Published on

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. Prof.Mrs.K.N.Rode, Prof.Mr.R.T.Patil/ International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1925-1928 Texture Analysis Of Mri Using Svm & Ann For Multiple Sclerosis Patients Prof.Mrs.K.N.Rode (S.I.T.C.O.E.),Prof.Mr.R.T.Patil (R.I.T.C.O.E.)ABSTRACT Due to the complexity & variance of Statistics were extracted,& statistical analysis methodautomated MRI segmentation of brain MS GLCM(Gray-level co-occurance matrix ) was appliedbecame a complex task. In this paper a structural between LWM,NAWM & NWM.texture analysis method on MS segmentationscheme is presented, that emphasis on structural II. PRIOR WORKanalysis of MS as well as on normal tissues.. All the previous methods are classified into 4 types.Methods:- MRI of any number of MS patients & 1)Thresholding methodhealthy subjects were selected & that much ROI This method is frequently used for imagewere chosen from MS patient MRI & healthy segmentation, simple & effective method for imagessubject MRI(magnetic resonance imaging) for with different intensities. It has the drawback ofLWM(lesion white matter),NAWM(Normal generating only two classes therefore this methodappearing white matter) & NWM(Normal white fails to deal with multichannel images. It also ignoresmatter) respectively. All of the ROI were analysed the spatial characteristics due to which an imageby gray-level difference statistics & contrast, becomes noise sensitive which corrupts the histogramangular second moment, mean & entropy those 4 of the image.texture features were extracted .Finally statisticsignificance was tested among three groups. An 2) Region growing methodANN & SVM are used for classification. This This technique is not fully automatic. It isapproach is designed to investigate the differences also noise sensitive therefore extracted regions mightof texture features among macroscopic have holes or even some discontinuitites.LWM,NAWM in & MRI from patients with MS& NWM. 3) Shape based method This method has limited degree of freedom & semi-automatic.KEYWORDS:- MS,Gray level difference 4)Supervised & unsupervised segmentationstatistics, T.A.(texture analysis),MRI methods,LWM,NAWM,NWM.I. INTRODUCTION III. PROPOSED WORK:- Multiple sclerosis is a chronic idiopathic In this paper texture analysis method usingdisease resulted in multiple areas of inflammatory SVM & ANN is presented & applied to MRI braindemyelination in the CNS(Central nervous MS segmentation. Firstly sample MR images will besystem).MS lesion formation often leads to selected & then the MR image is divided into smallunpredictable cognitive decline & Physical elements to extract the different kinds of featuresdisability.Due to the sensitivity in detecting MS which quantify the intensity, symmetry & texturelesions,MRI has become an important tool for properties of different tissues.diagnosing MS & monitoring its progression. We used gray-level co-occurrence matrix Over the past decade,the demand for approach introduced by Harlick [7] which is well-accurate detection & quantification of MRI known statistical method for extracting second-orderabnormalities increase rapidly.Thus it is crucial to texture information for images. The assumption isuse image analysis techniques to extract that local texture of tumor cells is highly differentdiagnostically significant image features[3]. from local texture of other biological tissues. From Texture analysis refers to a set of processes GLCM we extracted the textures features likeapplied to characterize spatial variations of pixel’s contrast,angular second moment,mean & entropy.Forgray levels in an image.Texture analysis which is segmentation & classification we used ANN & SVMused to quantify pathological changes that may be to detect various Tissues like white matter grayundetectable by conventioalMRI techniques,has the matter & MS.potential to detect the subtle changes in tissues &supports early diagnosis of MS. In this paper,texture analysis was performedon brain MRI of MS patients & normalcontrols.Texture features from gray-level difference 1925 | P a g e
