Identification of Brain Tumor using Neural Network and Support

advertisement
Identification of Brain Tumor using Neural Network and Support Vector Machine
Bhawna Gupta, Shamik Tiwari
CSE, Mody University of Technology & Science, India
Abstract: Brain tumor is one of the deliberate diseases, causes death among the peopl e. Imaging techniques like CT
scan (Computed Tomography) and MRI (Magnetic Resonance Images) are used for diagnosis the brain tumor. Here
in this paper a computer based system is designed for the detection of brain tumor from the MRI images. This
system consists following steps feature extraction and classification. The extraction of features from the MRI images
has been achieved by curvelet transform. Then these extracted features are employed in neural network and support
vector machine respectively. This system gives efficient accuracy in the detection of brain tumor.
Keywords: Brain Tumor, Curvelet Transform, Neural Network, Support Vector Machine
I Introduction
Medical imaging is an active research area in image
processing which is play crucible role in the detection
of brain tumor. The motivation of medical imaging
provides information about the human system and
this information is used by doctor to detect the
diseases[1].Medical
imaging
provides
many
techniques like MRI (Magnetic Resonance Imaging),
CT scan (Computed Tomography), and MRA
(Magnetic Resonance Angiography) for diagnosis of
brain tumor [2]. From these techniques, MRI is a
most efficient technique for the detection of brain
tumor because it doesn’t use x rays and it provides
high resolution images of soft tissues [3]. But due to
the shortest of radiologist and large volume MRI data
to be analyzed makes it cost expensive and labor
intensive. So there is need a computer based system
which classify medical images accurately [4]. Image
processing techniques are used for designing the
computer based system. Various author used image
processing techniques for the detection of brain
tumor. Some of discussed below. Joshi et al. [5]
developed a classification system. The main stages of
this system are, first tumor segmentation done by
histogram equalization, second feature extraction,
gray level co- occurrence matrix is used for this
approach. Then classification is done by neuro fuzzy
classifier. Texture features are used in the training of
the artificial neural network. Gupta et al. [6]
proposed a technique in which feature extraction is
done using curvelet transform, then these features are
applied to support vector machine for successful
classification. Sridhar et al. [7] recommend a
technique which follows two steps process in first
step dimensionality reduction and feature extraction
which is done by discrete cosine transform and
second step is classification done by probabilistic
neural network. The rest paper is organized as
follows section II describe the proposed
methodology. Section III describes the experimental
results and conclusion describe in section IV.
II Methodology
In our proposed method we use two classifiers for the
detection of brain tumor. The whole procedure shows
in the flow chart.
Apply
Curvelet
transform
Input Image
Curvelet
Decomposition
Feature Extraction
Mean and
Standard
Deviation
Neural Network
Classification
Cancerous or not
cancerous
Support vector
Machine
Result
Fig1 Schematic diagram of Proposed
Methodology
Figure 1 shows the procedure of proposed method. It
incorporates three steps i) Input Database ii) Feature
Extraction iii) Classification. Feature extraction
extracted the features of MRI and classification
classify brain MRI is cancerous or non cancerous.
Fig 2 Implementation of fast digital curvelet
transform
2.1 Input database:
The input data of our project is MRI (Magnetic
Resonance Images) images. These data are taken
from TCIA (The Cancer Imaging Archive) [8]. From
the TCIA database we selected 50 MRI images in
which 25 are cancerous and other 25 are non
cancerous. The size of images is 512x512.
The implementation of FDCT to an image in
wrapping version follows the following procedure
[11].
i) The 2d image first converted into the frequency
domain by applying the Fourier transform.
2.2 Feature Extraction:
Feature extraction defines the relevant features that
are used for understanding the image. Classification
depends on the feature extraction process. In our
system feature are extracted using curvelet transform.
1) Curvelet Transform
Candes et.al [9] developed a new multiscale
transform which represent curve singularities along
the edges much more efficiently than the other
transform, is known as curvelet transform. In 2006,
the initial construction of curvelet transform was
redesigned and then it was later introduced as fast
digital curvelet transform (FDCT). This method is
easy to understand in comparison to the first method.
Curvelet transform is defined in both forms like in
continuous and digital. Image based feature
extraction is done by FDCT. There are two types of
digital implementations of FDCT: i) Curvelets via
USFFT (Unequally Spaced Fast Fourier Transform)
and Curvelets via wrapping [10]. Both take Cartesian
array as an input and produces digital coefficients as
an output.
