Detection and Classification of Brain Tumor in MRI Images

Transcription

1 PrahiGadpayle and Prof.P.S.Mahajani 45 Detetion and Classifiation of Brain Tumor in MRI Images PrahiGadpayleand Prof.P.S.Mahajani Abstrat Brain tumor detetion in Magneti Resonane Imaging (MRI) is important in medial diagnosis beause it provides information assoiated to anatomial strutures as well as potential abnormal tissues neessary for treatment planning and patient follow-up. In this paper a brain tumour Detetion and Classifiation System is developed. The image proessing tehniques suh as preproessing, image enhanement, image segmentation, morphologial operations and feature extration have been implemented for the detetion of brain tumor in the MRI images. In this paper, extration of texture features in the deteted tumor is ahieved using Gray Level Coourrene Matrix (GLCM).BPNN and K-NN lassifier is used to lassify MRI brain image into abnormal and healthy image. Keywords- BPNN, GLCM, K-NN, Morphologial operator, Segmentation 1. INTRODUCTION Brain tumour is one of the major auses of death among people. Brain tumor is luster of abnormal ells growing in the brain. It is possible that the hanes of survival an be inreased if the tumor is deteted and lassified orretly at its early stage. Detetion of these tumours from brain is very diffiult at the regions where a tumour is overlapped with dense brain tissues. Visual detetion of these abnormal tissues may result in misdiagnosis of volume and loation of unwanted tissues due to human errors aused by visual fatigue. Nowadays, automati brain tumour detetion in MRI images is very important in many diagnosti and therapeuti appliations.automated lassifiation and detetion of tumours in different medial images is motivated by the neessity of high auray when dealing with a human life. In this paper the designed system is developed for Detetion and Classifiation of Brain tumour from a given MRI image of tumour affeted patients. Magneti Resonane Imaging (MRI) is a medial imaging tehnique used to visualize the internal struture of the body and provide high quality images. MRI images do not involve exposure to radiation, so they an be safely used in people who may be vulnerable to the effets of radiation, suh as pregnant women and babies.brain tumours are lassified into primary brain tumour and seondary braintumour.primary tumours are tumours that originate in the brain itself, seondary brain tumours are the aner ells that originate from another part of the body and have spread to the brain.this paper presents approah for feature extration and lassifiation of brain tumor [1][2]. 2. METHODOLOGY The method involves proessing of MRI images that are affeted by brain tumor for detetion and lassifiation of brain tumors. The image proessing tehniques like preproessing, segmentation, morphologial operator are used for the detetion of tumor and then texture feature extration method is used for extrating features from the MRI image. Features are extrated using Gray Level Coourrene Matrix. After feature extration BPNN and K- NN Classifier is used for the lassifiation of brain into normal and abnormal images. The methodology used for MRI brain images is as shown in Fig. 1 Image aquisition Pre-proessing Segmentation Morphologial operator BPN Abnormal Feature extration K-NN Result of lassifiation Normal Fig 1: Blok diagram of the system 2.1MRI Image database[6]:mri image database onsists tumor brain images and normal brain images. These images are olleted from MRI san enter. The samples of MRI brain images are shown in following Fig 2 International Journal of Emerging Trends in Eletrial and Eletronis (IJETEE ISSN: )

