International Journal of Image Processing and Visual Communication

ISSN (Online)2319-1724 : Volume 1 , Issue 2 , October 2012

Automated Brain Tumor Detection and Identification

Using Image Processing and Probabilistic Neural
Network Techniques
Dina Aboul Dahab1, Samy S. A. Ghoniemy2, Gamal M. Selim3
Dep. of Computer Engineering, Arab Academy for Science, Technology & Maritime Transport
Cairo, Egypt
1

[email protected]
2
[email protected]
[email protected]

technology which allows cross sectional view of the body

with unprecedented tissue contrast [4-5]. MRI is an effective
tool that provides detailed information about the targeted brain
tumor anatomy, which in turn enables effective diagnosis,
treatment and monitoring of the disease. Its techniques have
been optimized to provide measures of change within and
around primary and metastatic brain tumors, including edema,
deformation of volume and anatomic features within tumors,
etc [6]. MRI provides a digital representation of tissue
characteristic that can be obtained in any tissue plane. The
images produced by an MRI scanner are best described as
slices through the brain. MRI has the added advantage of
being able to produce images which slice through the brain in
both horizontal and vertical planes. This makes the MRI-scan
images an ideal source for detecting; identifying and
classifying the right infected regions of the brain.
Most of the current conventional diagnosis techniques are
based on human experience in interpreting the MRI-scan for
judgment; certainly this increases the possibility to false
detection and identification of the brain tumor. On the other
hand, applying digital image processing ensures the quick and
precise detection of the tumor [7]. One of the most effective
techniques to extract information from complex medical
images that has wide application in medical field is the
segmentation process [5, 8]. The main objective of the image
segmentation is to partition an image into mutually exclusive
and exhausted regions such that each region of interest is
spatially contiguous and the pixels within the region are
homogeneous with respect to a predefined criterion. Widely
used homogeneity criteria include values of intensity, texture,
color, range, surface normal and surface curvatures. Colorbased segmentation using K-means clustering for brain tumor
detection has been proposed, in which better results were
obtained using the developed algorithm than that in other edge
detection algorithms [9]. A modified method was proposed
that additionally takes into account the symmetry analysis and
any significant prior information of the region of interest as
well as the region area and edge information in the tumor
location of pathological cases [10]. However, all of these
research efforts pushed the limit of tumor detection accuracy,
they have been based on edge detection and were employed to

Abstract In this paper, modified image segmentation

techniques were applied on MRI scan images in order to detect
brain tumors. Also in this paper, a modified Probabilistic Neural
Network (PNN) model that is based on learning vector
quantization (LVQ) with image and data analysis and
manipulation techniques is proposed to carry out an automated
brain tumor classification using MRI-scans. The assessment of
the modified PNN classifier performance is measured in terms of
the training performance, classification accuracies and
computational time. The simulation results showed that the
modified PNN gives rapid and accurate classification compared
with the image processing and published conventional PNN
techniques. Simulation results also showed that the proposed
system out performs the corresponding PNN system presented in
[30], and successfully handle the process of brain tumor
classification in MRI image with 100% accuracy when the
spread value is equal to 1. These results also claim that the
proposed LVQ-based PNN system decreases the processing time
to approximately 79% compared with the conventional PNN
which makes it very promising in the field of in-vivo brain tumor
detection and identification.
Keywords Probabilistic Neural Network, Edge detection,
image segmentation, brain tumor detection and identification

I. INTRODUCTION
The National Cancer Institute (NCI) estimated that 22,070
new cases of brain and other central nervous system (CNS)
cancers would be diagnosed in the United States in 2009. The
American Brain Tumor Association (ABTA) clarifies this
statistic further by estimating that 62,930 new cases of
primary brain tumors would be diagnosed in 2010 [1-3].
Today, tools and methods to analyze tumors and their
behaviour are becoming more prevalent. Clearly, efforts over
the past century have yielded real advances; however, we have
also come to realize that gains in survival must be enhanced
by better diagnosis tools [1, 3]. Although we have yet to cure
brain tumors, clear steps forward have been taken toward
reaching this ultimate goal, more and more researchers have
incorporated measures into clinical trials each advance injects
hope to the team of caregivers and, more importantly, to those
who live with this diagnosis [1-3]. Magnetic Resonance
Imaging (MRI) is the state-of the-art medical imaging
1

