See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/320649812

Brain tumor detection using SVM classifier

Conference Paper · May 2017

DOI: 10.1109/SSPS.2017.8071613

CITATIONS READS

57 5,420

5 authors, including:

Sathies kumar Thangarajan

Sri Sairam Engineering college
14 PUBLICATIONS 98 CITATIONS

SEE PROFILE

All content following this page was uploaded by Sathies kumar Thangarajan on 17 March 2021.

The user has requested enhancement of the downloaded file.

 2017 IEEE 3rd International Conference on Sensing, Signal Processing and Security (ICSSS)

Brain Tumor Detection Using SVM Classifier

Mr. T. Sathies Kumar K. Rashmi Sreevidhya Ramadoss
AP/ EIE EIE EIE
Sri Sai Ram Engineering College Sri Sai Ram Engineering College Sri Sai Ram Engineering College
Chennai, India Chennai, India Chennai, India
[email protected] [email protected] [email protected]

L.K. Sandhya T.J. Sangeetha

EIE EIE
Sri Sai Ram Engineering College Sri Sai Ram Engineering
Chennai, India College Chennai, India
[email protected] [email protected]

Abstract— Magnetic Resonance Imaging is a standard non-invasive Primary tumor: The origin of the tumor cell is brain
methodology used in medical field for the analysis, diagnosis and itself.
treatment of brain tissues. The early diagnosis of brain tumor helps
in saving the patients’ life by providing proper treatment. The
accurate detection of tumors in the MRI slices becomes a fastidious Secondary tumor: It is known as metastatic brain
task to perform and therefore, by this proposed system, the tumor. The origins of the tumor cell are other parts of
classification and segmentation the tumor region can be done the body where it finally reaches the brain. This is the
accurately. Segmentation and 3D reconstruction also uses the most common type of brain tumor.
detection of tumor from an MR image. The manual tracing and
visual exploration by doctors will be restrained in order to avoid time Based on the malignancy, it is classified as given below
consumption. The brain tumor detection allows localizing a mass of
abnormal cells in a slice of Magnetic Resonance (MR) using SVM
Benign tumor: The tumor cells are noncancerous that
Classifier and segmentation of the tumor cells to know about the size
of the tumor present in that segmented area. The extracted features are normal in size and growth. They don’t generally
of the segmented portion will be trained using artificial neural invade the other parts of the body.
network to display the type of the tumor. These features will also be Malignant tumor: The tumor cells are cancerous cells
used for comparing the accuracy of different classifiers in that are abnormal in size and growth. They tend to
Classification learner app. The scope of this project is helpful in post invade the neighboring healthy cells.
processing of the extracted region like the tumor segmentation.
Based on the appearance, the tumor cells are graded.
During the period of treatment, this grade measurement gives
Keywords— Segmentation; Brain tumor; SVM Classifier; the information about the rate of growth of tumor [2].
artificial neural network; Magnetic Resonance Imaging;
Classification learner app. TABLE 1. Grades of brain tumor

I. INTRODUCTION Grade Appearance Growth rate

The cell is the fundamental structural unit of our body. 1 Nearly normal Slow
There are about hundred trillion cells and they normally divide 2 Slightly abnormal Slow
in a controlled way. The uncontrolled or abnormal division
leads to the formation of tumor cells [1]. 3 Abnormal Active growth

4 Most abnormal Quick growth

Brain tumor also called intracranial neoplasm, is one such
kind of tumor where the abnormal cell division occurs inside
the brain. A benign tumor can become malignant later and a low
grade tumor can turn into a higher grade tumor. Thus, the
Classification of brain tumor: knowledge about the exact grade or stage of the tumor is
Brain tumor is classified on many bases. Based on the origin necessary for the proper treatment.
of the tumor, it is classified as given below
The diagnosis of brain tumor can be achieved through
various images such as CT (Compute Tomography),
ultrasound, PET (Positron Emission Tomography), MRI

978-1-5090-4929-5©2017 IEEE
318
 2017 IEEE 3rd International Conference on Sensing, Signal Processing and Security (ICSSS)

(Magnetic Resonance Imaging), Spectroscopy, Fusion, etc.

