researcharticle rank ...healthy tissues extraction can help to provide gbm structure segmentation,...

Research ArticleRank-Two NMF Clustering for Glioblastoma Characterization

Aymen Bougacha 1 Ines Njeh1 Jihene Boughariou1 Omar Kammoun2

KheireddineBenMahfoudh2MariemDammak3ChokriMhiri3 andAhmedBenHamida1

1ATMS-ENIS Advanced Technologies for Medicine and Signals Department of Electrical and Computer EngineeringNational Engineers School Sfax University Sfax Tunisia2Department of Radiology University Hospital Habib Bourguiba Sfax Tunisia3Department of Neurology University Hospital Habib Bourguiba Sfax Tunisia

Correspondence should be addressed to Aymen Bougacha aymenbougachagmailcom

Received 29 June 2018 Accepted 26 September 2018 Published 23 October 2018

Academic Editor Vincenzo Conti

Copyright copy 2018 Aymen Bougacha et al (is is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

(is study investigates a novel classification method for 3Dmultimodal MRI glioblastomas tumor characterization We formulateour segmentation problem as a linear mixture model (LMM) (us we provide a nonnegative matrixM from every MRI slice inevery segmentation processrsquo step(ismatrix will be used as an input for the first segmentation process to extract the edema regionfrom T2 and FLAIR modalities After that in the rest of segmentation processes we extract the edema region from T1c modalitygenerate the matrixM and segment the necrosis the enhanced tumor and the nonenhanced tumor regions In the segmentationprocess we apply a rank-two NMF clustering We have executed our tumor characterization method on BraTS 2015 challengedataset Quantitative and qualitative evaluations over the publicly training and testing dataset from the MICCAI 2015 multimodalbrain segmentation challenge (BraTS 2015) attested that the proposed algorithm could yield a competitive performance for brainglioblastomas characterization (necrosis tumor core and edema) among several competing methods

1 Introduction

Brain tumor represents 85 to 90 of all primary centralnervous system tumors It is one of the main sources for theincrease in death rate among children and adults in theworld Bauer et al [1] noted that glioma could be consideredas the largest common brain tumor with themaximum deathrate According to its severity such brain tumor could beclassified as low-grade glioblastomas (LGG) and high-gradeglioblastomas (HGG) (e low-grade tumors keep de-veloping for many years and could be designed as slowinvaders of brain safety tissue On the other hand the high-grade tumors known as glioblastomas multiform (GBM) areincurable with an average life of one year after its revelationSuch invasive tumors are very heterogeneous due to theirmorphological cytological and molecular variability Itmight have a variety of shapes might be of any size andmight appear at any location and in different imageintensities

In behalf on their frequency and severity glioblastomascontinue to be the major therapeutic issue for neurosur-geons neuro-oncologists and radiation therapists

(e magnetic resonance imaging (MRI) could be con-sidered as one of the main noninvasive modalities used toexplore glioblastomas brain tumor for diagnosis evaluationas well as for inspection of the addressed treatment effectSuch procedure offers the generation of different sequencesby modifying the excitation and the repetition times duringthe acquisition of the image Each sequence provides rele-vant structural information (e main four standard MRImodalities are the T1-weighted MRI (T1) the T2-weightedMRI (T2) the T1-weighted MRI with gadolinium contrastenhancement (T1-Gd) and the fluid-attenuated inversionrecovery (FLAIR) Conventionally T1 images are used todifferentiate healthy tissue while T2 images provide a lightsignal on the image which helps to delineate the region of theedema In T1-Gd images the hyperintense given by theaccumulated contrast agent (gadolinium ions) in the active

HindawiJournal of Healthcare EngineeringVolume 2018 Article ID 1048164 7 pageshttpsdoiorg10115520181048164

cellular region of the tumor tissue allows us to facilitate theobservation of the tumor boundary (e necrotic cells areobserved by a hypointense part of the tumor core as they donot interact with the contrast agent which makes themeasily distinguishable from the active cell region In FLAIRimages the suppression of the signal of molecule waterprovides a good observation of edema region from cere-brospinal fluid (CSF)

Glioblastomas segmentation is a challenging task thatcould be considered as an essential preprocessing task inbrain tumors diagnoses Manual segmentation is a tediousand a time-consuming process for radiologists

Consequently automatic segmentation algorithmswould be recommended in order to obtain accurate andreliable brain tumor delimitation but it remains a persistentchallenge due to the structural complexity of glioblastomastumors Furthermore such tumors present essentially fourdifferent zones edema which represents an excess accu-mulation of fluid in the intracellular or extracellular spacesof the brain nonenhancing solid core necroticcystic coreand enhancing core

Several research studies have been investigated to seg-ment different tumor zones in multiple MRI modalities (T1T1-Gd T2 and Flair) [1ndash3] Dupont et al [4] present intheir review four main classes in order to segment glio-blastomas tumor region-based approach edge-based ap-proach and classification-based algorithms approach