International Journal of Image Processing and Visual Communication

ISSN (Online)2319-1724 : Volume 1 , Issue 2 , October 2012

introduced for edge-based segmentation, the widely accepted

segmentation methods are edge-image thresholding which is
used to eradicate insignificant edges that occur due to factors
such as noise and improper lighting conditions [13]. Edgeimage thresholding leads to stray edges in presence of noise
where the actual edges are often missing [11].
Stray edges problem can be solved if the edge properties
are determined with respect to the mutual neighbours, while
presence of edge is substantiated depending on the strength of
edges in local neighbourhood [11].
Region-based segmentation is then used which is based on
finding similarity measures to merge and split regions in an
image so as to form semantic or useful division in the
processed image. Self Organizing Map, SOM, as part of
competitive learning neural network (CLNN) has been used to
implement the vector quantization process [14-16]. The
importance of SOM for vector quantization is primarily due to
the similarity between the competitive learning process
employed in the SOM and the vector quantization procedure.
Neural units in the competitive layer need to be approximately
equal to the number of regions desired in the segmented image.
It is not however, possible to determine a priori the correct
number of regions in the segmented image. This is the main
limitation of the conventional SOM for image segmentation.
The HSOM directly addresses the aforesaid shortcomings of
the SOM. HSOM is the combination of self organization and
topographic mapping technique. HSOM combines the idea of
regarding the image segmentation process as one of data
abstraction where the segmented image is the final domain
independent abstraction of the input image [16].
In this paper we integrated the last three approaches and
enhanced their accuracy by combining both Gaussian and
Canny edge detection filters as discuss in the next sections.

filter out less relevant information while preserving the basic

structural properties of an image which significantly reduces
the amount of data to be processed in the subsequent steps
such as feature extraction, image segmentation, registration,
and interpretation. This is why with the recent developments.
On computational intelligence; the design of computerized
medical diagnosis systems has received more and more
attention. These reasons motivated us to propose two
automated diagnosis systems; the first system is completely
based on modified classical image processing algorithms,
while the second system is based on probabilistic artificial
neural network classifier to interpret medical images obtained
from clinical tests.
The rest of the paper is organized as follows. In section II,
conventional image segmentation techniques are summarized.
The proposed approach that includes image segmentation
techniques, filters and a modified edge detection algorithm is
presented in section III. The neural network model employed
for this research as well as the simulation results are discussed
in section IV. Section V presents the paper conclusions and
summary.
II. CONVENTIONAL IMAGE SEGMENTATION TECHNIQUES
Image segmentation plays a critical role in all advanced
image analysis applications, a key purpose of segmentation is
to divide image into regions and objects that correspond to
real world objects or areas, and the extent of subdivision
depends on requirements of specific application. Complete
segmentation of an image scene, where objects correlate with
real world objects, cannot be usually achieved without inputs
from the user or specific knowledge of the problem domain.
Image feature selection is a significant prerequisite for most
image processing algorithms, depending on these features the
segmentation methods can be classified into three categories
namely, thresholding, edge-based, region-based segmentation
and classifier such as Hierarchical Self Organizing Map
(HSOM) [11-12].
Image thresholding is the most popular segmentation
method due to its intuitive properties and simple
implementation [11]. Threshold selection plays a very crucial
role for efficient segmentation results. Intuitively, the
thresholds for multimodal histograms should be the minima
between the two maxima. Some techniques sharpen the
histogram peaks in image enhancement stage so as to facilitate
the threshold detection. The main disadvantage of this method
is the difficulty to separate object from background if the
object and background are of the same intensity distribution or
texture as in MRI-scans.
Edge-based segmentation is described in terms of
discontinuities in image attributes as Gray level, texture, color
etc. These discontinuities are known as edges and are detected
using edge detection operators, some of the commonly used
operators are Sobel, Prewitt, Laplace, etc [13]. Segmentation
resulting from edge-based method cannot be used as partial
segmented output due to the presence of broken, stray, or
noise edges. Advanced processing is required to obtain edges
corresponding to meaningful objects. Several algorithms

