A New and Robust Segmentation Technique Based on Pixel Gradient and Nearest Neighbors for Efficient Classification of MRI Images

Transcription

1 A New and Robus Segmenaion Technique Based on Pixel Gradien and Neares Neighbors for Efficien Classificaion of MRI Images Sanchi Kumar, Sahil Dalal Absrac This paper proposes a new fully auomaed mehod for he segmenaion of umour in a umour affeced brain MRI image and is classificaion from a normal image. The echnique deecs brain umour on he basis of differences in pixel inensiies beween he umour and he surrounding issues. I also explois he conneced pixels labelling of he umour pixels which accuraely segmens brain umour. Classificaion is done using suppor vecor machine (SVM) which uses GLCM echnique o calculae feaure vecor. The classificaion parameers (i.e. accuracy, sensiiviy and specificiy) and oher parameers (i.e. speed and compuaional power used) show ha our mehod is beer in performance as compared o he oher sae of he ar mehods. Keywords GLCM, MRI, SVM, Solidiy and Thresholding. I. INTRODUCTION Brain umour [] [] is one of he major causes of deah among men hroughou he world. I is he growh of some abnormal issues wihin he brain. I s mainly of wo ypes: one is benign umour and oher is malignan umour. Malignan umour is more dangerous han benign umour as i is cancerous in naure and has he abiliy o grow inside he brain, hus malignan umour can resul in deah if no diagnosed a an early sage. MRI is one of he echniques of medical imaging which helps in deecing umour accuraely. Radiologiss can easily examine a brain MRI image o conclude wheher i is normal or cancerous. Bu he problem arises when here are numerous images o be examined leading o decreased accuracy and high ime consumpion. Hence, in order o increase he accuracy and minimize he ime consumpion, auomaed classificaion wih he help of compuer echnology is needed. Sanchi Kumar, Deparmen of Elecronics and Communicaions Engineering, Delhi Technological Universiy, Delh India. Sahil Dalal, Deparmen of Elecronics and Communicaions Engineering, Delhi Technological Universiy, Delh India. The mehodology includes he following seps: Image preprocessing, segmenaion, feaure exracion, and classificaion. Image pre processing is used o remove noise from an image o make i suiable for furher processes. I is needed for he accurae segmenaion of umour from he image. Segmenaion refers o he process where only he desired par of he image is kep and he unwaned pars removed. I is required o exrac he region of umour in he image making he classificaion easier and accurae. Feaure exracion is he processing of some calculaed parameers from he segmened images which acs as inpu o he classifier for classificaion. I furher reduces he informaion in image o a minimum level. For accurae classificaion, feaures which are exraced should be meaningful and useful. Classificaion process differeniaes a normal image from an image having a umour wih he help of algorihms known as classifiers. There are many ypes of classifiers which can be used for such purposes. Some of hem are suppor vecor machine (SVM) [], k neares neighbours (KNN) [3], Arificial neural nework (ANN) [4], Probabilisic neural nework (PNN) [5] ec. Each of he classifier has is own advanages and disadvanages. K-NN [3] has some limiaions such as high compuaion cos, difficuly in finding he proper value of K parameer ec. Even hough K-NN [3] has high accuracy in classificaion, is performance is degraded by is limiaions. PNN [5] has a major problem ha i is a very slow classifier. ANN [4] performs beer han he above menioned classifiers wih he high dimension of feaure vecor, bu i also has high compuaional cos problem. As feaure vecor is small in our mehod, SVM [] performs he bes amongs all he oher classifiers. Classificaion is divided ino wo pars: raining and esing. In raining, known daa is given o he classifier for is raining. Afer ha unknown images are given o he classifier for classificaion for esing. The efficiency of classificaion mainly depends upon he raining. This Paper is organized as follows. Secion II describes he mehodology uilized for he proposed mehod.secion III describes experimenal work of he proposed mehod. The umour classificaion and experimenal resuls are given in secion IV wih conclusion in secion V. 708