In the region-based approach we intend to implementsegmentation by merging neighbourhood pixels that havesimilar characteristics A region-based method presented byFranz et al [5] is used to differentiate the enhanced tumorportion the necrotic zone and the edema zone Only twomodalities (T1-Gd and Flair) have been used as an input forthis algorithm and only the image intensity has beenemployed as a feature in order to delimit different tumorregionrsquos zone As a consequence coherent intensity pixelshave been grouped into three classes tumor enhancementzone necrosis zone and edema zone Sachdeva et al [6]introduced an edge-based method based on image texturersquosintensity and a specific active contour to achieve semi-automatic segmentation (e authors used multimodal MRI(T1-weighted T2-weighted and T1-Gd MRI) to test theiralgorithm Essidike et al [7] proposed a two-step braintumor segmentation For the first step a numerical simu-lation of the optical correlation has been used to detect braintumor and an active contour model is used to detect regionfor the next step

Healthy tissues extraction can help to provide GBMstructure segmentation and atlas-based approaches havebeen used in this way Prastawa et al [8] introduced anautomatic brain tumor segmentation with edemarsquos de-tection (is algorithm used only T2 MRI Pixels classifi-cation of cerebrospinal fluid (CSF) white matter (WM) andgray matter (GM) was performed from atlas template (eunclassified pixels have been labelled as tumor or edema

Classification approaches are widely used in imagesegmentation (ese methods consist in clustering pixelsdepending on different features used as an input vector(intensity texture neighbours and spatial distribution in

the image) of a clustering algorithm(ere are two groups ofclassification approaches supervised approaches or un-supervised approaches Wu et al [9] applied a supervisedmethod Amultimodal MRI is segmented into superpixels tominimize the sampling problem (en features wereextracted from the superpixels using multilevel Gaborwavelet filters (ese features are used to power the supportvector machine (SVM) model (e theory of conditionalrandom fields has been applied to segment the tumor basedon the output of the SVMmodels Finally the marking noisewas removed using ldquostructural knowledgerdquo (is system wasapplied with 20 GBM cases Recent studies [10ndash12] have usedthe deep learning technique for the segmentation of GBMtumors (ese methods of segmentation differ according tothe training concept Havaei et al [10] performed a modifiedconventional neural network (CNN) and a two-phasetraining to touch on problems related to the unbalance ofGBM labels Zhao et al [12] used a three-segmentationmodel based on fully convolutional neural networks(FCNNs) conditional random fields (CRFs) and recurrentneural networks (RNNs) (ese models are trained with 2Dimage patches and slices acquired in axial coronal andsagittal views respectively andmixed them to segment braintumors Hussain et al [11] implemented a deep conventionalneural network (DCNN) where two networks are piled overone another to construct a new linear nexus architecture(e first network holds in parallel placing of layers whereasin the second network layers are structured linearly

Corso et al [13] proposed a Bayesian model classifica-tion (is unsupervised method has used two concepts classmodel and graph cuts (e objective was to fuse speed ofgraph cuts and statistical distribution efficiency of the classmodel (e proposed method was executed to twenty GBMcases with T1 T1-Gd T2 and FLAIR previously investigatedby experts Presented as one of the popular unsupervisedclustering methods Cordova et al [14] developed a fuzzyc-means GBM segmentation using T1-GD images (ismethod has been tested with thirty seven cases In [15 16]authors applied a hierarchical nonnegative matrix factor-ization (hNMF) on multiparametric MRI to provide tissuecharacterization (e specification of tissuersquos patterns wasobtained and the spatial distribution of each tissue type wasvisualized Li et al [17] also applied hNMF to brain MRSIdata for GBM tissuersquos differentiation

In this work we propose a novel classification methodfor 3D multimodal MRI glioblastomas tumor character-ization We formulate our segmentation problem as a linearmixture model (LMM) (us we provide a nonnegativematrix M from every MRI slice in every segmentationprocessrsquo step(is matrix will be used as an input for the firstsegmentation process to extract edema region from T2 andFLAIR modalities After that in the rest of segmentationprocesses we extract the edema region from T1c modalitygenerate the matrix M and segment the necrosis the en-hanced tumor and the nonenhanced tumor regions asdescribed in the methodrsquos flowchart (see Figure 1) In thesegmentation process we apply a rank-two NMF clusteringwhich could be defined as a blind source separation tech-nique [18] It consists in approximately the factorization of

2 Journal of Healthcare Engineering

a matrix M into the product of a source matrix W and anabundance matrix H (is method has been used as a braintumor segmentation with MRSI (magnetic resonancespectroscopy image) [17] and multiparametric MRI data[15 16] (e main contribution of this study and the dif-ferences between our work and others mentioned previouslylies on the application of the GLCM features for nonnegativematrix M and the use of a rank-two NMF instead of thehierarchical NMF (e proposed method does not requirea training dataset as is the case of the many existingmethods Quantitative assessment over the publicly existingtraining and testing dataset from the MICCAI MultimodalBrain Tumor Segmentation 2015 (BraTS 2015) challenge [19]confirm that the proposed method provides a competitiveperformance

(e remainder of this paper is arranged as follows thematerials and methods section where we define the Multi-modal Brain Tumor Segmentation Benchmark (BraTS 2015)data and illustrate the segmentation methodology (e re-sults and discussion shows the experimental results witha discussion Finally the conclusion section illustratesvarious perspectives of this work

2 Materials and Methods

(e obtained results were based on approved evaluationsusing the Multimodal Brain Tumor Segmentation Bench-mark (BraTS 2015) [19] In this section we present in detailsthe used dataset the evaluation metrics and the differentsteps of the proposed methodology the preprocessing stepthe feature extraction and the rank-two NMF segmentation(e proposed approach could be outlined according to theflowchart (see Figure 2)