III. IMAGE PROCESSING PROPOSED APPROACH AND

SIMULATION RESULTS
A. Image Acquisition
In our proposed approach we first considered that the MRI
scan images of a given patient are either color, Gray-scale or
intensity images herein are displayed with a default size of
220220. If it is color image, a Gray-scale converted image is
defined by using a large matrix whose entries are numerical
values between 0 and 255, where 0 corresponds to black and
255 to white for instance. Then the brain tumor detection of a
given patient consist of two main stages namely, image
segmentation and edge detection.
B. Image Segmentation
The objective of image segmentation is to cluster pixels
into prominent image region. In this paper, segmentation of
Gray level images is used to provide information such as
anatomical structure and identifying the Region of Interest i.e.
locate tumor, lesion and other abnormalities. The proposed
approach is based on the information of anatomical structure
of the healthy parts and compares it with the infected parts. It
starts by allocating the anatomical structure of the healthy
parts in a reference image of a normal candidate brain scan
2

International Journal of Image Processing and Visual Communication

ISSN (Online)2319-1724 : Volume 1 , Issue 2 , October 2012
image as shown in Fig. 1 then it allocates the abnormal parts
in the unhealthy patient brain. scan image by comparing it
with the reference image information as shown in Fig. 2.

3) Smoothing using Gaussian filter

In this paper, the proposed Gaussian smoothing filter, Gf, is a
nonnegative, real-valued column matrix defined by,
G f (x , y ) =

Figure 1: Normal Brain

x 2 + y 2
1
exp

2 2
c

(4)

in which c is expressed as c = 2 2 .
However this type of filters enhanced the noise reduction
level compared with the linear filters, it was observed that
these smoothing and noise filters did not completely satisfy
the noise removal level from the original image as shown in
Fig. 4. Thus, for these applications a set of cascaded filters are
recommended. We therefore proposed another stage of noise
filtering by using an average filter.

Figure 2: Abnormal Brain

1) Enhancement and Smoothing: There are different types of

noise encountered by different techniques, depending on the
noise nature and characteristics, namely Gaussian noise and
impulse noise. In this paper we assumed that the main image
noise is additive and random; that is spurious and random
signal, n(i, j), added to the true pixel value I(i, j):
II ( i , j ) = I ( i , j ) + n ( i , j )
(1)
In this algorithm the enhancement in spatial domain is
based on direct manipulation of pixels in a small
neighbourhood of pixels, it generally takes the form;
(2)
g (x , y ) = T [ f (x , y )]
in which f(x, y) is the input image, g(x, y) is the processed
image, and T is an operator on f, defined over some
neighbourhood of (x, y). Then we applied the next
enhancement in frequency domain which is based on the
concept of the convolution theorem and spatial filters. In this
paper, the proposed noise enhancement algorithm is based on
using spatial filters and includes the following:
Smoothing filters that are used to reduce or remove
Gaussian noise from the MRI image.
Sharpening filters that are used for highlighting edges
in an image, and are based on the use of first and
second order derivatives.

Figure 3: Applying Gaussian filter

Figure 4: applying Average Filter

Applying the average filter resulted in an acceptable noise

reduction level for such applications. The conclusion from this
part is cascaded filter array is recommended to reach an
acceptable noise reduction levels brain tumor detection.

C. Edge Detection
An edge is a property attached to an individual pixel and
is calculated from the image function behaviour in a
neighbourhood of the pixel. It is also considered as a vector
variable (magnitude of the gradient, direction of an edge). The
2) Smoothing by Linear filter: Linear operations calculate the purpose of edge detection in general is to significantly reduce
resulting value in the output image pixel IA(i,j) as a linear the amount of data in an image, while preserving the structural
combination of brightness in a local neighbourhood of the properties to be used for further image processing.
In this paper, other than filtering the region of interest
pixel I(i,j) in the input image. In this algorithm we assumed I
as an NM image, m is an odd number smaller than both N (ROI) is proposed to identify different tumor types and/or
and M, and A is the convolution kernel or the filter mask of different infected areas. It also introduced to enhance the
the linear filter that is an mm mask. The filtered version of I processing time by executing the features processing
algorithm in the identified areas instead of the whole image
is given by the discrete convolution as follows:
frame. In this research, we first applied a vector subtraction
n /2
m /2
algorithm then the ROI is determined by finding the related
I A (i , j ) = I * A =
A (h, k )I (i h, j k )
(3)
h = n /2 k = m /2
adjacent portions in the resultant image from the vector
Where i=1 to N and j=1 to M. This filter replaces the value subtraction. The area of each related adjacent portion is
I(i, j) with a weighted sum of I values in a neighbourhood of (i, computed and the irrelevant portions removed resulting in the
j). If all entries of A in Eq. (3) are non-negative, the filter desired tumor region as shown in Fig. 6.
performs average smoothing.
To enhance the results of the proposed edge detection
Then the matrix of the abnormal brain scan image is algorithm we found that the most important criteria that affect
subtracted from that of the normal brain image resulting in a the edge detection performance are by reducing the rate error
matrix of the region of interest accompanied with some noise of losing edges in an image and that edge points must be well
as illustrated in Fig. 3.
localized. Therefore, we successfully modeled and