2 PrahiGadpayle and Prof.P.S.Mahajani 46 to blak. Segmentation is aomplished by sanning the whole image pixel by pixel and labeling eah pixel as objet or bakground aording to its binarized gray level. Fig 2 Samples of MRI images These Images are stored in MATLAB and displayed as a gray sale image of size 256* Image Pre-proessing This step is arried out to improve the quality of the image to make it ready for further proessing. This improved and enhaned image will help in deteting edges and improving the quality of the overall image. Edge detetion will lead to finding the exat loation of tumor. Following steps are used in the preproessing stage: a) Noise Removal: In medial image proessing, medial images are orrupted by different type of noises. It is very important to obtain preise images to failitate aurate observations for the given appliation. In this paper noise redution has done on an image by filtering or by wavelet analysis. b) Image Enhanement: Image enhanement is basially improving the interpretability or pereption of information in images for human viewers and providing `better' input for other automated image proessing tehniques. The prinipal objetive of image enhanement is to modify attributes of an image to make it more suitable for a given task and a speifi observer. During this proess, one or more attributes of the image are modified.filtering is tehnique for enhaning the image. 2.3 Segmentation: Image segmentation is based on the division of the image into regions. Division is done on the basis of similar attributes. Similarities are separated out into groups. Basi purpose of segmentation is the extration of affeted regions from the image, from whih information an easily be pereived. Thresholding is used for segmentation as it is most suitable for the present appliation in order to obtain a binarized image with gray level 1 representing the tumor and gray level 0 representing the bakground. Threshold Segmentation: Thresholding often provides an easy and onvenient way to perform the segmentation on the basis of the different intensities or olors in the foreground and bakground regions of an image. A thresholding operation is applied typially on a greysale or olor image.thresholding method is based on a liplevel or a threshold value to turn a gray-sale image into a binary image. Blak pixels orrespond to bakground and white pixels orrespond to foreground. If the pixel s intensity is higher than the threshold, the pixel is set to white, in the output. If it is less than the threshold, it is set 2.4 Morphologial Operators: After onverting the image in the binary format, morphologial operations are applied on the onverted binary image. The purpose of the morphologial operators is to separate the tumor part of the image. Now only the tumor portion of the image is visible, shown as white olor. This portion has the highest intensity than other regions of the image.the erosion operator is used to shrinks objet in images. A struturing element of disk operator is used to perform erosion operation. 2.5 Feature Extration [1],[3]: Features are said to be properties that desribe the whole image. The purpose of feature extration is to redue the original dataset by measuring ertain features. GLCM matrix features are used to distinguish between normal and abnormal brain tumors. GLCM is the gray-level oourrene matrix (GLCM), also known as the gray-level spatial dependene matrix. A GLCM is a matrix where the number of rows and olumns is equal to the number of gray levels, in the image.. Five o-ourrene matries are onstruted in four spatial Orientations horizontal, right diagonal, vertial and left diagonal (0, 45, 90, and 135 ). A fifth matrix is onstruted as the mean of the preeding four matries. Texture features (Grey Level Co-ourreneMatrix Features)from eah o-ourrene matrix, a set of sixfeatures areextrated in different orientations for the training of thelassifier. Let p be the N*N o-ourrene matrix, alulated for eah sub-image, then the following statistial texture features are alulated: 1. Energy (E): Energy measures textural uniformity i.e. Itmeasures pixel pairs repetitions E= p(i, j)^2.(1), 2. Contrast: It a measure of the intensity ontrast between a pixel and its neighbor over the whole image. Contrast is 0 for a onstant image. Contrast=, mod(i j) p(i, j ).. (2) Where, p(i,j) is pixel at loation (i,j) 3. Entropy (EN): It is a measure of randomness. EN=, p(i, j)log p (i,j)... (3) Where, N is no. of different values whih pixels an adopt. 4. Homogeneity (HOM):Itmeasures the loseness of the distribution of elements in the GLCM to the GLCM diagonal. HOM=, p(i, j) / (1+mod (i-j))... (4) International Journal of Emerging Trends in Eletrial and Eletronis (IJETEE ISSN: )

3 PrahiGadpayle and Prof.P.S.Mahajani Inverse Differene Moment (IDM): It is the measure of loal homogeneity. IDM=, * p (i, j).. (5) ( )^ 6. Dissimilarity: Dissimilarity is similar to GLCM ontrast and it is high if the loal region has high ontrast. Dissimilarity =, i j p(i, j) (6) Where G is number of gray levels used in image. 2.6Classifiation: Classifiation is a omputational method used to findpatterns and develops lassifiation shemes for data in very huge datasets. Classifiation is the proess where a given test sample is assigned a lass on the basis of knowledge gained by the lassifier during training. Its task is to assign an input pattern represented by a vetor to one of many pre-speified lasses. In this paper BPNNand K- NN lassifier is used for the lassifiation of brain MRI image into healthy brain or Tumour brain K-Nearest neighbour lassifier The K-nearest neighbor (KNN) lassifiation rule is one of the most well-known and widely used nonparametri pattern lassifiation methods. The k- nearest neighbor lassifier is a simple supervised lassifier that has yield good performane for optimal values of k. This lassifier omputes the distane from the unlabeled data to every training data point and selets the best k neighbors with the shortest distane. In this work, the Eulidean distane is used for distane metri K-NN estimation is based on searhing for the K losest (nearest) samples within a set of training samples (neighbours) to a test sample from the same type. K-NN lassifier omputes distanes between a test sample (feature vetor) and all training samples, and then Ksamples, out of n training samples, that are losest to test sample are subjeted to majority voting to hoose the lass. Eulidean distane is the measure of distane between a test sample and samples of a training set. For N-dimensional spae, Eulidean distane between any two samples or vetors P and Q is given by. Fig. 3 Neural network arhiteture Algorithm stages for BPNN: 1. Initialization of weights 2. Feed forward 3. Bak propagation of Error 4. Updating of weights and biases During the first stage whih is the initialization of Weights, some small random values are assigned. During feed forward stage eah output unit reeives an input signal and transmits this signal to eah of the hidden units. Eah hidden unit then alulates the ativation funtion and sends its signal to eah output unit. The output unit alulates the ativation funtion to form the response of the net for the given input pattern. In this paper the input layer onsistsof six neurons orresponding to the six features. Theoutput layer onsists of one neuron indiating whether themri is a andidate irumsribed tumor or not, and thehidden layer hanges aording to the number of rules thatgive best reognition rate for eah group of features [5]. 3. RESULTS AND DISCUSSION The Methodology lassifies the input MRIimage of brain into normal and abnormal images. The extrated part of tumor of brain images is as shown in following Fig 4 D = (Pi Qi) (7) Where Pi and Qi are the oordinates of Pand Qin Dimension Bak propagation neural network BPNN (Bak propagation neural network) onsist of an interonnetion of simple omponents referred to as neurons, whih are programming onstruts that mimi the properties of biologial neurons. BPNN onsist of one or more layers. Eah layer has one or more neurons. The neuron (pereptron) an be defined simply as a devie with many inputs, one output, and an ativation funtion. Following Fig 3 shows the neural network arhiteture. Fig 4: Tumor deteted from a Tumor brain MRI image International Journal of Emerging Trends in Eletrial and Eletronis (IJETEE ISSN: )

