MRI brain image analysis for tumour diagnosis using hybrid MB-MLMpattern classification technique.

Shenbagarajan A1*, Ramalingam V1, Balasubramanian C2, Palanivel S1

1Department of Computer Science and Engineering, Annamalai University, Chidambaram, India2Department of Computer Science and Engineering PSR Rengasamy College of Engineering for Women, Sivakasi,India

Abstract

Brain tumours are rooted by atypical and abandoned enlargement of brain cells, which are thesubsequent source of death associated to cancer in less than 30 years age of people in recent years. Earlystage diagnosing of these brain tumours will reduce the unconditional deaths of young people. For thatthe most suggested one of the finest expertises is Magnetic Resonance Imaging (MRI). In this work,proposed a brain MRI image based medical image analysis process, which consists of Modified BatAlgorithm with Modified Levenberg Marquardt (MB-MLM) classification with Active Contour Method(ACM) segmentation method to identify or classify tumor or non-tumor at earlier stage. For optimalresults, this work also proposes the methods like advanced median filter pre-processing method forenhance the input image, parallelized clustering method for surface feature extraction and Intensity inHomogeneity (IIH) for high segmentation accuracy, hybrid wavelet and Sobel and Canny featureextraction method and Fast Independent Component Analysis (Fast ICA) feature selection method fordimensionality reduction, these proposed methods are increase the efficiency of the proposed MRI brainimage based tumor diagnosis process. The performance of this proposed work is measured by standardparameters such as sensitivity, specificity and accuracy.

Keywords: Brain tumours, Magnetic resonance imaging (MRI), Modified levenberg marquardt (MB-MLM), Activecontour method (ACM), Hybrid wavelet and sobel and canny feature extraction, Fast independent component analysis(Fast ICA).

Accepted on August 31, 2016

IntroductionRecent years medical imaging is extensively utilized forexamination and analysis of image based applications. Theimprovement on the medical imaging systems at reduced costcan solve many current problems in medical field such asimage-guided surgery, therapy evaluation and diagnostic tools[1]. Thus, Magnetic Resonance Imaging (MRI) has beenwidely used due to its excellent spatial resolution, tissuecontrast and non-invasive character [2,3]. Typically MRIsupported imaging equipment utilized in brain imagesgrounded by diagnosis procedure, since it is distinguished thatthe brain has a problematical structure; thus, truthfulexamination of brain is highly significant for identifyingtumors founded on MRI method, in sort to offer appropriatetreatment [4]. A method to categorize tissues into thesegrouping is an essential stride in quantitative morphology ofbrain since the largest part of brain organizations is distinct byrestrictions of these tissue modules [5]. Contrasting additionaldiagnostic processes, MRI schemes generate recurrent images,where varied essential factors of interior anatomicalorganization in the identical body segment are tinted by every

image with numerous contrasts. In the largest part of the MRIsupported medical imaging investigations, primarily spotlightson pre-processing method. The main popular pre-processingapproach is noise correcting or suppression for non-uniformities. There are numerous algorithms anticipated forthis duty that adjacent to their profits, they may have off-putting effects on auxiliary processing phase [6]. Previous toseveral examinations on an exact objective in the image, it isessential to section or categorize that from further divisions inthe image.

In general segmentation in medical images is dividing thepixels to distinguish and divide the intended area generally atissue or an injury from the surroundings and strong tissues. Insome study meadows, segmentation of a definite tumor ortissue is the major principle. In other words, segmentation is anintermediary stage for auxiliary investigation such ascategorization or added measurements. In case of brain tumors,it is a tricky task concerning to the distinctiveness of thetumour in the MRI brain descriptions [7]. Image segmentationprocedures make use of region, edge, or intensity possessionsof the goal tissue in the image to divide them from the

Biomedical Research 2016; Special Issue: S191-S203 ISSN 0970-938Xwww.biomedres.info

Biomed Res- India 2016 Special Issue S191Special Section: Computational Life Science and Smarter Technological Advancement

surroundings [8]. The plan of edge-based segmentationtechniques is to discover the margins of two neighbouringregions that have dissimilar individuality. The most admiredalgorithms for recognition of tumour edges in MRIdescriptions is utilizing level-sets [9] and/or coalesce it withcategorization or grouping techniques [10]. One moregradually more significant process in the medical field iscategorization of MRI brain images because it is vital forsurgical preparation and intrusion. Physical categorization ofMagnetic Resonance (MR) brain tumor descriptions is ademanding and sustained job [11]. In a large amount ofsurroundings the assignment is completed by blotting thetumor area slice-by-slice, which restrictions the humanobserver’s sight and produces jaggy images. Physicalcategorization is also characteristically prepared with intensityimprovement offered by an inserted distinction agent [12]. Incase, categorization based on the human spectator is extremelyprone to fault. As a product, the categorization outcomes areextremely substandard which directs to severe results. Thus, aroutine or automatic medical image investigation usingcategorization with segmentation technique is extremelyattractive as it diminishes the consignment on the humanviewers and acquiesces better outcomes. In the recent times,several classification techniques have been evolved for brainMRI image analysis for the cause of tumor diagnosis. Howeverthese methods don’t consider noise removal and Intensity levelof the pixels is not focused by recent works. Here this proposedwork describes the most efficient techniques along withclassification process for tumor diagnosis obtained from thebrain MRI images. Removal of noises from the image is anearly stage in the image processing, which improves theimages for further proceedings, since noises in images resultsin the errors such as blurring effects.

In this proposed work uses following processes for MRI brainimage based tumor diagnosis, such as pre-processingtechniques Advanced Median Filter (AMF), parallelizedclustering method and Intensity in Homogeneity (IIH) forenhancing the given MRI images for further process in imageanalysis. Then feature extraction methods such hybrid wavelettransform and Sobel and Canny methods is proposed forextracting texture features. After features extraction, in order toreduce the dimensionality problem in this work uses the FastIndependent Component Analysis (Fast ICA) feature selectionapproach for best features selection. These two approaches aremainly significant and impartibly component of MB-MLMclassification task performed with region based ACMsegmentation for diagnosing process.

Related WorkXuan et al. proposed tumor segmentation method based onstatistical structure investigations [13]. At first, symmetry-based, intensity-based, and texture-based features are hauledout from structural rudiments. After that Ada boostcategorization method using that study by choosing themajority of distinguished features is anticipated to categorizethe structural components into usual tissues and irregular

tissues. This proposed method’s experimental results on morethan 100 tumor-contained brain MR descriptions attain anaverage accurateness of 96.82% on tumor division.

Shah et al. offered computationally intellectual methods tocategorize brain MRI images into usual and irregular (havingtumor) ones [14]. Initially Gabor filters technique used to haulout the consistency features from MRI brain images and thenexecutes categorization among usual and unusual images withthe use of Support Vector Machine (SVM). Subsequently newhistogram contrast technique is utilized in right and left halvesof brain grounded on Bhattacharya coefficient and thistechnique discovers bouncing box as Region of Interest (ROI).Then texture characters are hauled out with Gabor filters fromthe ROI. Lastly, the categorization of images was done usingartificial neural networks. A contrast of both the anticipatedways is given at closing stages.

