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

net/publication/328784457

Detection of Brain Tumor from MRI images by using Segmentation &SVM

Article · October 2016

CITATIONS READS

25 2,797

3 authors:

Swapnil R Telrandhe Ankita Kendhe

4 PUBLICATIONS   37 CITATIONS    5 PUBLICATIONS   34 CITATIONS   

SEE PROFILE SEE PROFILE

Amit Pimpalkar
G. H. Raisoni Academy of Engineering & Technology, Nagpur
43 PUBLICATIONS   110 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Detection and Extraction of Text and sign from Traffic Panel View project

Bug Triage System with Classification View project

All content following this page was uploaded by Amit Pimpalkar on 07 November 2018.

The user has requested enhancement of the downloaded file.

 2016 World Conference on Futuristic Trends in Research and Innovation for Social Welfare (WCFTR’16)

Detection of Brain Tumor from MRI images by using

Segmentation &SVM
Swapnil R. Telrandhe1 Amit Pimpalkar2 Ankita Kendhe3
Computer Science &Engineering Computer Science & Engineering Computer Science & Engineering
G.H.R.A.E.T, G.H.R.A.E.T, G.H.R.A.E.T,
Nagpur, India. Nagpur, India. Nagpur, India.
[email protected] [email protected] [email protected]

Abstract— In this paper we propose adaptive brain tumor detection and segmentation may be largely divided into three
detection, Image processing is used in the medical tools for groupings: region-based, edge-based and fusion of region and
detection of tumor, only MRI images are not able to identify the edge-based methods. Well known and broadly used
tumorous region in this paper we are using K-Means segmentation techniques are K-Means clustering algorithm,
segmentation with preprocessing of image. Which contains de- unsupervised method based on neural network classifier [4].
noising by Median filter and skull masking is used. Also we are
using object labeling for more detailed information of tumor
Also, the time spent to segment the tumor is getting condensed
region. To make this system an adaptive we are using SVM due to the detailed demonstration of the medical image by
(Support Vector Machine), SVM is used in unsupervised manner withdrawal of feature points. Region-based techniques look
which will use to create and maintain the pattern for future use. for the regions satisfying a given homogeneity standards and
Also for patterns we have to find out the feature to train SVM. edge based segmentation methods look for edges between
For that here we have find out the texture feature and color regions with different characteristics [5].
features. It is expected that the experimental results of the
proposed system will give better result in comparison to other
existing systems. II. LITRATURE SURVAY
Index Terms—K-Means algorithm, Object Labeling Algorithm,
R.Muthukrishnan, et.al [6] Proposed brain tumor
Image segmentation.
detection in which Segmentation separates an image into its
I. INTRODUCTION component regions or objects. Image segmentation it needs to
segment the object from the background to read the image
It is an important to find out tumor from MRI images but
properly and classify the content of the image carefully. In this
it is somewhat time-consuming and difficult task sometime
framework, edge detection was an important tool for image
performed manually by medical experts. Large amount of time
segmentation. In this paper their effort was made to study the
was spent by radiologist and doctors for identification of
performance of most commonly used edge detection
tumor and segmenting it from other brain tissues. However,
techniques for image segmentation and also the comparison of
exact labeling brain tumors is a time-consuming task, and
these techniques was carried out with an experiment.
considerable variation is observed between doctors [2].
M.Saritha et.al, [7] Proposed approach by integrating
Subsequently, over the last decade, from various research
wavelet entropy based spider web plots and probabilistic
results it is being observed that it is very time consuming
neural network for the classification of Brain MRI. The
method but it will get faster if we use image processing
proposed technique uses two steps for classification i.e.
techniques [3]. Primary brain tumors do not spread to other
Wavelet entropy based spider web plot for feature withdrawal
body parts and can be malignant or benign and secondary
and probabilistic neural network for classification. The
brain tumors are always malignant. Malignant tumor is more
obtained brain MRI, the feature extraction was done by
dangerous and life threatening than benign tumor.
wavelet transform and its entropy value was calculated and
The benign tumor is easier to identify than the malignant
spider web plot area calculation was done. With the help of
tumor. Also the first stage tumor may be malignant of benign
entropy value classification using probabilistic neural network
but after first stage it will change to dangerous malignant
was calculated. Probabilistic neural network provides a
tumor which is life threatening [12].
general solution for pattern classification problem and its
Different brain tumor detection algorithms have been
classification accuracy is about 100%.
developed in the past few years. Normally, the automatic
P.NandaGopal et.al, [8] in their paper they presented
segmentation problem is very challenging and it is yet to be
a combination of wavelet statistical features (WST) and co-
fully and satisfactorily solved. The main aim of this system is
occurrencewavelet texture feature (WCT) obtained from two
to make an automated system for detecting and identifying the
level distinct wavelet transform was used for the organization
tumor from normal MRI. It takes into account the statistical
of abnormal brain matters in to benign and malignant. The
features of the brain structure to represent it by significant
planned system was consists of four stages: segmentation of
feature points. Most of the early methods obtainable for tumor

