An Analysis the Brain Tumor using the Hybrid Classification Approach

The digital image processing can play an important role in medical side to identify the disease and tumors. A digital image can be collected of fixed number of elements known as image element or pixels, every one of which contains an intensity value and exacting position. In the medical area, the MRI can be mostly utilized for finding the idea of details in the body‘s inner structure. In common, it has been applied to identify the variations in the tissues of body which can be measured to be the enhanced method as contrasted to estimated tomography. Consequently, this method is used as a unique method particularly for the cancer imaging & brain tumor recognition. In this chapter, the brain tumors are detected and segmented using HC approach. The HC approach stated in this chapter detects the tumor affected brain MRI image and then segmentation approach is applied on the detected tumor affected brain MRI image in order to segment the tumor regions. Finally, the performance of the proposed tumor detection method is associated with the other conventional methods. Figure 1(a) shows the proposed brain tumor detection system using HC approach in training mode and Figure 1(b) shows the proposed brain tumor detection system using HC approach in testing mode.

During medical image acquisition process, the noises are generated which affects the performance of the tumor detection and segmentation process. Hence, the noises in the source brain MRI images are detected and reduced before the tumor detection process starts in order to improve the tumor segmentation accuracy. In this chapter, Gaussian noises in the source brain MRI images are detected and reduced using bilateral filter (Yunlong He et al. 2017). Even though the conventional noise reduction filters such as Mean filter and Median filters detects and reduces the Gaussian noises. But they degrade the edge pixels during noise reduction process. In order to achieve edge preserving during noise reduction process, bilateral filter is selected in this chapter to detect and reduce the noises in source brain MRI images.

The pixels belonging to noise reduced brain MRI image are spatial or time domain. The pixel behaviour can be improved in the multi resolution domain format. In order to transform the time domain pixels in the noise reduced image into multi resolution domain format, Gabor transform is used in this chapter (Palm et al. 2002 and Kumar et al. 2018). Time-domain pixels are multiplied by the Gabor kernel to produce multi-resolution images using the Gabor transform, which is a type of Gabor filter.

In this chapter, GLCM features and Pixel metric features are computed from the Gabor transformed

When analyzing a multi-resolution brain image, GLCM features are used to calculate the texture qualities of each individual pixel. In this section, we compute these characteristics using the GLCM matrices with the pixels oriented at a ninety-degree angle. In this chapter, 135 number of brain MRI images (70 images belonging to normal category and 65 images are belongs to abnormal category) are trained and the features are computed from all of these images. Each brain MRI image produces seven number of feature values. In this chapter, 70 normal brain images are used. Hence, the total number of features computed from the normal category is about 520 (70 normal images * 7 features). In this chapter, 65 abnormal brain images are used. Hence, the total number of features computed from the abnormal category is about 490 (65abnormal images * 7 features). Hence, the total number of features is computed by adding the total number of abnormal and normal features extracted from the brain image which is about 1015. All these computed feature values are stored in a feature vector.

The Gabor transformed brain MRI image is decomposed into four number of sub band metrics (sub band coefficients are represented as (𝑖,)) using DWT in this chapter. The size of each decomposed sub band metric is equals to the one by fourth of the source brain MRI image. Each brain MRI image produces four decomposed sub bands and each decomposed sub band produces three features. Hence, the total number of features computed is about 10 for a single brain MRI image. Thus the total number of features computed from the normal category is about 895 and the total number of features computed from the abnormal category is about 840. Hence, the total number of features computed is about 1690. All these computed feature values are stored in a feature vector. The feature vector size of the GLCM feature is 1010 and the feature vector size of the DWT feature is 1690. The total number of features computed in this proposed work is about 2750. Table 1 shows the extracted GLCM and DWT feature values for both normal and abnormal case brain images. These values show the average values of the features which are computed from the brain images. These values are used by the classifier to classify the source brain images into either normal or abnormal based on these features.