Gopal et al. designed an intellectual method to identify braintumor in the course of MRI using image dispensation, fuzzy-cmeans grouping algorithms along with intellectual optimizationapparatus, similar to Genetic Algorithm (GA), and ParticleSwarm Optimization (PSO) [15]. The tumor detection processis performed in two stages such as pre-processing and thesecond stage consists of segmentation and classification inprocesses. Wu et al. proposed the k-means grouping methodgrounded on color segmentation technique that is utilized totrail tumor substances in Magnetic Resonance (MR) brainimages [16]. This color-based segmentation procedure with k-means is the key idea, in which it renovate into a providedgray-level MR image to a color space image and then dividethe spot of tumor substances from additional items of an MRimage with the use of k-means grouping and histogram-grouping. The outcomes of this anticipated technique revealsthat these techniques can effectively attain segmentation forMR brain images to assist pathologists differentiate preciselyregion and lesion size.

Kharrat et al. introduced a well-organized revealing of braintumor from cerebral MRI descriptions [17]. In this proposedmethod comprises of following vital three steps such assegmentation, enhancement, and classification. In order toimprove the image quality and risk limitation of differentsections combination in the partitioning stage an improvementprocedure is functional. In this proposed work adoptsnumerical morphology to amplify the dissimilarity in MRIimages. Then wavelet transform is applied in the divisionprocedure to decay MRI images. Finally the k-means algorithmis employed to haul out the distrustful tumors or regions. Theproposed method’s performance results showed the feasibilityof brain image process.

Proposed MethodologyIn MRI brain medical image investigation research methodsusually comprises of numerous parts, which use diverseprocesses in a sequence. In this proposed work at first,advanced median filter pre-processing method is used toprepare the image for optimal results. Then these medical

Shenbagarajan/Ramalingam/Balasubramanian/Palanivel

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S192

image examination study methods consist of following processsuch as feature extraction, segmentation, feature selection andclassification. Region based Active Contour Method (ACM)segmentation method is proposed to segmenting the MRI brainimage grounded on visual demonstrations, which areassociated to the grey-levels. In features are arithmeticinformation and measurements of the image that can be hauledout from a preferred division of the image. The Figure 1 showsthe block diagram of the proposed methodology.

Figure 1. Flow of the proposed methodology.

In this work most significant shape and texture features arehauled out from the MRI image using hybrid wavelet methodand Sobel and Canny method. Finally the classification processis used to categorization of the input MRI brain imagegrounded on their characteristics which is an essential phasefor evaluating the tumors. For classification MB-MLMMmethod is proposed. The final evaluation results show theefficacy of proposed methods for tumour diagnosis process inmedical image analysis.

Pre-processing step for MRI brain image processThe proposed MRI brain image processing method using pre-processing step, which plays a vital role in enhancing the inputMRI brain images, which is also a crucial and essential stepthat tunes the MRI image for further processes such assegmentation, feature extraction, feature selection andclassification.

Advanced median filter for noise removalA novel denoising algorithm namely known as advancedmedian filter is presented in this work to remove noises fromthe input MRI images. The proposed advanced median filteringmethod is customized description of hybrid median filteringtechnique, which is discussed in detail as follows.

Hybrid median filter: Assume the input representation aspixels in figure of 3 × 3 matrix, the 45˚ neighbours medianvalues shapes an ‘x’ and the neighbours of the representationpixels forming a ‘+’ are evaluated with the median value andthe central pixel of that set is then kept as the novel pixelstandards. The three pace position process does not inflict asolemn computational consequence as in the case of medianstrain. Each of the position process is greatly smaller number

of standards than utilized in a square region of the similar size.In the hybrid technique, each one of the two clusters holds only5 pixels, and the concluding assessment engrosses only threestandards. Even with the supplementary manipulation and logicof values, the hybrid median filter technique is quicker than theconservative median. This median strain results the propensityof median and shortened median filters to rub out lines whichare slender than the half width of the neighbour and toencircling corners.

Hybrid median filter preserves edges better than other medianfilter since it processes a three-step ranking operation, in whichinformation from diverse spatial instructions are categorizedindependently. There are three median values computed byusing hybrid median filtering method such as HV signifies themedian of horizontal h, vertical v pixels, D represents themedian of oblique d pixels, the strained value is the median ofthe two median standards, the central pixel C, median of thevertical and horizontal, diagonal and central is represents asHV, D, C.� • �• � •� • �

• �� •�� � ��• �� •• • •• � •• • • − � �� ��� ���� ��� �� � (1)

Proposed advanced median filter: It is also known aswindowed filter. Advanced median filter has better edgepreserving characteristics compare than the other median filterversions. In this proposed pre-processing technique, advancedmedian filters are used to remove impulse noises whilepreserving edges of the MRI brain images. The fundamentalsuggestion at the rear filter is for neighbourhood pixels of theMRI image apply median technique numerous periods,unstable window shape and then obtain the median to attainmedian values. This anticipated filter is the customizeddescription of the hybrid median filter explicated above. Itfunctions on the sub windows comparable to hybrid medianfilter. The maximum value of the neighbouring pixels formingan and the median value of the neighbours forming a arecompared with the central pixel and the median value of thatset is then saved as the new pixel value.

Algorithm for advanced median filter

Step 1. Find the median HV of the horizontal and verticalpixels marked as HV and the central pixel C in the 3 × 3 matrix

Step 2. Find the maximum Dmax of the diagonal pixels markedas D and the central pixel C in the 3 × 3 matrix

Step 3. Finally compute new median value Mnew

Mnew=median (HV, Dmax, C)

Step 4. Filter median out value Ii, i =Mnew; here are indices ofthe spatial locations.

From the process of advanced median filter obtained a noisefree pixel, these are used in the further MRI brain imageprocess.

MRI brain image analysis for tumour diagnosis using hybrid MB-MLM pattern classification technique

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S193

Image surface extractionHere this work present parallel methods for surface featureextraction represented in the MRI brain image. The proposedclustering based methods are more memory-efficient and fastprocessing algorithm, iterative simplification puts moreemphasis on high surface quality. The proposed method forimage surface feature extraction is discussed in detail asfollows:

Parallelized method for surface feature extraction in MRIbrain image: In this proposed work used parallelized featureextraction and clustering method. The first phase of theparallelized method is feature extraction method, which isusing Hidden Markov model (HMM), which is described asfollows,

Hidden markov model (HMM): The method for surfacefeature extraction using HMMs can be summarized by thefollowing algorithm. Consider a given a set of N surfaces ofMRI images to be clustered; the algorithm performs thefollowing steps [18]:

1. Train one HMM profile λi for surfaces si.

2. Calculate the distance matrix D={D (si; si)} representing asimilarity measure between surface features; this is typicallyobtained from the forward probability that gained by the HMMprofile.