  2. 2. Prof.Mrs.K.N.Rode, Prof.Mr.R.T.Patil/ International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1925-1928 gΔ(x,y)=g(x,y)-g(x+Δx,y+Δy) (1) g Δ(x,y) in (1) is defined as gray difference. gΔ(x,y) of each pixel is caculated by moving (x,y) throught the image.Suppose gray difference level is m , the number of each gΔ(x,y) is counted up,and histogram of gΔ(x,y) is drawn.Finally occurred probability P Δ (i) of each gΔ(x,y) is caculated based on histogram Texture characteristic is closely related to PΔ(i). If PΔ(i) changes rapidly along with variation of i, texture of image is coarse, whereas it is fine if PΔ(i) distributes smoothly. In this study , texture features including contrast,angular secondmoment,Fig. 1. ROI selection from MRI of MS patients and mean and entropy were extracted from the graylevelnormal controls. difference statistics. Their mathematical expressionsLeft: MRI from a MS patient; Right: MRI from a are as belowhealthy subject. 1.Contrast:- CON=Σ i² PΔ(i)A: ROI of LWM ; B: ROI of NAWM; C: ROI of iNWM 2. Angular second moment:- ASM= Σ (PΔ (i) )² Input Image Feature i Extraction 3.Mean:- Mean=1/m Σ iPΔ(i) i 4. Entropy:- Ent= - Σ PΔ (i) log PΔ (i) Pre-processing i These 4 texture features describe the texture characteristics of image from 4 different texture Classification Classification analysis aspects. Using ANN Using SVM B. ARTIFICIAL NEURAL NETWORK The Neural networks [3] developed from the theories of how the human brain works. Many Result modern scientists believe the human brain is a large collection of interconnected neurons. These neurons Fig.2 Block diagram of proposed work are connected to both sensory and motor nerves.Methods:- Scientists believe, that neurons in the brain fire byA. Need for Texture Analysis:- emitting an electrical impulse across the synapse to In the early Multiple sclerosis (MS) stages other neurons, which then fire or dont depending onwhen only subtle pathological changes occur, it is certain conditions. Structure of a neuron is given invery hard to detect MS damage in MRI by visual fig.2examination or conventional measures such as lesionvolume & number[5].Thus tissue discriminationbetween Lesion white matter(LWM) & NormalWhite Matter(NWM) is of critical importance to theearly detection of MS. Texture analysis on MRI inessence is the analysis of Gray-level variationsbetween image pixels in ROIs which capture thespatial & intensity information from the possiblemicroscopic MS lesions. Texture features that areextracted by texture analysis techniques quantify bothmacroscopic lesions & microscopic abnormalitiesthat may be undetectable to conventional measures. Fig-2 Structure and functioning of a single neuronTerms Used in Ttexture Analysis: The Artificial neural network [3] is basically Gray-level difference statistics is commonly having three layers namely input layer, hidden layerused in texture analysis. Suppose g(x,y) is a pixel and output layer. There will be one or more hiddenpoint of image, the gray difference between this pixel layers depending upon the number of dimensions ofand its neighborhood pixel g(x+Δx,y+Δy) is as the training samples. Neural network structure usedbelow: in our experiment is consist of only two hidden layers 1926 | P a g e
  3. 3. Prof.Mrs.K.N.Rode, Prof.Mr.R.T.Patil/ International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1925-1928having 14 neurons in the input layer and 1 neuron inthe output layer as shown in Figure 3. Block division Block division Feature Feature extraction extractionFig-3 Simple Neural network Structure A learning problem with binary outputs (yes Feature selection & Feature selection/ no or 1 / 0) is referred to as binary classification & classifier classifier designproblem whose output layer has only one neuron. A designlearning problem with finite number of outputs isreferred to multi-class classification problem whoseoutput layer has more than one neuron. The examples Fig-4 (Left) the training process flow(Right)Theof input data set (or sets) are referred to as the tumor segmentation process flowtraining data. The algorithm which takes the trainingdata as input and gives the output by selecting best C.SUPPORT VECTOR MACHINE:-one among hypothetical planes from hypothetical MS slices based on visual examination byspace is referred to as the learning algorithm. The radiologist/physician may lead to missing diagnosisapproach of using examples to synthesize programs is when a large number of MRIs are analyzed. To avoidknown as the learning methodology. When the input the human error, an automated intelligentdata set is represented by its class membership, it is classification system is proposed which caters thecalled supervised learning and when the data is not need