Research ArticleImage Analysis for MRI Based Brain Tumor Detection andFeature Extraction Using Biologically Inspired BWT and SVM
Nilesh Bhaskarrao Bahadure1 Arun Kumar Ray1 and Har Pal Thethi2
1School of Electronics Engineering KIIT University Bhubaneswar Odisha India2Department of Electronics amp Telecommunication Engineering Lovely Professional University Jalandhar Punjab India
Correspondence should be addressed to Nilesh Bhaskarrao Bahadure nbahaduregmailcom
Received 16 January 2017 Accepted 16 February 2017 Published 6 March 2017
Academic Editor Guowei Wei
Copyright copy 2017 Nilesh Bhaskarrao Bahadure et al This is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited
The segmentation detection and extraction of infected tumor area from magnetic resonance (MR) images are a primary concernbut a tedious and time taking task performed by radiologists or clinical experts and their accuracy depends on their experienceonly So the use of computer aided technology becomes very necessary to overcome these limitations In this study to improve theperformance and reduce the complexity involves in themedical image segmentation process we have investigated Berkeley wavelettransformation (BWT) based brain tumor segmentation Furthermore to improve the accuracy and quality rate of the supportvector machine (SVM) based classifier relevant features are extracted from each segmented tissue The experimental results ofproposed technique have been evaluated and validated for performance and quality analysis on magnetic resonance brain imagesbased on accuracy sensitivity specificity and dice similarity index coefficient The experimental results achieved 9651 accuracy942 specificity and 9772 sensitivity demonstrating the effectiveness of the proposed technique for identifying normal andabnormal tissues from brainMR imagesThe experimental results also obtained an average of 082 dice similarity index coefficientwhich indicates better overlap between the automated (machines) extracted tumor region with manually extracted tumor regionby radiologists The simulation results prove the significance in terms of quality parameters and accuracy in comparison to state-of-the-art techniques
1 Introduction
In recent times the introduction of information technologyand e-health care system in the medical field helps clinicalexperts to provide better health care to the patientThis studyaddresses the problems of segmentation of abnormal braintissues and normal tissues such as gray matter (GM) whitematter (WM) and cerebrospinal fluid (CSF) from magneticresonance (MR) images using feature extraction techniqueand support vector machine (SVM) classifier [1 2]
The tumor is basically an uncontrolled growth of can-cerous cells in any part of the body whereas a brain tumoris an uncontrolled growth of cancerous cells in the brain Abrain tumor can be benign or malignant The benign braintumor has a uniformity in structure and does not containactive (cancer) cells whereas malignant brain tumors havea nonuniformity (heterogeneous) in structure and contain
active cells The gliomas and meningiomas are the examplesof low-grade tumors classified as benign tumors and glioblas-toma and astrocytomas are a class of high-grade tumorsclassified as malignant tumors
According to theWorld Health Organization and Ameri-can Brain Tumor Association [3] the most common gradingsystem uses a scale from grade I to grade IV to classify benignand malignant tumor types On that scale benign tumorsfall under grade I and II glioma and malignant tumors fallunder grade III and IV glioma The grade I and II glioma arealso called low-grade tumor type and possess a slow growthwhereas grade III and IV are called high-grade tumor typesand possess a rapid growth of tumors If the low-grade braintumor is left untreated it is likely to develop into a high-grade brain tumor that is a malignant brain tumor Patientswith grade II gliomas require serial monitoring and obser-vations by magnetic resonance imaging (MRI) or computed
HindawiInternational Journal of Biomedical ImagingVolume 2017 Article ID 9749108 12 pageshttpsdoiorg10115520179749108
2 International Journal of Biomedical Imaging
tomography (CT) scan every 6 to 12 months Brain tumormight influence any individual at any age and its impact onthe body may not be the same for every individual
The benign tumors of low-grade I and II glioma areconsidered to be curative under complete surgical excursionwhereas malignant brain tumors of grade III and IV categorycan be treated by radiotherapy chemotherapy or a combina-tion thereof The term malignant glioma encompasses bothgrade III and IV gliomas which is also referred to as anaplas-tic astrocytomas An anaplastic astrocytoma is a mid-gradetumor that demonstrates abnormal or irregular growth andan increased growth index compared to other low-gradetumors Furthermore the most malignant form of astrocy-toma which is also the highest grade glioma is the glioblas-toma The abnormal fast growth of blood vessels and thepresence of the necrosis (dead cells) around the tumor aredistinguished glioblastoma from all the other grades of thetumor class Grade IV tumor class that is glioblastoma isalways rapidly growing and highly malignant form of tumorsas compared to other grades of the tumors
To detect infected tumor tissues from medical imagingmodalities segmentation is employed Segmentation is nec-essary and important step in image analysis it is a process ofseparating an image into different regions or blocks sharingcommon and identical properties such as color texturecontrast brightness boundaries and gray level Brain tumorsegmentation involves the process of separating the tumortissues such as edema and dead cells from normal braintissues and solid tumors such asWM GM and CSF [4] withthe help of MR images or other imaging modalities [5ndash8]
In this study different magnetic resonance imaging(MRI) sequence images are employed for diagnosis includingT1-weighted MRI T2-weighted MRI fluid-attenuated inver-sion recovery- (FLAIR) weighted MRI and proton density-weighted MRI The detection of a brain tumor at an earlystage is a key issue for providing improved treatment Oncea brain tumor is clinically suspected radiological evaluationis required to determine its location its size and impact onthe surrounding areas On the basis of this information thebest therapy surgery radiation or chemotherapy is decidedIt is evident that the chances of survival of a tumor-infectedpatient can be increased significantly if the tumor is detectedaccurately in its early stage [9] As a result the study of braintumors using imaging modalities has gained importance inthe radiology department
The rest of the paper is organized as follows Section 2presents the related works Section 3 presents the materialsand methods with the steps used in the proposed techniqueSection 4 presents the results and discussion Section 5presents the comparative analysis and finally Section 6contains the conclusions and future work
2 Related Works
Medical image segmentation for detection of brain tumorfrom the magnetic resonance (MR) images or from othermedical imaging modalities is a very important process fordeciding right therapy at the right time Many techniqueshave been proposed for classification of brain tumors in MR
images most notably fuzzy clusteringmeans (FCM) supportvector machine (SVM) artificial neural network (ANN)knowledge-based techniques and expectation-maximization(EM) algorithm technique which are some of the populartechniques used for region based segmentation and so toextract the important information from the medical imagingmodalities An overview and findings of some of the recentand prominent researches are presented here Damodharanand Raghavan [10] have presented a neural network basedtechnique for brain tumor detection and classification In thismethod the quality rate is produced separately for segmenta-tion ofWM GM CSF and tumor region and claims an accu-racy of 83 using neural network based classifier Alfonseand Salem [11] have presented a technique for automaticclassification of brain tumor fromMR images using an SVM-based classifier To improve the accuracy of the classifierfeatures are extracted using fast Fourier transform (FFT)and reduction of features is performed using Minimal-Redundancy-Maximal-Relevance (MRMR) technique Thistechnique has obtained an accuracy of 989
The extraction of the brain tumor requires the separationof the brain MR images to two regions [12] One regioncontains the tumor cells of the brain and the second containsthe normal brain cells [13] Zanaty [14] proposed amethodol-ogy for brain tumor segmentation based on a hybrid type ofapproach combining FCM seed region growing and Jaccardsimilarity coefficient algorithm to measure segmented graymatter and white matter tissues from MR images Thismethod obtained an average segmentation score S of 90at the noise level of 3 and 9 respectively Kong et al [7]investigated automatic segmentation of brain tissues fromMR images using discriminative clustering and future selec-tion approach Demirhan et al [5] presented a new tissuesegmentation algorithm using wavelets and neural networkswhich claims effective segmentation of brain MR images intothe tumor WM GM edema and CSF Torheim et al [15]Guo et al [1] and Yao et al [16] presented a technique whichemployed texture features wavelet transform and SVMrsquosalgorithm for effective classification of dynamic contrast-enhanced MR images to handle the nonlinearity of real dataand to address different image protocols effectively Torheimet al [15] also claim that their proposed technique gives betterpredictions and improved clinical factors tumor volume andtumor stage in comparisonwith first-order statistical features
Kumar and Vijayakumar [17] introduced brain tumorsegmentation and classification based on principal compo-nent analysis (PCA) and radial basis function (RBF) kernelbased SVM and claims similarity index of 9620 overlapfraction of 95 and an extra fraction of 0025 The clas-sification accuracy to identify tumor type of this method is94 with total errors detected of 75 Sharma et al [18] havepresented a highly efficient technique which claims accuracyof 100 in the classification of brain tumor fromMR imagesThis method is utilizing texture-primitive features with arti-ficial neural network (ANN) as segmentation and classifiertool Cui et al [19] applied a localized fuzzy clustering withspatial information to form an objective of medical imagesegmentation and bias field estimation for brain MR imagesIn this method authors use Jaccard similarity index as a
International Journal of Biomedical Imaging 3
measurement of the segmentation accuracy and claim 83to 95 accuracy to segment white matter gray matter andcerebrospinal fluid Wang et al [20] have presented a med-ical image segmentation technique based on active contourmodel to deal with the problem of intensity inhomogeneitiesin image segmentation Chaddad [21] has proposed a tech-nique of automatic feature extraction for brain tumor detec-tion based on Gaussian mixture model (GMM) using MRimages In this method using principal component analysis(PCA) and wavelet based features the performance of theGMM feature extraction is enhanced An accuracy of 9705for the T1-weighted and T2-weighted and 9411 for FLAIR-weighted MR images are obtained
Deepa and Arunadevi [22] have proposed a technique ofextreme learning machine for classification of brain tumorfrom 3D MR images This method obtained an accuracyof 932 the sensitivity of 916 and specificity of 978Sachdeva et al [23] have presented a multiclass braintumor classification segmentation and feature extractionperformed using a dataset of 428MR images In this methodauthors used ANN and then PCA-ANN and observed theincrement in classification accuracy from 77 to 91
The above literature survey has revealed that some of thetechniques are invented to obtain segmentation only some ofthe techniques are invented to obtain feature extraction andsome of the techniques are invented to obtain classificationonly Feature extraction and reduction of feature vectors foreffective segmentation of WM GM CSF and infected tumorregion and analysis on combined approach could not beconducted in all the published literatureMoreover only a fewfeatures are extracted and therefore very low accuracy intumor detection has been obtained Also all the above liter-atures are missing with the calculation of overlap that is dicesimilarity index which is one of the important parametersto judge the accuracy of any brain tumor segmentationalgorithm
In this study we perform a combination of biologicallyinspired Berkeley wavelet transformation (BWT) and SVMas a classifier tool to improve diagnostic accuracy The causeof this study is to extract information from the segmentedtumor region and classify healthy and infected tumor tissuesfor a large database of medical images Our results lead toconclude that the proposed method is suitable to integrateclinical decision support systems for primary screening anddiagnosis by the radiologists or clinical experts
3 Materials and Methods
This section presents the materials the source of brain MRimage dataset and the algorithm used to perform brain MRtissue segmentation Figure 1 provides the flow diagramof thealgorithm As test images different MR images of the brainwere used including T1-weightedMR images with RepetitionTime (TR) of 1740 and Echo Time (TE) of 20 T2-weightedMR images with Repetition Time (TR) of 5850 and EchoTime (TE) of 130 and FLAIR-weightedMR images with Rep-etition Time (TR) of 8500 and Echo Time (TE) of 130 Thesetest images were acquired using a 3 Tesla Siemens Magnetom
SpectraMRmachineThe total numbers of slices for all chan-nels were 15 which leads to total of 135 images at 9 slices orimages per patientwith a field of viewof 200mm an interslicegap of 1mm and voxel of size 078mm times 078mm times 05mmThe proposed methodology is applied to real dataset includ-ing brainMR images of 512 times 512 pixel size and was convertedinto grayscale before further processing The following sec-tions discuss the implementation of the algorithm
31 Preprocessing The primary task of preprocessing is toimprove the quality of the MR images and make it in a formsuited for further processing by human or machine visionsystem In addition preprocessing helps to improve certainparameters of MR images such as improving the signal-to-noise ratio enhancing the visual appearance of MR imageremoving the irrelevant noise and undesired parts in thebackground smoothing the inner part of the region andpreserving its edges [5] To improve the signal-to-noise ratioand thus the clarity of the rawMR images we applied adaptivecontrast enhancement based on modified sigmoid function[24]
32 Skull Stripping Skull stripping is an important process inbiomedical image analysis and it is required for the effectiveexamination of brain tumor from the MR images [25ndash28]Skull stripping is the process of eliminating all nonbraintissues in the brain images By skull stripping it is possible toremove additional cerebral tissues such as fat skin and skullin the brain imagesThere are several techniques available forskull stripping some of the popular techniques are automaticskull stripping using image contour skull stripping basedon segmentation and morphological operation and skullstripping based on histogram analysis or a threshold valueFigure 2 provides the stages of the skull stripping algorithmThis study uses the skull stripping technique that is based ona threshold operation to remove skull tissues
33 Segmentation and Morphological Operation The seg-mentation of the infected brain MR regions is achievedthrough the following steps In the first step the preprocessedbrain MR image is converted into a binary image with athreshold for the cut-off of 128 being selectedThepixel valuesgreater than the selected threshold are mapped to whitewhile others are marked as black due to this two differentregions are formed around the infected tumor tissues whichis cropped out In the second step in order to eliminatewhite pixel an erosion operation ofmorphology is employedFinally the eroded region and the original image are bothdivided into two equal regions and the black pixel regionextracted from the erode operation is counted as a brain MRimage mask In this study Berkeley wavelet transformation isemployed for effective segmentation of brain MR image
A wavelet is a function that is defined over a finiteinterval of time and has an average value of zero The wavelettransformation technique is employed to develop functionsoperators data or information into components of differentfrequency which enables studying each component sepa-rately All wavelets are generated from a basic wavelet Ψ(119905)
4 International Journal of Biomedical Imaging
Enhancement
Skull stripping
Removal of skull
Separation of GMWM CSF and tumor
Segmentation
Feature extraction
Mean contrastentropy and energy
Morphologicaloperation
Area extraction ampdecision making
Classification usingSVMMR image
dataset
Normal tissue Abnormal tissue
Pre processing
Figure 1 Steps used in proposed algorithm
Input image
Convert image to grayscale
Convert image to binary image by thresholding
Find the number of connected objects
Find mask by assigning 1 to inside and 0 to outsideof the object that show brain region
Multiply the mask with T1 T2 and FLAIR MR imagesto get their skull-stripped MR image
Figure 2 Steps used in the skull stripping algorithm
by using the scaling and translation process defined by (1) abasic wavelet is also referred to as a mother wavelet becauseit is the point of origin for other wavelets
Ψ119904120591 = 1radic119904Ψ(119905 minus 120591119904 ) (1)
where 119904 and 120591 are the scale and translation factors respec-tively
The Berkeley wavelet transform (BWT) [29 30] isdescribed as a two-dimensional triadic wavelet transformand can be used to process the signal or image Just like themother wavelet transformation or other families of wavelettransformation the BWT algorithm will also perform data