978-1-4673-9214-3/16/$31.00 © 2016 IEEE

region of interest, discrete wavelet disintegration, feature objects, keeping those that interest us and removing those that
abstraction, feature selection, organization and evaluation. The do not. Also thresholding converts the image from a grayscale
support vector machine was employed for brain tumor image, with pixel values ranging from 0 to 255, to a binary
segmentation. A grouping of WST and WCT was used for image, with pixel values of 0 or l. The processing window in
feature extraction of tumor region extracted from two level vision assistant displays a preview of the threshold operation
discrete wavelet transform. Genetic algorithm was used to using the current set of parameters. The pixels showed in red
select the optimal texture features from the set of mined have strengths that fall inside the threshold range. The
features. The probabilistic neural network was used to classify threshold operator sets their values to 1. The pixels depicted in
abnormal brain tissue in to benign and malignant and the gray have values outside the threshold range. The threshold
performance evaluation was done by comparing the operator sets their values to 0 [9].
classification result of PNN with other neural network Filters can smooth, sharpen, transform, and remove noise
classifier. The classification accuracy of the proposed system from an image so that we can extract the information needed
is 97.5%. to sharpen edges, counting the edges of any holes inside a
A. Laxami et.al, [9] proposed the work on information particle, and create contrast between the elements and the
(region of interest) in the medical image and thereby vastly background.
improve upon the computational speed for tumor segmentation When applied Convolution-Highlight Details with size
results. Significant feature points based approach for primary 7x7 the image is as shown in Fig 4. Preprocessing of brain MR
brain tumor segmentation was proposed. Axial slices of T1- image is the first step in our projected technique.
weighted Brain MR Images with contrast enhancement have Preprocessing includes image filtering and skull masking for
been analyzed. In order to extract significant feature points in further processing. Our intention behind performing
the image, applied a feature point extraction algorithm based preprocessing is improvement of the image quality to get more
on a fusion of edge maps using morphological and wavelet surety and ease in spotting the tumor. Steps for preprocessing
methods. Evaluation of feature points thus obtained has been are as follows:
done for geometric transformations and image scaling. A 1) Image is converted to gray scale.
region growing algorithm was then employed to isolate the 2) A 3x3 median filter is applied on brain MR image in order
tumor region. Preliminary results show that our approach has to remove the noise.
achieved good segmentation results. Also this approach was 3) The obtained image is then passed through a high pass filter
reduces a large amount of calculation. Future work will to detect edges. The high pass filter mask is used [1].
involve an investigation of the method in automatic 3D tumor The addition of edge noticed image and the original
segmentation, segmentation of ROI’s in other medical images, image in order to obtain the enhanced image will give the
as well as the importance of implemented technique in resulted image.
medical image retrieval applications. Median filter should only change the intensity of
corrupted pixels on the damaged image in order to preserve
III. PROPOSED METHOD the local details of the image it has to be done. However, for
The main purpose of this paper is to identify the region of fixed-valued impulse noise (i.e. salt-and-pepper noise) it is
tumor and to do the detailed diagnosis of that tumor which very difficult to detect the tainted pixels from this image
will used in treating the cancer patient the detailed about the correctly.
proposed system is given below.
Threshold is a specific intensity value which contents a Skull Masking
predefined intensity value; it is used to separate object or Detection of skull is used to control the boundaries of
Region of Interest (ROI) from the image background, chosen the object. The edge information helps to find out the region
in the range of 0 to 255 [13]. But it is detected that clustering of interest (ROI) i.e. the portion of the imagecovering the
methods followed by threshold cannot notice tumor correctly tumor. This work is done with the help of the calculating the
from MRI image, because the image consist of several non- centroid in the image [16]. Extraction of brain tissue from
brain tumor tissue. For this reason we express the proposed non-brain tissues in MR images which is referred to as skull
method using K-Means algorithm followed by Object stripping is an important step in many neuron imaging studies.
Labeling algorithm also, some preprocessing steps (median In this, we used automatic threshold value selector to
filtering and morphological operation) is used for tumor automatically choose threshold value. Then, mathematical
detection purpose [14]. morphology operations on a binaries image are applied stage
by stage to achieve acceptable skull stripped brain images.
A. Preprocessing The proposed skull stripping method comprises four steps.
In the image processing the gray scale image is processed Initially image binarisation is completed using threshold value