21 Multimodal Brain Tumor Segmentation Benchmark(BraTS 2015) (e Multimodal Brain Tumor Segmentationdataset (BraTs 2015) is in continuation of BraTS 2012 BraTS2013 and BraTS 2014 It has been organized by B Menze MReyes K Farahani and J Kalpathy-Cramer in conjunctionwith the MICCAI 2015 conference (is available publiclytraining and testing dataset could be considered as very

useful to compare the existing method and to gauge thecurrent state of the art in brain tumor segmentation Itconsists in comparing and evaluating 3D MRI brain tumorregions obtained by segmenting multimodal imagingdataset Such task could be considered as a challenging taskin medical image analysis due to the unpredictable ap-pearance and shape of glioblastomas tumor (e coregis-tered the skull-stripped and the annotated training datasetare available via the Virtual Skeleton Database (VSD) [20]

Training dataset testing dataset and the ground truthare stored as signed 16-bit integers but only positive valuesare used FourMRImodalities are proposed for each case T1modality T2 modality T1c modality and FLAIR modality(e manual segmentations (ground truth) of the patientimages have the following five different labels (1) for ne-crosis (2) for edema (3) for nonenhancing tumor (4) forenhancing tumor and (0) for everything else

(e evaluation is done for 3 different tumorsubcompartments

(i) Region 1 Complete tumor (labels 1 + 2 + 3 + 4 forpatient data and labels 1 + 2 for synthetic data)

(ii) Region 2 Tumor core (labels 1 + 3 + 4 for patientdata and label 2 for synthetic data)

(iii) Region 3 Enhancing tumor (label 4 for patient dataand na for synthetic data)

(e total case of training data is 274 patients (220 high-grade tumors and 54 low-grade tumors) while the testingdataset contains 110 subjects with low-grade glioma (LGG)and high-grade glioma (HGG)

22 EvaluationMetric In this study the dice (DM) [21] andthe sensitivity metrics were used to evaluate the quantitativeperformance and the quality of segmentation (is requirescomputing the similarities between ground-truth segmen-tations provided with BraTS 2015 dataset and the obtainedresults (ese metrics take values within the interval [01]where 1 indicates a perfect match and 0 a complete mis-match An automated segmentation should be uploadeddirectly to the evaluation page to obtain the dice metric score[20]

23 Proposed Algorithm In this work we present a per-formed algorithm in order to segment the different tumorregions (necrosis edema nonenhancing tumor enhancingtumor and everything else) (e proposed methodology istested and validated using BraTS 2015 We present theflowchart that describes the segmentation process (seeFigure 1)

We formulate our segmentation problem as a linearmixture model (LMM) As depicted in Figure 1 from aninput MRI scan we generate a nonnegative matrixM that ismatrix M isin Rmlowastn

+ with m features and n voxels as follows(see Figure 3) the (ij)th entry M(i j) of the matrix M is theith GLCM feature of the jth voxel Hence each column ofMis equal to the feature signature of a voxel while each row isa vectorized image at a given feature (e linear mixing

Input MRI slice

Generate a non negativematrix M

Rank-two NMF

Region of interest[W(i 1) H(i 1)]

Otherwise[W(i 2) H(i 2)]

ClusteringW(i 1) H(i 1) W(i 2) H(i 2)

Figure 1 Segmentation process

Journal of Healthcare Engineering 3

model (LMM) assumes that the feature signature of eachvoxel is a linear combination of the feature signatures of theconstitutive pattern in image (endmembers) where theweights in the linear combination are the abundances ofeach endmember in this voxel [22]

Supposing the image encloses r endmembers and des-ignating W( k) isin Rm(1le kle r) the feature signatures ofthe endmembers we can write the LMM as

M( j) 1113944r

k1W( k)H(k j) 1le jle n (1)

where H(k j) is the abundance of the kth endmember in thejth voxel so 1113936

rk1H(k j) 1 for all j which is specified to as

the abundance sum-to-one constraint As all matricesconcerned M W and H are nonnegative the LMM iscorresponding to nonnegative matrix factorization (NMF)Having a nonnegative matrix M isin Rmlowastn

+ and a factorizationrank r discover two nonnegative matrices W isin Rmlowastr

+ andH isin Rrlowastn

+ such that M asympWH However having anMRI slicewith r endmembers it consists in to cluster the pixels into rclusters and each cluster is equivalent to one endmemberMathematically having a matrix M isin Rmlowastn

+ we aim to find rdisjoint clusters Ck sub 1 2 n for 1le kle r so thatcup k12rCk 1 2 n and so that all pixels in Ck aremonopolized by the same endmember

MRI imaging systems provide images with high reso-lution and high tissue contrast (ese images are defined

2D windows extraction

K = mlowast n

GLCM-feature matrixOriginal image

n

M(1 1)M(1 2)

M(1 k)

M(2 1)M(2 2)

M(2 k)

M(14 1) M(14 2)

M(14 k)

M(3 1)m

Figure 3 GLCM-feature matrix m generation

Input MRI scans

Oedema segmentation(Segmentation process with FLAIR and T2

modality)