conversion from a spatial form into temporal domain fre-quency The BWT presents an effective way of representationof image transformation and it is a complete orthonormal[30] The mother wavelet transformation 120573120593
120579is piecewise
constant function [29 31] The substitute wavelets from themother wavelet 120573120593
120579are produced at various pixels positions in
the two-dimensional plane through scaling and translation ofthe mother wavelet and it is shown in
120573120593120579 (120591 119904) = 11199042120573120593119909 (3119904 (119909 minus 119894) 3119904 (119910 minus 119895)) (2)
where 120591 and 119904 are translation and scale parameter of thewavelet transformation respectively and 120573120593
120579is the trans-
forming function and it is called the mother wavelet ofBerkeley wavelet transformation The only single constantterm is sufficient to represent the mean value of an image thecoefficient value of the single term is shown in
1205730 = 1radic9 [119906 (1199093 1199103 )] (3)
The morphological operation is used for the extractionof the boundary areas of the brain images Conceptuallythe morphological operation is only rearranging the relativeorder of pixel values not on theirmathematical values and sois suitable to process only binary images Dilation and erosionare the two most basic operations of morphology Dilationoperations are intended to add pixels to the boundary regionof the object while erosion operations are intended to removethe pixels from the boundary region of the objects The oper-ation of addition and removing pixels to or from boundaryregion of the objects is based on the structuring element ofthe selected image
The experimented results produced by the proposedtechnique depicted for the segmented outcome for the threeclasses of WM GM and CSF and for the extracted tumor
International Journal of Biomedical Imaging 5
(a) (b) (c) (d) (e) (f)
(g) (h) (i) (j) (k) (l)
Figure 3 Segmented and area extracted result of brain MR image (a) Original image (b) Enhanced image (c) Skull-stripped image (d)Wavelet transpose image (e) Intense segmented image (f) Inverse intense image (g) Gray matter (h) White matter (i) CSF (j) Dice overlapimage (k) Eroded image (l) Area extracted image
region are given in Figure 3 The experimental results alsofind dice overlap image indicating the comparison betweenthe algorithm output and ground truth
34 Feature Extraction It is the process of collecting higher-level information of an image such as shape texture colorand contrast In fact texture analysis is an important parame-ter of human visual perception andmachine learning systemIt is used effectively to improve the accuracy of diagnosissystem by selecting prominent features Haralick et al [32]introduced one of the most widely used image analysisapplications of Gray Level Cooccurrence Matrix (GLCM)and texture feature This technique follows two steps forfeature extraction from the medical images In the first stepthe GLCM is computed and in the other step the texturefeatures based on the GLCM are calculated Due to theintricate structure of diversified tissues such asWMGM andCSF in the brain MR images extraction of relevant featuresis an essential task Textural findings and analysis couldimprove the diagnosis different stages of the tumor (tumorstaging) and therapy response assessment The statisticsfeature formula for some of the useful features is listed below
(1) Mean (M) The mean of an image is calculated by addingall the pixel values of an image divided by the total number ofpixels in an image
119872 = ( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) (4)
(2) Standard Deviation (SD) The standard deviation is thesecond central moment describing probability distributionof an observed population and can serve as a measure of
inhomogeneity A higher value indicates better intensity leveland high contrast of edges of an image
SD (120590) = radic( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
(119891 (119909 119910) minus119872)2 (5)
(3) Entropy (E) Entropy is calculated to characterize therandomness of the textural image and is defined as
(4) Skewness (119878119896) Skewness is a measure of symmetry or thelack of symmetry The skewness of a random variable 119883 isdenoted as 119878119896(119883) and it is defined as
119878119896 (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)310038161003816100381610038161003816SD3
(7)
(5) Kurtosis (119878119896)The shape of a random variablersquos probabilitydistribution is described by the parameter calledKurtosis Forthe randomvariable119883 theKurtosis is denoted as119870urt(119883) andit is defined as
119870urt (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)410038161003816100381610038161003816SD4
(8)
(6) Energy (En) Energy can be defined as the quantifiableamount of the extent of pixel pair repetitions Energy is aparameter to measure the similarity of an image If energy is
6 International Journal of Biomedical Imaging
defined by Haralicks GLCM feature then it is also referred toas angular second moment and it is defined as
En = radic119898minus1sum119909=0
119899minus1sum119910=0
1198912 (119909 119910) (9)
(7) Contrast (119862119900119899) Contrast is ameasure of intensity of a pixeland its neighbor over the image and it is defined as
119862on = 119898minus1sum119909=0
119899minus1sum119910=0
(119909 minus 119910)2 119891 (119909 119910) (10)
(8) Inverse DifferenceMoment (IDM) orHomogeneity InverseDifference Moment is a measure of the local homogeneity ofan image IDMmay have a single or a range of values so as todetermine whether the image is textured or nontextured
IDM = 119898minus1sum119909=0
119899minus1sum119910=0
11 + (119909 minus 119910)2119891 (119909 119910) (11)
(9) Directional Moment (DM) Directional moment is atextural property of the image calculated by considering thealignment of the image as ameasure in terms of the angle andit is defined as
DM = 119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) 1003816100381610038161003816119909 minus 1199101003816100381610038161003816 (12)
(10) Correlation (119862119900119903119903) Correlation feature describes thespatial dependencies between the pixels and it is defined as
where119872119909 and 120590119909 are the mean and standard deviation in thehorizontal spatial domain and119872119910 and 120590119910 are the mean andstandard deviation in the vertical spatial domain
(11) Coarseness (119862119899119890119904119904) Coarseness is a measure of roughnessin the textural analysis of an image For a fixed window sizea texture with a smaller number of texture elements is saidto be more coarse than the one with a larger number Therougher texture means higher coarseness value Fine textureshave smaller values of coarseness It is defined as
Apart from the above textural feature extraction thefollowing quality assessment parameters are also needed toensure better result analysis on brain MR images
(1) Structured Similarity Index (SSIM) The Structural Simi-larity Index (SSIM) is a perceptual metric that signifies thatthe degradation in image quality may be caused by data
compression or losses in data transmission or by any othermeans of the image processing It is defined as
A Higher value of SSIM indicates better preservation ofluminance contrast and structural content
(2)Mean Square Error (MSE) Mean square error is ameasureof signal fidelity or image fidelity The purpose of signal orimage fidelity measure is to find the similarity or fidelitybetween two images by providing the quantitative scoreWhen MSE is calculated then it is assumed that one of theimages is pristine original while the other is distorted orprocessed by some means and it is defined as
(3) Peak Signal-to-Noise Ratio (PSNR) in dB Peak signal-to-noise ratio is a measure used to assess the quality ofreconstruction of processed image and it is defined as
PSBR in dB = 20 log10 (2119899 minus 1)MSE (17)
Lower value ofMSE and higher value of PSNR indicate bettersignal-to-noise ratio
(4) Dice Coefficient Dice coefficient or dice similarity index isa measure of overlap between the two images and it is definedas
Dice (119860 119861) = 2 times 10038161003816100381610038161198601 and 11986111003816100381610038161003816(100381610038161003816100381611986011003816100381610038161003816 + 100381610038161003816100381611986111003816100381610038161003816) (18)
where 119860 isin 0 1 is tumor region extracted from algorithmicpredictions and 119861 isin 0 1 is the experts ground truth Theminimum value of dice coefficient is 0 and the maximumis 1 a higher value indicates better overlap between the twoimages
Tables 1 and 2 show some of the prominent features forthe first-order statistical and second-order statistical analysisTable 2 also indicates the measure of coarseness and numberof key values present in the segmented image
35 Support Vector Machine (SVM) The original SVM algo-rithmwas contributed by Vladimir N Vapnik and its modernversion was developed by Cortes and Vapnik in 1993 [33]The SVM algorithm is based on the study of a supervisedlearning technique and is applied to one-class classificationproblem to n-class classification problems [1 34ndash36] Theprinciple aim of the SVM algorithm is to transform a non-linear dividing objective into a linear transformation using a
International Journal of Biomedical Imaging 7
Table 1 First-order statistical features for few images
Images Mean Standard deviation Skewness Kurtosis Energy EntropyImage 1 866 4399 000553 289041119864 minus 06 1094 065Image 2 1181 4911 000655 274079119864 minus 06 1637 094Image 3 3940 7559 001054 18506119864 minus 06 6599 303Image 4 683 3945 000517 333685119864 minus 06 811 045Image 5 1190 3881 002002 135422119864 minus 05 3317 209Image 6 533 2895 001647 205493119864 minus 05 1387 112
Table 2 Second-order textural features with coarseness and key points for few images
function called SVMrsquos kernel function In this study we usedthe Gaussian kernel function for transformation By using akernel function the nonlinear samples can be transformedinto a high-dimensional future space where the separation ofnonlinear samples or datamight becomepossiblemaking theclassification convenient [16] The SVM algorithm defines ahyperplane that is divided into two training classes as definedin
where119885 and 119879 are hyperplane parameters and 120601(119910) is a func-tion used to map vector 119910 into a higher-dimensional spaceEquation (20) provides the Gaussian kernel function ofnonlinear SVM [16 34] used for the optimal solution of clas-sification and generalization and its advanced classificationfunction is shown in (21)
119896 (119910119894 119910119895) = exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172] (20)
(exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172]) (21)
where 119910119894 and 119910119895 are objects 119894 and 119895 respectively and 120574 is acontour parameter used to determine the smoothness of theboundary region [4 15]
The features selectionwith kernel class separabilitymakesSVM the default choice for classification of a brain tumorThe SVM algorithmrsquos performance can be evaluated in termsof accuracy sensitivity and specificity The confusion matrixdefining the terms TP TN FP and FN from the expectedoutcome and ground truth result for the calculation ofaccuracy sensitivity and specificity are shown in Table 3
Where TP is the number of true positives which is used toindicate the total number of abnormal cases correctly clas-sified TN is the number of true negatives which is used toindicate normal cases correctly classified FP is the number
Table 3 Confusion matrix defining the terms TP TN FP and FN
Expected outcome Ground truth Row totalPositive Negative
Table 4 Accuracy sensitivity and specificity calculation
Quality parameter Formula
Accuracy TP + TNTP + TN + FP + FN
Sensitivity TPTP + FN
Specificity TNTN + FP
of false positive and it is used to indicate wrongly detectedor classified abnormal cases when they are actually normalcases and FN is the number of false negatives it is used toindicate wrongly classified or detected normal cases whenthey are actually abnormal cases [15] all of these outcomeparameters are calculated using the total number of samplesexamined for the detection of the tumor The quality rateparameter accuracy is the proportion of total correctly classi-fied cases that are abnormally classified as abnormal andnormally classified as normal from the total number of casesexamined [37 38] Table 4 shows the formulas to calculateaccuracy sensitivity and specificity
4 Results and Discussion
To validate the performance of our algorithm we used twobenchmark datasets and one dataset collected from expertradiologists which included sample images of 15 patients
8 International Journal of Biomedical Imaging
Table 5 Performance analysis parameters for segmented tissues
Images MSE PSNR SSIM Dice scoreImage 1 186 5545 dB 08944 083Image 2 058 6821 dB 09025 087Image 3 495 5628 dB 09702 082Image 4 123 5879 dB 08801 079Image 5 506 5965 dB 07978 090
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 4 Experimental results of image 1 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
with 9 slices for each patient The first dataset is the DigitalImaging and Communications in Medicine (DICOM) data-set [39] For the purpose of the analysis we considered 22images from the DICOM dataset all of which included aretumor-infected brain tissues However this dataset did nothave any ground truth imagesThe second dataset is the BrainWeb dataset [40] which consists of full three-dimensionalsimulated brain MR data obtained using three sequences ofmodalities namely T1-weightedMRI T2-weightedMRI andproton density-weightedMRIThis dataset included a varietyof slice thicknesses noise levels and levels of intensitynonuniformity The images used for our analysis are mostlyincludedT2-weightedmodality with 1mm slice thickness 3noise and 20 intensity nonuniformity In this dataset 13out of 44 images included are tumor-infected brain tissuesThe last dataset collected from expert radiologists consisted
of 135 images of 15 patients with all modalities This datasethad ground truth images that helped to compare the resultsof our method with the manual analysis of radiologists
This section presents the results of our proposed imagesegmentation technique which are obtained by using realbrain MR images The proposed algorithm was carried outusing Matlab 7120 (R2011a) which runs on the Windows 8operating system and has an Intel core i3 processor and a4GB RAM The sample experimental results obtained fromthe proposed technique that are depicted in Figures 4 5 and6 show the original image along with enhanced image skull-stripped image wavelet decompose image cluster (intense)segmented image dice overlap image and the tumor regionwith extracted area mark
Table 5 provides the details of the different performanceparameters such as mean squared error (MSE) and peak
International Journal of Biomedical Imaging 9
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
tomography (CT) scan every 6 to 12 months Brain tumormight influence any individual at any age and its impact onthe body may not be the same for every individual
The benign tumors of low-grade I and II glioma areconsidered to be curative under complete surgical excursionwhereas malignant brain tumors of grade III and IV categorycan be treated by radiotherapy chemotherapy or a combina-tion thereof The term malignant glioma encompasses bothgrade III and IV gliomas which is also referred to as anaplas-tic astrocytomas An anaplastic astrocytoma is a mid-gradetumor that demonstrates abnormal or irregular growth andan increased growth index compared to other low-gradetumors Furthermore the most malignant form of astrocy-toma which is also the highest grade glioma is the glioblas-toma The abnormal fast growth of blood vessels and thepresence of the necrosis (dead cells) around the tumor aredistinguished glioblastoma from all the other grades of thetumor class Grade IV tumor class that is glioblastoma isalways rapidly growing and highly malignant form of tumorsas compared to other grades of the tumors
To detect infected tumor tissues from medical imagingmodalities segmentation is employed Segmentation is nec-essary and important step in image analysis it is a process ofseparating an image into different regions or blocks sharingcommon and identical properties such as color texturecontrast brightness boundaries and gray level Brain tumorsegmentation involves the process of separating the tumortissues such as edema and dead cells from normal braintissues and solid tumors such asWM GM and CSF [4] withthe help of MR images or other imaging modalities [5ndash8]
In this study different magnetic resonance imaging(MRI) sequence images are employed for diagnosis includingT1-weighted MRI T2-weighted MRI fluid-attenuated inver-sion recovery- (FLAIR) weighted MRI and proton density-weighted MRI The detection of a brain tumor at an earlystage is a key issue for providing improved treatment Oncea brain tumor is clinically suspected radiological evaluationis required to determine its location its size and impact onthe surrounding areas On the basis of this information thebest therapy surgery radiation or chemotherapy is decidedIt is evident that the chances of survival of a tumor-infectedpatient can be increased significantly if the tumor is detectedaccurately in its early stage [9] As a result the study of braintumors using imaging modalities has gained importance inthe radiology department
The rest of the paper is organized as follows Section 2presents the related works Section 3 presents the materialsand methods with the steps used in the proposed techniqueSection 4 presents the results and discussion Section 5presents the comparative analysis and finally Section 6contains the conclusions and future work
2 Related Works
Medical image segmentation for detection of brain tumorfrom the magnetic resonance (MR) images or from othermedical imaging modalities is a very important process fordeciding right therapy at the right time Many techniqueshave been proposed for classification of brain tumors in MR