GLCM

Contrast (C) 2.28×102 4.19×102 Energy (E) 12.72×104 75.18×104 Homogeneity

(H) 0.01837 1.7283

Entropy (EN) 1.28 17.29 Maximum Probability

(MP) -0.172 1.823

Mean Metric 1

12.37 762.12

Mean Metric 2

11.29 673.19

DWT Energy Metric (EM)

128.87×1012 17.872×1012

Index Metric

(IM) 1.72 78.20

Invariance Feature (IF)

-0.0928 2.182

In this research work, the NN is designed with single input and output layer and with 5 numbers of hidden layers. The number of neurons in input layer of this proposed design is equipped with 15 and the number of neurons in output layer of this proposed design is equipped with single neuron. Each hidden layer of this proposed NN design consists of 25 neurons. In this study, the optimal classification rate is achieved by selecting the number of hidden layers & number of neurons in each hidden layer over a series of iterations. During training process of the hybrid classifier (NN stage), the edge pixels are detected using ‗Canny‘ edge detector in both normal and abnormal category. These detected edge pixels are trained by NN classifier which produces the edge trained pattern. So that, the input of the NN classifier is edge pixels of the brain image to be tested. During training process of the hybrid classifier (ANFIS stage), the features GLCM and DWT are computed from the both normal and abnormal category of images. These computed features along with the edge The architecture of the ANFIS classifier used in this proposed work is depicted in the following Figure 3. In this proposed architecture, ANFIS is designed with five numbers of layers as input layer, fuzzification layer, defuzzification layer, normalization layer and output layer. The input layer is connected with the extracted feature vector. The extracted features from the normal brain MRI images are connected with the input function X and the extracted features from the abnormal brain MRI images are connected with Y in this proposed architecture. The fuzzification layer applies fuzzy logic on each feature set and the defuzzification layer eliminates the negative values in the responses. The normalization layer performs unit normalization on the computed responses and the final output layer is connected with the normalization layer.

During testing process of hybrid classifier (NN stage), the edge pixels are detected using ‗Canny‘ edge detector in the test brain MRI image and the detected pixels are combined with the extracted features from the test brain MRI image. These computed pixels and features are classified using ANFIS classifier which produces binary values as 0 and 1. The normal brain MRI image is identified with 0 binary response values and the abnormal brain MRI image is identified with 1 binary response value.

Tumor pixels in an identified abnormal MRI of the brain are the focus of the segmentation procedure. In this chapter, morphological segmentation approach (dilation followed by erosion) is used to segment the tumor pixels.

The procedure presented in this study is tested in a simulated environment utilizing MATLAB R2018 & corresponding hardware configurations of 4GB RAM & 2.4GHz processor. Within the context of this chapter, the hybrid classifier system is trained using a dataset consisting of 70 normal brain MRI pictures & 65 abnormal brain MRI images (BRATS 2015 dataset). Table 3.2 also shows the results of testing the proposed hybrid classifier system with 147 examples of normal brain MRI images & 99 examples of abnormal brain MRI images. This results in a total of 222 normal brain MRIs & 169 abnormal ones.

Training images 70 65 Testing images 147 99 Total images 222 169 In this section, we evaluate the success of the constructed framework by looking at the classification rate, which is the percentage of correctly identified photos relative to the total number of images. The suggested system uses a NN classification strategy to label 112 out of 147 brain MRI pictures as normal. The suggested system uses a NN classification strategy to label 76 out of 99 brain MRI pictures as abnormal. Over 152 photos, the proposed system labels 134 as normal, utilizing an ANFIS classification strategy. Over a total of 99 pictures, the suggested system uses an ANFIS classification approach to label 87 as abnormal. Out of 147 brain MRIs, the suggested algorithm deems 141 to be normal. Over 99 brain MRI pictures, the proposed system correctly labels 92 as abnormal, thanks to the use of the HC method.

