# Introduction umor is one of the most common brain diseases; hence, its diagnosis and treatment are compulsory. In recent years, developments in medical imaging techniques have aided experts in numerous domains of medicine, such as computeraided diagnosis, follow-up of pathologies, surgical planning and surgical guidance. Among the available medical imaging techniques, MRI is a popular neuroimaging technique for the evaluation and treatment of brain tumors. MRI devices generate images in sagittal, axial, and coronal planes that give a better localization of a lesion in a 3D space of the brain. The segmentation of MR brain images is a complex task, as it involves a large amount of data. Few artifacts might arise due to a patient's motion or limited acquisition time; also, soft-tissue boundaries of tumor are vaguely defined. There exist a numerous classes of tumors with different shapes and sizes that appear at any location with dissimilar image intensities. A few of them might also deform the nearby structures or be associated to necrosis or edema that affects the image intensities in and around the tumor. A number of popular brain tumor segmentation methods are briefly mentioned here. Clark et al. [1] proposed a tumor segmentation method by a set of rules with fuzzy classification and a knowledge base. Dou et al. [3] proposed a fuzzy information fusion frame work for tumor segmentation. Gordillo et al. [4] developed a fully automatic and unsupervised brain tumor segmentation method. Hassan Khotanlou et al. [5,6] introduced a tumor segmentation using symmetry analysis, fuzzy classification, and spatially constrained deformable models. Jingxin Nie et al. [9] proposed expectation maximization and a spatial accuracyweighted hidden markov random field for tumor segmentation. Vida Harati et al. [15] proposed an automatic tumor segmentation technique based on improved fuzzy connectedness algorithm. Despite various efforts and promising results in the medical imaging area, exact segmentation and characterization of tumors are still difficult and challenging. This paper proposes a method that combines coclustering and morphological operations such as erosion, dilation and hole filling [12]. Firstly, morphological operations [13,14] are performed on contrast-enhanced axial T1-weighted MR Head scans to remove non-brain data (skull, fat, skin, and muscle) to enhance tumor segmentation efficiency. These external tissues often interfere with a brain tissue during segmentation which accounts for poor segmentation efficiency. Secondly, coclustering algorithm [2,10] is applied to segment brain tumor, and the volume of the tumor was evaluated by taking patient 1 of Table I. Performance of the proposed method was judged by comparing it with the most popular FCM algorithm [8]. # II. # Methodology Our proposed method was divided into two parts. The output obtained from one part was taken as an input to the next part. The process flow chart is given in Figure1.The skull region was removed in stage I by morphological operations, and the tumor was extracted in stage II using co-clustering algorithm. Abstract -Magnetic Resonance Imaging (MRI) is a significant imaging technology for brain tumor diagnosis because physicians can identify precise pathologies by studying the variations of tissue characteristics that occurs in various kinds of MR brain images. Segmentation of MRI is a pre-process in determining the volume of different brain tissues, but here tumor detection is of primary concern. We proposed a method to extract tumors as seen through MR brain images using coclustering and morphological operations and its volume estimation was done by Cavalier's estimator of morphometric volume method. Quantitative analysis showed that the proposed method yielded better results in comparison with Fuzzy C-Means algorithm (FCM). The conventional analysis of MR brain images with tumor by an expert is a complex and timely process. Thus, an automatic tumor segmentation method is desirable to give an adequate performance. As a result, tumor volume can be evaluated for follow-up of the disease effectively. The objective of the process was to generate the brain region of the axial T1-weighted MR head scans by morphological operations namely, erosion and dilation [13,14]. Erosion operation was performed on the selected image (I) and used to split the image's weakly attached regions. The image was eroded by an octagonal structuring element. The eroded image I 1 , would have several disconnected regions and was obtained using (1). ( )1 Where B is the octagonal structuring element with a size of The preceding process decomposed the given image into several isolated regions. Then, the eroded image was converted to a binary image, I 2 , using (2). (2) recapture the brain tissues that were lost in the process of erosion. It was done with the same structuring element that was used for erosion. The dilated image, I 3 , was obtained using (3). (3) Thereafter, the binary mask was obtained by filling the holes in the dilated image. The binary mask was, thus, convolved with the original image to get the final brain portion. The results of the skull removal process are shown in Figure 2. b) Volume estimation Cavalier's estimator of