images most notably fuzzy clusteringmeans (FCM) supportvector machine (SVM) artificial neural network (ANN)knowledge-based techniques and expectation-maximization(EM) algorithm technique which are some of the populartechniques used for region based segmentation and so toextract the important information from the medical imagingmodalities An overview and findings of some of the recentand prominent researches are presented here Damodharanand Raghavan [10] have presented a neural network basedtechnique for brain tumor detection and classification In thismethod the quality rate is produced separately for segmenta-tion ofWM GM CSF and tumor region and claims an accu-racy of 83 using neural network based classifier Alfonseand Salem [11] have presented a technique for automaticclassification of brain tumor fromMR images using an SVM-based classifier To improve the accuracy of the classifierfeatures are extracted using fast Fourier transform (FFT)and reduction of features is performed using Minimal-Redundancy-Maximal-Relevance (MRMR) technique Thistechnique has obtained an accuracy of 989
The extraction of the brain tumor requires the separationof the brain MR images to two regions [12] One regioncontains the tumor cells of the brain and the second containsthe normal brain cells [13] Zanaty [14] proposed amethodol-ogy for brain tumor segmentation based on a hybrid type ofapproach combining FCM seed region growing and Jaccardsimilarity coefficient algorithm to measure segmented graymatter and white matter tissues from MR images Thismethod obtained an average segmentation score S of 90at the noise level of 3 and 9 respectively Kong et al [7]investigated automatic segmentation of brain tissues fromMR images using discriminative clustering and future selec-tion approach Demirhan et al [5] presented a new tissuesegmentation algorithm using wavelets and neural networkswhich claims effective segmentation of brain MR images intothe tumor WM GM edema and CSF Torheim et al [15]Guo et al [1] and Yao et al [16] presented a technique whichemployed texture features wavelet transform and SVMrsquosalgorithm for effective classification of dynamic contrast-enhanced MR images to handle the nonlinearity of real dataand to address different image protocols effectively Torheimet al [15] also claim that their proposed technique gives betterpredictions and improved clinical factors tumor volume andtumor stage in comparisonwith first-order statistical features
Kumar and Vijayakumar [17] introduced brain tumorsegmentation and classification based on principal compo-nent analysis (PCA) and radial basis function (RBF) kernelbased SVM and claims similarity index of 9620 overlapfraction of 95 and an extra fraction of 0025 The clas-sification accuracy to identify tumor type of this method is94 with total errors detected of 75 Sharma et al [18] havepresented a highly efficient technique which claims accuracyof 100 in the classification of brain tumor fromMR imagesThis method is utilizing texture-primitive features with arti-ficial neural network (ANN) as segmentation and classifiertool Cui et al [19] applied a localized fuzzy clustering withspatial information to form an objective of medical imagesegmentation and bias field estimation for brain MR imagesIn this method authors use Jaccard similarity index as a
International Journal of Biomedical Imaging 3
measurement of the segmentation accuracy and claim 83to 95 accuracy to segment white matter gray matter andcerebrospinal fluid Wang et al [20] have presented a med-ical image segmentation technique based on active contourmodel to deal with the problem of intensity inhomogeneitiesin image segmentation Chaddad [21] has proposed a tech-nique of automatic feature extraction for brain tumor detec-tion based on Gaussian mixture model (GMM) using MRimages In this method using principal component analysis(PCA) and wavelet based features the performance of theGMM feature extraction is enhanced An accuracy of 9705for the T1-weighted and T2-weighted and 9411 for FLAIR-weighted MR images are obtained
Deepa and Arunadevi [22] have proposed a technique ofextreme learning machine for classification of brain tumorfrom 3D MR images This method obtained an accuracyof 932 the sensitivity of 916 and specificity of 978Sachdeva et al [23] have presented a multiclass braintumor classification segmentation and feature extractionperformed using a dataset of 428MR images In this methodauthors used ANN and then PCA-ANN and observed theincrement in classification accuracy from 77 to 91
The above literature survey has revealed that some of thetechniques are invented to obtain segmentation only some ofthe techniques are invented to obtain feature extraction andsome of the techniques are invented to obtain classificationonly Feature extraction and reduction of feature vectors foreffective segmentation of WM GM CSF and infected tumorregion and analysis on combined approach could not beconducted in all the published literatureMoreover only a fewfeatures are extracted and therefore very low accuracy intumor detection has been obtained Also all the above liter-atures are missing with the calculation of overlap that is dicesimilarity index which is one of the important parametersto judge the accuracy of any brain tumor segmentationalgorithm
In this study we perform a combination of biologicallyinspired Berkeley wavelet transformation (BWT) and SVMas a classifier tool to improve diagnostic accuracy The causeof this study is to extract information from the segmentedtumor region and classify healthy and infected tumor tissuesfor a large database of medical images Our results lead toconclude that the proposed method is suitable to integrateclinical decision support systems for primary screening anddiagnosis by the radiologists or clinical experts
3 Materials and Methods
This section presents the materials the source of brain MRimage dataset and the algorithm used to perform brain MRtissue segmentation Figure 1 provides the flow diagramof thealgorithm As test images different MR images of the brainwere used including T1-weightedMR images with RepetitionTime (TR) of 1740 and Echo Time (TE) of 20 T2-weightedMR images with Repetition Time (TR) of 5850 and EchoTime (TE) of 130 and FLAIR-weightedMR images with Rep-etition Time (TR) of 8500 and Echo Time (TE) of 130 Thesetest images were acquired using a 3 Tesla Siemens Magnetom
SpectraMRmachineThe total numbers of slices for all chan-nels were 15 which leads to total of 135 images at 9 slices orimages per patientwith a field of viewof 200mm an interslicegap of 1mm and voxel of size 078mm times 078mm times 05mmThe proposed methodology is applied to real dataset includ-ing brainMR images of 512 times 512 pixel size and was convertedinto grayscale before further processing The following sec-tions discuss the implementation of the algorithm
31 Preprocessing The primary task of preprocessing is toimprove the quality of the MR images and make it in a formsuited for further processing by human or machine visionsystem In addition preprocessing helps to improve certainparameters of MR images such as improving the signal-to-noise ratio enhancing the visual appearance of MR imageremoving the irrelevant noise and undesired parts in thebackground smoothing the inner part of the region andpreserving its edges [5] To improve the signal-to-noise ratioand thus the clarity of the rawMR images we applied adaptivecontrast enhancement based on modified sigmoid function[24]
32 Skull Stripping Skull stripping is an important process inbiomedical image analysis and it is required for the effectiveexamination of brain tumor from the MR images [25ndash28]Skull stripping is the process of eliminating all nonbraintissues in the brain images By skull stripping it is possible toremove additional cerebral tissues such as fat skin and skullin the brain imagesThere are several techniques available forskull stripping some of the popular techniques are automaticskull stripping using image contour skull stripping basedon segmentation and morphological operation and skullstripping based on histogram analysis or a threshold valueFigure 2 provides the stages of the skull stripping algorithmThis study uses the skull stripping technique that is based ona threshold operation to remove skull tissues
33 Segmentation and Morphological Operation The seg-mentation of the infected brain MR regions is achievedthrough the following steps In the first step the preprocessedbrain MR image is converted into a binary image with athreshold for the cut-off of 128 being selectedThepixel valuesgreater than the selected threshold are mapped to whitewhile others are marked as black due to this two differentregions are formed around the infected tumor tissues whichis cropped out In the second step in order to eliminatewhite pixel an erosion operation ofmorphology is employedFinally the eroded region and the original image are bothdivided into two equal regions and the black pixel regionextracted from the erode operation is counted as a brain MRimage mask In this study Berkeley wavelet transformation isemployed for effective segmentation of brain MR image
A wavelet is a function that is defined over a finiteinterval of time and has an average value of zero The wavelettransformation technique is employed to develop functionsoperators data or information into components of differentfrequency which enables studying each component sepa-rately All wavelets are generated from a basic wavelet Ψ(119905)
4 International Journal of Biomedical Imaging
Enhancement
Skull stripping
Removal of skull
Separation of GMWM CSF and tumor
Segmentation
Feature extraction
Mean contrastentropy and energy
Morphologicaloperation
Area extraction ampdecision making
Classification usingSVMMR image
dataset
Normal tissue Abnormal tissue
Pre processing
Figure 1 Steps used in proposed algorithm
Input image
Convert image to grayscale
Convert image to binary image by thresholding
Find the number of connected objects
Find mask by assigning 1 to inside and 0 to outsideof the object that show brain region
Multiply the mask with T1 T2 and FLAIR MR imagesto get their skull-stripped MR image
Figure 2 Steps used in the skull stripping algorithm
by using the scaling and translation process defined by (1) abasic wavelet is also referred to as a mother wavelet becauseit is the point of origin for other wavelets
Ψ119904120591 = 1radic119904Ψ(119905 minus 120591119904 ) (1)
where 119904 and 120591 are the scale and translation factors respec-tively
The Berkeley wavelet transform (BWT) [29 30] isdescribed as a two-dimensional triadic wavelet transformand can be used to process the signal or image Just like themother wavelet transformation or other families of wavelettransformation the BWT algorithm will also perform data
conversion from a spatial form into temporal domain fre-quency The BWT presents an effective way of representationof image transformation and it is a complete orthonormal[30] The mother wavelet transformation 120573120593
120579is piecewise
constant function [29 31] The substitute wavelets from themother wavelet 120573120593
120579are produced at various pixels positions in
the two-dimensional plane through scaling and translation ofthe mother wavelet and it is shown in
120573120593120579 (120591 119904) = 11199042120573120593119909 (3119904 (119909 minus 119894) 3119904 (119910 minus 119895)) (2)
where 120591 and 119904 are translation and scale parameter of thewavelet transformation respectively and 120573120593
120579is the trans-
forming function and it is called the mother wavelet ofBerkeley wavelet transformation The only single constantterm is sufficient to represent the mean value of an image thecoefficient value of the single term is shown in
1205730 = 1radic9 [119906 (1199093 1199103 )] (3)
The morphological operation is used for the extractionof the boundary areas of the brain images Conceptuallythe morphological operation is only rearranging the relativeorder of pixel values not on theirmathematical values and sois suitable to process only binary images Dilation and erosionare the two most basic operations of morphology Dilationoperations are intended to add pixels to the boundary regionof the object while erosion operations are intended to removethe pixels from the boundary region of the objects The oper-ation of addition and removing pixels to or from boundaryregion of the objects is based on the structuring element ofthe selected image
The experimented results produced by the proposedtechnique depicted for the segmented outcome for the threeclasses of WM GM and CSF and for the extracted tumor
International Journal of Biomedical Imaging 5
(a) (b) (c) (d) (e) (f)
(g) (h) (i) (j) (k) (l)
Figure 3 Segmented and area extracted result of brain MR image (a) Original image (b) Enhanced image (c) Skull-stripped image (d)Wavelet transpose image (e) Intense segmented image (f) Inverse intense image (g) Gray matter (h) White matter (i) CSF (j) Dice overlapimage (k) Eroded image (l) Area extracted image
region are given in Figure 3 The experimental results alsofind dice overlap image indicating the comparison betweenthe algorithm output and ground truth
34 Feature Extraction It is the process of collecting higher-level information of an image such as shape texture colorand contrast In fact texture analysis is an important parame-ter of human visual perception andmachine learning systemIt is used effectively to improve the accuracy of diagnosissystem by selecting prominent features Haralick et al [32]introduced one of the most widely used image analysisapplications of Gray Level Cooccurrence Matrix (GLCM)and texture feature This technique follows two steps forfeature extraction from the medical images In the first stepthe GLCM is computed and in the other step the texturefeatures based on the GLCM are calculated Due to theintricate structure of diversified tissues such asWMGM andCSF in the brain MR images extraction of relevant featuresis an essential task Textural findings and analysis couldimprove the diagnosis different stages of the tumor (tumorstaging) and therapy response assessment The statisticsfeature formula for some of the useful features is listed below
(1) Mean (M) The mean of an image is calculated by addingall the pixel values of an image divided by the total number ofpixels in an image
119872 = ( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) (4)
(2) Standard Deviation (SD) The standard deviation is thesecond central moment describing probability distributionof an observed population and can serve as a measure of
inhomogeneity A higher value indicates better intensity leveland high contrast of edges of an image
SD (120590) = radic( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
(119891 (119909 119910) minus119872)2 (5)
(3) Entropy (E) Entropy is calculated to characterize therandomness of the textural image and is defined as
(4) Skewness (119878119896) Skewness is a measure of symmetry or thelack of symmetry The skewness of a random variable 119883 isdenoted as 119878119896(119883) and it is defined as
119878119896 (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)310038161003816100381610038161003816SD3
(7)
(5) Kurtosis (119878119896)The shape of a random variablersquos probabilitydistribution is described by the parameter calledKurtosis Forthe randomvariable119883 theKurtosis is denoted as119870urt(119883) andit is defined as
119870urt (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)410038161003816100381610038161003816SD4
(8)
(6) Energy (En) Energy can be defined as the quantifiableamount of the extent of pixel pair repetitions Energy is aparameter to measure the similarity of an image If energy is
6 International Journal of Biomedical Imaging
defined by Haralicks GLCM feature then it is also referred toas angular second moment and it is defined as
En = radic119898minus1sum119909=0
119899minus1sum119910=0
1198912 (119909 119910) (9)
(7) Contrast (119862119900119899) Contrast is ameasure of intensity of a pixeland its neighbor over the image and it is defined as
119862on = 119898minus1sum119909=0
119899minus1sum119910=0
(119909 minus 119910)2 119891 (119909 119910) (10)
(8) Inverse DifferenceMoment (IDM) orHomogeneity InverseDifference Moment is a measure of the local homogeneity ofan image IDMmay have a single or a range of values so as todetermine whether the image is textured or nontextured
IDM = 119898minus1sum119909=0
119899minus1sum119910=0
11 + (119909 minus 119910)2119891 (119909 119910) (11)
(9) Directional Moment (DM) Directional moment is atextural property of the image calculated by considering thealignment of the image as ameasure in terms of the angle andit is defined as
DM = 119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) 1003816100381610038161003816119909 minus 1199101003816100381610038161003816 (12)
(10) Correlation (119862119900119903119903) Correlation feature describes thespatial dependencies between the pixels and it is defined as
where119872119909 and 120590119909 are the mean and standard deviation in thehorizontal spatial domain and119872119910 and 120590119910 are the mean andstandard deviation in the vertical spatial domain
(11) Coarseness (119862119899119890119904119904) Coarseness is a measure of roughnessin the textural analysis of an image For a fixed window sizea texture with a smaller number of texture elements is saidto be more coarse than the one with a larger number Therougher texture means higher coarseness value Fine textureshave smaller values of coarseness It is defined as
Apart from the above textural feature extraction thefollowing quality assessment parameters are also needed toensure better result analysis on brain MR images
(1) Structured Similarity Index (SSIM) The Structural Simi-larity Index (SSIM) is a perceptual metric that signifies thatthe degradation in image quality may be caused by data
compression or losses in data transmission or by any othermeans of the image processing It is defined as
A Higher value of SSIM indicates better preservation ofluminance contrast and structural content
