Page 1 of 6
European Journal of Business &
Social Sciences
Available at https://ejbss.org/
ISSN: 2235-767X
Volume 07 Issue 03
March 2019
Available online: https://ejbss.org/ P a g e | 441
Segmentation and Classification of Brain Tumor from MRI Brain
Images for Medical Diagnosis
Dr. B.Subrahmanyeswara Rao1
, K. Gayathri2
, S.Harshitha3
, L.V.S.Ganesh4
, P.Ramesh5
boyinasrao@gmail.com,gayathrisadhana68@gmail.com,harshithasayala@gmail.com
venkatasivaganesh111@gmail.com, rameshpechetti1998@gmail.com
Department of ECE, Swarnandhra College of Engineering and Technology, Narsapur,W.G.Dst,A.P
Abstract
This work aimed was to initiate an
automated tumor diagnostic system based on T1
and T2 weighted magnetic resonance images
(MRI) as the detection of brain tumor is
complicated procedure in medical field. This
system incorporates steps for pre-processing, image
segmentation with morphological operations in
multiple steps to segment benign and malignant
tumor or tissue, feature extraction and image
classification. This can be done by SVM (support
vector machine) and k-means algorithm. The
textural and shape based features were extracted by
discrete wavelet transform moments from the
segmented tumor regions. Finally, the
implementation of the suggested system was
evaluated with extreme learning machine (ELM)
algorithm to discriminate two types of tumor on
MR images. The Diagnostic efficiency of the
proposed system was evaluated with sensitivity,
specificity and accuracy.
Keywords
Pre-processing, SVM, Image
segmentation, morphological operations, benign
and malignant tumor.
Introduction
In human nervous system, brain is in the
center. Automated identification of brain tumor
based on magnetic resonance (MR) images is of
excessively beneficial in neurosurgery and
treatment designing. The brain tumors are varied in
size and characteristic properties, which are
important elements in the segmentation of for
anatomical structures. Hence automatic detection of
size and location of tumor on MR image is very
difficult task, due to the intensity variations and
magnetic field in homogeneities. By using semi- automated [5] and automated techniques, number
of computational brain tumor segmentation has
been increased for tumor structures over the past
two decades. To control the outcome of the
magnetic field in homogeneities and cavities, most
of the existing approaches utilized bias-field
correction [4] and wavelet transformation [2] as
preprocessing techniques for the segmentation
process.
There are various learning algorithms such as
neural networks were applied for MRI object
identification. The methods were introduced for
brain tumor and tissue surface localization.
Recently, K-means clustering [3] have shown its
potential in tumor segmentation, however it
increase the complexity in extract of features in
deep hierarchy network.
Hence, we suggest a successful and efficient
technique for segmentation, classification of benign
and malignant brain tumor identification form MR
based images. Textural and shape based features
were obtained by using wavelet transform and
Zernike moments. Furthermore, extreme machine
learning algorithm was applied for discriminating
benign and malignant tumor.
Existing method:
The proposed method consists of number
of steps such as enhancement to sharpen the edges,
segmentation of region of interest (ROI), extracting
texture and shape based features and finally,
classifying by machine learning technique to
identify brain tumor or tissue is described in fig.1.
Fig. 1. Illustration of brain tumor segmentation and
classification.
Page 2 of 6
European Journal of Business &
Social Sciences
Available at https://ejbss.org/
ISSN: 2235-767X
Volume 07 Issue 03
March 2019
Available online: https://ejbss.org/ P a g e | 442
MATERIALS AND METHODS
Dataset description
We conducted experiment on real patient data
obtained from the Tamil Nadu Government Multi
super specialty hospital from January to August
2016. It consists of 17 benign and 11 malignant
brain tumor or tissue MR images with an image
matrix of 500 X 540 pixels. Structural data
included both of high resolution T1 and T2
weighted MRI scans received on a GE Sigma HDxt
1.5T with 5mm slice thickness and display field of
view in the range from 24.0 – 28.0cm.
Enhancement
The gray scale image conversion whose
entries are between 0 and 255, with 0 to black and
255 to white. After the image conversion, the
enhancement technique was essential to remove
low frequency noise and sharpening the edges of
the of the brain image. Median filter was utilized to
remove noise by estimating the new neighboring
pixels. To emphasize the fine details in the image
we applied high pass filtering technique by
removing the low
frequency noise. Then the resultant high pass
filtering image is converted into the binary image
by using the mean pixel values of all the pixel
image.
