Brain Tumor Classification via Statistical Features and Back-Propagation Neural Network

do not necessarily reflect the views of UKDiss.com.

Brain Tumor Classification via Statistical Features and Back-Propagation Neural Network

Abstract— Classification of brain tumor is the heart of the computer-aided diagnosis (CAD) system designed to aid the radiologist in the diagnosis of such tumors using Magnetic Resonance Image (MRI). In this paper, we present a framework for classification of brain tumors in MRI images that combines statistical features and neural network algorithms. This algorithm uses region of interest (ROI), i.e. the tumor segment that is identified either manually by the technician/radiologist or by using any of the ROI segmentation techniques. We focus on feature selection by using a combination of the 2D Discrete Wavelet Transform (DWT) and 2D Gabor filter techniques. We create the features set using a complete set of the transform domain statistical features. For classification, back propagation neural network classifier has been selected to test the features selection impact. To do so, we used a large dataset consisting of 3,064 slices of T1-weighted MRI images with three types of brain tumors, Meningioma, Glioma, and Pituitary tumor. We obtained a total accuracy of 91.9 %, and specificity of 96 %, 96.29 %, and 95.66 % for Meningioma, Glioma, and Pituitary tumor respectively. Experimental results validate the effectiveness of the features selection method and indicate that it can compose an effective feature set to be used as a framework that can be combined with other classifications technique to enhance the performance.

Keywords— Brain tumor classification; Gabor filter; Wavelet Transform; statistical features

I.     Introduction

Brain tumors are the most common brain disease that affect the central nervous system (CNS), the brain and spinal cord [1]. During the last two decades, the computer-aided diagnosis system (CAD) has been employed to improve the accuracy of the diagnostic ability of radiologists in detecting, segmenting, and identifying the type of brain tumor [2].

Magnetic Resonance Imaging (MRI) is the most common modality that is exploited to diagnose abnormal tissue growth in the brain. Usually, radiologists depend on manual diagnosis to identify brain tumors [3]. In manual diagnosis, the amount of MRI images to be analyzed is large enough to make readings based on visual interpretation expensive, inaccurate, and intensive. In addition; the human eye is sensitive to any changes in the MRI image, this sensitivity may be decreased with the increase in the number of images [4]. Furthermore, manual diagnosis is time consuming [3], and it depends on the subjective decisions of the radiologists which is hard to quantify [5], thus it may lead to misclassification [6]. Radiologists used invasive techniques such as biopsy and spinal tap methods for a brain tumor diagnosis. Both approaches, biopsy and spinal tap, are painful, time consuming, risky, and expensive [7]. Consequently, developing an accurate, robust, practical, and automatic diagnosis system for this purpose is significant and essential to overcome the drawbacks of manual diagnosis and invasive techniques.

Many researchers have proposed different methods to develop a CAD system that can detect or classify abnormal tissues in brain MRI images. Feature extraction is a crucial step in building a capable CAD system [2]. The most widely used techniques for feature extraction are texture features and Discrete Wavelet Transform (DWT). Texture is a pattern that appears in an image frequently and repeatedly [8]. Texture features can be specified by statistical features [9] or Gabor filter [10]. There are two types of statistical features, first order and second order [8]. First order features are extracted from the image histogram [11], while second order features are extracted from Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRM) [8], or Neighborhood Gray Tone Difference Matrix (NGTDM) [12]. Gabor filter is a linear and local filter that is based on the convolution between Gaussian function and cosine function [10]. Texture features based on Gabor filter have been employed to extract the required features for the classification of brain tumor [13]. DWT is a beneficial tool for image analysis since it provides significant information with reduced time [14]. 2-D DWT has been implemented to extract features from brain MRI images for tumor classification [15]. Different classification techniques have been exploited by many authors for identifying tumor types from brain images. The most common used classifiers are neural network and support vector machines (SVM). Neural network was implemented using different architectures like feedforward neural network [6], multilayer perceptron neural network trained with backpropagation algorithm [8], and Probabilistic neural network (PNN) [7]. Support vector machine (SVM) was implemented using three kernels, homogenous polynomial, inhomogeneous polynomial, and Gaussian radial basis function [12], [16]. Other types of classifiers were employed like the ensemble learning method based on AdaBoost classifier [17], neuro-fuzzy classifier [5], dictionary learning and sparse coding [18], and k-nearest neighbor (KNN) [19].

