AUTOMATED CLASSIFICATION OF MRI

BRAIN IMAGES
CN Savithri R Carol Praveen
Department of Electronics and Communication Department of Electronics and Communication
Sri Sairam Engineering College SSM Institute of Engineering and Technology
Chennai, India Dindigul, India
[email protected] [email protected]

N Lavanya Devi Varunapriyan K

Department of Electronics and Communication Department of Electronics and Communication
PSNA College of Engineering and Technology Sri Sairam Engineering College
Dindigul, India Chennai, India
[email protected] [email protected]

Abstract: The death rate due to brain tumor is increasing called ischemic stroke in which the blockage in the
rapidly. The accumulation of cells in an uncontrolled blood vessels will cause hindrance to the supply of
manner leads to brain tumor. The life taking tumor can blood to the blood vessel. The second one is called
be cured if it is detected in the early stage. This paper hemorrhagic stroke in which there will be leakage of
deals with analyzing the performance of the different
kernels functions on SVM classification algorithm. The
blood from the blood vessel. The hemorrhagic stroke is
performance of the proposed algorithm is being measured more dangerous than the ischemic stroke as it causes
by parameters like accuracy, sensitivity, specificity and sudden death. For detecting the brain tumor, Magnetic
precision. The kernels like Laplace RBF kernel, Resonance Imaging (MRI) and Computer Tomography
Gaussian kernel, polynomial kernel, Hyperbolic tangent (CT) can be used. But, due to the clarity and the low
kernel, and Sigmoid kernel are applied on SVM. Among radiation, MRI imaging is more preferred. This paper
the various kernel functions, Gaussian kernel SVM gave deals with the automatic detection of tumor in the MRI
good results. Otsu segmentation is applied on the brain image by segmenting the abnormal region.
abnormal images to high light the tumor region in the
brain MRI images.
A. Related Work
The research related to the brain MRI images is
Key words- MRI brain image, SVM classifier, Gray discussed in this section. Research paper [1] proposed
level co-occurrence matrix, principle component analysis, a method for the automatic segmentation of brain
OTSU Segmentation vessel image. MIP calculation is applied to reduce the
amount of mixing the elements. Then, at that point, the
I. INTRODUCTION Gaussian Mixture Model is used to fix the stochastic
Tumor in the brain is the development of dissemination of the brain vessels and brain tissue. At
abnormality in cells of the brain. The spreading of the last, the SEM calculation is embraced to compute the
cancerous cells will depend on the type and size of the Gaussian Mixture Model's hyper parameters. The
cancerous cells. The growth rate of the benign tumors achievability and legitimacy of the model is also
is slow compared to the malignant tumors. The primary checked. With the model, minute parts of the blood
type of brain tumor, the abnormal cells is developed in vessel can be segmented.
the brain itself. In secondary types of brain tumors, the
cancerous cells would have developed somewhere else Paper [2] suggested an algorithm for the
and would have extended to the brain. Among adults, automatic segmentation of MRI brain tissue. The
the most common type of cancer is secondary. authors have used graph cut method for the
segmentation of the MRI brain tissue images. The steps
The brain tumor causes an interruption in the followed in the segmentation of the brain MRI images
flow of blood in the blood vessels and will lead to are brain masking, preprocessing for enhancing the
stroke. There are two types of strokes. The first one is MRI brain image and finally segmentation algorithm
based on graph cut with RBF kernel are used. The  Texture and statistical features are extracted and
specialty of the segmentation algorithm suggested in only the significant features alone are considered
this paper is multi-region graph cut method. The input for further analysis.
brain image is segmented into Gray Matter (GM),  The performance of different kernel functions
White Matter (WM) and Cerebrospinal Fluid (CSF). applied on SVM is studied.
Discrete and continuous formulations are suggested for
segmentation and masking of the brain MRI images. This research paper includes proposed method in
The performance of the suggested methodology is second section and conclusion in the third section.
evaluated by comparing it with the state of art methods
and ensured the better performance of the suggested II PROPOSED METHODOLOGY
algorithm.
The method proposed for the automated
Paper [3] put forward a novel method for the classification of the brain tumor is shown in fig 1. As
segmentation of MRI brain images which is based on shown in the figure the processes involved in the
the modified Fuzzy means algorithm. The researchers classification are preprocessing, feature extraction,
have implemented Gaussian filtering to remove the classification and segmentation. Preprocessing step
noise. The input brain images are chosen in such a way involves converting the RGB input image into gray
that they have 5% and 9% noise interference. The gray scale image and image resizing. The proposed
histogram is used to initialize the cluster center. The algorithm is implemented on brain MRI images with
terminating condition for the algorithm is fixed by 98 normal images and 155 abnormal images and these
determining the difference between the values obtained images are available in kaggle.
in the previous iteration and the current iteration. The
membership functions are also initialized. It the
terminating conditions are achieved, the membership
functions are updated. The maximum membership
functions are used to remove the blur in the image. To
Normal brain Abnormal
evaluate the proposed algorithm, area overlap measure
images brain images
and misclassified error are computed for both
conventional FCM and Improved FCM.
DATABASE