Input = FLAIR and T2 modality

Segmentation process

MRI with tumour characterization

Necrosis enhanced and non-enhancedtumour segmentation

Input = edema region extracted from T1c




Output 2 = IT1c (mask of necrosis region)necrosis

Output 1 = IT2Flair (mask of oedema region)edema

Output 3 = I enhanced tumour (mask of enhanced tumour region)

T1c

I1= I edema T2Flair-I

necrosisT1c

I2 = I edema T2Flair- Inecrosis

T1c -I enhanced tumourT1c

Output = Inon-enhanced tumour (mask of non-enhanced tumour region)

T1c

Figure 2 Flowchart of the proposed GBM characterization


with a depth of up to 16 bits corresponding to 65535 in-tensity levels In order to simplify the calculation all in-tensity values were rearranged to a gray level values witha maximum of 255 and texture features are computed usingthe grayscale co-occurrence matrix (GLCM) GLCM hasbeen useful in various image processing fields It is a squaredmatrix G(NN) where N represents the number of gray levelexisting in the window (is matrix is a structure thatrepresents the co-occurring intensity values at a given offset(is is defined by the fact that the GLCM gives informationon how often a gray level arise at different directions anda distance d Usually four directions are looked up in the 2Dcase ϕ 0deg ϕ 45deg ϕ 90deg and ϕ 135deg(e structure of the2D GLCM is shown in Figure 4 where nij is the number ofco-occurrences of gray levels i and j at a specific direction ϕand a distance d Haralick et al [23] defined texture featurescalculated using the GLCM Moreover Haralick recom-mended utilizing the average value of the features calculatedfor the four directions to ensure rotation invariance

(e GLCM features (Table 1) used in this study wereextracted at a distance d 1 with mean value for fourdirections

As illustrated in Figure 2 the MRI modalities are used asfollows fromT2 and Flairmodalities we apply the segmentationprocess in order to obtain the edema mask region Iedema

T2Flair (enwe apply the obtained mask on T1C modality and we apply thesegmentation process in order to obtain the necrosis regionrsquosmask InecrosisT1c We calculate after that the intermediary image I1 Iedema

T2Flair-InecrosisT1c which will be used as an input to the seg-

mentation process in order to obtain the enhanced tumor re-gionlsquos mask Ienhanced tumour

T1c We also calculate a secondintermediary image I2 Iedema

T2Flair-InecrosisT1c -Ienhanced tumour

T1c and weapply the segmentation process in order to obtain the non-enhanced tumor regionrsquos mask Inonenhanced tumour

T1c As we can see in Figure 2 that in every segmentation

process we aim to cluster the inputMRI slice into two clustersHowever we propose to use a rank-two NMF clusteringHaving a nonnegative matrix M isin Rmlowastn

+ rank-two NMFsearches two nonnegative matrices W isin Rmlowast2

+ and H isin R2lowastn+

where M asympWH (is factorization is a two-dimensionaldescription of the data more literally it conceives the col-umns ofM onto a two-dimensional pointed cone developed bythe columns of W (erefore the approach to segment theMRI slice in other words to cluster the columns of M is toselecting the clusters like this C1 i|H(i 1)geH(i 2) which represents the region of interest andC2 i|H(i 1)ltH(i 2) represents the otherwise zone

3 Results and Discussion

In this section we report the segmentation result obtainedby the proposed method over the publicly training datasetfrom BraTS 2015 We also present the quantitative evalu-ation by computing the dice metric and the sensitivity forrespectively complete tumor tumor core and the en-hancing tumor (is section is supported by illustration thatdepict typical example of the obtained results

Figure 5 depicts the segmentation obtained on two high-grade gliomas from the training dataset (e green zonecorresponds to the edema region the yellow zone representsthe enhanced tumor the red zone is the necrosis and theblue color represents the nonenhanced tumor Table 2 atteststhe performance of the proposed algorithm by a greet scorefor dice and sensitivity metric

(e segmentation methodology proposed in this papercan process an immense diversity of tumors because it doesnot depend on contrast enhancement It segments the wholebrain including healthy tissue types and automaticallyidentifies edema nonenhanced enhanced tumor and ne-crosis region Delineating the edema region can be valuablefor surgical planning and description of radiation therapyfields and since the edema region demonstrates the volumeover which the tumor applies obvious chemical effectsrecognition of areas of interest to various investigators isinvolved in tumor growth and treatment Delineating theedema region can also be valuable for surgical planning andradiation therapy Often edema regions need to be treated tominimize the risk of recurrence

We have carried out the proposed method to MR datafrom patients with glioblastoma tumors (ese images in-clude tumors with different intensities sizes locations andshapes (is authorizes us to demonstrate the large field ofapplication of our algorithm

We have executed our tumor characterization method onBRATS 2015 challenge dataset Two cases have been selectedrandomly in this experiment Definitely there are four labeltypes in this dataset including necrosis edema enhanced andnonenhanced tumor As pointed out in Table 2 the dice ratiois superior to 085 illustrating good overlap with groundtruth Moreover the sensitivity is superior to 08 whichmeansthat the segmentation results are reliable enough

(e results of this table also illustrate that the quality of thesegmentation for whole tumor is better than for core tumorsbecause of their well-defined boundaries Enhancement of theapproach for segmenting core tumors could still be valuable

Gϕd = nij

n

n

45deg90deg135deg

0deg

Figure 4 2D GLCM computation for nlowast n window Main directions (0deg 45deg 90deg and 135deg) and a distance d are used Mean value is affectedto central voxel


