Brain Tumour Segmentation using CNN and WT

 

B. Rajasekar

Associate Professor, Department of ECE, Sathyabama Institute of Science and Technology, Chennai-119

 

ABSTRACT:

The abnormal development of cells in brain leads to the formation of tumours in brain. In this paper, image segmentation based brain tumour detection and segmentation methodology is proposed using convolutional neural networks (CNN). This proposed methodology consists of image segmentation, feature extraction, classification, and segmentation. Wavelet transform (WT) is used for image segmentation and enhanced brain image is obtained by fusing the coefficients of the WT transform. Further, Grey Level Co-occurrence Matrix features are extracted and fed to the CNN classifier for glioma image classifications. Then, morphological operations with closing and open-ing functions are used to segment the tumour region in classified glioma brain image.

 

 

1. INTRODUCTION:

As per the report 2016 of American Brain Tumour Association (ABTA), 61 200 persons in United States of America (USA) are affected by brain tumour. WHO (2016) reported that the persons affected by brain tumour is 6.2 million worldwide. The tumour is formed in brain due to the abrupt development of the cells in brain region. The initial stage of the tumour formation in brain region is primary brain tumours (PBT) and they are further treated as benign. The tumours in this category are having homogeneous structures. The severe stage of the tumour in brain region is secondary brain tumours (SBT) which is also called as malignant. The brain tumours are differentiated with their shape, size, intensity, and density. The benign tumours can be controlled and cured by medication. The untreatable benign tumours become malignant tumours. The malignant tumours are the uncontrollable tumours which can be cured only by surgery. The brain tissues in brain are gray matter, white matter, and cerebro spinal fluid. These brain tissues play an important role in the process of brain tumour classification. The tumours are having different shape, size, and located in differ-end areas in brain region which makes the tumour detection and segmentation process more complex.

 

The brain tumour segmentation by physician is time consuming and error probe process. The manual detection of brain tumour is not suitable for large population countries. Hence, there is a need for auto-mastic detection and segmentation of brain tumours. The  malignant tumours in brain image are classified as glioma and meningioma. Figure 1A shows the glioma tumour brain magnetic resonance imaging (MRI) and Figure 1B shows the meningioma tumour affected brain image. Section 2 explains about the survey of tumor and 3 explain the overall description and 4 explain the result of the proposed system.

 

2. LITERATURE SURVEY:

Bahadure et al. used wavelet transform method to detect abnormal lesions in brain region. The authors used Berkeley type wavelet kernels which were symmetric with its decom-position level. This multi level decomposition sub-bands were trained properly with support vector machine (SVM) classifier. The sensitivity rate about 97.72%, specificity rate about 94.2%, and accuracy rate about 96.51% were achieved by authors on BRATS 2015 dataset [1].

 

Sreedhanya and Pawar applied hybrid classification technique on pre-processed brain MRI image. Gaussian mixture model was developed by authors to represent the texture features of the brain image for improving the segmentation accuracy [2].

 

 

Fast Fourier trans-form was used by Alfonse and Salem to transform the spatial domain features into frequency domain features. Maximal-relevance classifier was used to train the extracted features from the brain image. The classification accuracy of 98.9% was obtained by optimizing the extracted features [3].

 

FIG 1: Tumour affected brain MRI image:

(A) Glioma, (B) meningioma

 

Perez et al. analyzed the different meta stages of the various regions in brain images. Based on this analysis, the mass formation on brain tissues were detected and diagnosed. The shape and size of this extracted mass region were used to identify the severity of the tumour level. Su et al.⁵ detected abruptly developed cells in histopathological cell images [4].

 

The extracted features were trained by adaptive dictionary algorithm. The authors applied their proposed algorithm on large histopathological dataset images to validate their proposed experimental results. Rajesh Chandra et al.⁶ proposed brain tumour segmentation approach using Genetic Algorithm (GA) [5].

 

The disadvantages of the Ppaper Swarm Optimization meth-otology was overcome by proposing GA in classification process of normal and abnormal patterns in brain image. The mutation and crossover stages in GA process were well con-strutted by selecting proper features from brain image [6].

Isselmou et al. used threshold method for segmenting abnormal regions from normal regions in brain MRI image. Then, tumour region was detected from these segmented abnormal regions using morphological operations.