In [4], a method for the reducing the noise

while segmenting the brain tissue in MRI images using
modified Fuzzy C means is proposed. The authors
Feature
suggest that the conventional FCM method does not
take into consideration the spatial information and this Extraction
probably makes the image a noisy one. As an initial
step to this algorithm the researchers have fixed the
plane of symmetry of the brain image. This plane is
called as the mid-sagittal plane. The authors have also Classifier
considered the position of the pixel from the
symmetrical axis also where as in the conventional
fuzzy C means algorithm, this is not considered. The
number of classes, membership functions, cluster
centers and the terminating conditions are also Normal Abnormal
initialized. Finally, the last membership function is
used to perform the segmentation. The performance of
the proposed algorithm is validated by both
qualitatively and quantitatively. Segmentation

B. Novelty and Contribution

Fig. 1. Proposed methodology
 The aim of the proposed algorithm is to classify
the input brain images into abnormal and normal
images. A. Feature Extraction
 The classification of the input image is based on Feature extraction is the process of extracting
the features extracted. the features that will be useful for the classification of
the input image. Feature extraction methods: GLCM
(for texture feature) and statistical features are used in 2) Statistical Features
the proposed method.
Statistical features like skewness, mean, standard
deviation, and kurtosis are also computed to make an
1) GLCM: analysis of distribution of gray level values and their
equations are mentioned from equations (5) to equation
Gray Level Co-occurrence Matrix (GLCM) (8).
computes the frequency of occurrence of a
pair of pixel within a particular distance and x
angle [5, 6]. GLCM is capable of calculating Mean   N (5)
the co-occurrence at angles 00, 450, 900 and
1350. Fig. 2 shows the GLCM calculated for
( xi   ) 2
5 gray level values (0, 1, 2, 3 and 4) for the
distance [0 1] and angle 450 (for both +450
S tan dard Deviation    N
and -450). Texture features like energy,
(6)
entropy, correlation and contrast are computed
from the GLCM and their equations are
3( Mean  Median)
mentioned from equation (1) to (4). Skewness  (7)
S tan dard Deviation

Kurtosis 
 (x i  )4
(8)
n 4
B. Principal
Component Analysis (PCA)

Principal Component Analysis is applied on the

features extracted to compute the most significant
features. After this point, only the significant features
are considered for further computation. The steps
followed in finding the principal components [7, 8, 9]
are shown in the Fig 3. As the initial step for the
Fig. 2. GLCM calculated for 5 gray level values with distance [0 1]
and angle 450 calculation of the significant component, the mean
values of the input features are computed. The
covariance value is found from the deviation from the
mean value. Finally, the Eigen values and eigen vectors
Contrast  i , j | i  j | 2 p (i, j )
are also computed. The top 5 significant values are
(1) taken into consideration for the further process. Among
the features fed as the input to the PCA, all the texture
(i  i )( j  j ) p(i, j ) features and kurtosis turned out to be the important
Correlation   features for the classification of the brain MRI images.
i, j  i j
(2)