2 II. METHODOLOGY The mehodology used for he classificaion of brain MRI image of a normal person and a person having brain umour is shown in fig.. This sysem consiss of he following modules: pre-processing (RGB o grayscale conversion, resizing and filering), segmenaion, feaure exracion, raining and esing. Firsly, raining daase of brain MRI images is given o he sysem for raining of classifier. Then esing daa se is given o he sysem which classifies i ino wo caegories normal and abnormal based on he raining. (a) (b) Fig.. Brain MRI (a) Normal and (b) Abnormal B. Pre-processing of image Pre-processing includes filering of image using median filer. A he ime of acquisiion, he image ges disored wih noise and he conras is also low which leads o lower efficiency in furher seps of he process. Filering is needed o remove noise and o enhance image conras. Advanage of using median filer is ha i removes he noise while preserving he edges which helps in improving conras of he image. Accuracy of umour segmenaion is also increased by his sep. Filered image is shown in fig 3. Fig.3. Filered image Fig.. Block Diagram A. MRI Image The MRI (magneic resonance imaging) [6] is one of he bes ools used in hospials and clinics for medical diagnosis of various diseases. Some of is advanages are: (i) I prevens exposure of body from ionizing radiaions. (ii) I gives very high conras beween he issues. Fig shows he MRI image of normal and abnormal brain respecively. I is clearly seen ha he issues affeced from umour are of higher inensiy han he normal issues. When he image is aken, i is an RGB image. Before giving i o he sysem, image is firs convered o gray scale and resized o a size of 500 X 500. C. Segmenaion In he field of biomedical imaging, umour deecion is a crucial ask as brain umour can be a severe disease if no properly diagnosed. In he acquired MRI image, i can be seen ha generally he inensiy values where umour is presen is greaer in comparison o he surrounding inensiies. In our work, segmenaion is done using inensiy values and solidiy of he umour o deec he umour from is surrounding issues. The firs sep in segmenaion is hresholding, where he inensiy values greaer han predefined hresholds are kep and he values below he hreshold level are ignored. The hreshold level is found by hi and rial mehod. We ge a value of 64 as our hreshold. Afer hresholding, he image sill conains unwaned regions along wih he umour region. For removing hese unwaned regions, solidiy and area [7] of he segmened regions are calculaed. Only hose regions having solidiy values greaer han 0. are labelled as umour 709

3 inensiy values. The second crieria is area where he max area of he region is considered as umour based region. Afer exploiing hese properies, he MRI image is segmened o deec he umour as foreground and surrounding issues as background. CORRELATION: Correlaion is compued ino wha is known as he correlaion coefficien, which ranges beween - and +. Correlaion P i j ij () j 0 HOMOGENEITY: Homogeneiy is defined as he qualiy or sae of being homogeneous. Pij Homogenei y (3) ( i j) ENTROPY: Enropy is a measure of he uncerainy in a random variable. Enropy ln( P ij ) P (4) ij (a) (b) Fig.4. (a) Normal Brain MRI wih is segmened image and (b) Abnormal Brain MRI wih is segmened image The segmened resuls are shown in fig.4 in which he wo images are segmened where firs image conains a umour and second image is a normal MRI image. The segmened resuls are also shown as here counerpar images. D. Feaure exracion Feaure exracion plays a very imporan role in image processing /compuer vision for paern recogniion and classificaion. A raw segmened image canno be given direcly o he classifier as i conains unwaned daa which increases compuaional ime