International Journal of Image Processing and Visual Communication

ISSN (Online)2319-1724 : Volume 1 , Issue 2 , October 2012
implemented Cannys mathematical formulas [17] to increase
the performance of the proposed edge detection algorithm.
Even though it is quite old, it has become one of the standard
edge detection methods and it is still used in research [18].

where gx and gy are the gradients in the x- and y-directions

respectively and represents the results of the original image
filtered along rows and lines. is the gradient direction.
3. Non-maximum suppression: Only local maxima should
be marked as edges. If the gradient amplitude of the
pixel is no less than the gradient amplitude between two
adjacent pixels in the gradient direction, the point can be
judged as the edge point possibly. The purpose of this
step is to convert the blurred edges in the image of the
gradient magnitudes to sharp edges. Basically this is
done by preserving all local maxima in the gradient
image, and deleting everything else. The algorithm is for
each pixel in the gradient image.
a. Round the gradient direction to nearest 45 ,
corresponding to the use of an 8-connected
neighbourhood.
b. Compare the edge strength of the current pixel with
the edge strength of the pixel in the positive and
negative gradient direction. I.e. if the gradient

D. Canny Edge detection

The Canny algorithm can be used an optimal edge detector
based on a set of criteria which include finding the most edges
by minimizing the error rate, marking edges as closely as
possible to the actual edges to maximize localization, and
marking edges only once when a single edge exists for
minimal response. According to Canny [29], the optimal filter
that meets all three criteria above can be efficiently
approximated using the first derivative of a Gaussian function
in Eq. (4). These derivatives are used to calculate gradient
magnitude (edge strength) and gradient direction of most rapid
change in intensity.
E. The Modified Canny Edge Detection Algorithm
The algorithm runs in 5 separate steps as shown in Fig. and
described as follows:

direction is north ( =90), compare with the pixels

to the north and south.
c. If the edge strength of the current pixel is largest;
preserve the value of the edge strength. If not,
suppress (i.e. remove) the value.
4. Double thresholding: Potential edges are determined by
thresholding. Edge pixels stronger than the high
threshold are marked as strong; edge pixels weaker than
the low threshold are suppressed and edge pixels
between the two thresholds are marked as weak.
5. Edge tracking by hysteresis: Final edges are determined
by suppressing all edges that are not connected to a very
certain (strong) edge. Strong edges are interpreted as
certain edges, and can immediately be included in the
final edge image. Weak edges are included if and only if
they are connected to strong edges. The logic is of course
that noise and other small variations are unlikely to result
in a strong edge (with proper adjustment of the threshold
levels). Thus strong edges will (almost) only be due to
true edges in the original image. The weak edges can
either be due to true edges or noise/color variations. The
latter type will probably be distributed independently of
edges on the entire image, and thus only a small amount
will be located adjacent to strong edges. Weak edges due
to true edges are much more likely to be connected
directly to strong edges. In this paper edge tracking is
implemented by iterative BLOB-analysis (Binary Large
OBject). The edge pixels are divided into connected
BLOBs using 8-connected neighbourhood. BLOBs
containing at least one strong edge pixel is then
preserved, while other BLOBs are suppressed.