by using different techniques like brightness, threshold and and narrow connections are removed from binarised image
Filtering, Brightness makes the image by which white objects using morphological opening. Then, largest connected
are distinguished from gray and light items from dark objects. component from binarised image is selected by considering
Hence by changing the brightness of the image the tumor the fact that brain is the largest connected structure inside the
detection in the MRI image is easier. Thresholding isolates head.
 (i) By finding limits between regions based on discontinuities
in intensity levels
(ii) Thresholds based on the distribution of pixel properties,
such as intensity values
(iii) Based on finding the regions directly.
Thus the choice of image segmentation technique is
depends on the problem being considered. Region based
methods are based on continuity. These techniques split the
entire image into sub regions depending on some rules like all
the pixels in one region must have the same gray level.
Region-based techniques rely on common patterns in intensity
values within a cluster of neighboring pixels. The cluster is
referred to as the region in addition to group the regions
according to their anatomical or functional roles are the goal
of the image segmentation [6].
Threshold is the simplest way of segmentation. Using
thresholding technique regions can be classified on the basis
range values, which is applied to the strength values of the
image pixels. Thresholding is the transformation of an input
image to an output that is segmented binary image.
Segmentation Methods based on finding the regions directly
find for abrupt changes in the intensity value. These methods
are called as Edge or Boundary based methods. Edge detection
Figure 1: Flow diagram of proposed system is the problem of fundamental importance in image analysis.
Edge detection techniques are generally used for finding
Thirdly, a mathematical morphology operation such as: discontinuities in gray level images. To detect consequential
filling holes and dilation is carried out on selected largest discontinuities in the gray level image is the important
binarised image. Finally, we found skull stripped brain image. common approach in edge detection. Image segmentation
Create images and masks: methods for detecting discontinuities are boundary based
In this, the region properties of an image are used. By methods [1].
using centroid property of an image, a line is drawn in the
center of stripped skull. This divides the skull into two equal K-Means based segmentation
parts. One part is referred as a test image and other is referred A cluster is a collection of objects which are similar
as reference image. between them and are dissimilar to the objects belonging to other
Locating bounding box around tumor: In each input MR clusters. It deals with finding a structure in a collection of
slice (axial view), there is a left–right axis of symmetry of the unlabeled data. A loose description of clustering could be the
brain. A tumor, which is considered adeviation in the brain, process of organizing objects into groups whose members are
typically perturbs this symmetry. Parallel rectangle on the left similar in some way. K-Means clustering is an algorithm to group
side that is very unlike from its replication about the axis of objects based on attributes/features into k number of groups
symmetry on the right side. i.e., the intensity histograms of where k is a positive integer. The grouping (clustering) is done by
minimizing the Euclidean distance between the data and the
two rectangles are most different, but the intensity histograms
corresponding cluster centroid. Thus the function of K-Means
of the outside of the rectangles are comparatively similar [18].
clustering is to cluster the data. Commonly used initialization
We assume that one of the two rectangles will restrict the
methods are Random Partition [20]. The Forgy method
tumor appearing in one hemisphere of the brain.
randomly chooses k observations from the data set and uses
these as the initial means. The Random Partition method first
B. Segmentation
randomly assigns a cluster to each observation and then
Image Segmentation is the procedure of partition a digital
proceeds to the update step, thus computing the initial mean to
image into numerous regions or sets of pixels. Essentially, in
be the centroid of the cluster's randomly assigned points. The
image partitions are different objects which have the same
Forgy method tends to spread the initial means out, while
texture or color. The image segmentation results are a set of
Random Partition places all of them close to the center of the
regions that cover the whole image together and a set of
data set. According to Hamerly [10] the Random Partition
contours extracted from the image [11]. All of the pixels in a
method is generally preferable for algorithms such as the K-
region are similar with respect to some characteristics such as
harmonic means and fuzzy k-means. For expectation
color, intensity, or texture. Neighboring regions are
maximization and standard K-Means algorithms, the Forgy
considerably different with respect to the same individuality.
method of initialization is preferable [17].
The different approaches are
 the same operation independently within each partitioned
block of images.