3. The Log-Likelihood (LL) of each model, given surfaces thatcomputed in step 2 used to build an LL matrix. This matrix hasN × N numeric value. Where each row is a feature for onesurface. That includes the similarity measure from all othersurfaces of image.

4. At this stage where parallelized by distribute all the surfaceareas of images to the available process.

K-mean clustering: This proposed k-means clustering method[19] algorithm uses an iterative refinement technique,summarized in the following algorithm.

1. Place C points into the space represented by the surfacefeature vectors that are being clustered. These points representinitial group centroids C’

2. Assign each sequence to the group that has the closestcentroid C’.

3. Update and recalculate the positions of the C centroids.

4. Repeat steps 2 and 3 until the centroids no longer move ormax iteration is reached.

The sequential procedures/algorithms expend most of its timecomputing novel centroids (step 3) and computing thedistances among n image pixels and C centroids (step 2).Execution time can be cut down by parallelizing these twosteps. The distance computation process can be performedasynchronously and equivalent for every surfacecharacteristics. Thus this proposed methodology use aparallelized clustering approach, where clusters are building bycollecting neighbouring pixels of the image while regarding

extracted surfaces using HMM. This method creates a set ofclusters, each of which is replaced by a representative surfacefeature vector, typically its centroid C. Thus the MRI brainimage is enhanced based on the surface feature extractionprocess proposed in this work. In order increase thesegmentation accuracy for classification process, this work alsoproposed another pre-processing IIH method, which isdiscussed in detail as follows.

Intensity in homogeneity (IIH) correction in MRIbrain imagesIntensity in homogeneity modification is frequently anessential pre-processing phase facilitating improved imagesegmentation procedure. In difference, correct segmentationcreates intensity in homogeneity modification quiteunimportant. In this MRI image processing method, thepresence of Intensity in homogeneities can significantly reducethe MRI image segmentation accuracy, if it also affects theclassification accuracy. The IIH or intensity non-uniformityusually refers to the slow, non-anatomic intensity variations ofthe same pixels over the MRI image domain. It can be occurdue to imaging instrumentation such as radio-frequency non-uniformity, static field in-homogeneity, etc. or the patientmovement [20]. In accumulation to intensity in homogeneity,the MRI image configuration replica ought to include high-frequency noise. On the other hand, as elongated the Signal-to-Noise Ratio (SNR) is not low, noise can be estimated by aGaussian distribution [17]. This estimation is suitable forimage regions equivalent to tissues. Let consider the inhomogeneity-free image , intensity in homogeneity field y’,and noise ξ interact. There two sources of noise were describedsuch as biological noise, this is a type of noises correspond tothe within tissue in homogeneity, and second one is a scannernoise that arises from MR device imperfections [16,18,19].However, usually only one of these sources is modelled. Themost general replica of MRI image configuration presumes thatthe noise, estimated by Gaussian probability distribution,happens from the scanner and is consequently self-determiningof the intensity in homogeneity field y’. In accordance to thisreplica, the obtained image y is gained as follows, Let y denotethe measured intensity and y’ the true intensity. Then the mostpopular model denoted as

Y= y’+ξ → (2)

To simplify the computation, one often ignores the noise andtakes the logarithmic transform of intensity

xi=log yi=log y’+log i=xi’+βi → (3)

Where y’ is the intensity at pixel i (i=1, …n). Here, only pixelswith low intensities should be taken for avoid numericalproblems when it occurring in computation process, which areusually excluded from computation.

In the log-transformed model, noise is tranquil implicit to beGaussian, which is practical suitable but conflicting with theinitial replica from Equation 2 that presumes the noise isGaussian in the unique non-logarithmic field. This discrepancy

Shenbagarajan/Ramalingam/Balasubramanian/Palanivel

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S194

was actually not measured pertinent enough. Nonetheless, inthe majority of in homogeneity modification techniques, thenoise is gripped by uncomplicated filtering, smooth replicafitting, or some outline of regularization and is consequentlymeasured quite immaterial. In segmentation based intensity inhomogeneity correction method, both of these two intensity inhomogeneity modification technique and segmentationmethods are amalgamated, so that they profit from each oneother, concurrently acquiescent improved in homogeneitycorrection and segmentation. This intensity in homogeneitymodification techniques are furthermore categorized accordingto the image partition process exploited. The statisticalmethods such as segmentation method of this proposed workmay assume that the IIH follows the Gaussian distribution, toestimate the IIH map; frequently the Bayes’ rule has beenemployed when the IIH is modelled by a distribution likeGaussian. Let β be a random vector (1, …, n) with probabilitydensity p (β). To estimate β, one can maximize the conditionalprobability of β given x (the log-transform of y) as follows:� = ���� �   � � (4)This is called the maximum a posterior estimate and, by theBayes rule, is equivalent to� = ���� �   � � �   � (5)Wells et al. [21] used the Gaussian distribution to model theentire log-transformed bias field and the observed intensity atpixel i:�   � = ��� � (6)�   (�� ��,   ��) = ����(��− �   �� − �� (7)Where Γi is the tissue class at pixel i with mean value μ (Γi),and

���   � = 2� −�2 �� − 12 exp − 12����−1�   (8)With Ψy as the covariance matrix. By assuming the statisticalindependence of pixel intensities and from Equation 7 andequation can derive

�   �|� =∏�∑�� �   (�� ��,   ��)   �   �� (9)When the MRI image is not polluted by IIH, the above methodis simply the tissue classification using a mixture Gaussianmodel. Hence, this method essentially interleaves the IIHcorrection with a Gaussian classifier. Frequently finiteGaussian mixture models and finite mixture are utilized andcustomized to include intensity in homogeneity for increasingthe segmentation accuracy of this proposed MRI imageprocessing method, in order increasing the accuracy of thepartitioning, final MRI brain representation categorization also

increased and which will provide an efficient tumour diagnosisresults.

Segmentation of brain MRI using region based activecontour model (ACM)Segmentation by means of Active Contour Model ACM(Snakes) was introduced [22]. The fundamental concept of theactive contours, or deformable models, for image segmentationis relatively uncomplicated. The user indicates an initial guessfor the contour, which is subsequently moved by image drivenforces to the boundaries of the preferred objects. Mostcommonly applied scheme of medical image segmentation islevel set scheme or ACM. This entrenched scheme ofsegmentation is applied in order to obtain promising results ofsegmentation with accuracy and provide closed and smoothcontour of object boundary, which assist in extracting thefeatures like texture and shape of the image data. This schemeof segmentation is categorized into two broad classes. Firstclass is edge-based scheme of segmentation uses gradient toshow contour evolution. Because of the use of gradient,segmentation is susceptible to noise and weak edges. At thesame time, second class is region based scheme make use ofthe region descriptor like intensity, texture, shape etc. toidentify RoI, to show curve evolution. Region based schemesegmentation provide enhance results even in the occurrence ofnoise and edge leakage of object boundary since this is lesssusceptible to initial contour location. Here, majorlyconcentrated on region-based ACM with level set formulationfor image segmentation.