Figure 5: Modified Canny edge detection algorithm

1. Smoothing: Blurring of the image to remove noise.

Therefore the image is first smoothed by applying a
Gaussian filter. Our proposed method uses 55
Gaussian template and the original image to weight
neighbourhood. Denote any point (x, y) of the image as
the center when processing and extracting 55
neighbourhood, the weighting neighbourhood can be
indicated as follows:
I A ( x, y ) =

1 2 2
I ( x + i, y + j ) M ( 2 + i, 2 + j )
5 5 i=2 j=2

(5)

where x=1,2,,m; y=1,2,,n , I(x,y) is the pixel value

of the original sub-image, M is the Gaussian template,
and IA(x,y) is the pixel value of the smoothed image.
2. Finding gradients: The edges should be marked where
the gradients of the image has large magnitudes (edge
strength). In this step we compute gradient direction
and amplitude of smoothed image IA(x,y) adopting
first order partial finite difference of 22 neighborhood.
M ( x, y ) =

g x2 ( x, y ) + g y2 ( x , y )

(6)

= arctan ( gy ( x, y ) gx ( x, y ) )

(7)

fx = 2
1
2

(8)

1
2,

1
2

1
fy =
2

1
2

F. Simulation Results
Proposed algorithm is implemented using MATLAB where
the source image and the thresholds can be chosen arbitrarily
and the implementation uses the correct Euclidean measure
for the edge strengths. Simulation results after applying

International Journal of Image Processing and Visual Communication

ISSN (Online)2319-1724 : Volume 1 , Issue 2 , October 2012
Canny-based edge detection algorithm on MRI scan images is linked with one of K classes. The transpose of the matrix
showed the ability of the proposed algorithm to accurately created from the Q training pairs, P`, determines the first layer
detect and identify the contour of the tumor as shown in Fig. 7. input weights, IW1,1. With each input introduced, the ||dist||
box produces a vector whose elements indicate the proximity
of the input to the vectors of the training set. These elements
are then multiplied, element by element, by the bias and sent
to the radbas transfer function given by,

radbas(n) = e n

Figure 6: Region of Interest

(9)

An input vector close to a training vector is represented by

a number close to 1 in the output vector a1. If an input is close
to several training vectors of a single class, it is represented by
several elements of a1 that are close to 1.
The matrix T of the target vectors however determines the
second-layer weights, IW2,1. Each vector contains a single 1
only in the row associated with that specific class of input, and
0's elsewhere. The multiplication Ta1 then sums the elements
of a1 resulting from each of the K input classes. Finally, the
second-layer transfer function, compete, produces a 1
corresponding to the largest element of n2, and 0's elsewhere.
Thus, the network classifies the input vector into a specific K
class since that class has the maximum probability of being
correct [29].

Figure 7: Canny Edge Detection

IV. PROBABILISTIC NEURAL NETWORK PROPOSED APPROACH

AND SIMULATION RESULTS
Recently there has been an increase of activities in the
application of neural networks to medical imaging [19-26].
The purpose is to make use of the parallel distributed
processing nature of neural networks to reduce computing
time and enhance the classification accuracy. Applications of
neural network to pattern classification have been extensively
studied in the past many years.
Various kinds of neural-network architecture including
multilayer perceptron (MLP) neural network, radial basis
function (RBF) neural network, self-organizing map (SOM)
neural network, and probabilistic neural network (PNN) have
been proposed [19]. An inherent statistical foundation in
Bayesian estimation theory and its ease of training make PNN
an effective tool for solving many classification problems [2026]. However, it requires a very large neural network to
analyze an entire image with huge number of interconnected
networks and its associated network size, the locations of
pattern layer neurons as well as the value of the smoothing
parameter.
A PNN is predominantly a classifier since it can map any
input pattern to a number of classifications. PNN is a fast
training process and an inherently parallel structure that is
guaranteed to converge to an optimal classifier as the size of
the representative training set increases and training samples
can be added or removed without extensive retraining. A
consequence of a large network structure is that the classifier
tends to be oversensitive to the training data and is likely to
exhibit poor generalization capacities to the unseen data [27].
The second problem is related to the smoothing parameter
which also plays a crucial role in PNN classifier, and an
appropriate smoothing parameter is often data dependent.
These two problems have been solved by other researchers
such as in [25-27]. In this paper, Probabilistic Neural Network
is used to classify a brain tumor in an MRI image according to
its proximity to the most relevant training vector. The PNN
network consists of three layers respectively: input layer,
pattern layer and competitive layer as shown in Fig 8. It is
presumed that there are Q input vector/target vector pairs
(number of neurons in layer 1) where each target vector has
number of classes K (number of neurons in layer 2). One of
these elements is 1 and the rest are 0. Thus, each input vector