Image in Fourier
Domain
2d image
ii) Then the transformed image convoluted with the
set of window function. The shapes of window
function are according to the requirements of the
ideal curvelet transform like parabolic scaling law.
iii) Now inverse Fourier transform (IFFT) is applied
for curvelet coefficient. But due to the frequency
response of curvelet is non rectangular wedge, for
IFFT wedge must be in rectangular form. So
wrapping is done.
iv) After wrapping, IFFT is applied and curvelet
coefficients are obtained.
At scale 1
one sub band
image
Scale 2 eight
sub band
images
Apply Curvelet
Transform
Scale 3
sixteen sub
band images
FFT
Product of image
with curvelet
window
Scale 4 sixteen
sub band
images
Wrapping
IFFT
Curvelet
Coefficients
Fig 3 Curvelet Coefficients at varied angles (scale=4)
Curvelet
coefficient in
Fourier domain
Figure 3 shows the schematic diagram of a proposed
MRI features extraction. At every scale and from
each sub band images we computed two features
(Mean and standard deviation). So we get total 82
features. Now these extracted features are employed
into both classifiers.
2.3 Classification
2) Support Vector Machine:
Classification means to classify between the
cancerous and non cancerous images. In our
proposed work we used two classifiers i) Neural
Network ii) Support Vector Machine. Extracted
features are fed up into both classifiers.
In 1992 a new learning process is developed by
Vapnik which known Support Vector Machine
(SVM) [13]. SVM provides higher accuracy in terms
of data classification and other data classification
algorithms. So it is used in many applications like
text categorization, hand-written digit recognition,
tone recognition, image classification and object
detection, micro-array gene expression data analysis,
data classification [14]. The concept behind SVM is
finding the optimal hyperplane that separates the
classes and has maximum distance between both
classes.
Artificial neural network is a simplified model of the
brain and also referred as neural network. It is made up of
biological neurons that are connected together and forms
a network. A particular weight is associated to each
neuron. It resembles brain in two aspects [12].i.)
Knowledge is acquired by the network through the
learning process. ii.) Interneuron connection strength
known as synaptic weight, one used to store this acquired
knowledge.
Wi1
Summing
Junction
X1
Wi
X2
∑
2
Activation
Function
Xn
Wi3
Inputs
F
y
Optimal
Hyperplane
Feature of class2
1) Artificial Neural Network
Output
bk
Synaptic
weights
Bias
Fig 3 Basic Neuron Model
Feature of class1
The block diagram of figure 3 shows the basic neuron
model. Three basic elements of neural network are
Fig 4 A linear classifier



Input: Each input is associated with a numeric
weight
An adder is there which is used for summing the
inputs.
Activation function, which is used to limit the
amplitude of the output of a neuron. The output
of a neuron is a function of the weighted sum of
the inputs plus a bias. Bias is a fixed input.
yi = f ( ∑𝑗=0 𝑤 ij xj ) + bk
(1)
Neural network is used for solving the variety of
problems like classification, segmentation etc.
Figure 4 shows the example of linear two class
classifier. Here we have two classes’ cancerous and
non cancerous. There is a number of hyperplanes that
separates the both classes. So SVM find optimal
hyperplane which have maximum distance between
both classes and no points between them. SVM
works on maximum margin principle.
III Results
As explained earlier 50 images are selected from the
database out of which 25 are cancerous and 25 are
not cancerous. We have used all the images for
training and testing of classification framework. We
used 30 images for training and all images are used
for testing. In both classifiers positive samples
(cancerous MRI) are trained by 1 and negative
samples (Non cancerous MRI) by 0. To estimate the
performance of both classifiers we calculate accuracy
(AC), precision (P) and recall (R) of each classifier.
Accuracy defines how many samples are correctly
classified. Precision is the percentage of the samples
which actually have cancerous class with all those
which were classified as cancerous class. Recall is
the percentage of the samples which were classified
as cancerous class, with all samples which truly have
cancerous class.
AC= (TP+TN)/ (TP+TN+FP+FN)*100
(2)
P = TP/TP+FP*100
(3)
R= TP/TP+FN*100
(4)
draw between the true positive rate (TPR) and false
positive rate. In ROC the diagonal line shows the
random classification, if curve is draw above the
diagonal line it shows better classification.
Where TP is the number of true positives, TN is the
number of true negatives, FN is the number of false
negatives, and FP is the number of false positives, are
defined as:
TP: Predicts cancerous as cancerous.