(2)Mean Square Error (MSE) Mean square error is ameasureof signal fidelity or image fidelity The purpose of signal orimage fidelity measure is to find the similarity or fidelitybetween two images by providing the quantitative scoreWhen MSE is calculated then it is assumed that one of theimages is pristine original while the other is distorted orprocessed by some means and it is defined as
(3) Peak Signal-to-Noise Ratio (PSNR) in dB Peak signal-to-noise ratio is a measure used to assess the quality ofreconstruction of processed image and it is defined as
PSBR in dB = 20 log10 (2119899 minus 1)MSE (17)
Lower value ofMSE and higher value of PSNR indicate bettersignal-to-noise ratio
(4) Dice Coefficient Dice coefficient or dice similarity index isa measure of overlap between the two images and it is definedas
Dice (119860 119861) = 2 times 10038161003816100381610038161198601 and 11986111003816100381610038161003816(100381610038161003816100381611986011003816100381610038161003816 + 100381610038161003816100381611986111003816100381610038161003816) (18)
where 119860 isin 0 1 is tumor region extracted from algorithmicpredictions and 119861 isin 0 1 is the experts ground truth Theminimum value of dice coefficient is 0 and the maximumis 1 a higher value indicates better overlap between the twoimages
Tables 1 and 2 show some of the prominent features forthe first-order statistical and second-order statistical analysisTable 2 also indicates the measure of coarseness and numberof key values present in the segmented image
35 Support Vector Machine (SVM) The original SVM algo-rithmwas contributed by Vladimir N Vapnik and its modernversion was developed by Cortes and Vapnik in 1993 [33]The SVM algorithm is based on the study of a supervisedlearning technique and is applied to one-class classificationproblem to n-class classification problems [1 34ndash36] Theprinciple aim of the SVM algorithm is to transform a non-linear dividing objective into a linear transformation using a
International Journal of Biomedical Imaging 7
Table 1 First-order statistical features for few images
Images Mean Standard deviation Skewness Kurtosis Energy EntropyImage 1 866 4399 000553 289041119864 minus 06 1094 065Image 2 1181 4911 000655 274079119864 minus 06 1637 094Image 3 3940 7559 001054 18506119864 minus 06 6599 303Image 4 683 3945 000517 333685119864 minus 06 811 045Image 5 1190 3881 002002 135422119864 minus 05 3317 209Image 6 533 2895 001647 205493119864 minus 05 1387 112
Table 2 Second-order textural features with coarseness and key points for few images
function called SVMrsquos kernel function In this study we usedthe Gaussian kernel function for transformation By using akernel function the nonlinear samples can be transformedinto a high-dimensional future space where the separation ofnonlinear samples or datamight becomepossiblemaking theclassification convenient [16] The SVM algorithm defines ahyperplane that is divided into two training classes as definedin
where119885 and 119879 are hyperplane parameters and 120601(119910) is a func-tion used to map vector 119910 into a higher-dimensional spaceEquation (20) provides the Gaussian kernel function ofnonlinear SVM [16 34] used for the optimal solution of clas-sification and generalization and its advanced classificationfunction is shown in (21)
119896 (119910119894 119910119895) = exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172] (20)
(exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172]) (21)
where 119910119894 and 119910119895 are objects 119894 and 119895 respectively and 120574 is acontour parameter used to determine the smoothness of theboundary region [4 15]
The features selectionwith kernel class separabilitymakesSVM the default choice for classification of a brain tumorThe SVM algorithmrsquos performance can be evaluated in termsof accuracy sensitivity and specificity The confusion matrixdefining the terms TP TN FP and FN from the expectedoutcome and ground truth result for the calculation ofaccuracy sensitivity and specificity are shown in Table 3
Where TP is the number of true positives which is used toindicate the total number of abnormal cases correctly clas-sified TN is the number of true negatives which is used toindicate normal cases correctly classified FP is the number
Table 3 Confusion matrix defining the terms TP TN FP and FN
Expected outcome Ground truth Row totalPositive Negative
Table 4 Accuracy sensitivity and specificity calculation
Quality parameter Formula
Accuracy TP + TNTP + TN + FP + FN
Sensitivity TPTP + FN
Specificity TNTN + FP
of false positive and it is used to indicate wrongly detectedor classified abnormal cases when they are actually normalcases and FN is the number of false negatives it is used toindicate wrongly classified or detected normal cases whenthey are actually abnormal cases [15] all of these outcomeparameters are calculated using the total number of samplesexamined for the detection of the tumor The quality rateparameter accuracy is the proportion of total correctly classi-fied cases that are abnormally classified as abnormal andnormally classified as normal from the total number of casesexamined [37 38] Table 4 shows the formulas to calculateaccuracy sensitivity and specificity
4 Results and Discussion
To validate the performance of our algorithm we used twobenchmark datasets and one dataset collected from expertradiologists which included sample images of 15 patients
8 International Journal of Biomedical Imaging
Table 5 Performance analysis parameters for segmented tissues
Images MSE PSNR SSIM Dice scoreImage 1 186 5545 dB 08944 083Image 2 058 6821 dB 09025 087Image 3 495 5628 dB 09702 082Image 4 123 5879 dB 08801 079Image 5 506 5965 dB 07978 090
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 4 Experimental results of image 1 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
with 9 slices for each patient The first dataset is the DigitalImaging and Communications in Medicine (DICOM) data-set [39] For the purpose of the analysis we considered 22images from the DICOM dataset all of which included aretumor-infected brain tissues However this dataset did nothave any ground truth imagesThe second dataset is the BrainWeb dataset [40] which consists of full three-dimensionalsimulated brain MR data obtained using three sequences ofmodalities namely T1-weightedMRI T2-weightedMRI andproton density-weightedMRIThis dataset included a varietyof slice thicknesses noise levels and levels of intensitynonuniformity The images used for our analysis are mostlyincludedT2-weightedmodality with 1mm slice thickness 3noise and 20 intensity nonuniformity In this dataset 13out of 44 images included are tumor-infected brain tissuesThe last dataset collected from expert radiologists consisted
of 135 images of 15 patients with all modalities This datasethad ground truth images that helped to compare the resultsof our method with the manual analysis of radiologists
This section presents the results of our proposed imagesegmentation technique which are obtained by using realbrain MR images The proposed algorithm was carried outusing Matlab 7120 (R2011a) which runs on the Windows 8operating system and has an Intel core i3 processor and a4GB RAM The sample experimental results obtained fromthe proposed technique that are depicted in Figures 4 5 and6 show the original image along with enhanced image skull-stripped image wavelet decompose image cluster (intense)segmented image dice overlap image and the tumor regionwith extracted area mark
Table 5 provides the details of the different performanceparameters such as mean squared error (MSE) and peak
International Journal of Biomedical Imaging 9
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
measurement of the segmentation accuracy and claim 83to 95 accuracy to segment white matter gray matter andcerebrospinal fluid Wang et al [20] have presented a med-ical image segmentation technique based on active contourmodel to deal with the problem of intensity inhomogeneitiesin image segmentation Chaddad [21] has proposed a tech-nique of automatic feature extraction for brain tumor detec-tion based on Gaussian mixture model (GMM) using MRimages In this method using principal component analysis(PCA) and wavelet based features the performance of theGMM feature extraction is enhanced An accuracy of 9705for the T1-weighted and T2-weighted and 9411 for FLAIR-weighted MR images are obtained
Deepa and Arunadevi [22] have proposed a technique ofextreme learning machine for classification of brain tumorfrom 3D MR images This method obtained an accuracyof 932 the sensitivity of 916 and specificity of 978Sachdeva et al [23] have presented a multiclass braintumor classification segmentation and feature extractionperformed using a dataset of 428MR images In this methodauthors used ANN and then PCA-ANN and observed theincrement in classification accuracy from 77 to 91
The above literature survey has revealed that some of thetechniques are invented to obtain segmentation only some ofthe techniques are invented to obtain feature extraction andsome of the techniques are invented to obtain classificationonly Feature extraction and reduction of feature vectors foreffective segmentation of WM GM CSF and infected tumorregion and analysis on combined approach could not beconducted in all the published literatureMoreover only a fewfeatures are extracted and therefore very low accuracy intumor detection has been obtained Also all the above liter-atures are missing with the calculation of overlap that is dicesimilarity index which is one of the important parametersto judge the accuracy of any brain tumor segmentationalgorithm
In this study we perform a combination of biologicallyinspired Berkeley wavelet transformation (BWT) and SVMas a classifier tool to improve diagnostic accuracy The causeof this study is to extract information from the segmentedtumor region and classify healthy and infected tumor tissuesfor a large database of medical images Our results lead toconclude that the proposed method is suitable to integrateclinical decision support systems for primary screening anddiagnosis by the radiologists or clinical experts
3 Materials and Methods
This section presents the materials the source of brain MRimage dataset and the algorithm used to perform brain MRtissue segmentation Figure 1 provides the flow diagramof thealgorithm As test images different MR images of the brainwere used including T1-weightedMR images with RepetitionTime (TR) of 1740 and Echo Time (TE) of 20 T2-weightedMR images with Repetition Time (TR) of 5850 and EchoTime (TE) of 130 and FLAIR-weightedMR images with Rep-etition Time (TR) of 8500 and Echo Time (TE) of 130 Thesetest images were acquired using a 3 Tesla Siemens Magnetom
SpectraMRmachineThe total numbers of slices for all chan-nels were 15 which leads to total of 135 images at 9 slices orimages per patientwith a field of viewof 200mm an interslicegap of 1mm and voxel of size 078mm times 078mm times 05mmThe proposed methodology is applied to real dataset includ-ing brainMR images of 512 times 512 pixel size and was convertedinto grayscale before further processing The following sec-tions discuss the implementation of the algorithm
31 Preprocessing The primary task of preprocessing is toimprove the quality of the MR images and make it in a formsuited for further processing by human or machine visionsystem In addition preprocessing helps to improve certainparameters of MR images such as improving the signal-to-noise ratio enhancing the visual appearance of MR imageremoving the irrelevant noise and undesired parts in thebackground smoothing the inner part of the region andpreserving its edges [5] To improve the signal-to-noise ratioand thus the clarity of the rawMR images we applied adaptivecontrast enhancement based on modified sigmoid function[24]
32 Skull Stripping Skull stripping is an important process inbiomedical image analysis and it is required for the effectiveexamination of brain tumor from the MR images [25ndash28]Skull stripping is the process of eliminating all nonbraintissues in the brain images By skull stripping it is possible toremove additional cerebral tissues such as fat skin and skullin the brain imagesThere are several techniques available forskull stripping some of the popular techniques are automaticskull stripping using image contour skull stripping basedon segmentation and morphological operation and skullstripping based on histogram analysis or a threshold valueFigure 2 provides the stages of the skull stripping algorithmThis study uses the skull stripping technique that is based ona threshold operation to remove skull tissues
33 Segmentation and Morphological Operation The seg-mentation of the infected brain MR regions is achievedthrough the following steps In the first step the preprocessedbrain MR image is converted into a binary image with athreshold for the cut-off of 128 being selectedThepixel valuesgreater than the selected threshold are mapped to whitewhile others are marked as black due to this two differentregions are formed around the infected tumor tissues whichis cropped out In the second step in order to eliminatewhite pixel an erosion operation ofmorphology is employedFinally the eroded region and the original image are bothdivided into two equal regions and the black pixel regionextracted from the erode operation is counted as a brain MRimage mask In this study Berkeley wavelet transformation isemployed for effective segmentation of brain MR image
A wavelet is a function that is defined over a finiteinterval of time and has an average value of zero The wavelettransformation technique is employed to develop functionsoperators data or information into components of differentfrequency which enables studying each component sepa-rately All wavelets are generated from a basic wavelet Ψ(119905)
4 International Journal of Biomedical Imaging
Enhancement
Skull stripping
Removal of skull
Separation of GMWM CSF and tumor
Segmentation
Feature extraction
Mean contrastentropy and energy
Morphologicaloperation
Area extraction ampdecision making
Classification usingSVMMR image
dataset
Normal tissue Abnormal tissue
Pre processing
Figure 1 Steps used in proposed algorithm
Input image
Convert image to grayscale
Convert image to binary image by thresholding
Find the number of connected objects
Find mask by assigning 1 to inside and 0 to outsideof the object that show brain region
Multiply the mask with T1 T2 and FLAIR MR imagesto get their skull-stripped MR image
Figure 2 Steps used in the skull stripping algorithm
by using the scaling and translation process defined by (1) abasic wavelet is also referred to as a mother wavelet becauseit is the point of origin for other wavelets
Ψ119904120591 = 1radic119904Ψ(119905 minus 120591119904 ) (1)
where 119904 and 120591 are the scale and translation factors respec-tively
The Berkeley wavelet transform (BWT) [29 30] isdescribed as a two-dimensional triadic wavelet transformand can be used to process the signal or image Just like themother wavelet transformation or other families of wavelettransformation the BWT algorithm will also perform data
conversion from a spatial form into temporal domain fre-quency The BWT presents an effective way of representationof image transformation and it is a complete orthonormal[30] The mother wavelet transformation 120573120593
120579is piecewise
constant function [29 31] The substitute wavelets from themother wavelet 120573120593
120579are produced at various pixels positions in
the two-dimensional plane through scaling and translation ofthe mother wavelet and it is shown in
120573120593120579 (120591 119904) = 11199042120573120593119909 (3119904 (119909 minus 119894) 3119904 (119910 minus 119895)) (2)
where 120591 and 119904 are translation and scale parameter of thewavelet transformation respectively and 120573120593
120579is the trans-
forming function and it is called the mother wavelet ofBerkeley wavelet transformation The only single constantterm is sufficient to represent the mean value of an image thecoefficient value of the single term is shown in
1205730 = 1radic9 [119906 (1199093 1199103 )] (3)
The morphological operation is used for the extractionof the boundary areas of the brain images Conceptuallythe morphological operation is only rearranging the relativeorder of pixel values not on theirmathematical values and sois suitable to process only binary images Dilation and erosionare the two most basic operations of morphology Dilationoperations are intended to add pixels to the boundary regionof the object while erosion operations are intended to removethe pixels from the boundary region of the objects The oper-ation of addition and removing pixels to or from boundaryregion of the objects is based on the structuring element ofthe selected image
The experimented results produced by the proposedtechnique depicted for the segmented outcome for the threeclasses of WM GM and CSF and for the extracted tumor
International Journal of Biomedical Imaging 5
(a) (b) (c) (d) (e) (f)
(g) (h) (i) (j) (k) (l)
Figure 3 Segmented and area extracted result of brain MR image (a) Original image (b) Enhanced image (c) Skull-stripped image (d)Wavelet transpose image (e) Intense segmented image (f) Inverse intense image (g) Gray matter (h) White matter (i) CSF (j) Dice overlapimage (k) Eroded image (l) Area extracted image
region are given in Figure 3 The experimental results alsofind dice overlap image indicating the comparison betweenthe algorithm output and ground truth
34 Feature Extraction It is the process of collecting higher-level information of an image such as shape texture colorand contrast In fact texture analysis is an important parame-ter of human visual perception andmachine learning systemIt is used effectively to improve the accuracy of diagnosissystem by selecting prominent features Haralick et al [32]introduced one of the most widely used image analysisapplications of Gray Level Cooccurrence Matrix (GLCM)and texture feature This technique follows two steps forfeature extraction from the medical images In the first stepthe GLCM is computed and in the other step the texturefeatures based on the GLCM are calculated Due to theintricate structure of diversified tissues such asWMGM andCSF in the brain MR images extraction of relevant featuresis an essential task Textural findings and analysis couldimprove the diagnosis different stages of the tumor (tumorstaging) and therapy response assessment The statisticsfeature formula for some of the useful features is listed below