A region-based active contour model: ACM are categorizedinto two, namely: edge-based and region based models. Theedge-based model makes use of the incline of therepresentation to terminate the contour during evolution for thepurpose of boundary detection of the foreground object. Aregion-based ACM makes use of statistical details of regionsboth inside and outside the curve for contour evolution, forinstance, the Chan-Vese (C-V) model.

This model grounded on the supposition that the pixel regionsof the representation are statistically homogenous. It worksextremely well even with noisy, blur images and images thathave multiple holes, disconnected areas etc. In MRI brainimage analysis, the region based ACM since takes globalproperties of images like contour lengths and MRI image pixelregions as alongside local properties like gradients. The energyminimizing function can be given as follows:

ln�(�� �) =∫∫�0 ��(�,�)�� (10)Where IS (x, Y) is the intensity at the pixel location (x, y) in theimage, and the integral gives the total area enclosed by thecurve p. As is evident, the region-based information visuallyimproved the segmentation quality compared to the one usingonly gradient information.

MRI brain image analysis for tumour diagnosis using hybrid MB-MLM pattern classification technique

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S195

Feature extraction method for brain MRI imageanalysisAfter segmentation, the following features are extracted fromthe brain MRI images such as texture and shape. The feature ofthese images can be stored for further analysis. Featureextraction is a one of the important process in MRI medicalimage based tumor analysis. The feature extraction process isused for creating a representation, or transformation from theoriginal image. In the MRI medical images have the primitivefeatures like texture, shape, edge, darkness, etc. From thesefeatures, the most promising features like texture and shape/edge are extracted for classification accuracy in the proposedmethodology. The variation of each pixel with respect to itsneighbouring pixels of the images defined as a texture. Hencethese textural details of MRI image regions can be comparedwith a texture template and extracted using hybrid wavelettransformation method. The shape/edge is simply a large andfrequently changed. The shape features of the MRI images areextracted using Sobel and Canny edge detection method. Bothof these two types of feature descriptors are mainly used mostoften during feature extraction process in MRI brain imagebased tumour analysis process. Feature extraction is a one ofthe important process in MRI medical image based tumouranalysis. The feature extraction process is used for creating arepresentation, or transformation from the original image. Inthe MRI medical images have the primitive features liketexture, shape, edge, darkness, etc. From these features, themost promising features like texture and shape/edge areextracted for classification accuracy in the proposedmethodology. The variation of each pixel with respect to itsneighbouring pixels of the images defined as a texture. Hencethese textural details of MRI image regions can be comparedwith a texture template and extracted using hybrid wavelettransformation method. The shape/edge is simply a large andfrequently changed. The shape features of the MRI images areextracted using Sobel and Canny edge detection method. Thesetwo types of feature descriptors are mainly used most oftenduring feature extraction process in MRI brain image basedtumour analysis process. The texture and shape featureextraction technique are described in detail as follows.

Texture feature extraction (Hybrid wavelet transform): Inbrain MRI image analysis, an initial assumption ofdistinguishing image consistency is that all the texture data isenclosed in the gray-level value matrices of MRI image. Henceall these textural features are extracted from these gray-levelvalue matrices. The energy measures are mostly transmitting todetailed textural distinctiveness of the MRI image.Supplementary measures distinguish the difficulty andenvironment of gray level change which happens in the MRIrepresentation. Although these characteristics enclose dataabout the textural distinctiveness of the image, it is firm torecognize which exact textural feature is signifies by everyfeatures. Texture feature extraction is a most important stage inMRI image analysis in which texture features of each image isseparately extracted from MRI by wavelets which considersbetter method to extract most emphasizing pixels present in

images to improve results. To decompose data into differentfrequency components wavelets mathematical functions areused and then each element is studied having resolutionmatched to its degree. For analysis of complex datasets waveletis a new powerful mathematical tool. Fourier transform failsdue to the redundant set of features and only give frequencyinformation that is not spatially localized but wavelet providestime frequency information which is localized in space andperfect tool for pattern recognition tasks.

Hybrid discrete wavelet-cosine modulated wavelettransform: The recommended scheme makes use of theDiscrete Wavelet Transform (DWT) coefficients ascharacteristic vector. The wavelet is an influential arithmeticaltool for feature mining, and has been utilized to haul out thewavelet coefficient from MR representation. Wavelets arerestricted rooted functions, which are shifted and scaledaccounts of certain unchanging mother wavelets. The mostimportant benefit of wavelets is that they offer localizedfrequency data about a function of a signal, which is chieflyhelpful for categorization. An appraisal of essential basic ofwavelet decomposition is initiated as given below: Thecontinuous wavelet transform of an image square-integrablefunction, is distinct as:

����(�,�)∫−∞∞ �(�) *��, �(�)�� (11)Where, φu,v (t)=1/|√μ| and the wavelet φu,v is calculated fromthe mother wavelet by dilation and conversion, wavelet, a →dilation factor and b → translation factor (both being realpositive numbers). Underneath certain gentle assumptions, themother wavelet gratifies the restraint of containing zero mean.The Equation 6 can be discretized by preventive u and v to adiscrete lattice (u=2 v; u ε R+: v ε R) to provide the discretewavelet transform. The Discrete Wavelet Transform (DWT) isa linear conversion that functions on a data vector whoselength → integer power of two, converting it into anumerically diverse vector of the similar length. It is a meansthat divides data into diverse frequency elements, and theninvestigates every part with declaration coordinated to its level.DWT can be uttered as.

����(�) = ��, � = ∑(�(�)ℎ * �(� − 2��))��, � = ∑(�(�)� * �(� − 2��)) (12)The feature components in signal y (n) and match up to thewavelet function, where di, j → the approximation componentsin the representation signal. The functions h (n) and g (n) theequation symbolize the coefficients of the high-pass and low-pass filters, correspondingly, at the same time as factors i and j→ wavelet scale and translation factors respectively. The mostimportant characteristic of DWT is multiscale illustration offunction. With the use of wavelets, provided function can beinvestigated at diverse stages of resolution. In texturecharacteristic mining with the use of DWT, the segmentedunique representation is procedure along the x and y directionby h (n) and g (n) filters in which is the row demonstration of

Shenbagarajan/Ramalingam/Balasubramanian/Palanivel

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S196

the unique image. As a consequence of this change shows 4sub band (LL, LH, HH, and HL) representations at every scale.The sub bands of the representations are utilized for cosineadapted WT computation at the subsequent level. Thus thediverse sub bands are produced with the use of DWT.