Figure 8: Probabilistic Neural Network Architecture

A research trial as presented in [30] implemented an

automated brain tumor classification system using PNN. Their
data set images were divided into 20 training images and 15
testing. The Principle component analysis was applied on
MRI images to extract the features of the images. In this
research, neither image enhancement nor segmentation was
applied.
A. The Proposed System
In this paper, the proposed system is a modified version of
the conventional PNN. The modification is based on
automatic utilization of specified regions of interest (ROIs)
within the tumor area in the MRI images. Form each ROI, set
of extracted features include tumor shape and intensity
characteristics are extracted and normalized. Each ROI is then
given a weight to estimate the PDF of each brain tumor in the
MR image. These weights are used as a modeling process to
modify the conventional PNN. This method is based on
learning vector quantization (LVQ) which is a supervised
competitive learning technique that obtains decision
boundaries in input space based on training sets to reduce the
size of the hidden layer. It defines class boundaries prototypes,
a nearest-neighbour rule and a winner-takes-it-all paradigm.

International Journal of Image Processing and Visual Communication

ISSN (Online)2319-1724 : Volume 1 , Issue 2 , October 2012
LVQ is composed of three layers: input layer, competitive
layer and output layer. The input data is classified in the
competitive layer and those classes or patterns are mapped to
target class in the output layer. In the learning phase weights
of neurons are adjusted based on training data. The winner
neuron is calculated based on the Euclidean distance, and then
the weight of the winner neuron is adjusted.
B. Methodology
There are four major steps in the proposed approach for
brain tumor classification. The first step is ROI segmentation
in which the boundary of the tumor (ROI) in an MR image is
identified; the second step is the feature extraction of the
meaningful features of the ROI; the third step is the feature
selection; the last step is the classification process in which
learning a classification model using the features.
The proposed algorithm starts by reading the input image,
converting it to grey scale image then applying image
segmentation techniques for extracting the Region of Interest
(ROI). A set of reference MRIs is taken as the training
database. Feature vectors are extracted for each image in the
training set during the training phase. In the testing phase, the
feature vector of the test image is computed. Figure 9
illustrates the sequence of the proposed approach.
The proposed approach is evaluated on real images, and the
results are compared with other algorithms, in particular
conventional PNN algorithm presented by [30].
During the segmentation process, each image region
confined by a rectangular window is represented by a feature
vector of length R. These vectors computed for Q selected
regions are organized in the pattern matrix PR,Q and form
clusters in the R-dimensional space. The Q pattern vectors in
P are fed into the input NN layer, while the number C of the
output layer elements represents the desired number of
segmentation classes [31].

Figure 9: The Proposed PNN-Based Brain Tumor Classification System

The algorithm comprises of the following successive steps:

1. Feature vectors computation to create the feature
matrix P using the sliding window.
2. Initialization of the learning process coefficients and
the network weights matrix W.

3. Iterative application of the competitive process and the

Kohonen learning rule [32] for all feature vectors
during the learning stage.
4. NN simulation to assign class numbers to individual
feature vectors.
5. Evaluation of the regions classification results.
C. Model learning
Probabilistic Neural Networks (PNN) is a supervised feed
forward networks derived from Bayesian Decision Networks.
Typically PNN is built with four layers. Training and testing
vectors are normalized prior to input into the network. First
layer consists of neurons for each input feature. The second
layer has one neuron (Pattern unit) for each pattern in the
training dataset. Each neuron in the pattern unit computes the
dot product of training pattern and testing pattern then
performs Gaussian transformation.
The proposed architecture for PNN is based on using fmean instead of dot product of training pattern with the testing
pattern, which results in reduced computational time to
classify an unknown pattern. The modified PNN comprises
only three layers instead of four. First layer consists of
neurons for each input feature. The second layer has one
neuron for each class of the training dataset. Each neuron in
the second layer computes the f-mean between input vector
and the weight vector. The third layer is the decision layer
which compares values passed from the neurons of second
layer. The neuron of the second layer at which maximum
value occurs decides the class of the input pattern. In
conventional probabilistic neural network, the entire training
set must be stored and used during testing and the amount of
computation necessary to classify an unknown point is
proportional to the size of the training set. But in the modified
PNN, second layer contains one neuron for each class of the
training dataset leading to much less computational time for
classification. A three layer neural network was created with
500 nodes in the first (input) layer, 1 to 50 nodes in the hidden
layer, and 1 node as the output layer. We varied the number of
nodes in the hidden layer in a simulation in order to determine
the optimal number of hidden nodes. But in the final
simulation experiment we elected only ten nodes in the hidden
layer. The 500 data points extracted from each subject were
then used as inputs of the neural networks. The output node
resulted in either a 0 or 1, for control or patient data
respectively. The weights in the hidden node needed to be set
using training sets. Therefore, subjects were divided into
training and testing datasets.
D. Experimental Results
A set of MRI-scan Gray-scale image database was used in
this experiment each image size is 220220 pixels. A group of
64 MRI images were used that were categorized into 6 classes
respectively. Out of the 64 subjects a group of 18 random
patients MRI images were selected as a test set, while the rest
of the dataset was used for training. Training data was used to
feed into the neural networks as inputs and then knowing the
output, the weights of the hidden nodes were calculated. Many
trials were performed on the same Neural Network, selecting