(1) Mean (M) The mean of an image is calculated by addingall the pixel values of an image divided by the total number ofpixels in an image
119872 = ( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) (4)
(2) Standard Deviation (SD) The standard deviation is thesecond central moment describing probability distributionof an observed population and can serve as a measure of
inhomogeneity A higher value indicates better intensity leveland high contrast of edges of an image
SD (120590) = radic( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
(119891 (119909 119910) minus119872)2 (5)
(3) Entropy (E) Entropy is calculated to characterize therandomness of the textural image and is defined as
(4) Skewness (119878119896) Skewness is a measure of symmetry or thelack of symmetry The skewness of a random variable 119883 isdenoted as 119878119896(119883) and it is defined as
119878119896 (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)310038161003816100381610038161003816SD3
(7)
(5) Kurtosis (119878119896)The shape of a random variablersquos probabilitydistribution is described by the parameter calledKurtosis Forthe randomvariable119883 theKurtosis is denoted as119870urt(119883) andit is defined as
119870urt (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)410038161003816100381610038161003816SD4
(8)
(6) Energy (En) Energy can be defined as the quantifiableamount of the extent of pixel pair repetitions Energy is aparameter to measure the similarity of an image If energy is
6 International Journal of Biomedical Imaging
defined by Haralicks GLCM feature then it is also referred toas angular second moment and it is defined as
En = radic119898minus1sum119909=0
119899minus1sum119910=0
1198912 (119909 119910) (9)
(7) Contrast (119862119900119899) Contrast is ameasure of intensity of a pixeland its neighbor over the image and it is defined as
119862on = 119898minus1sum119909=0
119899minus1sum119910=0
(119909 minus 119910)2 119891 (119909 119910) (10)
(8) Inverse DifferenceMoment (IDM) orHomogeneity InverseDifference Moment is a measure of the local homogeneity ofan image IDMmay have a single or a range of values so as todetermine whether the image is textured or nontextured
IDM = 119898minus1sum119909=0
119899minus1sum119910=0
11 + (119909 minus 119910)2119891 (119909 119910) (11)
(9) Directional Moment (DM) Directional moment is atextural property of the image calculated by considering thealignment of the image as ameasure in terms of the angle andit is defined as
DM = 119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) 1003816100381610038161003816119909 minus 1199101003816100381610038161003816 (12)
(10) Correlation (119862119900119903119903) Correlation feature describes thespatial dependencies between the pixels and it is defined as
where119872119909 and 120590119909 are the mean and standard deviation in thehorizontal spatial domain and119872119910 and 120590119910 are the mean andstandard deviation in the vertical spatial domain
(11) Coarseness (119862119899119890119904119904) Coarseness is a measure of roughnessin the textural analysis of an image For a fixed window sizea texture with a smaller number of texture elements is saidto be more coarse than the one with a larger number Therougher texture means higher coarseness value Fine textureshave smaller values of coarseness It is defined as
Apart from the above textural feature extraction thefollowing quality assessment parameters are also needed toensure better result analysis on brain MR images
(1) Structured Similarity Index (SSIM) The Structural Simi-larity Index (SSIM) is a perceptual metric that signifies thatthe degradation in image quality may be caused by data
compression or losses in data transmission or by any othermeans of the image processing It is defined as
A Higher value of SSIM indicates better preservation ofluminance contrast and structural content
(2)Mean Square Error (MSE) Mean square error is ameasureof signal fidelity or image fidelity The purpose of signal orimage fidelity measure is to find the similarity or fidelitybetween two images by providing the quantitative scoreWhen MSE is calculated then it is assumed that one of theimages is pristine original while the other is distorted orprocessed by some means and it is defined as
(3) Peak Signal-to-Noise Ratio (PSNR) in dB Peak signal-to-noise ratio is a measure used to assess the quality ofreconstruction of processed image and it is defined as
PSBR in dB = 20 log10 (2119899 minus 1)MSE (17)
Lower value ofMSE and higher value of PSNR indicate bettersignal-to-noise ratio
(4) Dice Coefficient Dice coefficient or dice similarity index isa measure of overlap between the two images and it is definedas
Dice (119860 119861) = 2 times 10038161003816100381610038161198601 and 11986111003816100381610038161003816(100381610038161003816100381611986011003816100381610038161003816 + 100381610038161003816100381611986111003816100381610038161003816) (18)
where 119860 isin 0 1 is tumor region extracted from algorithmicpredictions and 119861 isin 0 1 is the experts ground truth Theminimum value of dice coefficient is 0 and the maximumis 1 a higher value indicates better overlap between the twoimages
Tables 1 and 2 show some of the prominent features forthe first-order statistical and second-order statistical analysisTable 2 also indicates the measure of coarseness and numberof key values present in the segmented image
35 Support Vector Machine (SVM) The original SVM algo-rithmwas contributed by Vladimir N Vapnik and its modernversion was developed by Cortes and Vapnik in 1993 [33]The SVM algorithm is based on the study of a supervisedlearning technique and is applied to one-class classificationproblem to n-class classification problems [1 34ndash36] Theprinciple aim of the SVM algorithm is to transform a non-linear dividing objective into a linear transformation using a
International Journal of Biomedical Imaging 7
Table 1 First-order statistical features for few images
Images Mean Standard deviation Skewness Kurtosis Energy EntropyImage 1 866 4399 000553 289041119864 minus 06 1094 065Image 2 1181 4911 000655 274079119864 minus 06 1637 094Image 3 3940 7559 001054 18506119864 minus 06 6599 303Image 4 683 3945 000517 333685119864 minus 06 811 045Image 5 1190 3881 002002 135422119864 minus 05 3317 209Image 6 533 2895 001647 205493119864 minus 05 1387 112
Table 2 Second-order textural features with coarseness and key points for few images
function called SVMrsquos kernel function In this study we usedthe Gaussian kernel function for transformation By using akernel function the nonlinear samples can be transformedinto a high-dimensional future space where the separation ofnonlinear samples or datamight becomepossiblemaking theclassification convenient [16] The SVM algorithm defines ahyperplane that is divided into two training classes as definedin
where119885 and 119879 are hyperplane parameters and 120601(119910) is a func-tion used to map vector 119910 into a higher-dimensional spaceEquation (20) provides the Gaussian kernel function ofnonlinear SVM [16 34] used for the optimal solution of clas-sification and generalization and its advanced classificationfunction is shown in (21)
119896 (119910119894 119910119895) = exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172] (20)
(exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172]) (21)
where 119910119894 and 119910119895 are objects 119894 and 119895 respectively and 120574 is acontour parameter used to determine the smoothness of theboundary region [4 15]
The features selectionwith kernel class separabilitymakesSVM the default choice for classification of a brain tumorThe SVM algorithmrsquos performance can be evaluated in termsof accuracy sensitivity and specificity The confusion matrixdefining the terms TP TN FP and FN from the expectedoutcome and ground truth result for the calculation ofaccuracy sensitivity and specificity are shown in Table 3
Where TP is the number of true positives which is used toindicate the total number of abnormal cases correctly clas-sified TN is the number of true negatives which is used toindicate normal cases correctly classified FP is the number
Table 3 Confusion matrix defining the terms TP TN FP and FN
Expected outcome Ground truth Row totalPositive Negative
Table 4 Accuracy sensitivity and specificity calculation
Quality parameter Formula
Accuracy TP + TNTP + TN + FP + FN
Sensitivity TPTP + FN
Specificity TNTN + FP
of false positive and it is used to indicate wrongly detectedor classified abnormal cases when they are actually normalcases and FN is the number of false negatives it is used toindicate wrongly classified or detected normal cases whenthey are actually abnormal cases [15] all of these outcomeparameters are calculated using the total number of samplesexamined for the detection of the tumor The quality rateparameter accuracy is the proportion of total correctly classi-fied cases that are abnormally classified as abnormal andnormally classified as normal from the total number of casesexamined [37 38] Table 4 shows the formulas to calculateaccuracy sensitivity and specificity
4 Results and Discussion
To validate the performance of our algorithm we used twobenchmark datasets and one dataset collected from expertradiologists which included sample images of 15 patients
8 International Journal of Biomedical Imaging
Table 5 Performance analysis parameters for segmented tissues
Images MSE PSNR SSIM Dice scoreImage 1 186 5545 dB 08944 083Image 2 058 6821 dB 09025 087Image 3 495 5628 dB 09702 082Image 4 123 5879 dB 08801 079Image 5 506 5965 dB 07978 090
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 4 Experimental results of image 1 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
with 9 slices for each patient The first dataset is the DigitalImaging and Communications in Medicine (DICOM) data-set [39] For the purpose of the analysis we considered 22images from the DICOM dataset all of which included aretumor-infected brain tissues However this dataset did nothave any ground truth imagesThe second dataset is the BrainWeb dataset [40] which consists of full three-dimensionalsimulated brain MR data obtained using three sequences ofmodalities namely T1-weightedMRI T2-weightedMRI andproton density-weightedMRIThis dataset included a varietyof slice thicknesses noise levels and levels of intensitynonuniformity The images used for our analysis are mostlyincludedT2-weightedmodality with 1mm slice thickness 3noise and 20 intensity nonuniformity In this dataset 13out of 44 images included are tumor-infected brain tissuesThe last dataset collected from expert radiologists consisted
of 135 images of 15 patients with all modalities This datasethad ground truth images that helped to compare the resultsof our method with the manual analysis of radiologists
This section presents the results of our proposed imagesegmentation technique which are obtained by using realbrain MR images The proposed algorithm was carried outusing Matlab 7120 (R2011a) which runs on the Windows 8operating system and has an Intel core i3 processor and a4GB RAM The sample experimental results obtained fromthe proposed technique that are depicted in Figures 4 5 and6 show the original image along with enhanced image skull-stripped image wavelet decompose image cluster (intense)segmented image dice overlap image and the tumor regionwith extracted area mark
Table 5 provides the details of the different performanceparameters such as mean squared error (MSE) and peak
International Journal of Biomedical Imaging 9
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
Find mask by assigning 1 to inside and 0 to outsideof the object that show brain region
Multiply the mask with T1 T2 and FLAIR MR imagesto get their skull-stripped MR image
Figure 2 Steps used in the skull stripping algorithm
by using the scaling and translation process defined by (1) abasic wavelet is also referred to as a mother wavelet becauseit is the point of origin for other wavelets
Ψ119904120591 = 1radic119904Ψ(119905 minus 120591119904 ) (1)
where 119904 and 120591 are the scale and translation factors respec-tively
The Berkeley wavelet transform (BWT) [29 30] isdescribed as a two-dimensional triadic wavelet transformand can be used to process the signal or image Just like themother wavelet transformation or other families of wavelettransformation the BWT algorithm will also perform data
conversion from a spatial form into temporal domain fre-quency The BWT presents an effective way of representationof image transformation and it is a complete orthonormal[30] The mother wavelet transformation 120573120593
120579is piecewise
constant function [29 31] The substitute wavelets from themother wavelet 120573120593
120579are produced at various pixels positions in
the two-dimensional plane through scaling and translation ofthe mother wavelet and it is shown in
120573120593120579 (120591 119904) = 11199042120573120593119909 (3119904 (119909 minus 119894) 3119904 (119910 minus 119895)) (2)
where 120591 and 119904 are translation and scale parameter of thewavelet transformation respectively and 120573120593
120579is the trans-
forming function and it is called the mother wavelet ofBerkeley wavelet transformation The only single constantterm is sufficient to represent the mean value of an image thecoefficient value of the single term is shown in
1205730 = 1radic9 [119906 (1199093 1199103 )] (3)
The morphological operation is used for the extractionof the boundary areas of the brain images Conceptuallythe morphological operation is only rearranging the relativeorder of pixel values not on theirmathematical values and sois suitable to process only binary images Dilation and erosionare the two most basic operations of morphology Dilationoperations are intended to add pixels to the boundary regionof the object while erosion operations are intended to removethe pixels from the boundary region of the objects The oper-ation of addition and removing pixels to or from boundaryregion of the objects is based on the structuring element ofthe selected image
The experimented results produced by the proposedtechnique depicted for the segmented outcome for the threeclasses of WM GM and CSF and for the extracted tumor
International Journal of Biomedical Imaging 5
(a) (b) (c) (d) (e) (f)
(g) (h) (i) (j) (k) (l)
Figure 3 Segmented and area extracted result of brain MR image (a) Original image (b) Enhanced image (c) Skull-stripped image (d)Wavelet transpose image (e) Intense segmented image (f) Inverse intense image (g) Gray matter (h) White matter (i) CSF (j) Dice overlapimage (k) Eroded image (l) Area extracted image
region are given in Figure 3 The experimental results alsofind dice overlap image indicating the comparison betweenthe algorithm output and ground truth
34 Feature Extraction It is the process of collecting higher-level information of an image such as shape texture colorand contrast In fact texture analysis is an important parame-ter of human visual perception andmachine learning systemIt is used effectively to improve the accuracy of diagnosissystem by selecting prominent features Haralick et al [32]introduced one of the most widely used image analysisapplications of Gray Level Cooccurrence Matrix (GLCM)and texture feature This technique follows two steps forfeature extraction from the medical images In the first stepthe GLCM is computed and in the other step the texturefeatures based on the GLCM are calculated Due to theintricate structure of diversified tissues such asWMGM andCSF in the brain MR images extraction of relevant featuresis an essential task Textural findings and analysis couldimprove the diagnosis different stages of the tumor (tumorstaging) and therapy response assessment The statisticsfeature formula for some of the useful features is listed below
(1) Mean (M) The mean of an image is calculated by addingall the pixel values of an image divided by the total number ofpixels in an image
119872 = ( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) (4)
(2) Standard Deviation (SD) The standard deviation is thesecond central moment describing probability distributionof an observed population and can serve as a measure of
inhomogeneity A higher value indicates better intensity leveland high contrast of edges of an image
SD (120590) = radic( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
(119891 (119909 119910) minus119872)2 (5)
(3) Entropy (E) Entropy is calculated to characterize therandomness of the textural image and is defined as
(4) Skewness (119878119896) Skewness is a measure of symmetry or thelack of symmetry The skewness of a random variable 119883 isdenoted as 119878119896(119883) and it is defined as
119878119896 (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)310038161003816100381610038161003816SD3
(7)
(5) Kurtosis (119878119896)The shape of a random variablersquos probabilitydistribution is described by the parameter calledKurtosis Forthe randomvariable119883 theKurtosis is denoted as119870urt(119883) andit is defined as
119870urt (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)410038161003816100381610038161003816SD4
(8)
(6) Energy (En) Energy can be defined as the quantifiableamount of the extent of pixel pair repetitions Energy is aparameter to measure the similarity of an image If energy is
6 International Journal of Biomedical Imaging
defined by Haralicks GLCM feature then it is also referred toas angular second moment and it is defined as
En = radic119898minus1sum119909=0
119899minus1sum119910=0
1198912 (119909 119910) (9)
(7) Contrast (119862119900119899) Contrast is ameasure of intensity of a pixeland its neighbor over the image and it is defined as