4 PrahiGadpayle and Prof.P.S.Mahajani 48 A u r a y ( % ) K=1 K=5 K=10 K=15 K=20 Seen images Unseen image Fig 6: Auray using K-NN lassifier for different K nearest samples Fig 5: Tumor notdeteted from a normal brain MRI image Six features areextrated of normal and abnormal brain images using GLCM. Extrated features valuesabnormal images are shown in following Table 1 GLCM Parameter Abnormal image Normal Image Energy 3.73E E+09 Contrast Entropy Homogeneity Inverse differene moment Dissimilarity Table 1: GLCM Statistial feature values The Extrated feature is used to train the BPNN and K- NN Classifier.20 images (10 abnormal and 10 normal) are used to trainthe Bak propagation neural network and K- NN Classifier whih lassify the MRI brain images into normal and abnormal images. 3.1 Classifiation using K-NN lassifier: The auray by using K-NN lassifier based on searhing for the different K losest (nearest) samples is as shown in following Fig 6.The experimental results show that maximum auray level ahieved for seen images are 100% and for unseen images are 70%. 3.1 Classifiation using BPNN lassifier: Theauray by using BPNN lassifier at the different epohs as shown in following Fig 7.The experimental results show that maximum auray level ahieved for seen images are 100% and for unseen images are 72.5%. A u r a y ( % ) epoh 500 epoh epoh epoh Fig 7: Auray by using BPNN lassifier for different epohs 4. CONCLUSION Seen images Unseen images In this paper, lassifiation tehnique is developed to lassify between normal and abnormal brain images.morphologial operator is used for the extration of tumour part. Textures features are used in the training of lassifier.the extrated features of MRI are used as input to the BPNN and K-NN.Twentyimages (10 abnormal and 10 normal) are used to train the Bak propagation neural network and K-NN Classifier. Forty unseen images are tested using BPNN and K-NN lassifier. The experimental results show that Auray ahieved by using K-NN lassifier is 70 %.and by using BPNNlassifier is72.5%. International Journal of Emerging Trends in Eletrial and Eletronis (IJETEE ISSN: )

5 PrahiGadpayle and Prof.P.S.Mahajani 49 REFERENCES [1] Dipali M. Joshi, Dr.N. K. Rana, V. M. Misra, Classifiation of Brain Caner Using Artifiial Neural Network, 2010 IEEE,Eletroni Computer Tehnology (ICECT), 2010 International Conferene,Conferene Loation :Kuala Lumpur,Date of Conferene:7-10 May 2010,Page(s): [2] Ehab F. Badran, EsraaGalal Mahmoud, and NadderHamdy, An Algorithm for Deteting Brain Tumours in MRI Images,2010 IEEE,Computer Engineering and Systems (ICCES),2010 International Conferene, Loation: Cairo, Publiation Year: 2010,Date of Conferene: Nov De Page(s): [3] S.N.Deepa&B.Aruna Devi, Artifiial Neural Networks design for Classifiation of Brain Tumour, 2012 IEEE, 2012 International Conferene on Computer Communiation and Informatis (ICCCI -2012), Date of Conferene :Jan , 2012, Coimbatore, INDIA. [4] Gonzalez, R.C. Rihard, E.W; Digital Image Proessing, (2004), II Indian Edition, Pearson Eduation, New Delhi, India. [5] YawarRehman,FahadAzim, Comparison of Different Artifiial Neural Networks for Brain Tumour Classifiation via Magneti Resonane Images, 2012 IEEE, th International Conferene on Modelling and Simulation, Conferene Loation :Cambridge, Date of Conferene: Marh 2012,Page(s): [6] PIONEER diagnosti entre. International Journal of Emerging Trends in Eletrial and Eletronis (IJETEE ISSN: )