Subsequent to that the cosine-modulated wavelettransformation [23], is relating a number of sub-bands, whichare produced using DWT in MRI representation. Themagnitudes of wavelet coefficients in certain sub-bands aresuperior for images with a burly textural contented at theorientation and frequency represented by that sub-band.Consequently, the textural characteristics of MRI image can besymbolized by feature vectors that enclose the averagecoefficient magnitude, recognized as averaged energy function.The energy distribution has significant biased possessions forimages and as such can be utilized as characteristics for MRIimage categorization. This work makes use of energy signaturefor mining of texture characters as it imitates the allocation ofenergy along the incidence axis over orientation and scale. Thebiased possessions of the energy distribution in sub-bandsconsequence in texture character that have been experimentalto acquiesce good description of textures for MRI imagecategorization. The energy characteristic of the image isprovided by,�� = 1��� ∑� = 1� ∑� = 1� �(�, �) ���� = 1, 2, ...� (13)Where x → wavelet decomposed image for any sub-band ofdimension M × N. For K-level decomposition of therepresentation, the dimension of the feature vector is Q=(3 × K+1) as diverse characters have diverse variety of probablevalues and the whole characteristic may not have the similarthe same level of implication for the reason that afterdecomposition of image, the sub-bands with superiorresolution communicates to noise and may not precious forcategorization. So all these feature standards are regularized inthe variety of 0 and 1 by the utmost value in the characteristicspace prior to categorization of these MRI representation.

Shape feature extraction: For shape feature hauling out inbrain MRI image analysis, edge finding is a most importantprocess. Shape feature mining in image investigationnecessitates the hauled out edges to be associated in order toimitate the limitations of substance available in therepresentation. Then this shape feature extraction based on theedges in MRI brain image process is very useful furthercomputational process. The diverse gradient operatives utilizedfor edge mining, here this proposed shape feature extractionmethod used for Sobel and Canny method.

Edge detection methods-Sobel and Canny: One of the mostpopular features for image classification is edges. Edges referto boundaries of an object surface where the intensities changesharply. This anticipated edge detection technique used Cannyand Sobel beside with the gradient operatives to haul out theshape characteristics in form of linked boundaries. The Sobeloperative is one of the most generally utilized edge detectors.

The amount of the gradient of the Sobel operator is computedby,�� = �����2+ �����2 (14)where the partial derivatives are computed by����� = (�2+ ��3+ �4)− (�0+ ��7+ �6) (15)����� = (�0+ ��1+ �2)− (�6+ ��5+ �4) (16)with the constant c=2. The other gradient operators, Gradx andGrady can be implemented using following convolution masks:

����� = 1 2 10 0 0−1 −2 −1 ,����� =1 0 −12 0 −21 0 −1 (17)

that this operator places an emphasis on image pixels that arecloser to the center of the mask.

Canny edge detector make use of a filter grounded on the firstimitative of a Gaussian, since it is vulnerable to noise availableon raw unrefined image information, so to start with, theuncooked image is convolved with a Gaussian filter. Cannyoperator is nonentity but incline of Gaussian filteredrepresentation. The convolution masks of canny operators aregiven as follows:����� = −2 −22 2 ,����� = 2 2−2 −2 (18)The edge detection using canny operator choose thresholds toget the edges of the images, the chosen higher threshold alwaysis three times of lower threshold. Handle the MRI image pixelsbetween the lower threshold area and higher threshold area.This will give a thin line in the output MRI image, which ismore efficient for final classification process in MRI brainimage analysis.

Feature selection using fast independent componentanalysisAim of this proposed feature selection based on ICA isselection of the best features for classification purposes.Independent component analysis is an unsupervised learningmethod based on high order information. The basic ICAsystem model of is given as

X=AS or BX=S → (19)

Where X be the extracted feature vectors, S are theindependent source features, A is the mixing matrix and B isthe un-mixing matrix. Unmixing B model is more correct thanthe mixing matrix A, then only B is used to project the featurevectors representing input image onto independentcomponents, so B must be calculated from the given input.Consider n is the number of source features of the MRI image,m is the number of extracted features of the MRI image, andthe hauled out characteristics are linear mixtures of the sourcefeatures. Then, the extracted features X=(x1, …, xm) and the

MRI brain image analysis for tumour diagnosis using hybrid MB-MLM pattern classification technique

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S197

source features are represented in the following ICA unmixingmodel,�� *� =�� *� .�� *� (20)Where W=(w1, …, wm) is the unmixing matrix, or weightmatrix, and wi=wi1, …, win: i=1, …, n.

In the ICA transformation model, the features si are assumed tobe statistically independent and no more than one feature isGaussian distributed. If information on a feature si in S doesnot give any information on the other features, then si isconsidered independent of these features. According to theassumption of non-gaussianity, the desired independentfeatures si contains the least Gaussian features.

The main work of ICA is to estimate the weight matrix W anda measure of non-gaussianity is the key. The classical measureof non-gaussianity is kurtosis, which is the fourth orderstatistics and has zero value for Gaussian distribution.However, kurtosis is sensitive to outliers. Since a Gaussianvariable has the largest entropy among all random variables ofequal variance, the reverse entropy namely known asnegentropy, which can be utilized as a gauge of non-gaussianity. The negentropy is defined as Equation 21.

J (X) = H (pxgs)-H (px)

Where H (pxgs) is the entropy of a Gaussian random variablewith the same covariance matrix as X and H (px) is thedifferential entropy. Since the negentropy is difficult tocompute, an approximation is used instead.�(�) ≈ �[�(�)]− �[�(���)] 2 (22)where (G (X)) is a non-quadratic function. In order tomaximize the objective function shown in Equation 22, theproposed Fast ICA consists of two processes, the one unitprocess and the de-correlation process.

Wi+=E {Xg (WT

iX)}-E {g (WTi)} Wi, g (.) = tan (.) → (23)

While the one unit process estimates individual weight featurevectors, the decorrelation process keeps different weightvectors from converging to the same maximal. Given pdecorrelated weight feature vectors and the p+1 th weightfeature vector estimated by the one unit process, the p+1 thweight feature vector is decorrelated from p weight featurevectors by Equations 24 and 25,��+ 1+ =��+ 1− ∑� = 1� ��+ 1� ���� (24)��+ 1 = ��+ 1+��+ 1+ (25)ICA-based feature selection: Main aim of this work is theselection of the best feature subset for classification purposesgiven a high-dimensional extracted feature vector. When ICAis used to reduce dimensionality of MRI images, the number ofextracted features X is the original dimensionality. Theextracted feature X is the vector values of all pixels in the MRI