International Journal of Image Processing and Visual Communication

ISSN (Online)2319-1724 : Volume 1 , Issue 2 , October 2012
18 subjects randomly every time for testing and the remaining
subjects for retraining to find accuracy of neural network
prediction. Table I shows the network performance results
compared with the results presented in [30]. The results
presented in this table show that the proposed system out
perform the presented system in [30], and successfully handle
the process of MRI image classification with 100% accuracy
when the spread value is equal to 1. It is also noted that the
proposed LVQ-based PNN system decreased the processing
time to approximately 79% compared with the conventional
PNN.
TABLE I
COMPARISON OF THE CLASSIFICATION PERFORMANCE USING OUR PROPOSED
MODEL WITH THAT PRESENTED IN [30].

Results presented
in [30]
Number of training
Images
Number of testing
Images
Spread Value
Network performance
for correct images

Our proposed
system

20

64

15

18

73%

100%

of the hidden nodes were calculated. Many trials were

performed on the same Neural Network, selecting 18 subjects
randomly every time for testing and the remaining subjects for
retraining to find accuracy of neural network prediction.
Simulation results showed that the proposed system out
perform the presented system in [30], and successfully handle
the process of MRI image classification with 100% accuracy
when the spread value is equal to 1. It was also concluded that
the proposed LVQ-based PNN system decreased the
processing time to approximately 79% compared with the
conventional PNN and despite considerable progress in
probabilistic neural networks, there has been a room for
improvement as far as network structure determination is
concerned.
REFERENCES
[1]
[2]

V. CONCLUSION
In this paper, we proposed two approaches for Brain tumor
detection, identification and classification. The first approach
is based on an integrated set of image processing algorithms,
while the other is based on a modified and improved
probabilistic artificial neural networks structure.
The proposed integrated image processing algorithm is
based on a modified Canny edge detection algorithm and
implemented using MATLAB. However, simulation results
using this algorithm showed its ability to accurately detect and
identify the contour of the tumor, its computational time and
accuracy were much less than its corresponding algorithms
that use the parallel distributed processing nature of neural
networks to reduce computing time and enhance the
classification accuracy. This leaded us to propose a modified
and improved probabilistic artificial neural networks structure.
The modification is based on automatic utilization of
specified regions of interest (ROIs) within the tumor area in
the MRI images. Form each ROI, set of extracted features
include tumor shape and intensity characteristics are extracted
and normalized. Each ROI is then given a weight to estimate
the PDF of each brain tumor in the MR image. These weights
are used as a modeling process to modify the conventional
PNN. This method is based on learning vector quantization
(LVQ) which is a supervised competitive learning technique.
This model is successfully tested by using a set of infected
brain MRI-scan images to classify brain tumor. In our
experiments, a database of 64 MRI-scan Gray-scale image
was used, each image size is 220220 pixels. Out of the 64
subjects a group of 18 random patients MRI images were
selected as a test set, while the rest of the dataset was used for
training. Training data was used to feed into the neural
networks as inputs and then knowing the output, the weights

[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]

[11]
[12]

[13]
[14]
[15]

[16]
[17]

American Brain Tumor Association. (2010). Facts and statistics, 2010.