morphometric volume method [11].The Cavalier's method was formulated on equally spaced slices of images using equation ( 11). Where'd' is the distance between each slice, 'y i ' is the area of slice 'i', 'n' is the total number of slices, 't' is the slice thickness, and 'y max ' is the maximum value of 'y'. IV. # Results and discussions The proposed method was verified on MRI brain image data sets of five patients named as patient 1 to patient 5, and one slice was selected from each patient's data set in random to evaluate the performance of the proposed method. Volume estimation has been done by taking all necessary slices of patient 1. These data sets have been acquired on a Philips Achieva 1.5T apparatus by an axial T1-weighted sequence. The slice thickness for patient 1 was 1.5mm, and the spacing between the adjacent slices was 0.9mm. The details of the data sets used are given in Table I. Data sets were collected from the Department of Radiology and Imaging Science, Apollo Health City, Hyderabad, India. The evaluated segmentation results at pixel level are shown in Table II, and the extracted result of a meningeal tumor is shown in Figure .4 for the slice 105 of patient 3. It can be observed that the results exhibit close proximity to the manually segmented images by the experts and are superior to FCM. Quantitative results obtained by the proposed method in comparison with FCM are given in Table III. From the Table III, the SI of the proposed method varies from 87.73% to 94.08% but for FCM it is The evaluation of tumor extraction results by the proposed method and the FCM algorithm was compared with the manually segmented tumors to measure its effectiveness. The manual segmentations were provided by medical experts, which might include abnormal tissues, like edema, along the tumor region. Let M be the manually segmented tumor and A be the segmented tumor by the proposed method or the FCM algorithm asshown in Figure 3. Here, the Similarity Index (SI), Correct Detection Ratio (CDR), Under Segmentation Error (USE), Over Segmentation Error (OSE), Hausdorff Distance (HD) and Average Surface Distance (ASD) were used for efficiency evaluation. SI is a measurement that gives the true segmented region, which is relative to the total segmented region in both the segmentations.CDR value indicates the degree of trueness of the actual tumor. USE is the ratio of the number of voxels falsely identified as tumor portion by the proposed method to the manually segmented tumor. OSE is the ratio of number of voxels falsely identified non tumor region by the proposed method to the manual segmented tumor. Total Segmentation Error (TSE) is the sum of USE and OSE. HD is the largest difference between two surfaces and ASD illustrates how much the two surfaces differ on average. All these performance metrics also apply to FCM algorithm. The evaluation metrics SI, CDR, USE, OSE, HD, and ASD were obtained by using equations ( 5), ( 6), ( 7), ( 8), ( 9) and ( 10) respectively [5,6,7,15]. Where True Positive (TP) is the number of pixels detected correctly, False Positive (FP) is the number of pixels detected falsely as tumor and False Negative (FN) is the number of pixels detected falsely as non tumor. # HD (M,A) = max (h(M,A), h (A,M)) Where ?(??, ??) = ?????? ?? ??? ?????? ????? ??(??, ??) , and ??(??, ??)denotes the Euclidean distance between ?? and ?? (?? and ?? are points of M and A respectively) The volume of tumor V t was estimated by 72.16% to 87.76% .The CDR of the proposed method ranges from 85.52% to 95.86%, and for FCM it is 85.69% to 91.61%. The TSE of the proposed method changes from 12.07% to 26.24%, and for FCM it is 24.75% to 70.69%.HD ranges from 5.4772 to 6.8557 for the proposed method and 5.6569 to 6.8557 for FCM, which proves a good position of the periphery of the tumors of the proposed method. ASD changes from 0.6375 to 1.3081 for the proposed method and 0.8761 to 1.5530 for FCM. The volume of tumor for patient 1 is shows close proximity as shown in Table IV. V. # Conclusion The presented tumor segmentation method was tested on five abnormal MRI brain slices of different patients, and the volume was evaluated for one of the patients. It was observed that integrating co-clustering with morphological operators minimizes segmentation error when compared with FCM. The qualitative evaluation of the obtained results for the proposed tumor extraction method achieved a good performance. A future scope in this area intends in finding the type of tumor based on ontology of tumors and segmentation of edema. 1![Figure 1 : Proposed methodology for tumor extraction a) Skull removal](image-2.png "Figure 1 :") ![-clustering algorithmClustering is a collection of similar image gray levels divided into segments. Co-clustering[2,10] is the simultaneous partitioning of the rows and columns of an image matrix. This algorithm could be used to find algorithm can be summarized as follows:1. Form 2. Compute [L = log 2 k] singular vectors of I n 3. Apply singular value decomposition (SVD) technique on to obtain (4) Where U and V represent the left and right eigenvector of 2 nd to (L+1) th eigenvalues. 