images. The source features S exist in a lesser dimensionalroom consequent to the available equipment in the MRI image,and each independent component i.e. independent features si isdistinctive for one material. Since the number of presentmaterials may be unknown, this process evades changing theunique MRI descriptions to the source Features S which isassociated to the amount of materials. Instead, here this workrandomly assume but evaluate the weight matrix W to observehow each original features contributes to the ICAtransformation described above. Suppose the number ofmaterials in an m feature MRI image is n, obtain thecorresponding weight matrix Wn-m using Fast ICA. In the ICAunmixing procedure, to estimate the source S (pure materials)from the extraction X (pixels in the MRI image) with theweight matrix W, after that this work can estimate theimportance of each feature vector for all materials bycalculating the average absolute weight coefficient , which isshown in Equation 26,�� = 1/� ∑� = 1� ��� , � = 1....� (26)By sorting the average absolute weight coefficients for allfeature vectors obtain a feature weight sequence[��, .........,��, ...............,��] (27)Where. In this series, the characteristics with superior averageabsolute weight coefficients that are also recognized as finestfeatures, which supplies additional ICA transformation thanother characteristics do. That means these features containmore feature vector information than other features. Therefore,this best feature selects the vectors with the highest averageabsolute weight coefficients for the purpose of featuresselection. Because typical MRI image analysis methods treateach image features as an independent variable, and call theseselected features as independent features.

Classification process using modified LM withmodified BAT algorithmThe MRI image classification has remained largely new workwith different approaches, one of them is, Modified LevenbergMarquardt (MLM). The MLM is a one of the most efficientclassifier to increase the classification performance in the MRIbrain image based tumor diagnosis process and that classifiesand to assign labels to extracted features from the featureextraction and selection steps through MRI images. This workis an attempt to use MLM to automatically classify brain MRIimages under two categories, either tumorous or non-tumorouswith use of Modified Bat (MB) algorithm optimizationmethod. Here this proposed MB-MLM classifier using featurevector gained from the MRI images and classify the MRI brainimages. The proposed classifier methodology is described indetail as in follows.

Bat algorithm: Generally BA is an extremely influentialalgorithm and creates healthy explanations on lowerdimensional functions but its presentation diminishes as themeasurement of difficulty increases [24]. Bat Algorithm (BA)

Shenbagarajan/Ramalingam/Balasubramanian/Palanivel

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S198

is a heuristic algorithm anticipated by Yang. It is grounded onthe echolocation capability of micro bats directing them ontheir foraging behaviour [25]. Here this proposed work used tooptimize the search direction space with Modified Batalgorithm (MB), the result of the optimized search directionused in the MLM for efficient classification results in MRIbrain images analysis process. Here the following sectiondescribe about the basic structure of Bat optimizationalgorithm.

Echolocation capability of bats: Bat utilizes a kind of sonaridentified as echolocation to distinguish diverse sorts ofinsects, communicate, intellect distance to their victim andshift without striking to any obstruction still in whole darkness.The echolocation features are put on a pedestal within thestructure of the subsequent rules by assistance suchcharacteristics of bats:

1. Echolocation process is used by all bats for sense distance,and the bats also know the difference between food andprey.

2. They are flying arbitrarily with velocity v at location with afrequency f, unreliable wavelength and volume L0 to lookfor prey. They can mechanically regulate the wavelength(or frequency) of their produced pulses and regulate thevelocity of pulse emission r ε (0, 1) based on thepropinquity of their object.

3. Even though the loudness can differ in numerous ways, andpresume that the loudness differs from a bigger (positive) toa least constant value Lmin

The structure of BAT algorithm: (a) Initialization of batpopulation: The search space is assumed as a region thatcontains many prey sources on it. The algorithm tends to findthe high or optimum quality food in the search space. Becauselocations of food sources are not known, initial inhabitants isarbitrarily produced from real-valued vectors withmeasurement D and number N, by enchanting into report upperand lower boundaries. Then, quality of food sources locatedwithin the population is evaluated.

Pij=plow+φ (phigh-plow) → (28)

where i=1, 2, …, N, j=1, 2, …, D, Phigh and are lower andupper boundaries for measurement correspondingly ϕ is aarbitrarily produced value assortment from 0 to 1.

(b) Generation of velocity, frequency, and novel solutions:Calculated fitness values of all bats authority their actions. Batstake off with velocity v which is exaggerated by an arbitrarilypredefined frequency f. Lastly they situate their novel positionxi in the investigated space.

Fi=flow+β (fhigh-flow) → (29)

Vt=vt+1 + (pt-p*) fi → (30)

Pti=pt-1+vt → (31)

Where fi is a frequency value belonging to the ith bat, flow arelower and higher frequency standards, correspondingly, βindicates a randomly generated value, p• is the obtained global

best result after contrast of all results amongst bats so far and vt

implies the velocity of the bat at tth time step.

(c) Local search capability of the algorithm: Theimprovement of local search ability of the algorithm, Yang hascreated a structure in sort that the bat can advance the resultnear the attained one [26].���� = ����+ ∈ �� (32)Where pold a higher excellence result is selected by a fewmechanisms (e.g. roulette wheel), ̂Lt is average loudness rate ofall bats at tth instant step and ε is a randomly generated valueranging from −1 to 1.

(d) Loudness and pulse emission rate: The pulse emission rate rand loudness L are rationalized as a bat gets nearer to itsobject, namely its prey. Loudness L is decreased while pulseemission rate r is increased with admiration to Equations 33and 34, respectively

Lt+1=α Lt → (33)

rt+1=r0 (1-eγt) → (34)

where α and γ are constraints, ro is the initial pulse emissionrate value of the ith bat.

Proposed modified BAT with modified LM algorithm forclassification: Considering general Levenberg-Marquardtwhich is trained in Artificial Neural Network (ANN) and usesapproximated Hessian matrix in the subsequent Newton-likeinform:

Xk+1=xk - [Jk (α) + τI]-1Jk (α)Tek → (35)

Where, xk+1 and xk points of Newton method at kth iteration, τis the learning rate of set-up, α is a vector, J is the Jacobianmatrix, and e is the error vector. There is utilized incline can becalculated as, gk=j (α)T e and the Levenberg-Marquardtalgorithm was intended to draw near second-order trainingspeed devoid of having to calculate the Hessian matrix. Whenthe presentation purpose has the outline of a sum of squaresthen the Hessian matrix can be estimated as

H=J (α)T J (α) → (36)

In Newton’s technique, hessian approximation process step isnot well conditioned. The vector error analysis definitelyestablishing that the Cholesky stable algorithm, that assumesthe Hessian matrix is sufficiently positive definite in LM. Themain drawback of the previous work is when the value of n islarge, here n is a parameter of vector; it is expensive both incomputational effort and memory to compute the errorreduction process. When H is sufficiently positive definite it isalso unnecessary. In order to overcome aforementionedproblem, the Hessian Matrix is modified to compute a searchdirection based on modifying the Cholesky factorization. Let �denote the Cholesky factor generated by this algorithm, then αis chosen such that �− �,   � ≤ � where α2=max {γ, ξ/n, α},γ and ξ are the largest in modulus of the diagonal and off-diagonal elements of H respectively, and as before is somesmall positive scalar.It can be shown that � ≡ �+ �,�is a

MRI brain image analysis for tumour diagnosis using hybrid MB-MLM pattern classification technique

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

S199

diagonal matrix; ||E|| is bounded and �has a bounded conditionnumber. The algorithm requires almost the same computationaleffort as the Cholesky algorithm.Let �� denote the Choleskyfactor of��. It follows a suitable sequence of sufficient descentsearch directions in network, {pk} are obtained by solving(�����)�� = − ��, and this obtained search direction isoptimized using with proposed modified BAT procedure foroptimal classification results.