Among all the Imaging techniques, MRI is preferred as it
effectively detects the following given conditions: cysts,
tumor, bleeding, swelling or infections.

One of the main problem in medical analysis is the

segmentation of obtained image where the boundaries of
objects such as abnormal region or organs are identified in the
images. Results obtained from the segmentation process is
useful in obtaining features of segmented tumor region such as
area, eccentricity, bounding box, orientation [3].

The image segmentation mainly relies on the followings:

threshold, edge, pixels, cluster, and neural network. Fig. 1 Classification using SVM

Processing of image based on shapes is called morphology. It builds a hyper plane based on a kernel function (K). As
Structuring element is applied to the preprocessed input image shown in Figue.1, feature vectors on left side of hyper plane
which creates output image of the same size. Every pixel of the belong to class -1 and the feature vectors on the right side of
input image is compared with all the neighboring pixels and the hyper plane belong to class +1.
compared result is the values of the corresponding output pixel.
In neural network, the dataset of MRI images are trained
Dilation and erosion are the basic operations of by iteration method. Therefore, when a new image is given to
morphology. Pixels are added to the boundaries of the objects the neural network, it is trained by neural network by
in dilation operation while they are removed in the erosion comparing the iteration value of the previously trained image.
operation depending on the size and shape of the structuring After the training is complete, if any changes need to be made
elements. During the comparing process, dilation operation to the training process, the network architecture, or the data
returns the highest value of the surrounding nearby pixels to sets can be determined using the network performance. The
the output whereas the erosion operation returns the lowest Fig. 2 shows that the test curves and the validation curves are
value of the surrounding nearby pixels to the output. very similar. If the test curve gets increased significantly
before the validation curve is increased, then it is evident that
There are different types of clustering that includes an over fitting might have occurred. The relationship between
mountain, K- means, fuzzy C- means and subtractive the targets and the resulting output of the network can be
clustering method. The k-means clustering is the common compared by plotting the regression plot [6]. If the training
clustering technique that is used. It is a simple, straightforward was perfect, the network output and the target would be equal
and quicker than the hierarchical clustering that can work for as shown in Fig. 3.
large number of variables [4].

The process of separating an image into various parts or

regions is called image segmentation. The major goal is to
simplify or change the representation of the image into
something else which is easier and more meaningful to
analyze. It is used to locate objects and boundaries in images.
Every pixel in an image is assigned a label such that the pixels
having the same label share certain characteristics in common.

One of the classical problems in image processing is image

classification. The major goal of image classification is to
predict the input image categories by using the features. There
are various classifiers such as ANN (Artificial Neural Fig. 2 Performance of training process
Network), SVM (Support Vector Machine), Random Forest,
Decision Forest k-NN (k Nearest Neighbor), Adaptive boost
(Ada boosted) etc.[5].

One of the best methods for classifying any image or pattern

is SVM. SVM is used to split a set of images into two various
classes. The classification is done by finding the hyper-plane that
differentiate the two classes very well as given in Fig. 1.

978-1-5090-4929-5©2017 IEEE
319
 2017 IEEE 3rd International Conference on Sensing, Signal Processing and Security (ICSSS)

B. Convolution filtering:

The MRI Images acquired from various sources contain

noises due to random movement of patient during scanning
and external noise, etc. The above mentioned noises should be
removed and the image perceptual quality have to be enhanced
for efficient brain tumor detection [8].

Various steps of pre-processing is as follows:

1) RGB to Gray-level conversion: The image is converted to

Fig. 3 Regression plot gray scale image from RGB model.