Fig.2. Benign image (A) original (B) Original (C)
highpass filter (D) highpass filter (E) binary and
malignant (F) binary
Labeling procedure
High density pixels were estimated from
the proportion of the pixels in the region by
applying the labeling of objects to the binarized
image and extract the objects composed greater
than 50 pixels. The object location of the maximum
value pixels of the extracted highly dense object
pixels was considered for further processing of the
segmentation of the brain tumor process. The
morphological operation of the dilation using a
square structuring element of the extracted object
removed small holes and determined exact
boundary of the tumor or tissue in the image as
shown in fig. 3
Fig.3. Tumor boundary of (A) benign and
(B) malignant
Segmentation
It separate image objects into number of
discrete regions such that pixels sharing high
similarity belong to the same region. We
incorporated traditional k-means clustering
algorithm [3] because of its simple and fast
computation than the other hierarchical clustering.
It targets to group number of evaluations into k
clusters in which each evaluation associates to the
cluster with the nearest mean. First k cluster centers
were chosen to accept with k arbitrarily chosen
model inside the hyper volume containing the
model C. Then allocate each model to the nearest
cluster center. Again the cluster centers are
recomputed based on current cluster memberships U.
If the convergence of criterion is not satisfied then
the process is repeated with new cluster centers
until to get minimal squared error. The resultant
image of the benign and the malignant and the
tumor shown in the following below figures.
Fig. 4. Segmentation using k-mean algorithm (A)
benign and (B) malignant
Page 3 of 6
European Journal of Business &
Social Sciences
Available at https://ejbss.org/
ISSN: 2235-767X
Volume 07 Issue 03
March 2019
Available online: https://ejbss.org/ P a g e | 443
Feature extraction
In feature extraction, multi-resolution
transform called wavelet transform followed by
Zernike moments were applied by which textural
and shape feature vectors were extracted from the
identified tumor or tissue region from the MR
images. The wavelet extraction of an image is
acquired to examine the different frequencies of an
image using different scales. Zernike moments
were used to explain the properties of an image
with no redundancy or overlap information
between the moments. Thus, four frequency band
utilizing the hard features from wavelet and two
rotational invariant features from Zernike were
included.
Classification
Extreme learning machines (ELM) are
introduced for classification based on single layer
of hidden nodes, with good generalization
performance. The weights between hidden nodes
and outputs are learned using a linear model. Since
it is gradient based algorithm, it analytically solves
the problem by calculating the optimal weights of
the single-hidden Layer feed-forward Neural
Networks (SLFN). Hence β of the linear system
Hβ=T:
||H(w1,,,,wn,q1,,,,qn)β-T||=
minβ||H(w1,,,,wn,q1,,,,qn)β-T
...................(1)
In many occurrences the number of hidden nodes
are quite lesser than the number of training, thus H
is not a square matrix, and there may not remain wi
,
q i
,
β i
, such that Hβ=T. The smallest norm least
square solution of the linear system is(2)
B = H*T.............................(2)
Where H*
is the Moore-Penrose
generalized inverse of matrix.
The classification of the extracted features
from the tumor or tissue was evaluated using
sigmoidal kernel function. The performance of the
suggested system can be validated by sensitivity,
specificity and accuracy.
Sensitivity = TP/(TP+FN)........(3)
Specificity = TN/(FP+TN).......(4)
Accuracy=(TP+TN)/(TP+TN+FN+FP)
.............(5)
Accuracy = (TP+TN)/(TP+TN+FN+FP)
.............(6)
These are the previous algorithms which are
different from us.
Proposed system
In this proposed system, there obtained the
number of steps such as enhancement of the
sharpened edges, segmentation by using k-means
algorithm and segmentation with multistage of
morphological operations, feature extraction and
then classification. There by here also consists of
segmentation. We are aiming to present the
different MRI images segmentation methods by
using k-means algorithm and morphological
operations in multistages, discrete wavelet and
SVM (support vector machine) based on the
features of MRI (magnetic resonance image)
Fig-1. Stages in brain tumor segmentation.
The below figure shows the stages in brain
tumor. In this paper we are presenting to take
review on different methods of brain tumor image
segmentation. We are aiming to present the
different MRI images segmentation methods by
using k-means algorithm and morphological
operations in multistages.
As compared with the above exist method;
the illustration of brain tumor segmentation and