D. SVM
The SVM is a supervised learning method. It is a good
tool for data analysis and classification. SVM classifier has a
fast learning speed even in large data. SVM is used for two or
more class classification problems. Support Vector Machine is
based on the conception of decision planes. A decision plane
is one that separates between a set of items having dissimilar
class memberships. The Classification and detection of brain
tumor was done by using the Support Vector Machine
technique. Classification is done to identify the tumor class
present in the image. The use of SVM involves two basic steps
of training and testing.

Linear SVM:
Figure 2: K-Means work flow The training patterns are linearly divisible. That is, there
In above figure we are showing K-Means segmentation exists a linear function of the form f(x) = w T x + b (1) such
algorithm work flow. The first block shows the input image that for each training example xi, the function yields (f(x)) • 0
then it will convert that image into numbers of clusters, then it if x for =+1, y and f (xi) <0 for yi=-1.
will find out the cluster centroid and find out the distance of In other words, training examples from the two different
each pixel from the centroid, then it will continue to grouping classes are separated by the hyper plane
of pixels till it will reached the last pixel and stop. F(x) = w T x + b =0, where w is the unit vector and b is a
invariable. For a given training set, while there may exist
C. Object Labeling many hyper planes that maximize the separating margin
The object labeling will performed after between the two classes, the SVM classifier is based on the
segmentation of an image, actually for fine result we are using hyper plane that maximizes the separating margin between the
both segmentation as well as object labeling for tumor two classes In other words, SVM finds the hyper plane that
detection. Usually used initialization methods are Random causes the largest separation between the decision function
Partition. The Forgy method arbitrarily chooses k annotations values for the “borderline” examples from the two classes.
from the data set and uses these as the initial means. The SVM classification with a hyper plane that minimizes the
Random divider method first arbitrarily assigns a cluster to separating margin between the two classes. Support vectors
each observation and then proceeds to the update step, thus are elements of the training set that lie on the boundary hyper
computing the original mean to be the centroid of the cluster's planes of the two classes.
randomly assigned points. The Forgy method tends to spread
the initial means out, while Random divider places all of them Non-Linear SVM
close to the center of the data set. According to In the above discussed cases of SVM classifier also
Hamerly[10] the Random divider method is generally shown in figure 7. Straight line or hyper plane is used to
preferable for algorithms such as the k-harmonic means and distinguish between two classes. But datasets or data points
fuzzy k-means. For expectation maximization and standard k- are always not separated by drawing a straight line between
means algorithms, the Forgy method of initialization is two classes. For example the data points in the below figure 7.
preferable. It can’t be separable by using above SVMs discussed. So,
We used three scales of patch size for computing Kernel functions are used with SVM classifier. Kernel
HOG and LBP, namely, 16×16, 24×24 and 32×32. The function provides the bridge between from nonlinear to linear.
multiple patch sizes provide richer coverage of different scales Basic idea behind using kernel function is to map the low
and make the features more invariant to scale changes [19]. dimensional data into the high dimensional feature space
After extracting dense local image descriptors, we where data points are linearly separable [15].
perform the ‘coding’ and ‘pooling’ steps, where the coding A pattern gratitude network, which is a feed-backward
step encodes each local descriptor z via a nonlinear feature network with tan sigmoid transfer functions in both the hidden
mapping into a new space, then the pooling step aggregates layer and the output layer, is used. The network has only one
the coding results fallen in a local area into a single vector. We output neuron, as there are 24 input vectors. The hidden layer
apply two state-of-the-art ‘coding + pooling’ pipelines in our neurons are 100 and the learning rate is 0.1. The momentum
system, one is based on local coordinate coding (LCC), and factor is 0.9 and total numbers of epochs are 500. The error is
the other is based on super-vector coding (SVC). For minimized by 0.001 and the performance of the classifier is
simplicity, we assume the pooling is comprehensive. But evaluated by calculating accuracy.
spatial pyramid pooling is simply implemented by applying
 Figure 5 shows the noisy image and the result of whole
preprocessing is shown in figure 6