In this paper, we propose a classification algorithm for brain MRI images that exploits statistical features extracted from 2-D DWT and 2-D Gabor filter. Combining both techniques (2-D DWT and 2-D Gabor filter) can improve the classification accuracy as compared to using each method separately since it benefits from all the directional information of the input MRI image. This paper is organized as follows; section I is an introduction, section II presents the methodology used in this algorithm, section III shows the results and section IV is the conclusion.

II.    Material and Methodology

Feature extraction is implemented using 2-D DWT, 2-D Gabor filter, and statistical features as illustrated in figure 1. The input image is the tumor region that can be obtained using manual segmentation by radiologists or automatic segmentation. Back-propagation neural network (BPNN) is used for the classification.

Input Image

2-D

Gabor Filter

2-D DWT

Statistical Features

Statistical Features

Features Set

BPNN Classifier

Tumor Type

Figure 1: Framework of the proposed system

A.    Database

The dataset used to train and test our algorithm consists of 3064 brain MRI slices collected from 233 patients with three kinds of brain tumors, Meningioma, Glioma, and Pituitary.

      A picture containing indoor, white, photo

Description generated with high confidence

(a)                            (b)                            (c)

      

(d)                            (e)                            (f)

Fig. 2: Sample images from database, (a) Meningioma, (b) Glioma, and (c) Pituitary and its tumor region (d), (e), and (f) respectively

The total number of slices for each type of tumor is: 708 slices for Meningioma tumor, 1426 slices for Glioma tumor, and 930 slices for Pituitary tumor [20]. The tumor region was manually segmented by three radiologists. The original slices along with its tumor region are available online from Figshare website [21]. In figure 2, three different MRI images are presented along with its tumor region.

B.    Discrete Wavelet Transform

Wavelet Transform (WT) is a powerful tool that transforms the signal from the time domain into the wavelet domain to analyze the time and frequency contents at the same time [22]. For high frequency signals, Wavelet Transform gives high time resolution and low frequency resolution and for low frequency signals, it gives high frequency resolution and low time resolution. A basis function called “mother wavelet” is scaled and translated to achieve the time and frequency resolution. Wavelet basis function is generated from the mother wavelet as follows [23]:

ψa,bx=1aψx-ba, a ,b ∈Z  a > 0                   ( SEQ Equation * ARABIC 1)

where

ψx, is the mother wavelet, and (a, b) represent the dilation and translation parameters respectively.

By applying 2-D DWT, the image is decomposed into four subbands labeled LL which represent approximation image, LH, HL and HH that correspond to detail images as shown in figure 3. The approximation and detail coefficients are used for texture features representation. In this algorithm, we implement three levels of 2-D DWT using “symlet4” filter and all the subbands images (LL, LH, HL, and HH) are utilized for feature extraction results in 12 subband images. Combining the approximation and detail coefficient can improve the discrimination ability of the classification algorithm [24].

Fig. 3: One level filter bank for computation of 2-D DWT,

honis a low pass filter and

h1nis a high pass filter

C.    Gabor Filter

A Gabor filter is a linear filter that is obtained by modulating a sinusoidal wave with a Gaussian function. The frequency and orientation of Gabor filter are similar to the human visual system and it can be used for texture description. The Gabor function is a useful tool in computer vision and image processing, especially for texture analysis, due to its optimal localization properties in both the spatial and frequency domain [25]. 2-D Gabor filter is defined as [10]:

Gx,y,λ,θ, ϕ,σ,γ=e-x’2+γ2y’2σ2*ei 2πx’λ+ψ           ( SEQ Equation * ARABIC 2)

Where

x’=x cosθ+ysin⁡θ and y’=- x cosθ+ysin⁡θ.

(

λ) is the wavelength of the sinusoidal form, (

θ) is the orientation of the Gabor function, (

ϕ) is the phase offset, (

γ) is the spatial aspect ratio, and (

σ) is the standard deviation of the Gaussian envelope. To calculate the textural features of an image a set of Gabor filters are used with different frequencies and orientations [13]. We implement Gabor filter with three wavelengths (2, 4, and 8) and five orientations (0

°,45

°, 90

°, 135

°, and 180

°) which results in 15 filters, each filter is convolved with the input image generating 15 filtered images as shown in figure 4.

