimensionality diminishment. At the point when the info information to a calculation is too vast to possibly be prepared and it is suspected to be famously repetitive (much information, however very little data) at that point the information will be changed into a diminished portrayal set of highlights (additionally named highlights vector). Changing the information into the arrangement of highlights is called highlights extraction. On the off chance that the highlights extricated are precisely picked it is normal that the highlights set will remove the important data from the information keeping in mind the end goal to play out the coveted errand utilizing this decreased portrayal rather than the full size info. The extricated highlights give the qualities of the info sort to the classifier by considering the portrayal of the applicable properties of the picture into a component space. The vast majority of the tumor is heterogeneous tissues and the mean estimations of unwinding times are not in any manner adequate to portray the heterogeneity of the distinctive tumor sorts. Here GLCM and DWT techniques are utilized for highlight extraction.

 

3.3 Image Classification:

The point of characterization is to amass things that have comparable element esteems into gatherings. Classifier accomplishes this by settling on a characterization choice in light of the estimation of the straight mix of the highlights. For the most part SVM is utilized for order which is a double classifier that is the class marks can just take two esteems. Be that as it may, in proposed approach have more than two classes so Multi class SVM is an utilized for grouping.

 

4. IMPLEMENTATION:

The implementation of classification of MRI brain image is done using MATLAB. The methods and algorithms involved in each module are discussed in detail. The database for experimental verification is having 100 images out of which 70 images are used in training image and 30 image are used testing phase.

 

4.1   PREPROCESSING USING MEDIAN FILTERING:

The middle channel is one sort of preprocessing methods. It can play long tail dissemination to improve a de-noising impact is shown in (Fig.2). Middle Filter evacuate the commotion with high recurrence segments from MRI picture without exasperating the edges and it is utilized to lessen the salt and pepper clamor. This system computes the middle esteems i.e. set middle esteem pixels estimations of the encompassing pixels to decide the new de-noised estimation of the pixel. A middle is ascertained by arranging all pixel esteems by their size, at that point choosing the middle an incentive as the new incentive for the pixel. The fundamental capacity for middle picture is composed beneath in condition (1),

 

f^e(x; y) =^median_(i;j)eN fg(i; j)g          (1)

Where;

f(x,y) yield middle and

g(x,y) is the first esteems.

 

Middle channel is a nonlinear separating. The pixel esteems in the area window are positioned by power and the center value(median) turns into the yield an incentive for limits. Since the middle is less touchy than the mean esteems are all the more successfully evacuated. Middle sifting jelly the edges.

 

Fig. 2: Preprocessing using Median Filtering

 

4.2 FEATURE EXTRACTION USING DWT:

The two dimensional discrete wavelet change is basically a one dimensional examination of a two dimensional flag. It just works on one measurement at any given moment, by dissecting the lines and sections of a picture in a detachable manner. The initial step applies the investigation channels to the columns of a picture.

 

This produces two new pictures, where one picture is set or coarse column coefficients, and the other an arrangement of detail push coefficients. Next examination channels are connected to the segments of each new picture, to create four unique pictures called sub groups. Lines and segments broke down with a high pass channel are assigned with a H. In like manner lines and segments examined with a low pass channel are assigned with a L. Each subband gives diverse data about the picture. The LL subband is a coarse estimate of the picture and expels all high recurrence data. The LH subband expels high recurrence data along the lines and underlines high recurrence data along the sections. The outcome is a picture in which vertical edges are stressed. The HL subband underscores flat edges, and the HH subband accentuates askew edges. To figure the DWT of the picture at the following scale the procedure is connected a pick up to the LL subband.

 

Each level of the wavelet deterioration, four new pictures are made from the first N x N-pixel picture. The span of these new pictures is diminished to 1/4 of the first size. The new pictures are named by the channel (low-pass or high-pass) which is connected to the first picture in flat and vertical headings is shown in (Fig.3). Along these lines, the four pictures delivered from every decay level are LL, LH, HL, and HH. The LL picture is viewed as a decreased adaptation of the first as it holds generally subtle elements. The LH picture contains even edge highlights, while the HL contains vertical edge highlights.