4 Conclusion

In this paper we define a novel methodology for 3D mul-timodal MRI GBM tumor characterization Unlike from theclassic tumor segmentation methods in the proposed

method we observe the brain tumor segmentation task asa four-class (tumor (including necrosis enhanced andnonenhanced tumor) edema and normal tissue) classifi-cation problem regarding three modalities T2 FLAIR andT1c We formulate our segmentation problem as a linear

(a1) (b1) (c1)

(a2) (b2) (c2)

Figure 5 Examples of characterization results obtained from rank-two NMF methods on BRATS 2015 data T1c images with high-gradetumor case HG-02 (a1) HG-03 (a2) b1-b2 ground truth c1-c2 results using rank-two NMF edema (green) necrosis (red) enhanced tumor(yellow) and nonenhanced tumor (blue)

Table 1 GLCM features

Feature Formula Feature Formula

Contrast 1113936Np1(p21113936

Ni11113936

Nj1G

ϕd(i j)) p |iminus j| Sum average 1113936

2Nminus1i1 (ipx+y(i))

Energy 1113936Ni11113936

Nj1G

ϕd(i j)2 Cluster shade 1113936

Ni11113936

Nj1(i + jminus μx minus μy)3G

ϕd(i j)

Dissimilarity 1113936Ni11113936

Nj1(iminus j)G

ϕd(i j) Cluster prominence 1113936

Ni11113936


ϕd(i j)

Entropy minus1113936Ni11113936

Nj1G

ϕd(i j)log(G

ϕd(i j) Maximum probability 1113936

Ni11113936

Nj1maxij G

ϕd(i j)1113966 1113967

Correlation 1113936Ni11113936

Nj1G

ϕd(i j)

(iminus μx)(iminus μy)

σxσyDifference variance 1113936

2Nminus1i1 (i2pxminusy(i))

μx 1113936Ni11113936

Nj1jG

ϕd(i j) μy 1113936

Ni11113936

Nj1iG

ϕd(i j)

σx 1113936Ni11113936

Nj1(jminus μx)2G

ϕd(i j)

σy 1113936Ni11113936

Nj1(iminus μy)2G

ϕd(i j)

Homogeneity 1113936Ni11113936

Nj1

11+(iminus j)2

Gϕd(i j) Autocorrelation 1113936

Ni11113936

Nj1(ilowastj)G

ϕd(i j)

Variance 1113936Ni11113936

Nj1(1minus μ)2G

ϕd(i j) Sum entropy minus11139362Nminus1

i1 (px+y(i)log(px+y(i)))

Difference entropy minus1113936Nminus1i0 (pxminusy(i)log(pxminusy(i))) Sum variance minus11139362Nminus1

i1 (1minus SumEnt)2px+y(i)


mixture model (LMM) (us we provide a nonnegativematrix M from every MRI slice in every step of the seg-mentation process (is matrix will be used as an input forthe first segmentation process to extract edema from T2 andFLAIR modality After that in the rest of segmentationprocesses we extract the edema region from T1c modalitygenerate the matrix M from this modality and segmentnecrosis enhanced tumor and nonenhanced tumor regionsIn the segmentation process we apply a rank-two NMFclustering Compared to the traditional tumor segmentationmethodologies the proposed method is easy to achieve andquite robust to high-intensity inhomogeneity imagesComparison results on BRATS 2015 challenge dataset il-lustrate the superior achievements of the proposed method

As a perspective we will apply the proposed methodthrough all training data and also the proposed testing datain order to attest the performance of the algorithm

Data Availability

(e BRATS 2015 data used to support the findings of thisstudy are included within the article

Conflicts of Interest

(e authors declare that they have no conflicts of interest

References

[1] S Bauer R Wiest L-P Nolte and M Reyes ldquoA survey ofMRI-based medical image analysis for brain tumor studiesrdquoPhysics in Medicine and Biology vol 58 no 13 pp R97ndashR1292013

[2] E-S A El-Dahshan H M Mohsen K Revett andA-B M Salem ldquoComputer-aided diagnosis of human braintumor through MRI a survey and a new algorithmrdquo ExpertSystems with Applications vol 41 no 11 pp 5526ndash5545 2014

[3] N Gordillo E Montseny and P Sobrevilla ldquoState of the artsurvey on MRI brain tumor segmentationrdquo Magnetic Reso-nance Imaging vol 31 no 8 pp 1426ndash1438 2013

[4] C Dupont N Betrouni N Reyns and M Vermandel ldquoOnimage segmentation methods applied to glioblastoma state ofart and new trendsrdquo IRBM vol 37 no 3 pp 131ndash143 2016

[5] A Franz S Remmele and J Keupp ldquoBrain tumour seg-mentation and tumour tissue classification based on multipleMR protocolsrdquo in Proceedings of Medical Imaging 2011 ImageProcessing Lake Buena Vista FL USA August 2011

[6] J Sachdeva V Kumar I Gupta N Khandelwal andC K Ahuja ldquoA novel content-based active contour model forbrain tumor segmentationrdquo Magnetic Resonance Imagingvol 30 no 5 pp 694ndash715 2012

[7] A Essadike E Ouabida and A Bouzid ldquoBrain tumor seg-mentation with Vander Lugt correlator based active contourrdquoComputer Methods and Programs in Biomedicine vol 160pp 103ndash117 2018