                          

(2, 0

°)      (2, 45

°)       (2, 90

°)     (2, 135

°)    (2, 180

°)

A picture containing indoor, white, photo

Description generated with high confidence             

(4, 0

°)        (4, 45

°)     (4, 90

°)     (4, 135

°)    (4, 180

°)

    

(8, 0

°)      (8, 45

°)       (8, 90

°)     (8, 135

°)    (8, 180

°)

Tumor segment

Fig. 4: Resulted images from 2-D Gabor filter with three values of wavelengths and five values of orientations, numbers in brackets are (wavelength, orientation)

D.   Statistical Features

Texture of an image can be described easily using statistical approach [26]. These features are widely used in the classification of biomedical images [24]. We consider four first order statistics (mean, variance, skewness, and kurtosis) that is calculated from the histogram of the image and six second order statistics (Contrast, Correlation, Energy, Homogeneity, Entropy, and Maximum probability) which is calculated from the Gray Level Co-occurrence Matrix (GLCM). The GLCM is a 2-D histogram that describes the frequency of occurrence of two pixels separated by a certain distance. Table 1 summarizes these features.

As a result, 10 statistical features were obtained from each subband image resulted from three levels of 2-D DWT so, the total number of features extracted from the wavelet transform is 120 features. In a similar way, the same 10 statistical features were extracted from the images generated using Gabor filters results in 150 features, these features are combined to generate the feature vector with the size of 270 features.

Table 1: A LIST OF SELECTED STATISTICAL FEATURES

Feature Description Formula
Mean (m) Average intensity of the image m=∑i=0L-1i*pi
Variance ( 

μ)

Second moment about the mean μ=∑i=0L-1i-m2*pi
Skewness ( 

SK)

Third moment about the mean SK=∑i=0L-1i-m3*pi
Kurtosis ( 

KU)

Fourth moment about the mean KU=∑i=0L-1i-m4*pi
Contrast Measures local variations in GLCM Contrast=∑i=1A∑j=1Ai-j2*pi,j
Correlation Measures correlation Correlation= 

∑i=1A∑j=1Ai-m1*j-m2*pi,jσ1σ2

Energy Measures smoothness Energy=∑i=1A∑j=1Api,j2
Homogeneity Measures the spatial closeness of GLCM elements to the diagonal Homogeneity=∑i=1A∑j=1Api,j1+ i-j
Entropy Measures randomness Entropy= 

-∑i=1A∑j=1Api,j*log2⁡pi,j

Maximum probability The maximum probability in GLCM Maximum probability=MAXi,j pi,j

In table 1, L is the image intensity levels, A is the size of GLCM matrix,

piis the image histogram,

pi,jthe element of GLCM, (m1, m2) are the mean along rows and columns respectively, and (

σ1,σ2) are the standard deviation along rows and columns respectively.

E.    Classification

For the classification, we used the neural network classifier that is implemented using the multilayer perceptron neural network and comprised of three layers, input layer of size 270 neurons, one hidden layer of size 90 neurons, and an output layer of size 3 neurons. The Back-propagation algorithm is used to adjust the weights and biases of the network to minimize the error. Figure 5 shows the architecture of the neural network where X represents the input vector of size M and C is the number of classes.

Figure 5: Architecture of neural network classifier

III.   Experimental results and Discussion

The database images are sorted randomly and divided into two sets, the training set (70 % of the whole dataset) which has 493 slices with Meningioma tumors, 1019 slices with Glioma tumors, and 633 slices with Pituitary tumors, and the testing set (30 % of the whole dataset) that results in 215 slices with Meningioma tumors, 407 slices with Glioma tumors, and 297 slices Pituitary tumors.

There are four primary outcomes that are used to test the classifier; true positives, true negatives, false positives, and false negatives. In the case of brain tumor classification with multiclass, these quantities should be calculated separately for each type of tumor with positive represent tumor output and negative case non-tumor output. The performance of the proposed algorithm is measured using three criteria based on the four outcomes [27]:

Accuracy =NTP+NTNNFP+NFN+NTP+NTN                                  (3)

Sensitivity =NTPNFN+NTP                                                          (4)

Specificity =NTNNFP+NTN                                                          (5)

Where,

NTPis the number of true positives,

NTNis the number of true negatives,

NFPis the number of false positives, and

NFNis the number of false negatives. Accuracy, sensitivity, and specificity for each type of tumor is described in table 2.

Table 2: Accuracy, sensitivity, and specificity for the three types of tumors

Classifier Performance Meningioma Glioma Pituitary
Accuracy 93.9 95.75 94.23
Sensitivity 86.97 95.32 90.94
Specificity 96.02 96.1 95.81

A confusion matrix is a table that describes the performance of a classification model based on a set of test data for which the true values are known [28]. Table 3 shows the confusion matrix for classification of three types of tumors using neural network classifier.