4. Form the matrix © 2012 Global Journals Inc. (US) Global Journal of Computer Science and Technology Volume XII Issue XII Version I brain image of size (I i,j ) and number of clusters (k) Output: Segmented MR brain image with k clusters where and and ] Tumor Extraction and Volume Estimation for T1-Weighted Magnetic Resonance Brain Images Erosion Binarization Dilation We performed a dilation operation on I 2 to k-clusters of MR brain images. The co-clustering 5. The k-means algorithm is applied on the L-dimensional data Z to get the k number of clusters. 6. Find out the mean of the centers and their indices from the obtained clusters.](image-3.png "") 273![Figure 2 : Skull removal process results (a) Original MR brain image I; (b) Eroded image I 1 ; (c) Binarized image I 2 ; (d) Dilated image I 3 ; (e) Brain mask ;(f) Segmented brain image](image-4.png "Figure 2 : 7 .Figure 3 :") ![Volume Estimation for T1-Weighted Magnetic Resonance Brain Images](image-5.png "") ![Volume Estimation for T1-Weighted Magnetic Resonance Brain Images](image-6.png "") 4![Figure 4 : Tumor extraction result of a meningeal tumor: (a) One axial slice of the selected tumor class; (b) Extracted tumor by the FCM; (c) Extracted tumor by the proposed method; (d) Manually segmented tumor](image-7.png "Figure 4 :") I II III IVTumor Extraction and Volume Estimation for T1-Weighted Magnetic Resonance Brain ImagesPatientGenderAgeTumor Type1Male54Meningioma2Male42Cystic glioma3Female48Meningioma4Male31Dural tumor201241,486mm 3 for the proposed method when compared with the value of 43,547mm 3 for manual segmentation5Male46MetastatisYear12Volume XII Issue XII Version IPatient 1 2 3 4 5Slice Number 83 95 105 112 120TP 420 988 1255 1201 797FCM FP 243 133 191 258 542FN 64 165 159 150 73Proposed Method TP FP FN 454 97 30 986 92 167 1239 12 175 1260 78 91 834 69 36D D D D ) D(Global Journal of Computer Science and TechnologyPatient 1 2 3 4 5Slice No. 83 95 105 112 120FCM OSE 72.23 86.78 50.21 13.22 63.43 5.6569 0.8761 87.73 93.80 20.04 Proposed Method SI CDR USE TSE HD ASD SI CDR USE OSE TSE 6.20 26.24 5.4772 0.6375 HD ASD 86.90 85.69 11.54 14.31 25.85 6.3246 1.1738 88.39 85.52 7.98 14.48 22.46 6.4807 1.1296 87.76 88.76 13.51 11.24 24.75 6.7823 1.4961 92.98 87.62 0.85 12.38 13.23 6.5574 1.3025 85.48 88.90 19.10 11.10 30.20 6.8557 1.5530 93.72 93.26 5.77 6.74 12.51 6.8557 1.3081 72.16 91.61 62.30 8.39 70.69 6.2450 1.4178 94.08 95.86 7.93 4.14 12.07 6.4031 0.9403 Volume of tumor using manual Volume of tumor using the segmentation (mm 3 ) proposed method (mm 3 )43,54741,486 © 2012 Global Journals Inc. (US) * Automatic tumor segmentation using knowledge-based techniques MCClark LOHall DBGoldgof RVelthuizen FRMurtagh MSSilbiger IEEE Transactions on 17 2 April 1998 Medical Imaging * Co-clustering documents and words using bipartite spectral graph partitioning InderjitSDhillon Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining -KDD '01 the seventh ACM SIGKDD international conference on Knowledge discovery and data mining -KDD '01San Francisco, California 2001 * A framework of fuzzy information fusion for segmentation of brain tumor tissues on MR images WDou SRuan YChen DBloyet JConstans Image and Vision Computing 25 2 2007 Februvary * A new fuzzy approach to brain tumor segmentation NGordillo EMontseny PSobrevilla Fuzzy Systems (FUZZ), 2010 IEEE International Conference on July 2010 * 3D brain tumor segmentation in MRI using fuzzy classification, symmetry analysisand spatially constrained deformable models HassanKhotanlou OlivierColliot JamalAtif IsabelleBloch Fuzzy Sets and Systems 16 May 2009 160 * Adaptive segmentation of internal brain structures in pathological MR images depending on tumor types HassanKhotanlou JAtif EAngelini HDuffau IBloch 4th IEEE International Symposium on 2007. ISBI 2007. April 2007 Biomedical Imaging: From Nano to Macro * Automatic volumetric liver segmentation from MRI data HaidiIbrahim MariaPetrou KevinWells SimonDoran OysteinOlsen International journal of computer theory and engineering 2 2 April, 2010 * Pattern Recognition with Fuzzy Objective Function Algorithms JCBezdek 1981 Plenum Press New York * JingxinNie ZhongXue TianmingLiu GeoffreySYoung KianSetayesh LeiGuo Stephen T.C * Automated brain tumor segmentation using spatial accuracy-weighted hidden Markov Random Field Wong Computerized Medical Imaging and Graphics 33 6 September 2009 * MRege ; MingDong * Co-clustering Documents and Words Using Bipartite Isoperimetric Graph Partitioning FFotouhi Data Mining, 2006. ICDM '06. Sixth International Conference on 18-22 Dec. 2006 * Brain volume estimation from serial section measurements: a comparison of methodologies GDRosen JDHarry J Neurosci Methods 35 2 1990 Nov * Automatic tumor extraction for contrast enhanced axial T1-weighted magnetic resonance brain images integrating co-clustering and morphological operations SSatheesh DrK V S VPrasad DrKReddy International Conference on Signal, Image and Video Processing (ICSIVP), IITPatna January -2012 * Fully automatic brain extraction algorithm for axial T2-weighted magnetic resonance images KSomasundaram Kalaiselvi Computers in Biology and Medicine 40 10 October 2010 * Automatic brain extraction methods for T1 magnetic resonance images using region labeling and morphological operations KSomasundaram Kalaiselvi Computers in Biology and Medicine 41 8 August 2011