Fig 5 ROC Curve for SVM
TN: Predicts noncancerous as noncancerous.
FN: Predicts cancerous as noncancerous.
FP: Predicts noncancerous as cancerous.

ROC for Support vector machine is shown in figure
5. In this curve blue line depicts the performance of
the classifier. This curve is draw above the diagonal
shows the good classification.
Performance of Neural Network
Accuracy
94.00%
Precision
100%
Recall
89.28
Table 1 Performance of Neural Network
Table 1 shows the results after testing the neural
network with all images.

Performance of Support Vector Machine
Accuracy
98.00%
Precision
96.15%
Recall
100%
Table 2 Performance of Support Vector Machine
Fig 6 ROC Curve for NN
Table 2 shows the results after testing the neural
network with all images. According to the tables
support vector machine gives better results than the
neural network.
ROC Curve
ROC curve (Receiver Operating Characteristics)
demonstrates the performance of the classifier. It is
ROC for Neural Network is shown in figure 6. In this
curve blue line depicts the performance of the
classifier. According to the ROC curve of both
classifier support vector machine give better
classification than the neural network.
IV Conclusion
A computer based method has been developed for
classify the brain tumor using image processing
techniques. The main steps of this system are feature
extraction which us done by curvelet transform. For
the classification task we used two classifier1) neural
network 2) support vector machine. From the results
it is found that support vector machine gives better
classification in comparison to neural network. In this
technique classification is done between the
cancerous and non cancerous images. In future this
technique makes more advances then tumor is
classified according to its type.
References
[1] M. G. Mostafa and M. F. Tolba, Medical Image
Segmentation Using Wavelet Based Multi resolution
EM Algorithm, IEEE trans on Industrial Electronics,
Technology and automation, IETA, 2001.
[2]Mubashir Ahmad, Mahmood ul-Hassan, Imran
Shafi, Abdelrahman Osman,"Classification of
Tumors in Human Brain MRI using Wavelet and
Support Vector Machine",in IOSR Journal of
Computer Engineering(2012)
[3] Islam, Atiq, Syed Reza, and K. Iftekharuddin.
"Multi-fractal texture estimation for detection and
segmentation of brain tumors." (2013)
[4] Bin Othman, M. F., Norarmalina Abdullah, and
N. A. Bin Ahmad Rusli. "An overview of MRI brain
classification
using
FPGA
implementation." Industrial
Electronics
&
Applications (ISIEA), 2010 IEEE Symposium on.
IEEE, 2010.
[5] Joshi, D.M., Rana, N.K. ; Misra, V.M.”
Classification of Brain Cancer Using Artificial
Neural Network” in “Electronic Computer
Technology
(ICECT),
2010
International
Conference”,
©IEEE,
doi:
10.1109/ICECTECH.2010.5479975
[6] Gupta, Bhawna, and Shamik Tiwari. "Brain
Tumor Detection using Curvelet Transform and
Support Vector Machine." (2014).
[7] Sridhar, D., Murali Krishna, I.V. “Brain Tumor
Classification Using Discrete Cosine Transform and
Probabilistic Neural Network” in “Signal Processing
Image Processing & Pattern Recognition (ICSIPR),
2013 International
Conference”,©IEEE,
10.1109/ICSIPR.2013.6497966
doi:
[8] http://www.cancerimagingarchive.net
[9] E.J. Candes, D.L. Donoho, Curvelets, multiresolution representation, and scaling laws, Wavelet
Applications in Signal and Image Processing VIII,
vol. 4119-01, SPIE, 2000
[10] Guha, Tanaya, and QM Jonathan Wu. "Curvelet
Based Feature Extraction." Face Recognition (2010):
35-46
[11] Gupta, Bhawna, and Shamik Tiwari. "Lung
Cancer Detection using Curvelet Transform and
Neural Network." International Journal of Computer
Applications 86 (2014).
[12] Haykin, Simon. Neural networks: a
comprehensive foundation. Prentice Hall PTR, 1994
[13] Boser, Bernhard E., Isabelle M. Guyon, and
Vladimir N. Vapnik. "A training algorithm for
optimal margin classifiers." Proceedings of the fifth
annual workshop on Computational learning theory.
ACM, 1992
[14] Srivastava, Durgesh K., and Lekha Bhambhu.
"Data Classification Using Support Vector Machine."
Journal of Theoretical & Applied Information
Technology 12 (2010).
Download