Table 3: Confusion matrix

Predicted 

Classes

Actual

Classes

Meningioma Glioma Pituitary
Meningioma 187 7 21
Glioma 14 387 6
Pituitary 14 12 271

Receiver operating characteristics (ROC) is defined as a trade-off visualization between a pair of criteria like sensitivity and specificity in the plane spanned by the two measures [29]. The plot of ROC curve is shown in figure 6. The area under ROC curve for each class is 0.92, 0.95, and 0.93 for Meningioma tumors, Glioma tumors, and Pituitary tumors respectively.

Figure 6: ROC Curve for the classification

To investigate the effect of feature extraction techniques, we train the classifier model on three types of features, wavelet features only, Gabor features only, and a combination of wavelet and Gabor features, then we compare the results in terms of sensitivity and specificity for each type of tumor as shown in table 4.

The 2-D DWT is a powerful tool for feature extraction represented by wavelet coefficients [7]. On the other hand, Gabor filter is suitable for texture representation [13]. Combining both techniques for the MRI tumor image representation improves the classification performance as described in table 4.

Table 4: Comparison between wavelet features, Gabor features, and combined features.

Classifier performance Type of tumor Wavelet Features Only Gabor Features Only Wavelet and Gabor Features
Sensitivity Meningioma 82.32 83.72 86.97
Glioma 94.1 91.64 95.1
Pituitary 82.49 85.52 91.24
Specificity Meningioma 92.47 94.03 96
Glioma 95.5 91.79 96.29
Pituitary 93.89 95.49 95.66

The only work that used the same dataset is the method suggested by Cheng et al. [20], they implemented an algorithm based on the three feature extraction methods, intensity histogram, gray level co-occurrence matrix (GLCM), and bag of words model and SVM model for classification. Figure 6 shows the performance evaluation of the proposed algorithm and the related work algorithm in terms of total accuracy, specificity for Meningioma, Glioma, and Pituitary. It can be seen from figure 7 that the overall accuracy of the proposed algorithm is 91.9 %, it is higher than that of the related work that achieved an accuracy of 91.28 %.

Figure 7: Comparison with related work

IV.   Conclusion

Brain tumor is a disease that affects the central nervous system (CNS), which is typically diagnosed by radiologists using either invasive or noninvasive techniques. Magnetic Resonance Imaging (MRI) is the most widely used modality for a noninvasive diagnosis of brain tumors. In this paper, we presented a new algorithm to classify three types of brain tumors, namely, the Meningioma, Glioma, and Pituitary tumors from MRI images. We used the 2D Discrete Wavelet transform (DWT), 2D Gabor filter, and first and second order statistics of transform domain data to generate a features pool that encompasses the tumors attributes. For classification, we used multilayer perceptron neural network trained using back-propagation. Simulation results show the effectiveness of extracting statistical features solely from wavelet and Gabor transforms with an accuracy of 91.9 %. For future work, we would like to use other classification methods including Support Vector Machine (SVM) or K-Nearest Neighbor (KNN) to validate the proposed framework effectiveness.

References

[1] M. S. H. Al-Tamimi and G. Sulong, “TUMOR BRAIN DETECTION THROUGH MR IMAGES: A REVIEW OF LITERATURE.,” J. Theor. Appl. Inf. Technol., vol. 62, no. 2, 2014.

[2] E.-S. A. El-Dahshan, H. M. Mohsen, K. Revett, and A.-B. M. Salem, “Computer-aided diagnosis of human brain tumor through MRI: A survey and a new algorithm,” Expert Syst. Appl., vol. 41, no. 11, pp. 5526–5545, 2014.

[3] M. K. Abd-Ellah, A. I. Awad, A. A. M. Khalaf, and H. F. A. Hamed, “Design and implementation of a computer-aided diagnosis system for brain tumor classification,” in Microelectronics (ICM), 2016 28th International Conference on, 2016, pp. 73–76.

[4] H. Selvaraj, S. T. Selvi, D. Selvathi, and L. Gewali, “Brain MRI slices classification using least squares support vector machine,” Int. J. Intell. Comput. Med. Sci. Image Process., vol. 1, no. 1, pp. 21–33, 2007.

[5] D. M. Joshi, N. K. Rana, and V. M. Misra, “Classification of brain cancer using artificial neural network,” in Electronic Computer Technology (ICECT), 2010 International Conference on, 2010, pp. 112–116.

[6] N. Zulpe and V. Pawar, “GLCM textural features for brain tumor classification,” IJCSI Int. J. Comput. Sci. Issues, vol. 9, no. 3, pp. 354–359, 2012.

[7] P. John and others, “Brain tumor classification using wavelet and texture based neural network,” Int. J. Sci. Eng. Res., vol. 3, no. 10, pp. 1–7, 2012.