119862on = 119898minus1sum119909=0
119899minus1sum119910=0
(119909 minus 119910)2 119891 (119909 119910) (10)
(8) Inverse DifferenceMoment (IDM) orHomogeneity InverseDifference Moment is a measure of the local homogeneity ofan image IDMmay have a single or a range of values so as todetermine whether the image is textured or nontextured
IDM = 119898minus1sum119909=0
119899minus1sum119910=0
11 + (119909 minus 119910)2119891 (119909 119910) (11)
(9) Directional Moment (DM) Directional moment is atextural property of the image calculated by considering thealignment of the image as ameasure in terms of the angle andit is defined as
DM = 119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) 1003816100381610038161003816119909 minus 1199101003816100381610038161003816 (12)
(10) Correlation (119862119900119903119903) Correlation feature describes thespatial dependencies between the pixels and it is defined as
where119872119909 and 120590119909 are the mean and standard deviation in thehorizontal spatial domain and119872119910 and 120590119910 are the mean andstandard deviation in the vertical spatial domain
(11) Coarseness (119862119899119890119904119904) Coarseness is a measure of roughnessin the textural analysis of an image For a fixed window sizea texture with a smaller number of texture elements is saidto be more coarse than the one with a larger number Therougher texture means higher coarseness value Fine textureshave smaller values of coarseness It is defined as
Apart from the above textural feature extraction thefollowing quality assessment parameters are also needed toensure better result analysis on brain MR images
(1) Structured Similarity Index (SSIM) The Structural Simi-larity Index (SSIM) is a perceptual metric that signifies thatthe degradation in image quality may be caused by data
compression or losses in data transmission or by any othermeans of the image processing It is defined as
A Higher value of SSIM indicates better preservation ofluminance contrast and structural content
(2)Mean Square Error (MSE) Mean square error is ameasureof signal fidelity or image fidelity The purpose of signal orimage fidelity measure is to find the similarity or fidelitybetween two images by providing the quantitative scoreWhen MSE is calculated then it is assumed that one of theimages is pristine original while the other is distorted orprocessed by some means and it is defined as
(3) Peak Signal-to-Noise Ratio (PSNR) in dB Peak signal-to-noise ratio is a measure used to assess the quality ofreconstruction of processed image and it is defined as
PSBR in dB = 20 log10 (2119899 minus 1)MSE (17)
Lower value ofMSE and higher value of PSNR indicate bettersignal-to-noise ratio
(4) Dice Coefficient Dice coefficient or dice similarity index isa measure of overlap between the two images and it is definedas
Dice (119860 119861) = 2 times 10038161003816100381610038161198601 and 11986111003816100381610038161003816(100381610038161003816100381611986011003816100381610038161003816 + 100381610038161003816100381611986111003816100381610038161003816) (18)
where 119860 isin 0 1 is tumor region extracted from algorithmicpredictions and 119861 isin 0 1 is the experts ground truth Theminimum value of dice coefficient is 0 and the maximumis 1 a higher value indicates better overlap between the twoimages
Tables 1 and 2 show some of the prominent features forthe first-order statistical and second-order statistical analysisTable 2 also indicates the measure of coarseness and numberof key values present in the segmented image
35 Support Vector Machine (SVM) The original SVM algo-rithmwas contributed by Vladimir N Vapnik and its modernversion was developed by Cortes and Vapnik in 1993 [33]The SVM algorithm is based on the study of a supervisedlearning technique and is applied to one-class classificationproblem to n-class classification problems [1 34ndash36] Theprinciple aim of the SVM algorithm is to transform a non-linear dividing objective into a linear transformation using a
International Journal of Biomedical Imaging 7
Table 1 First-order statistical features for few images
Images Mean Standard deviation Skewness Kurtosis Energy EntropyImage 1 866 4399 000553 289041119864 minus 06 1094 065Image 2 1181 4911 000655 274079119864 minus 06 1637 094Image 3 3940 7559 001054 18506119864 minus 06 6599 303Image 4 683 3945 000517 333685119864 minus 06 811 045Image 5 1190 3881 002002 135422119864 minus 05 3317 209Image 6 533 2895 001647 205493119864 minus 05 1387 112
Table 2 Second-order textural features with coarseness and key points for few images
function called SVMrsquos kernel function In this study we usedthe Gaussian kernel function for transformation By using akernel function the nonlinear samples can be transformedinto a high-dimensional future space where the separation ofnonlinear samples or datamight becomepossiblemaking theclassification convenient [16] The SVM algorithm defines ahyperplane that is divided into two training classes as definedin
where119885 and 119879 are hyperplane parameters and 120601(119910) is a func-tion used to map vector 119910 into a higher-dimensional spaceEquation (20) provides the Gaussian kernel function ofnonlinear SVM [16 34] used for the optimal solution of clas-sification and generalization and its advanced classificationfunction is shown in (21)
119896 (119910119894 119910119895) = exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172] (20)
(exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172]) (21)
where 119910119894 and 119910119895 are objects 119894 and 119895 respectively and 120574 is acontour parameter used to determine the smoothness of theboundary region [4 15]
The features selectionwith kernel class separabilitymakesSVM the default choice for classification of a brain tumorThe SVM algorithmrsquos performance can be evaluated in termsof accuracy sensitivity and specificity The confusion matrixdefining the terms TP TN FP and FN from the expectedoutcome and ground truth result for the calculation ofaccuracy sensitivity and specificity are shown in Table 3
Where TP is the number of true positives which is used toindicate the total number of abnormal cases correctly clas-sified TN is the number of true negatives which is used toindicate normal cases correctly classified FP is the number
Table 3 Confusion matrix defining the terms TP TN FP and FN
Expected outcome Ground truth Row totalPositive Negative
Table 4 Accuracy sensitivity and specificity calculation
Quality parameter Formula
Accuracy TP + TNTP + TN + FP + FN
Sensitivity TPTP + FN
Specificity TNTN + FP
of false positive and it is used to indicate wrongly detectedor classified abnormal cases when they are actually normalcases and FN is the number of false negatives it is used toindicate wrongly classified or detected normal cases whenthey are actually abnormal cases [15] all of these outcomeparameters are calculated using the total number of samplesexamined for the detection of the tumor The quality rateparameter accuracy is the proportion of total correctly classi-fied cases that are abnormally classified as abnormal andnormally classified as normal from the total number of casesexamined [37 38] Table 4 shows the formulas to calculateaccuracy sensitivity and specificity
4 Results and Discussion
To validate the performance of our algorithm we used twobenchmark datasets and one dataset collected from expertradiologists which included sample images of 15 patients
8 International Journal of Biomedical Imaging
Table 5 Performance analysis parameters for segmented tissues
Images MSE PSNR SSIM Dice scoreImage 1 186 5545 dB 08944 083Image 2 058 6821 dB 09025 087Image 3 495 5628 dB 09702 082Image 4 123 5879 dB 08801 079Image 5 506 5965 dB 07978 090
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 4 Experimental results of image 1 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
with 9 slices for each patient The first dataset is the DigitalImaging and Communications in Medicine (DICOM) data-set [39] For the purpose of the analysis we considered 22images from the DICOM dataset all of which included aretumor-infected brain tissues However this dataset did nothave any ground truth imagesThe second dataset is the BrainWeb dataset [40] which consists of full three-dimensionalsimulated brain MR data obtained using three sequences ofmodalities namely T1-weightedMRI T2-weightedMRI andproton density-weightedMRIThis dataset included a varietyof slice thicknesses noise levels and levels of intensitynonuniformity The images used for our analysis are mostlyincludedT2-weightedmodality with 1mm slice thickness 3noise and 20 intensity nonuniformity In this dataset 13out of 44 images included are tumor-infected brain tissuesThe last dataset collected from expert radiologists consisted
of 135 images of 15 patients with all modalities This datasethad ground truth images that helped to compare the resultsof our method with the manual analysis of radiologists
This section presents the results of our proposed imagesegmentation technique which are obtained by using realbrain MR images The proposed algorithm was carried outusing Matlab 7120 (R2011a) which runs on the Windows 8operating system and has an Intel core i3 processor and a4GB RAM The sample experimental results obtained fromthe proposed technique that are depicted in Figures 4 5 and6 show the original image along with enhanced image skull-stripped image wavelet decompose image cluster (intense)segmented image dice overlap image and the tumor regionwith extracted area mark
Table 5 provides the details of the different performanceparameters such as mean squared error (MSE) and peak
International Journal of Biomedical Imaging 9
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
Figure 3 Segmented and area extracted result of brain MR image (a) Original image (b) Enhanced image (c) Skull-stripped image (d)Wavelet transpose image (e) Intense segmented image (f) Inverse intense image (g) Gray matter (h) White matter (i) CSF (j) Dice overlapimage (k) Eroded image (l) Area extracted image
region are given in Figure 3 The experimental results alsofind dice overlap image indicating the comparison betweenthe algorithm output and ground truth
34 Feature Extraction It is the process of collecting higher-level information of an image such as shape texture colorand contrast In fact texture analysis is an important parame-ter of human visual perception andmachine learning systemIt is used effectively to improve the accuracy of diagnosissystem by selecting prominent features Haralick et al [32]introduced one of the most widely used image analysisapplications of Gray Level Cooccurrence Matrix (GLCM)and texture feature This technique follows two steps forfeature extraction from the medical images In the first stepthe GLCM is computed and in the other step the texturefeatures based on the GLCM are calculated Due to theintricate structure of diversified tissues such asWMGM andCSF in the brain MR images extraction of relevant featuresis an essential task Textural findings and analysis couldimprove the diagnosis different stages of the tumor (tumorstaging) and therapy response assessment The statisticsfeature formula for some of the useful features is listed below
(1) Mean (M) The mean of an image is calculated by addingall the pixel values of an image divided by the total number ofpixels in an image
119872 = ( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) (4)
(2) Standard Deviation (SD) The standard deviation is thesecond central moment describing probability distributionof an observed population and can serve as a measure of
inhomogeneity A higher value indicates better intensity leveland high contrast of edges of an image
SD (120590) = radic( 1119898 times 119899)119898minus1sum119909=0
119899minus1sum119910=0
(119891 (119909 119910) minus119872)2 (5)
(3) Entropy (E) Entropy is calculated to characterize therandomness of the textural image and is defined as
(4) Skewness (119878119896) Skewness is a measure of symmetry or thelack of symmetry The skewness of a random variable 119883 isdenoted as 119878119896(119883) and it is defined as
119878119896 (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)310038161003816100381610038161003816SD3
(7)
(5) Kurtosis (119878119896)The shape of a random variablersquos probabilitydistribution is described by the parameter calledKurtosis Forthe randomvariable119883 theKurtosis is denoted as119870urt(119883) andit is defined as
119870urt (119883) = ( 1119898 times 119899) sum (119891 (119909 119910) minus119872)410038161003816100381610038161003816SD4
(8)
(6) Energy (En) Energy can be defined as the quantifiableamount of the extent of pixel pair repetitions Energy is aparameter to measure the similarity of an image If energy is
6 International Journal of Biomedical Imaging
defined by Haralicks GLCM feature then it is also referred toas angular second moment and it is defined as
En = radic119898minus1sum119909=0
119899minus1sum119910=0
1198912 (119909 119910) (9)
(7) Contrast (119862119900119899) Contrast is ameasure of intensity of a pixeland its neighbor over the image and it is defined as
119862on = 119898minus1sum119909=0
119899minus1sum119910=0
(119909 minus 119910)2 119891 (119909 119910) (10)
(8) Inverse DifferenceMoment (IDM) orHomogeneity InverseDifference Moment is a measure of the local homogeneity ofan image IDMmay have a single or a range of values so as todetermine whether the image is textured or nontextured
IDM = 119898minus1sum119909=0
119899minus1sum119910=0
11 + (119909 minus 119910)2119891 (119909 119910) (11)
(9) Directional Moment (DM) Directional moment is atextural property of the image calculated by considering thealignment of the image as ameasure in terms of the angle andit is defined as
DM = 119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) 1003816100381610038161003816119909 minus 1199101003816100381610038161003816 (12)
(10) Correlation (119862119900119903119903) Correlation feature describes thespatial dependencies between the pixels and it is defined as
where119872119909 and 120590119909 are the mean and standard deviation in thehorizontal spatial domain and119872119910 and 120590119910 are the mean andstandard deviation in the vertical spatial domain
(11) Coarseness (119862119899119890119904119904) Coarseness is a measure of roughnessin the textural analysis of an image For a fixed window sizea texture with a smaller number of texture elements is saidto be more coarse than the one with a larger number Therougher texture means higher coarseness value Fine textureshave smaller values of coarseness It is defined as
Apart from the above textural feature extraction thefollowing quality assessment parameters are also needed toensure better result analysis on brain MR images
(1) Structured Similarity Index (SSIM) The Structural Simi-larity Index (SSIM) is a perceptual metric that signifies thatthe degradation in image quality may be caused by data
compression or losses in data transmission or by any othermeans of the image processing It is defined as
A Higher value of SSIM indicates better preservation ofluminance contrast and structural content
(2)Mean Square Error (MSE) Mean square error is ameasureof signal fidelity or image fidelity The purpose of signal orimage fidelity measure is to find the similarity or fidelitybetween two images by providing the quantitative scoreWhen MSE is calculated then it is assumed that one of theimages is pristine original while the other is distorted orprocessed by some means and it is defined as
(3) Peak Signal-to-Noise Ratio (PSNR) in dB Peak signal-to-noise ratio is a measure used to assess the quality ofreconstruction of processed image and it is defined as
PSBR in dB = 20 log10 (2119899 minus 1)MSE (17)
Lower value ofMSE and higher value of PSNR indicate bettersignal-to-noise ratio
(4) Dice Coefficient Dice coefficient or dice similarity index isa measure of overlap between the two images and it is definedas
Dice (119860 119861) = 2 times 10038161003816100381610038161198601 and 11986111003816100381610038161003816(100381610038161003816100381611986011003816100381610038161003816 + 100381610038161003816100381611986111003816100381610038161003816) (18)
where 119860 isin 0 1 is tumor region extracted from algorithmicpredictions and 119861 isin 0 1 is the experts ground truth Theminimum value of dice coefficient is 0 and the maximumis 1 a higher value indicates better overlap between the twoimages
Tables 1 and 2 show some of the prominent features forthe first-order statistical and second-order statistical analysisTable 2 also indicates the measure of coarseness and numberof key values present in the segmented image
35 Support Vector Machine (SVM) The original SVM algo-rithmwas contributed by Vladimir N Vapnik and its modernversion was developed by Cortes and Vapnik in 1993 [33]The SVM algorithm is based on the study of a supervisedlearning technique and is applied to one-class classificationproblem to n-class classification problems [1 34ndash36] Theprinciple aim of the SVM algorithm is to transform a non-linear dividing objective into a linear transformation using a
International Journal of Biomedical Imaging 7
Table 1 First-order statistical features for few images
Images Mean Standard deviation Skewness Kurtosis Energy EntropyImage 1 866 4399 000553 289041119864 minus 06 1094 065Image 2 1181 4911 000655 274079119864 minus 06 1637 094Image 3 3940 7559 001054 18506119864 minus 06 6599 303Image 4 683 3945 000517 333685119864 minus 06 811 045Image 5 1190 3881 002002 135422119864 minus 05 3317 209Image 6 533 2895 001647 205493119864 minus 05 1387 112
Table 2 Second-order textural features with coarseness and key points for few images
function called SVMrsquos kernel function In this study we usedthe Gaussian kernel function for transformation By using akernel function the nonlinear samples can be transformedinto a high-dimensional future space where the separation ofnonlinear samples or datamight becomepossiblemaking theclassification convenient [16] The SVM algorithm defines ahyperplane that is divided into two training classes as definedin