[8] M Prastawa E Bullitt S Ho and G Gerig ldquoA brain tumorsegmentation framework based on outlier detectionrdquoMedicalImage Analysis vol 8 no 3 pp 275ndash283 2004

[9] W Wu A Chen L Zhao and J Corso ldquoBrain tumor de-tection and segmentation in a CRF (conditional randomfields) framework with pixel-pairwise affinity and superpixel-level featuresrdquo International Journal of Computer AssistedRadiology and Surgery vol 9 no 2 pp 241ndash253 2013

[10] M Havaei A Davy D Warde-Farley et al ldquoBrain tumorsegmentation with deep neural networksrdquo Medical ImageAnalysis vol 35 pp 18ndash31 2017

[11] S Hussain S M Anwar and M Majid ldquoSegmentation ofglioma tumors in brain using deep convolutional neuralnetworkrdquo Neurocomputing vol 282 pp 248ndash261 2018

[12] X Zhao Y Wu G Song Z Li Y Zhang and Y Fan ldquoA deeplearning model integrating FCNNs and CRFs for brain tumorsegmentationrdquoMedical Image Analysis vol 43 pp 98ndash111 2018

[13] J J Corso E Sharon S Dube S El-Saden U Sinha andA Yuille ldquoEfficient multilevel brain tumor segmentation withintegrated bayesian model classificationrdquo IEEE Transactionson Medical Imaging vol 27 no 5 pp 629ndash640 2008

[14] J S Cordova E Schreibmann C G Hadjipanayis et alldquoQuantitative tumor segmentation for evaluation of extent ofglioblastoma resection to facilitate multisite clinical trialsrdquoTranslational Oncology vol 7 no 1 pp 40ndashW5 2014

[15] N Sauwen D M Sima S Van Cauter et al ldquoHierarchicalnon-negative matrix factorization to characterize brain tumorheterogeneity using multi-parametric MRIrdquo NMR in Bio-medicine vol 28 no 12 pp 1599ndash1624 2015

[16] N Sauwen M Acou D M Sima et al ldquoSemi-automatedbrain tumor segmentation on multi-parametric MRI usingregularized non-negative matrix factorizationrdquo BMC MedicalImaging vol 17 no 1 p 29 2017

[17] Y Li D M Sima S V Cauter A R Croitor Sava et alldquoHierarchical non-negative matrix factorization (hNMF)a tissue pattern differentiation method for glioblastomamultiforme diagnosis using MRSIrdquo NMR in Biomedicinevol 26 no 3 pp 307ndash319 2013

[18] D D Lee and H S Seung ldquoLearning the parts of objects bynon-negative matrix factorizationrdquoNature vol 401 no 6755pp 788ndash791 1999

[19] B H Menze A Jakab S Bauer et al ldquo(e multimodal braintumor image segmentation Benchmark (BRATS)rdquo IEEE Trans-actions on Medical Imaging vol 34 no 10 pp 1993ndash2024 2015

[20] M Kistler S Bonaretti M Pfahrer R Niklaus andP Buchler ldquo(e virtual Skeleton Database an open accessrepository for biomedical research and collaborationrdquo Journalof Medical Internet Research vol 15 no 11 pp e245ndash12 2013

[21] L R Dice ldquoMeasures of the amount of ecologic associationbetween speciesrdquo Ecology vol 26 no 3 pp 297ndash302 1945

[22] N Gillis D Kuang and H Park ldquoHierarchical clustering ofhyperspectral images using rank-two nonnegative matrixfactorizationrdquo IEEE Transactions on Geoscience and RemoteSensing vol 53 no 4 pp 2066ndash2078 2015

[23] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on Sys-tems Man and Cybernetics vol SMC-3 no 6 pp 610ndash6211973

Table 2 Quantitative results for the BRATS 2015 MRI images

Class Dice SensitivityComplete tumor 087 084Tumor core 077 064Enhancing tumor 074 061


International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom

cellular region of the tumor tissue allows us to facilitate theobservation of the tumor boundary (e necrotic cells areobserved by a hypointense part of the tumor core as they donot interact with the contrast agent which makes themeasily distinguishable from the active cell region In FLAIRimages the suppression of the signal of molecule waterprovides a good observation of edema region from cere-brospinal fluid (CSF)

Glioblastomas segmentation is a challenging task thatcould be considered as an essential preprocessing task inbrain tumors diagnoses Manual segmentation is a tediousand a time-consuming process for radiologists

Consequently automatic segmentation algorithmswould be recommended in order to obtain accurate andreliable brain tumor delimitation but it remains a persistentchallenge due to the structural complexity of glioblastomastumors Furthermore such tumors present essentially fourdifferent zones edema which represents an excess accu-mulation of fluid in the intracellular or extracellular spacesof the brain nonenhancing solid core necroticcystic coreand enhancing core

Several research studies have been investigated to seg-ment different tumor zones in multiple MRI modalities (T1T1-Gd T2 and Flair) [1ndash3] Dupont et al [4] present intheir review four main classes in order to segment glio-blastomas tumor region-based approach edge-based ap-proach and classification-based algorithms approach