 

In this work, sharpening filter was used in order to smooth the regions in brain image which improved the tumour segmental-tion accuracy. The authors achieved 98.55 of tumour segmentation accuracy. This methodology was tested on less number of brain images only as main limitation of this work [7].

 

3. MATERIALS AND METHODS:

3.1. Materials:

In this paper, BRATS 2015 database⁸ is used to validate the proposed tumour detection and segmentation methodology. BRATS 2015 database⁸ contains three sub datasets as Training, Leader board and Challenge. Each subset contains 65 MR brain images from different patients. The training subset consists of 20 high grade tumour MR images and 10 low grade tumour MR images, respectively. The leader board subset consists of 21 high grade tumour MR images and 4 low grade tumour MR images, respectively, where as Challenge subset consists of 25 high grade tumour MR images and 10 low grade tumour MR images, respectively. This dataset also includes manually segmented tumour images which are obtained from expert radiologist.

 

3.2. Methods:

In this paper, convolutional neural networks (CNN) classifier based brain tumour detection and segmentation methodology is proposed using image segmentation technique. The enhancement of brain image is achieved using image segmentation and enhanced brain image is obtained by fusing the coefficients of the WT trans-form.

 

FIG 2. Overview of the proposed method.

Further, Grey Level Co-occurrence Matrix (GLCM) features are extracted and fed to the CNN classifier for Glioma image classifications. Then, post processing steps are applied on the classified glioma brain image in order to segment the tumour regions. The overview of the proposed brain tumour detection and segmentation methodology is depicted in Figure 2. 

 

3.3. Image segmentation:

The accurate detection and segmentation of tumour region in brain image requires proper enhancement. In this paper, WT-based image segmentation technique is used to improve the contrast of the brain image.

 

The source brain MRI image is rotated at different orientations in bilinear interpolation method⁹ to obtain the tilted brain MRI image. In this paper, the source brain image is tilted at the orientation of 108 using bilinear transformation approach.

 

The source (non tilted) brain MRI image and tilted brain MRI image are decomposed into four sub-bands as “Approx-imate,” “Horizontal,” “Vertical,” and “Diagonal” using two-dimensional WT.

 

The low pass sub-bands of both source and tilted brain MRI images are fused in arithmetic manner in order to obtain unique low pass sub-band and high pass sub-bands of both source and tilted brain MRI images. Then, inverse WT is applied on both low and high pass sub-bands to produce the fused brain image. The fused brain MRI image is shown in Figure 3.

 

3.4. Feature extraction:

This feature is used to differentiate the glioma brain image from the non-glioma brain image. The GLCM features are extracted from GLCM matrix, which can be constructed at the orientation of 458. The elements in GLCM matrix repre-sent the number of counts which correlates each pixel with its adjacent pixels at the orientation of 458.

 

The GLCM features as contrast, energy, entropy, and correlation are extracted from the brain MR image for glioma image classifications (Bahadure et al.¹). Table 1 shows the GLCM features of both glioma and non-glioma brain images, respectively. It is very clear from this table that there is a significance difference between glioma and non-glioma brain images.

 

3.5. Classifications:

In this paper, CNN classifier is used to classify the brain MRI image into either normal or abnormal. In this classifier, the features which are extracted from the fused brain image are convolved with kernels in order to produce the feature map.

Kernels in each layer of the CNN classifier have differ-ent weights and each feature map is connected to the next layer of the CNN classifier through different weights of the kernels by back propagation methodology.

 

CNN functions are optimized in order to differentiate the normal and abnor-mal brain images using training mode of this classifier. CNN classifier outperforms to the other conventional classifiers in terms of classification rate.

 

TABLE 1: Extracted GLCM features

 

Generally, brain tumours are having large variability in their intra structures, which makes the tumour region segmentation process more complex.

 

The internal architecture of CNN classifier is designed and tuned in order to reduce such complexity for differentiating the abnormal tumour image from the normal brain image. The flow diagram for the architecture of CNN classifier is depicted in Figure 3.

 

FIG 3. Flow diagram of CNN architecture.