 

Fig.3: Original Image

 

Fig. 4: DWT picture in light of estimated picture detail(LL), even details(HL), vertical details(LH) and inclining details(HH) in one level.

 

From (Fig. 4) speak to inexact points of interest, even subtle elements, vertical points of interest and corner to corner points of interest of and distinctive pictures. Inexact points of interest are same as unique picture subtle elements. Level subtle elements develop just even data (edges). Vertical points of interest develop just vertical data (edges). Askew points of interest build not very many data of info picture is shown in (Fig.5).

 Input Query:

 

Fig.5: Input Image

 

Mean (Horizontal details) = 9.3269, Mean (Vertical details) =11.3904

 

4.3 TEXTURE FEATURE EXTRACTION USING GLCM:

Surface examination makes separation of ordinary and anomalous tissue simple. It even gives differentiate amongst harmful and typical tissue, which might be beneath the limit of human observation⁸. GLCM figures the co-event framework of a picture by registering how frequently a pixel with a specific force "I" happens in connection with other pixel "j" at a specific separation d and introduction. In the proposed approach, measurable highlights in light of picture power and highlights from dark level co-event lattice are utilized to recognize typical and irregular patient.

 

We have removed the highlights like Entropy, Contrast, Energy, Correlation, Skew, Kurt, Variance, Standard deviation, Median and Homogeneity for 100 MRI pictures with picture estimate 256 x 256 pixels. Vitality is additionally called Angular Second Moment (ASM) where it quantifies textural consistency. On the off chance that a picture is totally homogeneous its vitality will be most extreme. Entropy is a measure which is conversely connected to vitality. It quantifies the confusion or irregularity of a picture. Next difference is a measure of neighborhood dark level variety of a picture. This parameter takes low an incentive for a smooth picture and high incentive for a coarse picture. Consequence of this procedure is include vectors for all picture in database.

 

Highlight database contains include vectors that is separated for all preparation dataset picture in view of the low level highlights. In this strategy, Gray level co-event network was shaped and the factual surface highlights, for example, differentiate relationship, vitality, homogeneity and entropy.

 

This parameter takes low an incentive for a smooth picture and high incentive for a coarse picture. Aftereffect of this procedure is highlight vectors for all picture in database. Highlight database contains include vectors that is removed for all preparation dataset picture in view of the low level highlights.

 

Vitality is liked to entropy as its esteems have a place with standardized range. Balance is related with the normal dim level contrast between neighbor pixels⁹. It is like change however favored because of decreased computational load and its viability as a spatial recurrence measure. Vitality and differentiation are the most critical parameters regarding visual appraisal and computational load to segregate between various textural designs.

Input Query

Fig.6: Feature vector for Input query image:

 

Contrast = 20, Variance = 1.2377, Standard Deviation = 53.7012, Skew = 3.5125, Kurt = 41.2891, Entropy = 0, Contrast = 0.1884, Homogenity = 0.9966, Energy = 0.9917, Correlation = 0.1414

 

4.4 CLASSIFICATION USING MULTICLASS SVM:

The separated highlights are arranged by the classifier into two classes ordinary and strange alongside various sort of tumor by utilizing the multiclass SVM are shown in (Fig.7). The multiclass SVM recognizes whether the cerebrum picture is ordinary picture or strange picture or gliomas picture are shown in (Fig.8).

 

 

 

5. PERFORMANCE ANALYSIS

The performance of brain tumor classification is analyzed with the following parameters, Accuracy, sensitivity and specificity are calculated using the below formulas are shown in (Table 2).

 

Table 1: Specifications for calculating the accuracy, sensitivity and specificity

Symbol	Expansion	Interpretation
TP	True positives	Tumor diagnosed as Tumor
TN	True negatives	NonTumor diagnosed as NonTumor
FP	False positives	Tumor diagnosed as NonTumor
FN	False negatives	NonTumor diagnosed as Tumor

These parameters are assessed and recorded in (Table1), where, TP indicates genuine positive, FP signifies false positive, FN is false negative and TN is genuine negative. Genuine Positive alludes to the accurately distinguished tumor pixels, True Negative alludes to the wrongly recognized tumor pixels, False Positive alludes to the effectively distinguished non-tumor pixels and False Negative alludes to the wrongly recognized non-tumor pixels. The parameters, Se and Sp characterize the proportion of very much arranged tumor and non-tumor pixels, individually. Every one of these parameters help in characterizing the execution of our proposed characterization procedure and performance analysis report is shown in (Fig.10).