Energy   p (i, j )
i, j (3)

Entropy    P(i, j ) log( P(i, j ))

i, j (4)
 i) Polynomial kernel: SVM with this type of kernel
uses the similar features of the classes and it forms the
Input feature generalized form of linear SVM.
vector
(9)
k ( xi , x j )  ( xi .x j  1) d

Mean calculation ii) Gaussian Kernel: This type of SVM is applied on

the input data when there is no preliminary knowledge
about the input.
2
Compute deviation x y
k ( x, y )  exp ( ) (10)
2 2
iii) Hyperbolic tangent kernel: This type of kernel
Compute covariance
applied on the SVM is used in neural networks.

k ( xi , x j )  tanh(xi x j  c) (11)

Compute Eigen values and Eigen iv) Laplace RBF kernel: SVM with laplace RBF kernel
vectors is used as a general purpose kernel.

x y
Fig. 3. Steps for the computation of PCA k ( x, y )  exp ( ) (12)

C. CLASSIFICATION
v) Sigmoid kernel: The mathematical function used in
For the classification of the brain image, kernel SVM is this model is similar to the neural network with the two
employed. The conventional SVM algorithm classifies layer perceptron model. This type of mathematical
the input by optimizing the hyperplane. Normally, the model is usually applied on the artificial neural
classification algorithms feed the most outstanding network.
character of the class. SVM feeds the common features
of the two classes. k ( x, y )  tanh(xT y  c) (13)