for classification of image slices afterrepresented by class membership, the learning is identifying abnormal MRI volume, for MSknown as unsupervised learning. There are two identification. In this research work, advanceddifferent styles of training i.e., Incremental Training classification techniques based on MATLABand Batch training. In incremental training the Support Vector Machines(SVM) are proposed andweights and biases of the network are updated each app lied to brain image slices classification usingtime an input is presented to the network. In batch features derived from slices. Support vector machinestraining the weights and biases are only updated after are a set of related supervised learning methodsall of the inputs are presented. In this experimental which analyze data and recognize patterns. In thiswork; back propagation algorithm is applied for paper we are using it only to train the algorithm tolearning the samples, Tan-sigmoid and log-sigmoid recognize the lesions as well as the clean white cells.functions are applied in hidden layer and output layer We use only the in built MATLAB svm functions torespectively, Gradient descent is used for adjusting train and classify.the weights as training methodology. In this paper, each pixel together with asmall square neighborhood is defined as a structure COMPARISION OF PERFORMANCE OFelement, which is called „block‟. Further steps are all CLASSIFIERS:-based on the blocks. For training process, firstly In this paper we performed the classificationdifferent features are extracted block by block in one using two techniques SVM& ANN. Experimentsimage. When a new image comes, only those selected were performed on MR images. In this work, wefeatures are extracted and the trained classifier is have considered two different training and test sets asused to categorize the tumor in the image. The explained in the earlier section. The experimentstraining and detection process flow of the proposed were carried out on PC. The performance ofmethod is shown in figure 4. It should be noticed that classifier using Support Vector Machine ( SVM) &the input images are preprocessed beforehand, Artificial neural network (ANN) was evaluated. Forincluding skull stripping which eliminates the skull comparative analysis SVM and ANN classifiers wefrom the brain image and scale normalization to have considered two different training and test setsadjust the intensity scale of the input images. as explained in the earlier section. The results obtained from the SVM & ANN classifier .But SVM has a higher accuracy of classification over ANN classifier. 1927 | P a g e
  4. 4. Prof.Mrs.K.N.Rode, Prof.Mr.R.T.Patil/ International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1925-1928CONCLUSION In this study, texture analysis using SVM &ANN based on Gray-level difference statistics wasperformed on brain MRI of MS patients and normalcontrols. The results suggested that texture featuresof NAWM with MS patients were significantdifference from LWM and nomal controls, which isvaluable in supporting early diagnosis of MS.REFERENCES [1] M. Filippi, C. Tortorella, and M. Rovaris, “Magnetic resonance imaging of multiple sclerosis,” J Neuroimaging. vol.12, pp.289– 301, 2002 [2] D.H. Miller, R.I. Grossman, S.C. Reingold and F. McFarlan, “The role of magnetic resonance techniques in understanding and managing multiple sclerosis” Brain. Vol.121, pp.3–24, 1998. [3] Armspach,JP,Gounot,D,Rumbach L, “In vivo determination of multiexponential T2 relaxation in the brain of patients with multiple sclerosis”,Magn Reson Imaging, vol. 9, pp107-113, 1991. [4] Kurtzke JF. Rating neurological impairment in multiple sclerosis: an expanded disability status scale (EDSS). Neurology, 33, pp1444–52 ,1983 [5] Mathias JM, Tofts PS and Losseff NA, “Texture analysis of spinal cord pathology in multiple sclerosis”. Magn Reson Med, 42:pp929–935, 1999 [6] Jing Zhang, Longzheng Tong, Lei Wang and Ning Li, “Texture analysis of multiple sclerosis: a comparative study”,Magnetic Resonance Imaging, 26, pp1160–1166, 2008 [7] YU Chunshui, LIN Fuchun and LI Kuncheng,”Study of relapsing remitting multiple sclerosis by using magnetization transfer imaging”,Chin J Med Imaging Technol ,Vol 21 ,No 8,pp1202-1204, 2005 [8] YU Chunshui, LI Kuncheng and QIN Wen, “Study of multiple sclerosis using diffusion weighted imaging”,Chin J Med Imaging Technol ,Vol 21 ,No 5,pp687-689 ,2005 77846 1928 | P a g e