Search directions optimization with modified BAT: Fromthe detail description of the Bat algorithm in the above section,in this proposed work used modified version of the Batalgorithm for classification process. This new Modified Batalgorithm (MB) performed with MLM and provides optimizedsearch space for further MLM classification process.

The update processes of velocity and location in the Batalgorithm provides the best optimization result. The mainpurpose of this modification is to intensify the frequencyequation at the beginning of the optimization process and thenthe velocity toward the end of the optimization process in turn.The modified equation is used in this proposed work givenbelow.

vt = ω (vt-1) + (pt- p*) f → (37)

where ω → inertia weight factor that balances, search intensityof the ith result by calculating the amount of old velocity v.This alteration configuration was also exploited in MLM tooptimize the investigated direction pk at kth iteration forcategorization procedure.

From MLM, when H• is sufficiently positive definite then inthe neighbourhood of the solution �� = �� and Newton stepsare taken, an optimized search direction pk as defined avoidssaddle points.

For the Newton method, this approximation can be employedfor this modified Hessian and subsequently to solve,

J (α)T J (αk) pk=-J (αk)T F (αk) → (38)

Where F is the features vector-valued function, then tocalculate the optimized search direction pk and then let αk+1=αk+pk. More willingly than approximating the Hessian asin Equation 38, the outline term in Equation 37 can be given asby τI where τ ≥ 0. Following this the Hessian is approximatedas

where J → Jacobian matrix that comprises first derivatives ofthe system faults with esteem to the biases and weights, and eis a vector of set of connections faults. The Jacobian matrix canbe calculated during a typical back propagation method that ismuch lesser multifaceted than calculating the Hessian matrix.

Now to find descent search direction, p, and the followingequation is solved:

[J (α)T J (α) + τI] P = -J (α)T F (α) → (40)

After that the performance index of the back propagationneural network defined by,

F (w) = gi (α) gi (α)T → (41)

Where w=(w1, w2, …, wN) consists of all weights of thenetwork, gi (α) is the gradient error vector comprising the errorfor all the training examples.

When training with the LM method, the updating of weights∆w can be obtained as follows:

∆w = -[τI+∑J (α)T J (α)]-1 → (42)

Where J→ Jacobian matrix, I is the identity matrix τ is thelearning rate which is to be updated using α depending on theoutcome.

Based on LM classification algorithm, modernizes the ANNweights as given below:

∆w = -[τI+∑J (α)T J (α)]-1 → (43)

Where J → Jacobian matrix of the error gradient gi (α)assessed in vector α, and I → identity matrix. The errorgradient gi (α) → error of the network. The parameter τ isdecreased r increased at every step. If the error gradient isdecreased, then is alienated by a factor α and it is multiply by αin supplementary case, while updating the network weights.

It computes the error vectors, the set-up output, and theJacobian matrix for every pattern. Then, it calculates ∆w usingEquation 43 and recalculates the fault with w+ ∆w as set-upweights. If the fault has diminished, τ is alienated by α, thenovel weights are upholded, and the procedure initiates again;or else, is multiplied by α, ∆w is intended with a novel value,and it categorize again. The procedure is repetitive until thefault decreases. When this happens, the current iteration ends.

Algorithm for MB-MLM:

Begin

Initialize ANN Weights;

While not stop criterion

Calculates the gradient error gi (α) each i, jth element;

Solve search space optimization, to obtain the efficient optimalresult using Equation 36.

Recalculate the gradient error gi (α)

Using w+ ∆w as the trial w, and evaluate

J1=∑ J (w + ∆W)T gi (α) (w + ∆W);

Calculates j (α) for each element;

Repeat

Calculates ∆W;

J2=∑ gi (α)

if (J1<=J2) then

τ=t × a;

End if;

Until (J2<J1)

Shenbagarajan/Ramalingam/Balasubramanian/Palanivel

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

200

τ=τ/a;

w=w + ∆W;

End while;

The outcome of MB-MLM classifier decides whether an inputimage dataset is tumorous or non-tumorous. The proposedmethodologies performance is evaluated based on thefollowing performance parameters such as sensitivity,specificity and classification accuracy, which is described indetail in following section. Similarly some of the workproposed in the recent work [27] based on the featureextraction and feature selection to enhance classificationaccuracy.

Experimentation ResultsFor tumour diagnosis in MRI brain images, the proposedmethodology used different medical image analysis methods.Mainly the proposed classification algorithm is given theoptimal result of the tumour diagnosis in MRI image withsegmentation process. The pre-processing, feature extractionand feature selection processes are enhance and reduce thedimensionality of the given input MRI images, which will belead to the optimal results in tumour diagnosis process in MRIbrain images. The following section shows that the proposedclassification method’s performance evaluation with use of theparameters such as sensitivity, specificity and classificationaccuracy rate. The proposed classification method results arealso compared with the existing methods with and withoutsegmentation method. The proposed methodology results areevaluated with use of DICOM database MRI brain Images,namely known as DCI and obtained by using MATLAB. Theproposed classification result proven that it will classify thegiven input MRI brain images as tumorous and non-tumorousin most efficient manner than existing methods. In future workpre-processing of the brain images, feature extraction andfeature selection and classification is performed to pathologicaltissues (tumor), normal tissues (White Matter (WM) and GrayMatter (GM)) and fluid (Cerebrospinal Fluid (CSF)), extractionof the relevant features from each segmented tissues andclassification of the tumor images.

Performance evaluationThe proposed method’s results are evaluated by using thefollowing performance parameters, which are defined as

follows:

Sensitivity: Recall (or) sensitivity (or) true positive rate is thepossibility of the real optimistic classes which are recognizedproperly.����������� = ��(��+ ��) * 100 (44)The Figure 2 shows that the comparison result of sensitivityusing the existing four different methods with proposedclassification algorithm. The graph results shows in Figure 2proven that the proposed classification method given highsensitivity rate than the existing methods.

Figure 2. The comparison result of sensitivity rate.

Figure 3. The comparison result of specificity rate.

Specificity also called the true negative rate measures theproportion of negatives that are correctly identified asnegatives.