The proposed brain tumor detection system is tested in this chapter with NN, ANFIS and Hybrid classifier for evaluating and analyzing the performance of the proposed method in terms of classification rate (CR). The proposed brain tumor detection system using NN classifier achieves 73.6% of CR for normal category and achieves 73% of CR for abnormal category. Hence, the average CR for NN case is about 73.3%. The proposed brain tumor detection system using ANFIS classifier achieves 88.1% of CR for normal category and achieves 83.6% of CR for abnormal category. Hence, the average CR for ANFIS case is about 85.5%. The proposed brain tumor detection system using Hybrid classifier achieves 92.7% of CR for normal category and achieves 88.4% of CR for abnormal category. Hence, the average CR for Hybrid classifier case is about 90.5%. The proposed tumor detection system is also tested with respect to the extracted features. The proposed system with GLCM feature alone achieves 67.1% of CR for normal category and achieves 69.6% of CR for abnormal category. The proposed system with DWT sub band feature alone achieves 72.7% of CR for normal category and achieves 70.5% of CR for abnormal category. The proposed system with GLCM and DWT sub band feature achieves 92.7% of CR for normal category and achieves 88.4% of CR for abnormal category, as depicted in Table 5.

GLCM features 67.1 69.6

DWT 72.7 70.5 GLCM+ DWT 92.7 88.4

The evaluated characteristics & metrics of the simulation environment are then applied to the examined designed architecture or framework. Accuracy is defined as the sum of the numbers of tumor & non-tumor pixels correctly detected in the final tumor region segmented brain picture, while sensitivity & specificity determine the number of pixels that are appropriately connected with the tumor region. Well-detected non-tumor pixels are defined by precision, and the Disc Similarity Index indicates the total amount of tumor detected pixels that are comparable to ground truth images.

Table 8 is the simulation results of the proposed brain tumor detection system using HC approach. The proposed system using HC approach obtains 91.54% Sen, 96.2% Spe, 98.4% A, 88.5% Pre, 94.8% FS and 93.2% DSI. In addition, all of the brain MRI pictures in the open dataset are used to evaluate the proposed method and produces the similar simulation results on the entire number of brain MRI images in the dataset.

1 90.7 90.9 98.2 83.5 95.6 90.1 2 91.7 95.8 91.4 85.7 97.6 91.2 4 90.6 92.5 92.7 95.4 92.6 92.6 5 89.7 95.1 95.2 96.8 .95.6 98.4 6 90.5 92.3 91.5 97.3 94.6 95.8 7 91.8 94.1 93.5 88.5 93.5 96.6 8 98.6 98.5 96.6 84.6 98.5 97.7 9 94.5 97.4 93.8 93.6 91.0 95.2 10 93.5 91.5 91.2 96.8 92.4 93.8

Average 91.54 96.2 98.4 88.5 94.8 93.2

Table 9 is the comparison analysis of proposed method with respect to classification approaches. Based on the simulation findings, the proposed brain tumor detection system employing the HC technique achieves impressive tumor segmentation results in comparison to existing classification approaches.

Simulation results are compared to those of more traditional approaches in Table 10. Milica M Badža et al. (2020) and Tripathi et al. (2019) on same dataset images.Milica M Badža et al. (2020) obtained 96.5% Sen, 96.1% Spe and 92.1%. Tripathi et al. (2019) obtained 92.8% Sen, 94.5% Spe and 94.1% A. The conventional methods for brain tumor detection used single classification process to classify the brain images into either normal or abnormal. In this chapter, the integration of two classifiers is used to improve the tumor image detection accuracy. Comparisons with other traditional approaches on the same dataset of brain MRI images suggest that the suggested system employing the HC approach produces high simulation values.

task for clinical applications. This thesis develops brain tumor detection and diagnosis system using machine and deep learning algorithms. The approach detects and segments the brain tumors using machine learning based Hybrid Classification method. Finally, the performance of the proposed tumor detection method is compared with the other conventional methods. The proposed brain tumor detection system is tested in this chapter with NN, ANFIS and Hybrid classifier for evaluating and analyzing the performance of the proposed method in terms of classification rate.