and cos of he process. Some meaningful and imporan daa poins are exraced from he image (also called feaures) using feaure exracion process. Feaures can be of many ypes such as inensiy based feaures [], exure based feaures [3] or shape based feaures [4]. Several feaures are calculaed from an image which forms a feaure vecor which is given as inpu o a classifier for accurae classificaion of normal and abnormal MRI image. In our work, we are calculaing 5 exure feaures using Grey Level co-occurrence marix (GLCM) [8] echnique which consiues a feaure vecor of 5X dimension for each image. Feaures using GLCM are given by: CONTRAST: Conras is defined as he separaion beween he darkes and brighes area. Conras ij P ( i j) () ENERGY: I provides he sum of squared elemens in he GLCM.Also known as he uniformiy or he angular second momen. Where (GLCM). n P ij Energy (5) is he Grey Level Co-occurrence marix E. Classificaion using Suppor vecor machine Suppor vecor machine (SVM) [] is a classificaion algorihm based on he supervised learning model. In SVM, a high dimensional feaure space is divided ino wo halves by consrucing a hyperplane where each half belongs o a differen class. Hyperplane is given by he equaion: G( x) w x b (6) Where x is he feaure vecor, w gives he orienaion of he plane, b is he bias i.e. posiion of he hyper plane w.r.. origin in high dimensional space. Here values of w and b are opimized by giving raining samples from boh he classes of image which also maximizes he disance beween he wo classes, hus forming an accurae classifier. In SVM, feaure vecor of each class are on eiher sides of he hyper plane. Class vecors a boundary are called suppor vecors as he posiion and orienaion of hyper plane depends upon hem. G( x x C ) w x b 0, (7) G( x x C ) w x b 0, (8) Afer raining he classifier, a feaure vecor is given for esing i. G( ) is calculaed for he given vecor if G( ) 70

4 value comes ou o be posiive hen x belongs o class and if G(x) comes ou o be negaive hen belongs o class. F. Performance Measure Performance of a classificaion algorihm can be measured by a specific able layou which is called he Confusion Marix [9]. I is called so because i makes i easier o see wheher he mehod is confusing he resul of classificaion beween he wo classes. Each column in confusion marix shows insances in he prediced class and each row shows insances in he acual class. Confusion marix is shown in he fig.5. ACTUAL CLASS PREDICTED CLASS Abnormal Normal Abnormal TP FN Normal FP TN Fig. 5 Confusion Marix Where TP, TN, FP, and FN are given as: True Posiive (TP): Abnormal brain correcly idenified as abnormal. True Negaive (TN): Normal brain correcly idenified as normal. False Posiive (FP): Normal brain incorrecly idenified as abnormal. False Negaive (FN): Abnormal brain incorrecly idenified as normal. Parameers ha are calculaed using confusion marix are:. TP TN Accuracy = * 00 TP TN FP FN. TP Sensiiviy = * 00 TP FN 3. TN Specificiy = * 00 TN FP role as i makes he classificaion easier and faser. In feaure exracion, we have calculaed only 5 feaures for he classificaion. As he number of feaures in our feaure vecor is so small, our mehod is fas and akes less compuaional power compared o oher mehods. GLCM FEATURES TABLE I Feaures Exraced NORMAL IMAGE ABNORMAL IMAGE Conras.557 e-04.3 e-04 Correlaion Homogeneiy Enropy e e07 Energy The classificaion processing is divided ino wo pars i.e. raining and esing par. In raining par, known daa (i.e. 5 feaures*40 images) is given for he raining of classifier. Afer he compleion of raining, 50 unknown images are given o he classifier for classificaion, his is he esing par. The classificaion is performed using SVM. The accuracy of classificaion mainly depends upon is raining par. IV. RESULTS & DISCUSSION In his mehod, we are using brain MRI images of 90 paiens. Ou of hese 90 images, 40 images have been used for raining of classifier and oher 50 images for esing. Firsly, feaures calculaed from 40 raining images have been given o he classifier (based