Specifity = TP/(TP+FN) → (45)

Classification accuracy: Classification accuracy defined asthe percentage of correct classification of tumour and non-tumour classes from the MRI brain image. The accuracy rate isdenoted as follows,�������� = (��+ ��)(��+ ��+ ��+ ��) * 100 (46)The problems of segmentation and classification areintersecting each other because segmentation implies aclassification, while a classifier implicitly segments an image.In this proposed MRI image based brain tumour medical imageanalysis process the segmentation results are further used inclassification for efficient tumour diagnosis. The Figure 4demonstrates the influence of the segmentation process resultsin the proposed and existing classification approaches in theMRI brain image analysis. In order to increase theclassification accuracy for diagnosis process, here this workused IIH method for increase the segmentation accuracy. Inorder to high accuracy rate of the proposed region based ACMsegmentation method automatically increases the proposedMB-MLM classification algorithm’s accuracy rate. Figure 4illustrates the graph demonstration of the anticipatedclassification technique and existing methods results with and

MRI brain image analysis for tumour diagnosis using hybrid MB-MLM pattern classification technique

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

201

without ACM segmentation method proven that the final resultof this proposed methodology.

Figure 4. The comparison result of classification results of theproposed and existing methods with and without ACM.

ConclusionMRI Brain image analysis process for brain tumor diagnosisusing classification with segmentation, feature extraction andselection have been carried out in the past with limited success.The medical image analysis method suggested in this work forthe above work includes the steps, pre-processing,segmentation, shape and texture feature selection, featureextraction, and classification. In this work also introduced theIIH for increasing the segmentation accuracy. Here this workused extracted texture and shape features for classification. Themost significant features are selected by using Fast ICA, usedfor reduce the dimensionality of image for final classification.This maximizes the classification accuracy. Thus theanticipated technique executes enhanced than the obtainableworks.

References1. Chronic care networks for Alzheimer’s disease initiative

tools for early identification assessment and treatment forpeople with Alzheimer’s disease and dementia.Alzheimer’s Assoc Nat Chron Care Consort 2003.

2. Senthilkumaran N, Rajesh R. Brain image segmentationusing granular rough sets. Int J Arts Sci 2009; 3: 69-78.

3. Forghani N, Forouzanfar M, Forouzanfar E. MRI fuzzysegmentation of brain tissue using IFCM algorithm withparticle swarm optimization. Proc Int Symp Comp InformSci 2007; 1-4.

4. Maji P, Kundu MK, Chanda B. Second order fuzzy measureand weighted co-occurrence matrix for segmentation ofbrain MR images. Fundam Inform 2008; 88: 161-176.

5. Mostafa MG, Tolba MF, Gharib TF, Megeed MA. AGaussian multi-resolution algorithm for medical imagesegmentation. Proc IEEE Int Conf Intel Eng Sys 2003.

6. Diaz I, Boulanger P, Greiner R, Murtha A. A critical reviewof the effects of de-noising algorithms on MRI brain tumorsegmentation. Ann Int Conf IEEE Eng Med Biol Soc 2011;3934-3937.

7. Bauer S, Wiest R, Nolte LP, Reyes M. A survey of MRIbased medical image analysis for brain tumor studies. PhysMed Biol 2013; 58: 97.

8. Gonzalez RC, Woods RE. Digital image processing.Pearson Int (3rd edn.) 2002.

9. Ho S, Bullitt E, Gerig G. Level-set evolution with regioncompetition: automatic 3D segmentation of brain tumors.Proc Int Conf Pattern Recogn 2002; 1: 532-535.

10. Satheesh S, Prasad KVSR, Reddy KJ. Brain tumorextraction from T1-weighted MRI using co-clustering andlevel set methods. Int J Image Proc 2013; 7: 219.

11. Kaus MR, Warfield SK, Nabavi A, Black PM, Jolesz FA,Kikinis R. Automated segmentation of MR images of braintumors. J Radiol 2001; 218: 586-591.

12. Kao YH, Guo WY, Wu YT, Liu KC, Chai WY, Lin CY, YehTC. Hemodynamic segmentation of MR brain perfusionimages using independent component analysis,thresholding and Bayesian estimation. J Magn Reson Med2003; 49: 885-894.

13. Xuan X, Liao Q. Statistical structure analysis in MRI braintumor segmentation. IEEE Int Conf Imag Graph 2007;421-426.

14. Shah S, Chauhan NC. Classification of brain MRI imagesusing computational intelligent techniques. Int J CompAppl 2015; 124: 14.

15. Gopal NN, Karnan M. Diagnose brain tumor through MRIusing image processing clustering algorithms such as FuzzyC Means along with intelligent optimization techniques.IEEE Int Conf Comput Intel Comp Res 2010; 1-4.

16. Wu MN, Lin CC, Chang CC. Brain tumor detection usingcolor-based k-means clustering segmentation. Int Conf IntelInform Hid Multimed Sig Proc 2007; 2: 245-250.

17. Kharrat A, Benamrane N, Ben MM, Abid M. Detection ofbrain tumor in medical images. IEEE Int Conf Sig Circ Sys2009; 1-6.

18. Bicego M, Vittorio M, Mario ATF. Similarity-basedclustering of sequences using hidden Markov models. MacLearn Dat Mining Pattern recogn Spr 2003; 86-95.

19. Bicego M, Murino V, Figueiredo MAT. Similarity-basedclustering of sequences using hidden Markov models. MacLearn Dat Mining Pattern recogn Spr 2003; 86-95.

20. Othman F, Abdullah R, Addul Rashid N, Abdul Salam R.Parallel K-means clustering algorithm on DNA dataset.Parallel Distr Comp Appl Technol Spr 248-251.

21. Wells WM, Grimson WEL, Kikinis R, Jolesz FA. Adaptivesegmentation of MRI data. Teleimersion Res J PharmacBiol Chem Sci 1996; 15: 429-442.

22. Airouche M, Bentabet L, Zelmat M. Image segmentationusing active contour model and level set method applied todetect oil spills. Proc W Congr Eng 2009; 1: 1-3.

23. Mushrif MM, Dubey YK. Texture classification usingCosine modulated wavelets. Int Conf Sig Acqu Proc 2011.

24. Fister I, Dusan F, Xin SY. A hybrid bat algorithm. arXiv2013.

Shenbagarajan/Ramalingam/Balasubramanian/Palanivel

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

202

25. A human computer interfacing application. Int J Pharm BioSci 201; 2: 29-31.

26. Yang XS, Deb S. Cuckoo search via levy flights. W CongrNature Biol Insp Comput 2009; 210-214.

27. Rathi VGP, Palani DS. A novel approach for featureextraction and selection on MRI images for brain tumorclassification. Int Conf Comp Sci Eng Appl 2012; 225-234.

*Correspondence toShenbagarajan A

Department of Computer Science and Engineering

Annamalai University

India

MRI brain image analysis for tumour diagnosis using hybrid MB-MLM pattern classification technique

Biomed Res- India 2016 Special IssueSpecial Section: Computational Life Science and Smarter Technological Advancement

203