1. Cui, S, Shen, X& Lyu, Y 2019,‗Automatic segmentation of brain tumor image based on region growing with co-constraint‘, in International Conference on Multimedia Modeling, Proceedings of the MMM 2019: MultiMedia Modeling, Thessaloniki, Greece, Springer: Berlin/Heidelberg, Germany, vol. 11295,pp.178-188. 2. Cui, W, Wang, Y, Fan, Y, Feng, Y & Lei, T 2013, ‗Localized FCM clustering with spatial information for medical image segmentation and ias field estimation‘, Journal of Biomedical Imaging, pp. 1-8. 3. Da Cunha, AL, Zhou, JP & Do, MN 2006, ‗The non subsampled contourlet transform: Theory, design, and applications‘, Transactions on Image Processing, IEEE, vol. 15, pp. 3089-3101. 4. Dahab et al, ―Automated Brain Tumor Detection and Identification Using image Processing and Probabilistic Neural Network Techniques‖, Vol 1, Issue 2, pp 1-8, 2012 5. Daimary, D, Bora, MB, Amitab, K& Kandar, D 2020,‗Brain tumor segmentation from MRI images using hybrid convolutional neural networks‘, Procedia Comput. Sci.,vol. 167, pp. 2419-2428. 6. Das A,‖ A Study on Prognosis of Brain Tumors Using Fuzzy Logic and Genetic Algorithm Based Techniques―, International Joint Conference on Bioinformatics, Systems Biology and Intelligent Computing, IJCBS '09, pp 348 - 351, 2009. 7. Davy, A, Havaei, M, Warde-Farley,D& Biard,A 2014, ‗Brain tumor segmentation with deep neural networks‘, in: Proc MICCAI-BRATS. 8. Dawant, BM, Hartmann, SL, Pan, S& Gadamsetty, S 2002, ‗Brain atlas deformation in the presence of small and large space-occupying tumors‘, in: Comput. Aided Surg.,vol. 7, no. 1, pp. 1-10. 9. Dawngliana M., Deb D., Handique M., Roy S., "Automatic brain tumor segmentation in MRI: Hybridized multilevel thresholding and 10. Debelee, TG, Kebede, SR, Schwenker, F& Shewarega, ZM 2020,‗Deep learning in selected cancers‘, Image Analysis - A Survey. J. Imaging, vol. 6, P. 121. 11. Deepak Agrawal, Vyomkulshreshtha & Pankaj Sharma 2018, ‗Brain tumor detection using K-means clustering and threshold segmentation‘, International Journal of Advanced Research in Science, Engineering and Technology, vol. 5, no. 3. 12. Deepali Aneja, Tarun Kumar Rawat, ―Fuzzy Clustering Algorithms for Effective Medical Image Segmentation‖, International Journal of Intelligent Systems and Applications, Edition 11, pp 55-61, 2013 13. Deepthi Murthy, T.S, "Brain tumor segmentation using thresholding, morphological operations and extraction of features of tumor", IEEE International Conference on Advances in Electronics, Computers and Communications (ICAECC), 10-11 Oct. 2014, pp 1 -6 14. Demirhan, Törü, M & Güler, I 2015, ‗Segmentation of tumor and edema along with healthy tissues of brain using wavelets and neural networks‘, Journal of Biomedical Health and Information, IEEE, vol. 19, pp. 1451-1458. 15. Devkota, B, Alsadoon, A, Prasad, PWC, Singh, AK & Elchouemi, A 2018, 'Image segmentation for early stage brain tumor detection using mathematical morphological reconstruction', Procedia Computer Science, vol. 125, pp. 115-123. 16. Dhage P., Phegade M.R., Shah S.K., "Watershed segmentation brain tumor detection," 2015 International Conference on Pervasive Computing (ICPC), pp.1-5, 2015. 17. Dhamodharan & Raghavan 2015, ‗Combining tissue segmented and neural network for brain tumor detection‘, International Arab Journal of Information Technology, vol. 12, no. 1, pp. 45-52. 18. Dickson, S& Thomas, B 1997, ‗Using neural networks to automatically detect brain tumors in MR images‘, In: Int J Neural Syst.,vol. 4, no. 1, p. 91-99.

Research Scholar, Shri Krishna University, Chhatarpur M.P.