In the region-based approach we intend to implementsegmentation by merging neighbourhood pixels that havesimilar characteristics A region-based method presented byFranz et al [5] is used to differentiate the enhanced tumorportion the necrotic zone and the edema zone Only twomodalities (T1-Gd and Flair) have been used as an input forthis algorithm and only the image intensity has beenemployed as a feature in order to delimit different tumorregionrsquos zone As a consequence coherent intensity pixelshave been grouped into three classes tumor enhancementzone necrosis zone and edema zone Sachdeva et al [6]introduced an edge-based method based on image texturersquosintensity and a specific active contour to achieve semi-automatic segmentation (e authors used multimodal MRI(T1-weighted T2-weighted and T1-Gd MRI) to test theiralgorithm Essidike et al [7] proposed a two-step braintumor segmentation For the first step a numerical simu-lation of the optical correlation has been used to detect braintumor and an active contour model is used to detect regionfor the next step

Healthy tissues extraction can help to provide GBMstructure segmentation and atlas-based approaches havebeen used in this way Prastawa et al [8] introduced anautomatic brain tumor segmentation with edemarsquos de-tection (is algorithm used only T2 MRI Pixels classifi-cation of cerebrospinal fluid (CSF) white matter (WM) andgray matter (GM) was performed from atlas template (eunclassified pixels have been labelled as tumor or edema

Classification approaches are widely used in imagesegmentation (ese methods consist in clustering pixelsdepending on different features used as an input vector(intensity texture neighbours and spatial distribution in

the image) of a clustering algorithm(ere are two groups ofclassification approaches supervised approaches or un-supervised approaches Wu et al [9] applied a supervisedmethod Amultimodal MRI is segmented into superpixels tominimize the sampling problem (en features wereextracted from the superpixels using multilevel Gaborwavelet filters (ese features are used to power the supportvector machine (SVM) model (e theory of conditionalrandom fields has been applied to segment the tumor basedon the output of the SVMmodels Finally the marking noisewas removed using ldquostructural knowledgerdquo (is system wasapplied with 20 GBM cases Recent studies [10ndash12] have usedthe deep learning technique for the segmentation of GBMtumors (ese methods of segmentation differ according tothe training concept Havaei et al [10] performed a modifiedconventional neural network (CNN) and a two-phasetraining to touch on problems related to the unbalance ofGBM labels Zhao et al [12] used a three-segmentationmodel based on fully convolutional neural networks(FCNNs) conditional random fields (CRFs) and recurrentneural networks (RNNs) (ese models are trained with 2Dimage patches and slices acquired in axial coronal andsagittal views respectively andmixed them to segment braintumors Hussain et al [11] implemented a deep conventionalneural network (DCNN) where two networks are piled overone another to construct a new linear nexus architecture(e first network holds in parallel placing of layers whereasin the second network layers are structured linearly

Corso et al [13] proposed a Bayesian model classifica-tion (is unsupervised method has used two concepts classmodel and graph cuts (e objective was to fuse speed ofgraph cuts and statistical distribution efficiency of the classmodel (e proposed method was executed to twenty GBMcases with T1 T1-Gd T2 and FLAIR previously investigatedby experts Presented as one of the popular unsupervisedclustering methods Cordova et al [14] developed a fuzzyc-means GBM segmentation using T1-GD images (ismethod has been tested with thirty seven cases In [15 16]authors applied a hierarchical nonnegative matrix factor-ization (hNMF) on multiparametric MRI to provide tissuecharacterization (e specification of tissuersquos patterns wasobtained and the spatial distribution of each tissue type wasvisualized Li et al [17] also applied hNMF to brain MRSIdata for GBM tissuersquos differentiation

In this work we propose a novel classification methodfor 3D multimodal MRI glioblastomas tumor character-ization We formulate our segmentation problem as a linearmixture model (LMM) (us we provide a nonnegativematrix M from every MRI slice in every segmentationprocessrsquo step(is matrix will be used as an input for the firstsegmentation process to extract edema region from T2 andFLAIR modalities After that in the rest of segmentationprocesses we extract the edema region from T1c modalitygenerate the matrix M and segment the necrosis the en-hanced tumor and the nonenhanced tumor regions asdescribed in the methodrsquos flowchart (see Figure 1) In thesegmentation process we apply a rank-two NMF clusteringwhich could be defined as a blind source separation tech-nique [18] It consists in approximately the factorization of


















Input MRI slice


Rank-two NMF








M( j) 1113944r











K = mlowast n


n

M(1 1)M(1 2)

M(1 k)

M(2 1)M(2 2)

M(2 k)

M(14 1) M(14 2)

M(14 k)

M(3 1)m


Input MRI scans


modality)












T1c


necrosisT1c




T1c
























Gϕd = nij

n

n

45deg90deg135deg

0deg



4 Conclusion



(a1) (b1) (c1)

(a2) (b2) (c2)




Contrast 1113936Np1(p21113936

Ni11113936

Nj1G



Energy 1113936Ni11113936

Nj1G


Ni11113936


ϕd(i j)


Nj1(iminus j)G


Ni11113936


ϕd(i j)


Nj1G

ϕd(i j)log(G


Ni11113936

Nj1maxij G

ϕd(i j)1113966 1113967


Nj1G

ϕd(i j)




μx 1113936Ni11113936

Nj1jG

ϕd(i j) μy 1113936

Ni11113936

Nj1iG

ϕd(i j)

σx 1113936Ni11113936

Nj1(jminus μx)2G

ϕd(i j)