 

The flow diagram of CNN architecture consists of the following layers as Initialization, Activation function, Pool-ing, Regularization, and Data augmentation. The convergence of the layer is achieved by the process of initialization.

 

The activations and their gradients are initialized in this step. The input features are transformed into non-linear features which makes feature map through activation function. In this paper, sigmoid activation function is used to achieve high classification rate.

 

FIG 4 Clustered tumour pixels; B, tumour segmented brain image.

 

Next, feature values in constructed feature map are spatially correlated by pooling function. The computational complexity of the layer is minimized by pooling function.

 

The over fitting of the feature map is reduced by regularization process of CNN layer. The size of the training data-set and over fitting can be reduced using data augmentation process.

 

Further, morphological functions with opening and closing operations are now applied on the classified glioma image. This closed function image is subtracted from open function image to detect the tumour pixels in glioma brain image.

 

The detected tumour pixels in glioma brain image are shown in Figure 5A, B shows the tumor region segmented brain image.

 

 

 

Received on 14.03.2019           Modified on 21.05.2019

Accepted on 28.06.2019         © RJPT All right reserved

Figure 5, shows the brain MRI images from open access dataset, Figure 6B shows the ground truth images which are obtained from expert radiologist and Figure 6C shows the tumour segmented image using proposed methodology stated in this paper.

 

4. RESULTS AND DISCUSSIONS:

In this paper, MATLAB R2014b is used as simulation soft-ware for simulating the proposed brain tumour detection and segmentation methodology. The energy features alone achieve 87% of classification accuracy, while energy with entropy feature achieves 93% of classification accuracy. The proposed sys-tem achieves 97% of classification accuracy by using the integration features of energy, contrast, correlation, and entropy.

 

The following metrics which are widely used in many medical imaging applications are used in this paper to vali-date the performance of the proposed brain tumor detection and segmentation methodology as Sensitivity (Se), Specificity (Sp), Accuracy (Acc), and Dice Similarity Coefficient (DSC). The overlap between proposed brain tumor segmented image and ground truth image is determined by DSC parameter. The following expression is used to express DSC as

 

                 2*TP

DSC=––––––––––––––– ……………………………..(1)

             FP*2*TP*FN

 

Where, TP is true positive which counts the number of correctly detected tumour pixels and FP is false positive which counts the number of correctly detected non-tumour pixels and FN is false negative which counts the number of falsely detected non-tumour pixels, respectively.

 

The performance metrics sensitivity and specificity evaluates the strength of the proposed glioma tumour classification methodology. The high value of sensitivity and specificity shows that the strength of the proposed method-logy is high for clinical practice. The parameter accuracy correlates the metrics sensitivity and specificity. The following expressions are used to express sensitivity, spec-ificity, and accuracy as

 

                TP

Se= ––––––––––…………….………………………..(2)

             TP+FN

 

TABLE 2 Performance analysis of proposed methodology

 

Table 2 shows the performance metrics of the proposed brain tumor detection and segmentation methodology in terms of DSC, sensitivity, specificity, and accuracy. The methodology proposed in this paper achieves 97.3% of Sensitivity, 98.1% of specificity, 98.7% of accuracy, and 0.96 of DSC.

 

The proposed tumour segmentation method stated in this paper achieves 97.4%, 96.6%, and 98.1% of sensitivity on Training, Leader board, and Challenge dataset, respectively. The proposed tumour segmentation method stated in this paper achieves 98.1%, 97.7%, and 98.6% of specificity on Training, Leader board, and Challenge dataset, respectively.

Table 3: Performance comparison of SVM and CNN classification for glioma tumor segmentation

Datasets	Classifiers
	SVM classifier			CNN classifier
	Se (%)	Sp (%)	Acc (%)	Se (%)	Sp (%)	Acc (%)
Training	96.1	96.8	96.5	97.4	98.1	98.7
Leaderboard	95.2	96.1	96.4	96.6	97.7	98.4
Challenge	97.2	98.1	97.9	98.1	98.6	99.1
Average	96.1	97	96.9	97.3	98.1	98.7

 