where119885 and 119879 are hyperplane parameters and 120601(119910) is a func-tion used to map vector 119910 into a higher-dimensional spaceEquation (20) provides the Gaussian kernel function ofnonlinear SVM [16 34] used for the optimal solution of clas-sification and generalization and its advanced classificationfunction is shown in (21)
119896 (119910119894 119910119895) = exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172] (20)
(exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172]) (21)
where 119910119894 and 119910119895 are objects 119894 and 119895 respectively and 120574 is acontour parameter used to determine the smoothness of theboundary region [4 15]
The features selectionwith kernel class separabilitymakesSVM the default choice for classification of a brain tumorThe SVM algorithmrsquos performance can be evaluated in termsof accuracy sensitivity and specificity The confusion matrixdefining the terms TP TN FP and FN from the expectedoutcome and ground truth result for the calculation ofaccuracy sensitivity and specificity are shown in Table 3
Where TP is the number of true positives which is used toindicate the total number of abnormal cases correctly clas-sified TN is the number of true negatives which is used toindicate normal cases correctly classified FP is the number
Table 3 Confusion matrix defining the terms TP TN FP and FN
Expected outcome Ground truth Row totalPositive Negative
Table 4 Accuracy sensitivity and specificity calculation
Quality parameter Formula
Accuracy TP + TNTP + TN + FP + FN
Sensitivity TPTP + FN
Specificity TNTN + FP
of false positive and it is used to indicate wrongly detectedor classified abnormal cases when they are actually normalcases and FN is the number of false negatives it is used toindicate wrongly classified or detected normal cases whenthey are actually abnormal cases [15] all of these outcomeparameters are calculated using the total number of samplesexamined for the detection of the tumor The quality rateparameter accuracy is the proportion of total correctly classi-fied cases that are abnormally classified as abnormal andnormally classified as normal from the total number of casesexamined [37 38] Table 4 shows the formulas to calculateaccuracy sensitivity and specificity
4 Results and Discussion
To validate the performance of our algorithm we used twobenchmark datasets and one dataset collected from expertradiologists which included sample images of 15 patients
8 International Journal of Biomedical Imaging
Table 5 Performance analysis parameters for segmented tissues
Images MSE PSNR SSIM Dice scoreImage 1 186 5545 dB 08944 083Image 2 058 6821 dB 09025 087Image 3 495 5628 dB 09702 082Image 4 123 5879 dB 08801 079Image 5 506 5965 dB 07978 090
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 4 Experimental results of image 1 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
with 9 slices for each patient The first dataset is the DigitalImaging and Communications in Medicine (DICOM) data-set [39] For the purpose of the analysis we considered 22images from the DICOM dataset all of which included aretumor-infected brain tissues However this dataset did nothave any ground truth imagesThe second dataset is the BrainWeb dataset [40] which consists of full three-dimensionalsimulated brain MR data obtained using three sequences ofmodalities namely T1-weightedMRI T2-weightedMRI andproton density-weightedMRIThis dataset included a varietyof slice thicknesses noise levels and levels of intensitynonuniformity The images used for our analysis are mostlyincludedT2-weightedmodality with 1mm slice thickness 3noise and 20 intensity nonuniformity In this dataset 13out of 44 images included are tumor-infected brain tissuesThe last dataset collected from expert radiologists consisted
of 135 images of 15 patients with all modalities This datasethad ground truth images that helped to compare the resultsof our method with the manual analysis of radiologists
This section presents the results of our proposed imagesegmentation technique which are obtained by using realbrain MR images The proposed algorithm was carried outusing Matlab 7120 (R2011a) which runs on the Windows 8operating system and has an Intel core i3 processor and a4GB RAM The sample experimental results obtained fromthe proposed technique that are depicted in Figures 4 5 and6 show the original image along with enhanced image skull-stripped image wavelet decompose image cluster (intense)segmented image dice overlap image and the tumor regionwith extracted area mark
Table 5 provides the details of the different performanceparameters such as mean squared error (MSE) and peak
International Journal of Biomedical Imaging 9
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
defined by Haralicks GLCM feature then it is also referred toas angular second moment and it is defined as
En = radic119898minus1sum119909=0
119899minus1sum119910=0
1198912 (119909 119910) (9)
(7) Contrast (119862119900119899) Contrast is ameasure of intensity of a pixeland its neighbor over the image and it is defined as
119862on = 119898minus1sum119909=0
119899minus1sum119910=0
(119909 minus 119910)2 119891 (119909 119910) (10)
(8) Inverse DifferenceMoment (IDM) orHomogeneity InverseDifference Moment is a measure of the local homogeneity ofan image IDMmay have a single or a range of values so as todetermine whether the image is textured or nontextured
IDM = 119898minus1sum119909=0
119899minus1sum119910=0
11 + (119909 minus 119910)2119891 (119909 119910) (11)
(9) Directional Moment (DM) Directional moment is atextural property of the image calculated by considering thealignment of the image as ameasure in terms of the angle andit is defined as
DM = 119898minus1sum119909=0
119899minus1sum119910=0
119891 (119909 119910) 1003816100381610038161003816119909 minus 1199101003816100381610038161003816 (12)
(10) Correlation (119862119900119903119903) Correlation feature describes thespatial dependencies between the pixels and it is defined as
where119872119909 and 120590119909 are the mean and standard deviation in thehorizontal spatial domain and119872119910 and 120590119910 are the mean andstandard deviation in the vertical spatial domain
(11) Coarseness (119862119899119890119904119904) Coarseness is a measure of roughnessin the textural analysis of an image For a fixed window sizea texture with a smaller number of texture elements is saidto be more coarse than the one with a larger number Therougher texture means higher coarseness value Fine textureshave smaller values of coarseness It is defined as
Apart from the above textural feature extraction thefollowing quality assessment parameters are also needed toensure better result analysis on brain MR images
(1) Structured Similarity Index (SSIM) The Structural Simi-larity Index (SSIM) is a perceptual metric that signifies thatthe degradation in image quality may be caused by data
compression or losses in data transmission or by any othermeans of the image processing It is defined as
A Higher value of SSIM indicates better preservation ofluminance contrast and structural content
(2)Mean Square Error (MSE) Mean square error is ameasureof signal fidelity or image fidelity The purpose of signal orimage fidelity measure is to find the similarity or fidelitybetween two images by providing the quantitative scoreWhen MSE is calculated then it is assumed that one of theimages is pristine original while the other is distorted orprocessed by some means and it is defined as
(3) Peak Signal-to-Noise Ratio (PSNR) in dB Peak signal-to-noise ratio is a measure used to assess the quality ofreconstruction of processed image and it is defined as
PSBR in dB = 20 log10 (2119899 minus 1)MSE (17)
Lower value ofMSE and higher value of PSNR indicate bettersignal-to-noise ratio
(4) Dice Coefficient Dice coefficient or dice similarity index isa measure of overlap between the two images and it is definedas
Dice (119860 119861) = 2 times 10038161003816100381610038161198601 and 11986111003816100381610038161003816(100381610038161003816100381611986011003816100381610038161003816 + 100381610038161003816100381611986111003816100381610038161003816) (18)
where 119860 isin 0 1 is tumor region extracted from algorithmicpredictions and 119861 isin 0 1 is the experts ground truth Theminimum value of dice coefficient is 0 and the maximumis 1 a higher value indicates better overlap between the twoimages
Tables 1 and 2 show some of the prominent features forthe first-order statistical and second-order statistical analysisTable 2 also indicates the measure of coarseness and numberof key values present in the segmented image
35 Support Vector Machine (SVM) The original SVM algo-rithmwas contributed by Vladimir N Vapnik and its modernversion was developed by Cortes and Vapnik in 1993 [33]The SVM algorithm is based on the study of a supervisedlearning technique and is applied to one-class classificationproblem to n-class classification problems [1 34ndash36] Theprinciple aim of the SVM algorithm is to transform a non-linear dividing objective into a linear transformation using a
International Journal of Biomedical Imaging 7
Table 1 First-order statistical features for few images
Images Mean Standard deviation Skewness Kurtosis Energy EntropyImage 1 866 4399 000553 289041119864 minus 06 1094 065Image 2 1181 4911 000655 274079119864 minus 06 1637 094Image 3 3940 7559 001054 18506119864 minus 06 6599 303Image 4 683 3945 000517 333685119864 minus 06 811 045Image 5 1190 3881 002002 135422119864 minus 05 3317 209Image 6 533 2895 001647 205493119864 minus 05 1387 112
Table 2 Second-order textural features with coarseness and key points for few images
function called SVMrsquos kernel function In this study we usedthe Gaussian kernel function for transformation By using akernel function the nonlinear samples can be transformedinto a high-dimensional future space where the separation ofnonlinear samples or datamight becomepossiblemaking theclassification convenient [16] The SVM algorithm defines ahyperplane that is divided into two training classes as definedin
where119885 and 119879 are hyperplane parameters and 120601(119910) is a func-tion used to map vector 119910 into a higher-dimensional spaceEquation (20) provides the Gaussian kernel function ofnonlinear SVM [16 34] used for the optimal solution of clas-sification and generalization and its advanced classificationfunction is shown in (21)
119896 (119910119894 119910119895) = exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172] (20)
(exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172]) (21)
where 119910119894 and 119910119895 are objects 119894 and 119895 respectively and 120574 is acontour parameter used to determine the smoothness of theboundary region [4 15]
The features selectionwith kernel class separabilitymakesSVM the default choice for classification of a brain tumorThe SVM algorithmrsquos performance can be evaluated in termsof accuracy sensitivity and specificity The confusion matrixdefining the terms TP TN FP and FN from the expectedoutcome and ground truth result for the calculation ofaccuracy sensitivity and specificity are shown in Table 3
Where TP is the number of true positives which is used toindicate the total number of abnormal cases correctly clas-sified TN is the number of true negatives which is used toindicate normal cases correctly classified FP is the number
Table 3 Confusion matrix defining the terms TP TN FP and FN
Expected outcome Ground truth Row totalPositive Negative
Table 4 Accuracy sensitivity and specificity calculation
Quality parameter Formula
Accuracy TP + TNTP + TN + FP + FN
Sensitivity TPTP + FN
Specificity TNTN + FP
of false positive and it is used to indicate wrongly detectedor classified abnormal cases when they are actually normalcases and FN is the number of false negatives it is used toindicate wrongly classified or detected normal cases whenthey are actually abnormal cases [15] all of these outcomeparameters are calculated using the total number of samplesexamined for the detection of the tumor The quality rateparameter accuracy is the proportion of total correctly classi-fied cases that are abnormally classified as abnormal andnormally classified as normal from the total number of casesexamined [37 38] Table 4 shows the formulas to calculateaccuracy sensitivity and specificity
4 Results and Discussion
To validate the performance of our algorithm we used twobenchmark datasets and one dataset collected from expertradiologists which included sample images of 15 patients
8 International Journal of Biomedical Imaging
Table 5 Performance analysis parameters for segmented tissues
Images MSE PSNR SSIM Dice scoreImage 1 186 5545 dB 08944 083Image 2 058 6821 dB 09025 087Image 3 495 5628 dB 09702 082Image 4 123 5879 dB 08801 079Image 5 506 5965 dB 07978 090
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 4 Experimental results of image 1 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
with 9 slices for each patient The first dataset is the DigitalImaging and Communications in Medicine (DICOM) data-set [39] For the purpose of the analysis we considered 22images from the DICOM dataset all of which included aretumor-infected brain tissues However this dataset did nothave any ground truth imagesThe second dataset is the BrainWeb dataset [40] which consists of full three-dimensionalsimulated brain MR data obtained using three sequences ofmodalities namely T1-weightedMRI T2-weightedMRI andproton density-weightedMRIThis dataset included a varietyof slice thicknesses noise levels and levels of intensitynonuniformity The images used for our analysis are mostlyincludedT2-weightedmodality with 1mm slice thickness 3noise and 20 intensity nonuniformity In this dataset 13out of 44 images included are tumor-infected brain tissuesThe last dataset collected from expert radiologists consisted
of 135 images of 15 patients with all modalities This datasethad ground truth images that helped to compare the resultsof our method with the manual analysis of radiologists
This section presents the results of our proposed imagesegmentation technique which are obtained by using realbrain MR images The proposed algorithm was carried outusing Matlab 7120 (R2011a) which runs on the Windows 8operating system and has an Intel core i3 processor and a4GB RAM The sample experimental results obtained fromthe proposed technique that are depicted in Figures 4 5 and6 show the original image along with enhanced image skull-stripped image wavelet decompose image cluster (intense)segmented image dice overlap image and the tumor regionwith extracted area mark
Table 5 provides the details of the different performanceparameters such as mean squared error (MSE) and peak
International Journal of Biomedical Imaging 9
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
function called SVMrsquos kernel function In this study we usedthe Gaussian kernel function for transformation By using akernel function the nonlinear samples can be transformedinto a high-dimensional future space where the separation ofnonlinear samples or datamight becomepossiblemaking theclassification convenient [16] The SVM algorithm defines ahyperplane that is divided into two training classes as definedin
where119885 and 119879 are hyperplane parameters and 120601(119910) is a func-tion used to map vector 119910 into a higher-dimensional spaceEquation (20) provides the Gaussian kernel function ofnonlinear SVM [16 34] used for the optimal solution of clas-sification and generalization and its advanced classificationfunction is shown in (21)
119896 (119910119894 119910119895) = exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172] (20)
(exp [minus120574 10038171003817100381710038171003817119910119894 minus 119910119895100381710038171003817100381710038172]) (21)
where 119910119894 and 119910119895 are objects 119894 and 119895 respectively and 120574 is acontour parameter used to determine the smoothness of theboundary region [4 15]
The features selectionwith kernel class separabilitymakesSVM the default choice for classification of a brain tumorThe SVM algorithmrsquos performance can be evaluated in termsof accuracy sensitivity and specificity The confusion matrixdefining the terms TP TN FP and FN from the expectedoutcome and ground truth result for the calculation ofaccuracy sensitivity and specificity are shown in Table 3
Where TP is the number of true positives which is used toindicate the total number of abnormal cases correctly clas-sified TN is the number of true negatives which is used toindicate normal cases correctly classified FP is the number
Table 3 Confusion matrix defining the terms TP TN FP and FN
Expected outcome Ground truth Row totalPositive Negative
Table 4 Accuracy sensitivity and specificity calculation
Quality parameter Formula
Accuracy TP + TNTP + TN + FP + FN
Sensitivity TPTP + FN
Specificity TNTN + FP
of false positive and it is used to indicate wrongly detectedor classified abnormal cases when they are actually normalcases and FN is the number of false negatives it is used toindicate wrongly classified or detected normal cases whenthey are actually abnormal cases [15] all of these outcomeparameters are calculated using the total number of samplesexamined for the detection of the tumor The quality rateparameter accuracy is the proportion of total correctly classi-fied cases that are abnormally classified as abnormal andnormally classified as normal from the total number of casesexamined [37 38] Table 4 shows the formulas to calculateaccuracy sensitivity and specificity
4 Results and Discussion
To validate the performance of our algorithm we used twobenchmark datasets and one dataset collected from expertradiologists which included sample images of 15 patients
8 International Journal of Biomedical Imaging
Table 5 Performance analysis parameters for segmented tissues
Images MSE PSNR SSIM Dice scoreImage 1 186 5545 dB 08944 083Image 2 058 6821 dB 09025 087Image 3 495 5628 dB 09702 082Image 4 123 5879 dB 08801 079Image 5 506 5965 dB 07978 090