Figure 5: Noisy image Figure 6: Enhanced image

Figure 7: The classification process of SVM In figure 7 de-noising is done and in figure 8 skull
SVM are based on optimal hyper plane for linearly pair
masking is done the combining result of both will give result
able patterns but can be complete to patterns that are not
of preprocessing module.
linearly divisible by transformations of unique data to map
into new space. They are clearly based on a abstract model of
learning and come with theoretical guarantees about their
performance. They also have a modular design that allows one
to separately apply and design their components and are not
affected by local minima. Support vectors are the elements of
the training set that would change the location of the dividing
hyper plane if removed[19]. Support vectors are the grave
elements of the training set. The problem of discovering the
best hyperplane is an optimization problem and can be solved
by optimization techniques.
IV. RESULTS AND DISCUSSION Figure 7: Enhanced Image Figure 8: Skull MaskingImage
The test of projected technique to discover and segment
brain tumor is performed using MR images of diverse long- In figure 9 the segmentation by using K-Means
suffering. Each test image has brain tumor of diverse size, algorithm is done here. Segmentation is done by using some
shape and intensity. Manual examination is used to check the colors in preprocessed image on the basis of intensity we have
correctness of automated segmented tumor area. The divided brain MRI in to 3 colors as shown in resulted image.
experimental result for different MR images containing tumor
of different shapes, sizes and intensities.