Classification Learner app is used to train the database for 2) Edge detection: The edges present in the gray scale image are
classifying the obtained data by supervised machine learning. detected using canny method. The Canny edge detector
The tasks involved in this app includes: smoothens the gray scale image by convoluting it with a Gaussian
filter. Both the x and y direction derivatives are computed based
• uploading the data on the resulting smoothed image. The gradient magnitude of the
• selecting the features image is computed with these derivatives. If they do not
• training the database constitute a local maximum, then the pixels of the image are
• obtaining the results suppressed. The final step is performed using the hysteresis
operator, where the pixels are marked as either edges, non-edges
Types of classification includes SVM (Support Vector or between the edges and non- edges, based upon the threshold
values. If the in-between pixels are connected with the edged
Machine), K-nearest neighbors (k-NN), decision tree. This app
performs model assessment and comparisons using confusion pixels, they are also considered as edge pixels. A binary image is
matrix and scattering plots. There are two types of classifier obtained as the result, where the white pixels closely approximate
training: the true edges of the original image.

3) Image resizing: The gray scale image is resized to a size of 200ྶ200.

• Manual classifier training
• Parallel classifier training 4) 2D convolution masking: In this step, the value of the pixel
is given in matrix format (mostly 3x3 matrix) and a masking
Manual classifier training: It is used to explore individual matrix is created as shown below.
model types or if we already know what type of classifier we
need. Thus, Manual classifier training classifies one at a time hN = [0 1 0; 0 -1 0; 0 0 0]; (1)
or a group of the same type.
Parallel classifier training: Using Parallel computing tool
box we can train classifiers. Thus this app starts a parallel pool This masking matrix and the matrix form of image is
of works automatically until and unless we turn off the default convoluted as shown below.
parallel preference. Thus Parallel classifier can train multiple
classifiers at once. In this project we are using Parallel nablaN = imfilter(diff_im,hN,'conv'); (2)
classifier training.
Perona- Malik Diffusion equation and PDE direct method is
used for the filtering of image.
II. METHODOLOGY Perona-Malik Formula:

The flowchart of various steps involved in the proposed g=1/(1+(x/k)^2) .*a (3)
system is shown in Fig. 4. g=exp(-(x/k)/2).*a (4)
A. Acquisition of input image :
PDE Direct Method:
The dataset of MRI Images are acquired from publicly
Img= Img + dt*(CN.*GN + Cs.*Gs + Ce.*Ge + Cw.*Gw) (5)
available sources such as hospitals, laboratories and medical
institutions and colleges which is saved in .jpg format.
The ‘Result 1’ and ‘Result 2’ in the Fig. 4
respectively refer to the type of the tumor and the comparison
result of classifiers.

978-1-5090-4929-5©2017 IEEE
320
 2017 IEEE 3rd International Conference on Sensing, Signal Processing and Security (ICSSS)

boundaries are checked for the subscripts of the neighbors and

Image acquisition then the indices of the eight neighbors are found from the
central derivative. The derivatives of SDF at x, y are
ʌ_x = - ʌ (a)+ ʌ (b); (6)
Pre-processing ʌ _y = - ʌ (c)+ ʌ (d); (7)
ʌ _xx = ʌ (a)-2* ʌ (e)+ ʌ (b); (8)
ʌ _yy = ʌ (c)-2* ʌ (e)+ ʌ (d); (9)
Feature extraction ʌ _xy = -0.25* ʌ (f)-0.25* ʌ (g)...
+0.25* ʌ (h)+0.25* ʌ (i); (10)
ʌ _x2 = ʌ _x.^2; (11)
ʌ _y2 = ʌ _y.^2; (12)

Neural Classification The kappa value is calculated using central derivatives as

Morphology follows
network learner app
Kappa= ((ʌ _x2.* ʌ _yy+ ʌ _y2.* ʌ _xx-* ʌ _x.* ʌ _y.* ʌ
_xy)./..( ʌ _x2+ʌ_y2+eps).^(3/2).*(ʌ_x2
+ʌ_y2).^(1/2); (13)
K-NN
clustering
G. Classification learner app:

SVM For comparison between various classifiers such as K-NN,

classifier SVM, and Trees, classification learner app is used. While
training each classifier, both the correct and misclassified
results are plotted in the scattering plot. Similarly the
confusion matrix displays the result in matrix form. By the
Result 1 Result 2
above process, the efficiency of each classifier is found and
are compared with one another.