The performance of the above kernel functions on the

The important steps [10] carried out in the SVM are SVM classification algorithm are analyzed and the best
performing kernel function will be suggested based on
i) Select the hyperplanes the confusion matrix computed for each one of them.
Fig. 4 shows the confusion matrix.
ii) The margin/distance of the class and the hyperplane
is to be maximum
Predicted
iii) Determine the average line so that it becomes the
decision boundary. Negative Positive
SVM shows superior results only if the inputs are
linear. In order to make SVM suit well for the Actual Negative True Negative False
nonlinear application also, some kernels are applied on (TN) Positive
SVM. Kernels are some mathematical functions and Positive False Negative (FP)
True
these functions transform the input into the required (FN) Positive
form. The following kernels functions are being (TP)
implemented for the classification of the MRI brain
Fig. 4. Confusion Matrix
images.
TN- indicates that the predicted and the actual are Table 1 depicts the comparison of kernel
normal functions like polynomial, Gaussian, hyperbolic
tangent, laplace RBF, sigmoid applied on SVM and the
FN- indicates that the predicted is normal but actually performance of each kernel is being analyzed by
it is abnormal computing accuracy, sensitivity, precision and
specificity in the classification of brain MRI images
FP- indicates that the predicted is abnormal but into normal and abnormal images.
actually it is normal
D. SEGMENTATION
TP- indicates that the predicted and the actual are
abnormal Segmentation in image processing is done to
Based on the confusion matrix framed for each kernel identify the region of interest [10]. In the proposed
method, tumor region is considered as the region of
SVM accuracy, sensitivity, specificity and precision
interest. The identified tumor region is classified into
are determined to find the effectiveness of the proposed benign and malignant by extracting the features. In the
algorithm. Equation 1 to 4 shows the formula for proposed algorithm Otsu method are used as
accuracy, precision, sensitivity and specificity. segmentation algorithms.
TP  TN A. OTSU segmentation
Accuracy  (14)
TP  TN  FP  FN
Otsu algorithm computes a threshold which
TP will reduce the intra class variance and is interpreted as
Pr ecision  (15) sum of weighted variance of the two classes
TP  FP (foreground and background). It is given in (18).
TP  2 w (t )  0 (t )  02 (t )  1 (t )  12 (t )
Sensitivity  (16) (18)
TP  FN
 0 and  1 are the probabilities of the two classes; σ02
TN and σ12 are the variances of the two classes
Specificity  (17)
TN  FP
The following steps are followed for the
implementation of otsu algorithm
1. Compute the histogram
TABLE 1: COMPARISON OF PERFORMANCE OF VARIOUS
KERNEL FUNCTIONS ON SVM
2. Initialize ѡi(0) and μi(0)
3. Compute ѡi, μi and σb2(t) for all the possible values
Kernel Accu Precis Sensitiv Specific of the threshold.
Function racy ion ity ity
applied on
SVM
Polynomial 92.38 93.28 90.09 91.23
kernel
SVM
Gaussian 94.65 95.01 92.09 91.38
Kernel
SVM (a) (b)
Hyperbolic 93.27 93.9 90.12 92.64
tangent
kernel
SVM
Laplace 91.38 92.07 88.23 87.65
RBF kernel
SVM
Sigmoid 91.5 92.89 89.88 88.56 (c) (d)
Fig. 5. (a) and (b) input MRI brain images; (c) and (d) segmented
kernel tumor region
SVM
 [4] Jayasuriya, S. A., & Liew, A. W. C. (2013, December). A
modified Fuzzy C-means algorithm with symmetry information for
III. CONCLUSION MR brain image segmentation. In 2013 9th International Conference
on Information, Communications & Signal Processing (pp. 1-5).
An algorithm for the classification of brain IEEE.
MRI images into abnormal and normal is developed by
analyzing the performance of various kernels functions [5] S.Arivazhagan, W.Sylvia Lilly Jebarani, M.Shanmugaraj, “A
novel approach to Low Volume Generic Steganalysis”, 21st
on SVM. Among the various kernels applied, better International Conference on Systems Engineering, pp.153-158, 2011.
results are obtained for Gaussian Kernel SVM. After
the successful classification of the brain MRI images, [6] Arivazhagan S, Ganesan L, Savithri CN. Effective multi-
resolution transform identification for characterization and
segmentation is applied on abnormal images and the classification of texture groups. ictact journal on image and video
tumor region is highlighted. The main limitation of the processing. 2011 Nov;2(02).
proposed algorithm is the manual feature extraction. In
[7] Ahmad, I. (2015). Feature selection using particle swarm
the future, the features can be automatically extracted
optimization in intrusion detection. International Journal of
and i.e., the accuracy of the proposed system may be Distributed Sensor Networks, 11(10), 806954.
improved.
[8] Mohsen, F., Hadhoud, M. M., Moustafa, K., & Ameen, K.
REFERENCES (2012). A new image segmentation method based on particle swarm
optimization. Int. Arab J. Inf. Technol., 9(5), 487-493.
[1] Xu, F., Wang, X. C., Zhou, M. Q., Wu, Z., & Liu, X. Y. (2010,
August). Segmentation algorithm of brain vessel image based on [9] Ait-Aoudia, S., Guerrout, E. H., & Mahiou, R. (2014, July).
SEM statistical mixture model. In 2010 Seventh International Medical image segmentation using particle swarm optimization.
Conference on Fuzzy Systems and Knowledge Discovery (Vol. 4, In 2014 18th International Conference on Information
pp. 1830-1833). IEEE. Visualisation (pp. 287-291). IEEE.

[2] Messaoud, Z. B., Chnitti, S., & Njeh, I. (2016, March). An [10] Sajeena T A, Jereesh A S, “Automated Cervical Cancer
automated MRI brain tissue segmentation approach. In 2016 2nd Detection through RGVF segmentation and SVM Classification”,
International Conference on Advanced Technologies for Signal and IEEE Intl. Conference on Computing and Network Communications
Image Processing (ATSIP) (pp. 370-374). IEEE. (CoCoNet'15), pp. 663- 669, 2015.

[3] Wan, C., Ye, M., Yao, C., & Wu, C. (2017, October). Brain MR [11] Jemimma, T. A., & Vetharaj, Y. J. (2018, December).
image segmentation based on Gaussian filtering and improved FCM Watershed algorithm based DAPP features for brain tumor
clustering algorithm. In 2017 10th International Congress on Image segmentation and classification. In 2018 International Conference on
and Signal Processing, BioMedical Engineering and Informatics Smart Systems and Inventive Technology (ICSSIT) (pp. 155-158).
(CISP-BMEI) (pp. 1-5). IEEE. IEEE.