Figure 9: Segmentation resulted image

The reason behind using various colors in

segmentation is to identify an area of interest from the MRI
Figure 3: Image with skull Figure 4: Removing skull images; human eyes are more sensitive to color images than
(outer ring) tissues the grey scale images. So here we are using the intensities of
In figure 3 this is one of the MRI image with skull in MRI images to place different color on the image the resultant
next figure 4 the result showing after skull masking. In skull image of this process will give the idea of tumor region.
masking we have generated some horizontal, vertical, diagonal
and anti-diagonal mask and after dividing image in small V. CONCLUSION
segments we have applied this masks over image sub parts and Brain tumor detection is done by preprocessing which is
result image get combined is given this resultant image. first step in that median filter and by using diagonal, anti-
diagonal masks segmented images get preprocessed and skull
masking is done here. After skull masking fatty tissues and [9] A. Laxmi and P.Samata, “Optimization of visual
other unwanted details get smoothen. presentation of MRI image for accurate detection of
Preprocessed image is segmented with the K-Mean tumor in human brain using virtual instrument” in The
segmentation and Object Labeling with HOG, HOG is friendly Biomedical Engineering International Conference 2012.
with feature extraction. So the texture feature and color feature [10] G.Hamerly and C.Elkan, “Learning the k in K-Means” in
are extracted here in the system which is use to find out the region Proceedings 7th Annual Conference on Neural
of interest and SVM is use for pattern mapping and pattern Information Processing Systems (NIPS), pp. 281-288,
matching process. Also use to learn Neural Network. 2003.
Image processing has become a very important task in [11] B.C.Patel and G.R.Sinha, “An adaptive K-Means
today’s world. Today applications of image processing can be Clustering algorithm for breast image
originate in number of areas like medical, remote sensing, segmentation” International Journal of Computer
electronics and so on. If we focus on medical applications, and Applications, vol. 7, no. 4, pp. 35-38, 2010.
image segmentation is widely used for diagnosis purpose. [12] T.Kalaiselvi and K.Somasundaram, “Fuzzy c-means
In this paper, we have proposed a system that can be used technique with histogram based centroid initialization for
for segmentation of brain MR Images for Detection and brain tissue segmentation in MRI of head scans” in
identification of brain tumor. We find area of tumor and its Proceedings in IEEE-international Symposium on
type of tumor. Future scope for detection and segmentation of Humanities, Science and Engineering Research, pp. 149-
brain tumor is that if we obtained the three dimensional image 154, 2011.
of brain with tumor then we can also find out its tumor size [13] M.Shasidhar, Y.S.Raja and B.Y.Kumar, “MRI brain
and also can evaluate its tumor type and also its stage of image segmentation using modified fuzzy c-means
tumor. clustering algorithm,” in Proceedings in IEEE-
REFERENCES International Conference on Communication Systems and
Network Technologies, pp. 473-478, 2011.
[1] S.R.Telrandhe, A.Pimpalkar and A.Kendhe, “Brain [14] S.Koley and A.Majumder, “Brain MRI segmentation for
Tumor Detection using Object Labeling Algorithm & tumor detection using cohesion based self-merging
SVM”, in International Engineering Journal For Research algorithm,” in Proceedings in IEEE-3rd International
& Development Vol. 2, special issue, pp. 2-8, Nov 2015. Conference on Communication Software and Networks,
[2] SobriMuda and M. Mokji, “Brain Lesion Segmentation pp. 781-785, 2011.
from Diffusion-Weighted MRI Based On Adaptive [15] M.U.Akram and A.Usman, “Computer aided system for
Thresholding And Gray Level Co-Occurrence Matrix”, J. brain tumor detection and segmentation,” in Proceedings
Telecommunication Electronic and Computer in IEEE-International Conference on Computer Networks
Engineering, Vol. 3, No. 2, 2011. and Information Technology, vol. 1, pp. 299-302, 2011.
[3] S. Cha et.al, “Review Article: Update on Brain Tumor [16] K.Thapaliya and G.Kwon, “Extraction of brain tumor
Imaging: From Anatomy to Physiology”, Journal of based on morphological operations,” in Proceedings in
Neuro radiology, Vol.27, pp.475- 487, 2006. IEEE-8th international Conference on Computing
[4] R. Adams and L. Bischof, “Seeded region growing,” Technology and Information Management, pp. 515-520,
IEEE Trans. On Pattern Analysis and Machine 2012.
Intelligence, vol.16, no.6, pp.641–647, 1994. [17] M.Schmidt, I.Levner, R.Greiner, A.Murtha and Bistritz,
[5] J.Selvakumar, A.Lakshmi and T.Arivoli, “Brain tumor “Segmenting brain tumors using alignment based
segmentation and its area calculation in brain MR images features,” in Proceedings in IEEE-4th International
using k-mean clustering and fuzzy c-mean algorithm” in Conference on Machine Learning and Applications,
Proceedings in IEEE-International Conference On pp.567-578, 2005.
Advances In Engineering, Science And Management, pp. [18] Elamy and M. Hu, “Mining brain tumors and tracking
186-190, 2012. their growth rates,” in Proceedings in IEEE-Canadian
[6] R.Muthukrishnan and M.Radha, “Edge Detection Conference On Electrical and Computer Engineering,
Techniques for Image Segmentation” in International Seaside, CA, USA, pp. 872-875, 2007.
Journal of Computer Science & Information Technology, [19] S.H.Teoh et.al, “Median Filtering Frameworks for
Vol 3, No 6, Dec 2011. Reducing Impulse Noise from Grayscale Digital Images:
[7] M.Saritha and K.Paul Joseph, ”Classification of MRI A Literature Survey” in International Journal of Future
brain images using combined wavelet entropy based Computer and Communication, Vol. 1, No. 4, pp. 323-
spider web plot and probabilistic neural network” 333 December 2012.
ELSEVIER, pattern Recognition letters 34, pp. 2151- [20] J.Vijay and J.Subhashini, “An Efficient Brain Tumor
2159, 2013. Detection Methodology Using K- Means Clustering
[8] P.N.Gopal and R.Sukanesh,” wavelet statistical feature Algorithm” in International conference on
based segmentation and classification of brain computed Communication and Signal Processing, April 3-5, pp.653-
tomography images” IET Image Prosess Vol.7 pp 25 -32 658, 2013.
2013

View publication stats