[8] B. Sudha, P. Gopikannan, A. Shenbagaraj, and C. Balasubramanian, “Classification of Brain Tumor Grades using Neural Network,” in Proceedings of the World Congress on Engineering 2014, 2014, vol. 1.

[9] K. M. Priya, S. Kavitha, and B. Bharathi, “Brain tumor types and grades classification based on statistical feature set using support vector machine,” in Intelligent Systems and Control (ISCO), 2016 10th International Conference on, 2016, pp. 1–8.

[10] S. E. Grigorescu, N. Petkov, and P. Kruizinga, “Comparison of texture features based on Gabor filters,” IEEE Trans. Image Process., vol. 11, no. 10, pp. 1160–1167, 2002.

[11] G. L. Qurat-Ul-Ain, S. B. Kazmi, M. A. Jaffar, and A. M. Mirza, “Classification and segmentation of brain tumor using texture analysis,” Recent Adv. Artif. Intell. Knowl. Eng. data bases, pp. 147–155, 2010.

[12] U. Javed, M. M. Riaz, A. Ghafoor, and T. A. Cheema, “MRI brain classification using texture features, fuzzy weighting and support vector machine,” Prog. Electromagn. Res. B, vol. 53, pp. 73–88, 2013.

[13] A. Vidyarthi and N. Mittal, “Performance analysis of Gabor-Wavelet based features in classification of high grade malignant brain tumors,” in Systems Conference (NSC), 2015 39th National, 2015, pp. 1–6.

[14] N. Aggarwal and R. K. Agrawal, “First and second order statistics features for classification of magnetic resonance brain images,” J. Signal Inf. Process., vol. 3, no. 2, p. 146, 2012.

[15] V. Anitha and S. Murugavalli, “Brain tumour classification using two-tier classifier with adaptive segmentation technique,” IET Comput. Vis., vol. 10, no. 1, pp. 9–17, 2016.

[16] Y. Zhang and L. Wu, “An MR brain images classifier via principal component analysis and kernel support vector machine,” Prog. Electromagn. Res., vol. 130, pp. 369–388, 2012.

[17] S. Ghanavati, J. Li, T. Liu, P. S. Babyn, W. Doda, and G. Lampropoulos, “Automatic brain tumor detection in magnetic resonance images,” in Biomedical Imaging (ISBI), 2012 9th IEEE International Symposium on, 2012, pp. 574–577.

[18] S. D. S. Al-Shaikhli, M. Y. Yang, and B. Rosenhahn, “Brain tumor classification using sparse coding and dictionary learning,” in Image Processing (ICIP), 2014 IEEE International Conference on, 2014, pp. 2774–2778.

[19] E. I. Zacharaki et al., “Classification of brain tumor type and grade using MRI texture and shape in a machine learning scheme,” Magn. Reson. Med., vol. 62, no. 6, pp. 1609–1618, 2009.

[20] J. Cheng et al., “Enhanced performance of brain tumor classification via tumor region augmentation and partition,” PLoS One, vol. 10, no. 10, p. e0140381, 2015.

[21] J. Cheng, “brain tumor dataset,” 2017.

[22] Y. Liu, M. Muftah, T. Das, L. Bai, K. Robson, and D. Auer, “Classification of MR tumor images based on Gabor wavelet analysis,” J. Med. Biol. Eng., vol. 32, no. 1, pp. 22–28, 2012.

[23] P. S. Addison, The illustrated wavelet transform handbook: introductory theory and applications in science, engineering, medicine and finance. CRC press, 2017.

[24] S. Lahmiri and M. Boukadoum, “Classification of brain MRI using the LH and HL wavelet transform sub-bands,” Proc. – IEEE Int. Symp. Circuits Syst., pp. 1025–1028, 2011.

[25] J. Yang, L. Liu, T. Jiang, and Y. Fan, “A modified Gabor filter design method for fingerprint image enhancement,” Pattern Recognit. Lett., vol. 24, no. 12, pp. 1805–1817, 2003.

[26] R. C. Gonzalez and R. E. Woods, “Digital image processing,” Up. Saddle River, NJ, 2014.

[27] M. Kubat, An Introduction to Machine Learning. Springer, 2015.

[28] K. Markham, “Simple guide to confusion matrix terminology,” Data Sch. [online], 2014.

[29] T. Sing, O. Sander, N. Beerenwinkel, and T. Lengauer, “ROCR: visualizing classifier performance in R,” Bioinformatics, vol. 21, no. 20, pp. 3940–3941, 2005.

Professor

You must be logged in to post a comment