Also this method achieves 98.7%, 98.4%, and 99.1% of accuracy on Training, Leader board, and Challenge dataset, respectively. Table 3 shows the performance comparison of SVM and CNN classification for glioma tumour segmentation. The CNN classifier based brain tumour segmentation methodology improves sensitivity about 1.23%, specificity about 1.12%, accuracy about 1.82% than the SVM classifier based brain tumour detection method. Used SVM classifier for classifying the brain tumour images. The main limitation of this method is that this method was not suitable for detecting tumours in low resolution brain images. Sreedhanya and Pawar² used hybrid classifier to detect the brain tumour. The thick edges of the abnormal regions in the brain image were not detected by this method which degraded the performance of the brain tumour segmentation system. Alfonse and Salem³ used Maximal-Relevance approach for brain tumour detection and segmentation. The optimal map was constructed using texture features which were extracted from the brain image. The relevance classification was complex task due to the miss classified pixels of the extracted feature set. The comparisons of the experimental results of the proposed method with conventional methods are tabulated in Table 4. From Table 4, it is clear that the proposed methodology for brain tumour detection achieves high performance ratio in terms of sensitivity, specificity, and accuracy. These performance evaluation metrics are now compared with other conventional methods such as SVM classifier, Hybrid classifier, and Maximal Relevance classification approach. These conventional classifiers have no feedback section to improve the performance of the tumour segmentation system and to reduce the error rate of the tumour segmentation region. The proposed system stated in this paper uses ANFIS classifier which has feedback section in order to improve the performance metric of the tumour segmentation system and to reduce the error rate of the tumour segmentation region.

 

5. CONCLUSIONS:

In this paper, CNN classifier based brain tumour detection and segmentation methodology is proposed using image segmentation technique. The enhancement of brain image is achieved using image segmentation technique and then texture features are extorted. Further, these features are given as input to the CNN classifier. In training of CNN classifier, the features from glioma and non-glioma are extorted and fed to the CNN classifier. This CNN classifier classifies the test brain MRI image into normal (non-glioma) or abnormal (glioma) images using the trained patterns of the CNN in training phase. Then, post processing steps are applied on the classified glioma brain image in order to segment the tumour regions. The performance metrics of the proposed brain tumour detection and segmentation methodology is analyzed in terms of DSC, sensitivity, specificity, and accuracy. The methodology proposed in this paper achieves 97.3% of sensitivity, 98.1% of specificity, 98.7% of accuracy, and 0.96 of DSC. In future, the segmented tumour region will be diagnosed using Fuzzy logic and the extracted features can be optimized using Genetic Algorithm.

 

6. REFERENCS:

1.      Bahadure NB, Ray AK, Thethi HP. Image analysis for MRI based brain tumor detection and feature extraction using biologi cally inspired BWT and SVM. Int J Biomed Imaging. 2017; 2017: 1–12.

2.      Sreedhanya S, Pawar CS. An automatic brain tumor detection and Segmentation using hybrid method. Into J Apple Inform Syst. 2017; 11:6–11.

3.      Alfonse M, Salem ABM. An automatic classification of brain tumours through MRI using support vector machine. Egyptian Comput Sic J. 2016; 40:11–21.

4.      Perez U, Arana E, Mortal D. Brain metastases detection algorithms in magnetic resonance imaging. IEEE Latin Am Trans. 2016; 14:1109–1114.

5.      Su H, Xing F, Yang L. Robust cell detection of histopathological brain tumour images using sparse reconstruction and adaptive dictionary selection. IEEE Trans Med Imaging. 2016; 35: 1575–1586.

6.      Rajesh CG, Ramchand K, Rao H. Tumour detection in brain using genetic algorithm. Procedia Comput Sci. 2016; 79:449–457.

7.      Isselmou ZS, Xu G. A novel approach for brain tumour detection using MRI images. J Biomed Sic Eng. 2016; 9:44–52.

8.      BRATS database, 2015. Available at: http://braintumorsegmenta-tion.org/ (accessed 15 October 2016).

9.      Cai Z, Zhang M, Wang J, Zhao J. A reassigned bilinear transformation and Gaussian kernel function-based approach for detecting weak targets in sea clutter. Int J Remote Sensing 2016; 37:5668–5686.

 

 

 

Received on 14.03.2019           Modified on 21.05.2019

Accepted on 28.06.2019         © RJPT All right reserved