σy 1113936Ni11113936

Nj1(iminus μy)2G

ϕd(i j)


Nj1

11+(iminus j)2


Ni11113936

Nj1(ilowastj)G

ϕd(i j)

Variance 1113936Ni11113936

Nj1(1minus μ)2G








Data Availability




References





























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


















Input MRI slice


Rank-two NMF








M( j) 1113944r











K = mlowast n


n

M(1 1)M(1 2)

M(1 k)

M(2 1)M(2 2)

M(2 k)

M(14 1) M(14 2)

M(14 k)

M(3 1)m


Input MRI scans


modality)












T1c


necrosisT1c




T1c
























Gϕd = nij

n

n

45deg90deg135deg

0deg



4 Conclusion



(a1) (b1) (c1)

(a2) (b2) (c2)




Contrast 1113936Np1(p21113936

Ni11113936

Nj1G



Energy 1113936Ni11113936

Nj1G


Ni11113936


ϕd(i j)


Nj1(iminus j)G


Ni11113936


ϕd(i j)


Nj1G

ϕd(i j)log(G


Ni11113936

Nj1maxij G

ϕd(i j)1113966 1113967


Nj1G

ϕd(i j)




μx 1113936Ni11113936

Nj1jG

ϕd(i j) μy 1113936

Ni11113936

Nj1iG

ϕd(i j)

σx 1113936Ni11113936

Nj1(jminus μx)2G

ϕd(i j)

σy 1113936Ni11113936

Nj1(iminus μy)2G

ϕd(i j)


Nj1

11+(iminus j)2


Ni11113936

Nj1(ilowastj)G

ϕd(i j)

Variance 1113936Ni11113936

Nj1(1minus μ)2G








Data Availability




References





























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




M( j) 1113944r











K = mlowast n


n

M(1 1)M(1 2)

M(1 k)

M(2 1)M(2 2)

M(2 k)

M(14 1) M(14 2)

M(14 k)

M(3 1)m


Input MRI scans


modality)












T1c


necrosisT1c




T1c
























Gϕd = nij

n

n

45deg90deg135deg

0deg



4 Conclusion



(a1) (b1) (c1)

(a2) (b2) (c2)




Contrast 1113936Np1(p21113936

Ni11113936

Nj1G



Energy 1113936Ni11113936

Nj1G


Ni11113936


ϕd(i j)


Nj1(iminus j)G


Ni11113936


ϕd(i j)


Nj1G

ϕd(i j)log(G


Ni11113936

Nj1maxij G

ϕd(i j)1113966 1113967


Nj1G

ϕd(i j)




μx 1113936Ni11113936

Nj1jG

ϕd(i j) μy 1113936

Ni11113936

Nj1iG

ϕd(i j)

σx 1113936Ni11113936

Nj1(jminus μx)2G

ϕd(i j)

σy 1113936Ni11113936

Nj1(iminus μy)2G

ϕd(i j)


Nj1

11+(iminus j)2


Ni11113936

Nj1(ilowastj)G

ϕd(i j)

Variance 1113936Ni11113936

Nj1(1minus μ)2G








Data Availability




References





























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia























Gϕd = nij

n

n

45deg90deg135deg

0deg



4 Conclusion



(a1) (b1) (c1)

(a2) (b2) (c2)




Contrast 1113936Np1(p21113936

Ni11113936

Nj1G



Energy 1113936Ni11113936

Nj1G


Ni11113936


ϕd(i j)


Nj1(iminus j)G


Ni11113936


ϕd(i j)


Nj1G

ϕd(i j)log(G


Ni11113936

Nj1maxij G

ϕd(i j)1113966 1113967


Nj1G

ϕd(i j)




μx 1113936Ni11113936

Nj1jG

ϕd(i j) μy 1113936

Ni11113936

Nj1iG

ϕd(i j)

σx 1113936Ni11113936

Nj1(jminus μx)2G

ϕd(i j)

σy 1113936Ni11113936

Nj1(iminus μy)2G

ϕd(i j)


Nj1

11+(iminus j)2


Ni11113936

Nj1(ilowastj)G

ϕd(i j)

Variance 1113936Ni11113936

Nj1(1minus μ)2G








Data Availability




References





























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


4 Conclusion



(a1) (b1) (c1)

(a2) (b2) (c2)




Contrast 1113936Np1(p21113936

Ni11113936

Nj1G



Energy 1113936Ni11113936

Nj1G


Ni11113936


ϕd(i j)


Nj1(iminus j)G


Ni11113936


ϕd(i j)


Nj1G

ϕd(i j)log(G


Ni11113936

Nj1maxij G

ϕd(i j)1113966 1113967


Nj1G

ϕd(i j)




μx 1113936Ni11113936

Nj1jG

ϕd(i j) μy 1113936

Ni11113936

Nj1iG

ϕd(i j)

σx 1113936Ni11113936

Nj1(jminus μx)2G

ϕd(i j)

σy 1113936Ni11113936

Nj1(iminus μy)2G

ϕd(i j)


Nj1

11+(iminus j)2


Ni11113936

Nj1(ilowastj)G

ϕd(i j)

Variance 1113936Ni11113936

Nj1(1minus μ)2G








Data Availability




References





























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




Data Availability




References





























RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


researcharticle rank ...healthy tissues extraction can help to provide gbm structure segmentation,...

Documents