 

Table 2: Performance of Multiclass SVM

 

Fig.10: Analysis report of Multiclass SVM

 

6. CONCLUSION:

In this work, we have proposed a new way of medical image classification using the multiclass SVM. The medical system has been designed by the statistical features and texture based features such as DWT (Discrete Wavelet Transform), GLCM (Gray Level Co-occurrence Matrix) of MRI images and multiclass SVM which help we have got very promising results in classifying the normal and abnormal images along with different type of abnormality. MRI brain images undergoes some preprocessing step in order to filter the noise. Features are extracted from the images using feature extraction algorithm Gray Level Co-occurence Matrix (GLCM) and DWT. The extracted features are classified by multiclass SVM into three classes normal, abnormal and gliomas images. Hence the future work is to improve the classification accuracy by extracting more features and increasing the training data set.

 

7. REFERENCES:

1.     Noramalina Abdullah, Umi Kalthum Ngah, Shalihatun Azlin Azi. Image Classification of Brain MRI Using Support Vector Machine, IEEE International Conference on Imaging Systems and Techniques (IST), 2011, pp. 242 - 247.

2.     Vinod Kumar, Jainy Sachdeva, Indra Gupta, Niranjan Khandelwal, Chirag Kamal Ahuja. Classification of brain tumors using PCA-ANN, IEEE CONFERENCE PUBLICATIONS on Information and Communication Technologies (WICT), 2011, pp. 1079 - 1083.

3.     Jayashri Joshi, Mrs. A.C. Phadke. Feature Extraction and Texture Classification in MRI, Special Issue of IJCCT or International Conference [ICCT-2010], vol. 2, 2010, pp. 130-136.

4.     A. Jayachandran, R. Dhanasekaran. Brain Tumor Detection and Classification of MR Images Using Texture Features and Fuzzy SVM Classifier, Research Journal of Applied Sciences, Engineering and Technology, 2013, pp. 2264-2269.

5.      N.H. Rajini, R. Bhavani. Classification of MRI Brain Images using k-Nearest Neighbor and Artificial Neural Network, IEEE-International Conference on Recent Trends in Information Technology, 2011, pp. 563 – 568.

6.      Garima Singh, M. A. Ansari. Efficient detection of brain tumor from MRIs using K-means segmentation and normalized histogram 1st India International Conference on Information Processing (IICIP),2016,pp.1 – 6.

7.      Manu Gupta, B. V. V. S. N. Prabhakar Rao, Venkateswaran Rajagopalan. Brain Tumor Detection in Conventional MR Images Based on Statistical Texture and Morphological Features International Conference on Information Technology (ICIT), 2016, pp.129-13.

8.     C. Hemasundara Rao, P. V. Naganjaneyulu, K. Satya Prasad. Brain Tumor Detection and Segmentation Using Conditional Random Field, IEEE 7th International Advance Computing Conference (IACC), 2017, pp.807-810.

9.      Luxit Kapoor, Sanjeev Thakur. A survey on brain tumor detection using image processing techniques,7th International Conference on Cloud Computing, Data Science and Engineering - Confluence,2017,P.P-582 – 585.

10.   Bisio A. Fedeli, F. Lavagetto; G. Luzzati, M. Pastorino, M. Raffetto, A. Randazzo, E. Tavanti. Application of EMFs at microwave frequencies for brain stroke detection: Preliminary results International Applied Computational Electromagnetics Society Symposium - Italy (ACES), 2017, P.P-1 – 2.

 

 

 

 

 

 

 

 

Received on 22.10.2017         Modified on 24.11.2017

Accepted on 12.12.2017      © RJPT All right reserved