on SVM). Afer Training, 50 unknown images have been given o he classifier for classificaion ino normal and abnormal. Accuracy, sensiiviy and specificiy were calculaed based on he confusion marix shown in Table II and hey came ou be 96%, 9% and 00% respecively. Table III shows he comparison of accuracy of oher mehods and our mehod i.e. SVM wih segmenaion. I shows ha accuracy of our mehod is greaer han he mehod [0] bu slighly lower han mehod []. TABLE II Confusion Marix III. EXPERIMENTAL DISCUSSION Our mehod includes following seps: image preprocessing, segmenaion, feaure exracion, classificaion. In image pre-processing sep, MRI images are firs convered from RGB o grayscale and resized o a size of 500 X 500.Then median filer is used o remove noise from he image and o make segmenaion accurae. In segmenaion sep, Tumour region from an abnormal brain MRI image is exraced leaving behind all oher unwaned pars of he image. These seps play an imporan ACTUAL CLASS PREDICTED CLASS Abnormal Normal Accuracy Abnormal 3/5 /5 9% Normal 0/5 5/5 00% 7

5 TABLE III Comparison of proposed mehod wih oher mehods used. [9] Evangelia I. Zacharak Sumei Wang, Sanjeev Chawla, Dong Soo Yoo, Ronald Wolf, Elias R. Melhem, and Chrisos Davazikos, Classificaion of brain umour ype and grade using MRI exure and shape in a machine learning scheme, PMC 00 December. REFRENCES METHOD ACCURACY Hari Babu Nandpuru e al [0] Kean Machhale e al [] Skull Masking + SVM Skull Masking + SVM-KNN 84% 98% [0] Hari Babu Nandpuru, Dr. S. S. Salankar, Vibha Bora, MRI BRAIN CANCER CLASSIFICATION USING SUPPORT VECTOR MACHINE, 04 IEEE Sudens' Conference on Elecrical, Elecronics and Compuer Science. [] Kean Machhale, Hari Babu Nandpuru, Vivek Kapur, Laxmi Kosa, MRI Brain Cancer Classificaion Using Hybrid Classifier (SVM-KNN), 05 Inernaional Conference on Indusrial Insrumenaion and Conrol (ICIC) College of Engineering Pune, India. May 8-30,05. [] V.P.GladisPushparah S.Palan Linear Discriminan analysis for brain umour classificaion using Feaure Selecion,In. J. Communicaion and Engineering, vol 5, issue 4, pp PROPOSED METHOD SEGMENTATION + SVM 96% [3] Dr.S.Palan V.P. Gladis Pushpa Rah Deecion and characerizaion of brain umour using segmenaion based on HSOM, Wavele packe feaure spaces and ANN, Ieeeexplore.ieee.org/xpl/freeabsrac.jsp/ anumber=594097(00). V. CONCLUSION This research paper involves using a segmenaion echnique for he exracion of umour region in an abnormal image and is classificaion from normal image using SVM. Resuls obained from his mehod shows ha he accuracy aained for classificaion is 96%. Oher advanages of our mehod are ha i is fas and uses small compuaional power as compared o oher mehods. For fuure work, a hybrid SVM algorihm can be used o ge beer accuracy. References [] hp:// [] M. F. B. Ohman, N. B. Abdullah, N. F. B. Kamal, MRI brain classificaion using Suppor vecor machine, IEEE, Cenre for Arificial Inelligence & Roboics (CAIRO), Universiy Teknology Malaysia, 0. [3] Dr. R J. Rameke and Khachane Monali Y, Auomaic Medical Image Classificaion and Abnormaliy Deecion Using K Neares Neighbor, Inernaional Journal of Advanced Compuer Research, Volume- Number-4 Issue-6 December-0l. [4] E. A. El-Dihshan, T. Hosney, A B. M. Salem, Hybrid inelligence echniques for MRI Brain images classificaion, ELSEVIER, Digial Signal Processing, Volume 0, pp,433-44, 00. [5] Lashkari Amir Ehsan A Neural Nework-Based Mehod for Brain Abnormaliy Deecion in MR Images Using Zernike Momens and Geomeric Momens, UCA-IO. [6] Chinnu A, MRI Brain Tumour Classificaion Using SVM and Hisogram Based Image Segmenaion, Inernaional Journal of Compuer Science and Informaion Technologies, Vol. 6 (), 05, [7] K. Somasundaram. T. Kalaiselvi. Auomaic brain exracion mehods for TI magneic resonance images using region labeling and morphological operaions, ELSEVIER, Compuers in Biology and Medicine 4 (0) [8] Shwea Jain, "Brain Cancer Classificaion Using GLCM Based Feaure Exracion in Arificial Neural Nework", IJCSET, ISSN: , VoL 4 No. 07 JuI03. 7