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 4 Experimental results of image 1 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
with 9 slices for each patient The first dataset is the DigitalImaging and Communications in Medicine (DICOM) data-set [39] For the purpose of the analysis we considered 22images from the DICOM dataset all of which included aretumor-infected brain tissues However this dataset did nothave any ground truth imagesThe second dataset is the BrainWeb dataset [40] which consists of full three-dimensionalsimulated brain MR data obtained using three sequences ofmodalities namely T1-weightedMRI T2-weightedMRI andproton density-weightedMRIThis dataset included a varietyof slice thicknesses noise levels and levels of intensitynonuniformity The images used for our analysis are mostlyincludedT2-weightedmodality with 1mm slice thickness 3noise and 20 intensity nonuniformity In this dataset 13out of 44 images included are tumor-infected brain tissuesThe last dataset collected from expert radiologists consisted
of 135 images of 15 patients with all modalities This datasethad ground truth images that helped to compare the resultsof our method with the manual analysis of radiologists
This section presents the results of our proposed imagesegmentation technique which are obtained by using realbrain MR images The proposed algorithm was carried outusing Matlab 7120 (R2011a) which runs on the Windows 8operating system and has an Intel core i3 processor and a4GB RAM The sample experimental results obtained fromthe proposed technique that are depicted in Figures 4 5 and6 show the original image along with enhanced image skull-stripped image wavelet decompose image cluster (intense)segmented image dice overlap image and the tumor regionwith extracted area mark
Table 5 provides the details of the different performanceparameters such as mean squared error (MSE) and peak
International Journal of Biomedical Imaging 9
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
Table 5 Performance analysis parameters for segmented tissues
Images MSE PSNR SSIM Dice scoreImage 1 186 5545 dB 08944 083Image 2 058 6821 dB 09025 087Image 3 495 5628 dB 09702 082Image 4 123 5879 dB 08801 079Image 5 506 5965 dB 07978 090
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 4 Experimental results of image 1 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
with 9 slices for each patient The first dataset is the DigitalImaging and Communications in Medicine (DICOM) data-set [39] For the purpose of the analysis we considered 22images from the DICOM dataset all of which included aretumor-infected brain tissues However this dataset did nothave any ground truth imagesThe second dataset is the BrainWeb dataset [40] which consists of full three-dimensionalsimulated brain MR data obtained using three sequences ofmodalities namely T1-weightedMRI T2-weightedMRI andproton density-weightedMRIThis dataset included a varietyof slice thicknesses noise levels and levels of intensitynonuniformity The images used for our analysis are mostlyincludedT2-weightedmodality with 1mm slice thickness 3noise and 20 intensity nonuniformity In this dataset 13out of 44 images included are tumor-infected brain tissuesThe last dataset collected from expert radiologists consisted
of 135 images of 15 patients with all modalities This datasethad ground truth images that helped to compare the resultsof our method with the manual analysis of radiologists
This section presents the results of our proposed imagesegmentation technique which are obtained by using realbrain MR images The proposed algorithm was carried outusing Matlab 7120 (R2011a) which runs on the Windows 8operating system and has an Intel core i3 processor and a4GB RAM The sample experimental results obtained fromthe proposed technique that are depicted in Figures 4 5 and6 show the original image along with enhanced image skull-stripped image wavelet decompose image cluster (intense)segmented image dice overlap image and the tumor regionwith extracted area mark
Table 5 provides the details of the different performanceparameters such as mean squared error (MSE) and peak
International Journal of Biomedical Imaging 9
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
Figure 5 Experimental results of image 2 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
signal-to-noise ratio (PSNR) structured similarity index(SSIM) and dice score A lower value of MSE and a highervalue of PSNR indicate better signal-to-noise ratio in theextracted image Dice coefficient measures the overlap of theautomatic and manual segmentation for the given datasetIt is important to note that as some of the features do notcontribute to the classification it is around 8614 in anadaptive fuzzy inference system (ANFIS) 8029 in BackPropagation 9054 in SVM and 8455 in 119870-NearestNeighbors (119870-NN) without feature extraction Table 6 showsthe accuracy of the classification without feature extractionand with feature extraction and shows that it will increasethe performance of the classifiers on the diagnosis of thetumor from brain MR image with feature extractionThe testperformance of the SVM classifier determined by the compu-tation of the statistical parameters such as sensitivity speci-ficity and accuracy in comparison with different classifiertechniques is shown in Table 7 Furthermore higher valuesof accuracy and sensitivity and a lower value of specificityindicate better performance It can be seen from Table 7 thatthe performance of our segmentation algorithm is better thanthe state-of-the-art techniques Even a modest improvementin the sensitivity parameter is very important and critical fora radiologist or clinical doctors for surgical planning
Table 6 Classification accuracies based on feature extraction
The proposed algorithm performs segmentation featureextraction and classification as is done in human vision per-ception which recognizes different objects different texturescontrast brightness and depth of the image Moreover ifcertain agents are applied effectively the application of theproposed technique can be extended to a varying range oftumors and MR modalities In a future study we intendto investigate the application of the proposed method tomore realistic and more clinically bounded cases with a largevariety of scenarios covering different aspects by using largedataset Table 8 shows the area of the extracted brain tumorin square cm and pixels and its comparison with the areacalculated by expert radiologists
10 International Journal of Biomedical Imaging
(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 6 Experimental results of image 3 (a) Original image (b) Enhanced image (c) Skull-stripped image (d) Wavelet decompose image(e) Intense segmented image (f) Dice overlap image (g) Tumor region (h) Area extracted tumor region
Table 7 Comparison of accuracies in different classifiers
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
Area in squarecentimeters Area ratio Accuracy of the area compared to the
area calculated by expert radiologistImage 1 274 times 278 76172 9877 122 01296 998Image 2 257 times 256 65792 7064 058 01073 100Image 3 336 times 407 136752 6365 145 00465 100Image 4 200 times 198 39600 7608 023 01921 998 Image 5 336 times 204 68544 4494 179 01079 100
International Journal of Biomedical Imaging 11
9004 85579651
8706
0
20
40
60
80
100
120
ANFIS BackPropagation
SVM(proposedclassifier)
Specificity ()Sensitivity ()Accuracy ()
K-NN
Figure 7 Comparative analysis of classifiers
5 Comparative Analysis
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
Theresult obtained using the proposed brain tumor detectiontechnique based on Berkeley wavelet transform (BWT) andsupport vector machine (SVM) classifier is compared withthe ANFIS Back Propagation and 119870-NN classifier on thebasis of performance measure such as sensitivity specificityand accuracyThe detailed analysis of performance measuresis shown in Figure 7 and through the performance measureit is depicted that the performance of the proposed method-ology has significantly improved the tumor identificationcompared with the ANFIS Back Propagation and 119870-NNbased classification techniques
6 Conclusion and Future Work
In this study using MR images of the brain we segmentedbrain tissues into normal tissues such as white matter graymatter cerebrospinal fluid (background) and tumor-infectedtissues Fifteen patients infected with a glial tumor in benignand malignant stages assisted in this study We used prepro-cessing to improve the signal-to-noise ratio and to eliminatethe effect of unwanted noise We used a skull strippingalgorithm based on threshold technique to improve theskull stripping performance Furthermore we used Berkeleywavelet transform to segment the images and support vectormachine to classify the tumor stage by analyzing featurevectors and area of the tumor In this study we investigatedtexture based and histogram based features with a commonlyrecognized classifier for the classification of brain tumor fromMR brain images From the experimental results performedon the different images it is clear that the analysis for the braintumor detection is fast and accurate when compared withthe manual detection performed by radiologists or clinicalexperts The various performance factors also indicate thatthe proposed algorithm provides better result by improvingcertain parameters such as mean MSE PSNR accuracysensitivity specificity and dice coefficient Our experimental
results show that the proposed approach can aid in theaccurate and timely detection of brain tumor along withthe identification of its exact location Thus the proposedapproach is significant for brain tumor detection from MRimages
The experimental results achieved 9651 accuracydemonstrating the effectiveness of the proposed technique foridentifying normal and abnormal tissues from MR imagesOur results lead to the conclusion that the proposed methodis suitable for integrating clinical decision support systemsfor primary screening and diagnosis by the radiologists orclinical experts
In the future work to improve the accuracy of the clas-sification of the present work we are planning to investigatethe selective scheme of the classifier by combining more thanone classifier and feature selection techniques
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
The authors would like to thank Dr G Dhondse Sai ClinicBalaji Nagar Nagpur Maharashtra India and GovernmentHospital of State Reserve Police Force (SRPF) Nagpur Maha-rashtra India for providing the necessary guidance and helpin the analysis of the algorithm
References
[1] L Guo L Zhao Y Wu Y Li G Xu and Q Yan ldquoTumor detec-tion in MR images using one-class immune feature weightedSVMsrdquo IEEE Transactions on Magnetics vol 47 no 10 pp3849ndash3852 2011
[2] RKumari ldquoSVMclassification an approach ondetecting abnor-mality in brain MRI imagesrdquo International Journal of Engineer-ing Research and Applications vol 3 pp 1686ndash1690 2013
[3] American Brain Tumor Association httpwwwabtaorg[4] N Gordillo E Montseny and P Sobrevilla ldquoState of the art
[5] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[6] S Madhukumar and N Santhiyakumari ldquoEvaluation of k-Means and fuzzy C-means segmentation on MR images ofbrainrdquo Egyptian Journal of Radiology and Nuclear Medicine vol46 no 2 pp 475ndash479 2015
[7] Y Kong Y Deng and Q Dai ldquoDiscriminative clustering andfeature selection for brain MRI segmentationrdquo IEEE SignalProcessing Letters vol 22 no 5 pp 573ndash577 2015
[8] M T El-Melegy and H M Mokhtar ldquoTumor segmentation inbrain MRI using a fuzzy approach with class center priorsrdquoEURASIP Journal on Image and Video Processing vol 2014article no 21 2014
[9] G Coatrieux H Huang H Shu L Luo and C Roux ldquoA water-marking-based medical image integrity control system and an
12 International Journal of Biomedical Imaging
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
image moment signature for tampering characterizationrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 6 pp1057ndash1067 2013
[10] S Damodharan and D Raghavan ldquoCombining tissue segmen-tation and neural network for brain tumor detectionrdquo Interna-tional Arab Journal of Information Technology vol 12 no 1 pp42ndash52 2015
[11] M Alfonse and A-B M Salem ldquoAn automatic classificationof brain tumors through MRI using support vector machinerdquoEgyptian Computer Science Journal vol 40 pp 11ndash21 2016
[12] Q AinM A Jaffar and T-S Choi ldquoFuzzy anisotropic diffusionbased segmentation and texture based ensemble classification ofbrain tumorrdquo Applied Soft Computing Journal vol 21 pp 330ndash340 2014
[13] E Abdel-Maksoud M Elmogy and R Al-Awadi ldquoBrain tumorsegmentation based on a hybrid clustering techniquerdquo EgyptianInformatics Journal vol 16 no 1 pp 71ndash81 2014
[14] E A Zanaty ldquoDetermination of gray matter (GM) and whitematter (WM) volume in brain magnetic resonance images(MRI)rdquo International Journal of Computer Applications vol 45pp 16ndash22 2012
[15] T Torheim E Malinen K Kvaal et al ldquoClassification of dyna-mic contrast enhancedMR images of cervical cancers using tex-ture analysis and support vector machinesrdquo IEEE Transactionson Medical Imaging vol 33 no 8 pp 1648ndash1656 2014
[16] J Yao J Chen and C Chow ldquoBreast tumor analysis in dynamiccontrast enhanced MRI using texture features and wavelettransformrdquo IEEE Journal on Selected Topics in Signal Processingvol 3 no 1 pp 94ndash100 2009
[17] P Kumar and B Vijayakumar ldquoBrain tumour Mr image seg-mentation and classification using by PCA and RBF kernelbased support vectormachinerdquoMiddle-East Journal of ScientificResearch vol 23 no 9 pp 2106ndash2116 2015
[18] N Sharma A Ray S Sharma K Shukla S Pradhan and LAggarwal ldquoSegmentation and classification of medical imagesusing texture-primitive features application of BAM-type arti-ficial neural networkrdquo Journal of Medical Physics vol 33 no 3pp 119ndash126 2008
[19] W Cui Y Wang Y Fan Y Feng and T Lei ldquoLocalized FCMclustering with spatial information for medical image segmen-tation and bias field estimationrdquo International Journal of Bio-medical Imaging vol 2013 Article ID 930301 8 pages 2013
[20] G Wang J Xu Q Dong and Z Pan ldquoActive contour modelcouplingwith higher order diffusion formedical image segmen-tationrdquo International Journal of Biomedical Imaging vol 2014Article ID 237648 8 pages 2014
[21] A Chaddad ldquoAutomated feature extraction in brain tumor bymagnetic resonance imaging using gaussian mixture modelsrdquoInternational Journal of Biomedical Imaging vol 2015 ArticleID 868031 11 pages 2015
[22] S N Deepa and B Arunadevi ldquoExtreme learning machine forclassification of brain tumor in 3DMR imagesrdquo Informatologiavol 46 no 2 pp 111ndash121 2013
[23] J Sachdeva V Kumar I Gupta N Khandelwal and C KAhuja ldquoSegmentation feature extraction and multiclass braintumor classificationrdquo Journal of Digital Imaging vol 26 no 6pp 1141ndash1150 2013
[24] S Lal andM Chandra ldquoEfficient algorithm for contrast enhan-cement of natural imagesrdquo International Arab Journal of Infor-mation Technology vol 11 no 1 pp 95ndash102 2014
[25] C C Benson andV L Lajish ldquoMorphology based enhancementand skull stripping of MRI brain imagesrdquo in Proceedings of theInternational Conference on Intelligent Computing Applications(ICICA rsquo14) pp 254ndash257 Tamilnadu India March 2014
[26] S Z Oo and A S Khaing ldquoBrain tumor detection and seg-mentation using watershed segmentation and morphologicaloperationrdquo International Journal of Research in Engineering andTechnology vol 3 no 3 pp 367ndash374 2014
[27] R Roslan N Jamil and R Mahmud ldquoSkull stripping mag-netic resonance images brain images region growing versusmathematical morphologyrdquo International Journal of ComputerInformation Systems and Industrial Management Applicationsvol 3 pp 150ndash158 2011
[28] S Mohsin S Sajjad Z Malik and A H Abdullah ldquoEfficientway of skull stripping in MRI to detect brain tumor by applyingmorphological operations after detection of false backgroundrdquoInternational Journal of Information and Education Technologyvol 2 no 4 pp 335ndash337 2012
[29] B Willmore R J Prenger M C Wu and J L Gallant ldquoTheBerkeley wavelet transform a biologically inspired orthogonalwavelet transformrdquoNeural Computation vol 20 no 6 pp 1537ndash1564 2008
[30] P Remya Ravindran and K P Soman ldquoBerkeley wavelet trans-form based image watermarkingrdquo in Proceedings of the Inter-national Conference on Advances in Recent Technologies inCommunication and Computing (ARTCom rsquo09) pp 357ndash359IEEE Kerala India October 2009
[31] I M Alwan and E M Jamel ldquoDigital image watermarkingusing Arnold scrambling and Berkeley wavelet transformrdquo Al-Khwarizmi Engineering Journal vol 12 pp 124ndash133 2015
[32] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973
[33] J LiuM Li JWang FWu T Liu andY Pan ldquoA survey ofMRI-based brain tumor segmentation methodsrdquo Tsinghua Scienceand Technology vol 19 no 6 pp 578ndash595 2014
[34] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[35] V Anitha and S Murugavalli ldquoBrain tumor classification basedon clustered discrete cosine transform in compressed domainrdquoJournal of Computer Science vol 10 no 10 pp 1908ndash1916 2014
[36] Parveen and A Singh ldquoDetection of brain tumor in MRIimages using combination of fuzzy c-means and SVMrdquo in Pro-ceedings of the 2nd International Conference on Signal Processingand Integrated Networks (SPIN rsquo15) pp 98ndash102 February 2015
[37] K Dhanalakshmi and V Rajamani ldquoAn intelligent miningsystem for diagnosing medical images using combined texture-histogram featuresrdquo International Journal of Imaging Systemsand Technology vol 23 no 2 pp 194ndash203 2013
[38] P Rajendran and M Madheswaran ldquoPruned associative clas-sification technique for the medical image diagnosis systemrdquoin Proceedings of the 2nd International Conference on MachineVision (ICMV rsquo09) pp 293ndash297 Dubai UAE December 2009