Fig. 4 Flow chart of the proposed system

C. Neural Training:
III. EXPERIMENTALRESULTS
After the pre-process operation on the image, the dataset of The GUI menu option opens as in Fig. 5.
the image along with its extracted features are trained using
the neural network. The performance of the training operation
is validated and is plotted in a graph with mean square error in
y-axis and iteration values in x-axis.

D. Morphology:

The morphology operation is performed on the

preprocessed image using the ‘disk’ structuring element. The
image is first closed and the small objects on the image are
removed. Then a boundary is plotted around the area of tumor.
E. Clustering:

The morphed image is then clustered using the same ‘disk’ Fig. 5 GUI menu
structuring element. The clustering includes the process of
dilation and erosion followed by ‘open’ operation. After all The options are selected in order. On selecting the option
these operations, the connected components found on the “TESTING INPUT”, the input image can be choosen from any
image are returned. number of images. The input image is displayed as in Fig. 6.
F. Segmentation:
On selecting the option “NEURAL NETWORK AND
A signed distance map (SDF) is created for masking
PREPROCESSING”, the unknown input image is preprocessed
purpose. Interior and exterior points of the mask are found and
when it is simultaneously added as a dataset to the neural
using these points, their mean values are calculated. The kappa
value is then calculated by setting a threshold point. The

978-1-5090-4929-5©2017 IEEE
321
 2017 IEEE 3rd International Conference on Sensing, Signal Processing and Security (ICSSS)

network. The neural network training dailog box opens and “RESULT”, the result of the tumor either being normal,
the iterations starts as shown in Fig. 7. benign or malignant is displayed as in Fig. 9.

Fig. 6 Image acquisition Fig. 9 Final segmentation result

Then the classification learner app is opened and the data

for the given input image is plotted as a graph as in Fig.10.

Fig. 7 Neural network training

Fig. 10 Scatter plot
The performance of the validation can be viewed in the
form of graph as in Fig.8. The scatter plot and the confusion matrix of the image using
Linear SVM classifier is displayed as in Fig. 11.a and Fig.11.b.

Fig. 8 Performance of traing process

Fig. 11.a Scatter plot of Linear SVM classifier
After the process of morphology and clustering, the tumor
part is alone segmented and is displayed along with its
properties as in Figure.7. Also, on selecting the option

978-1-5090-4929-5©2017 IEEE
322
 2017 IEEE 3rd International Conference on Sensing, Signal Processing and Security (ICSSS)

[2] Jayashree. M. J, Charutha. S, “An Efficient Brain Tumor Detection By

Integrating Modified Texture Based Region Growing And Cellular
Automata Edge Detection”
[3] Ek Tsoon Tan, James V Miller, Anthony Bianchi, Albert Montillo- “Brain
Tumor Segmentation With Symmetric Texture And symmetric Intensity-
Based Decision Forests” GE Global Research, Niskayuna, NY, USA,
University of California Riverside, Riverside, CA USA.
[4] Subhashini. J, Vijay. J, “An Efficient Brain Tumor Detection
Methodology using k- means Clustering Algorithm”, International
conference on communication and Signal Processing, pg. 653– 657,
April 2013.
[5] Albert Singh. N, Amsaveni. V, “Detection of Brain Tumor using Neural
Network”, in ICCCNT, pg, 1-5, July 2013.
[6] Nguyen Thanh Thuy, Tran Son Hai, Le Hoang Thai, “Image
Classification using Support Vector Machine and Artificial Neural
Network”, Vietnam.
Fig. 11.b Classification matrix of Linear SVM classifier

The confusion matrix of Medium KNN is displayed as in

Fig. 12.

Fig.12 Confusion matrix of Medium KNN classifier

It is evident from the Fig.13, that the SVM classifier is

more efficient than other classifiers.

Fig.13 Comparison of classifiers

References
[1] Saptalakar. B. K, Rajeshwari. H, “Segmenttion Based Detection of Brain
Tumor,” International Journal of computer and Electronics Research,
Vol. 2, pp.20-23, February 2013.

978-1-5090-4929-5©2017 IEEE
View publication stats 323