Retrieved
from
http://www.abta.org/sitefiles/pdflibrary/ABTAFactsandStatistics2010.pdf
Central Brain Tumor Registry of the United States. (2010). CBTRUS
statistical report:Primary brain and central nervous system tumors
diagnosed in the United States in 20042006. Retrieved from
http://www.cbtrus.org/2010-NPCR-SEER/CBTRUS
-WEBREPORT-Final-3-2-10.pdf
Sidney Croul, Jessica Otte and Kamel Khalili, Brain tumors and
polyomaviruses, Journal of NeuroVirology, 9: 173182, 2003
S. M. Bhandarkar and P. Nammalwar, Segmentation of Multispectral
MR images Using Hierarchical Self-Organizing Map, Proceedings of
Computer-Based medical system CBMS 2001.
C. A. Parra, K. Iftekharuddin and R. Kozma, Automated Brain Tumor
segmentation and pattern Recoginition using ANN, Computional
Intelligence Robotics and Autonomous Systems, 2003.
Andreas
Rimner, Andrei I. Holodny and Fred H. Hochberg,
Perfusion Magnetic Resonance Imaging to Assess Brain Tumor
Responses to New Therapies, US neurological disease, 2006.
V. J. Nagalkar and S. S. Asole, Brain tumor detection using digital
image processing based on soft computing, Journal of Signal and
Image Processing , Vol. 3,No. 3, pp.-102-105, 2012
J. Alirezaie, M. E. Jernigan and C. Nahmias, Neural Network based
segmentation of Magnetic Resonance Images of the Brain, IEEE
Trans, Nuclear Science, Vol.18, No.2, pp:7-30, 2002.
S. Datta and M Chakraborty, Brain Tumor Detection from PreProcessed MR Images using Segmentation Techniques, CCSN, 2011.
Sudipta Roy and Samir K. Bandyopadhyay, Detection and
Quantification of Brain Tumor from MRI of Brain and its Symmetric
Analysis, International Journal of Information and Communication
Technology Research, Vol. 2, No. 6, June 2012.
D. Shevad, Multiple object segmentation and tracking in image
sequence, Lamar University -Beaumont, April 2005.
T. Logeswari and M. Karnan, An Improved Implementation of Brain
Tumor Detection Using Segmentation Based on Hierarchical Self
Organizing Map, International Journal of Computer Theory and
Engineering, Vol. 2, No. 4, August, 2010.
R. Gonzalez and R. Woods, Digital Image Processing, 3rd Edition.
Prentice Hall, 2008.
Alexander Statnikov, Automatic cancer diagnostic decision support
system for gene expression domain, Thesis, August, 2005.
M. M. Ibrahiem, E Emary and S. Ramakrishnan, On the Application
of Various Probabilistic Neural Networks in Solving Different Pattern
Classification Problems, World Applied Sciences Journal, Vol. 4 No.
6, pp. 772-780, 2008.
S. Asif Hussain and M. Praveen Raju, Neuro-Fuzzy System for
Medical Image Processing Proceedings of the Inter. Conf. on Comm.
and Computational Intellig., pp.382-385, India. Dec., 2010.
John Canny, A Computational Approach to Edge Detection, IEEE
Transactions On Pattern Analysis And Machine Intelligence, Vol.
Pami-8, No. 6, November 1986.

International Journal of Image Processing and Visual Communication

ISSN (Online)2319-1724 : Volume 1 , Issue 2 , October 2012
[18]
[19]
[20]
[21]

[22]

[23]
[24]

F. Mai, Y. Hung, H. Zhong, and W. Sze. A hierarchical approach for

fast and robust ellipse extraction, Pattern Recognition, Vol. 41, No. 8,
pp.25122524, August 2008.
K. Z. Mao, K.-C. Tan, and W. Ser, Probabilistic Neural-Network
Structure Determination for Pattern Classification, IEEE Transactions
on Neural Networks, Vol. 11, No. 4, JULY 2000.
B. D. Sekar, Ming Chui Dong, Jun Shi and Xiang Yang Hu, "Fused
Hierarchical Neural Networks for Cardiovascular Disease Diagnosis,"
IEEE Sensors Journal, Vol.12, No.3, pp.644-650, March 2012.
Huafen Yang and You Yang, "An Improved Probabilistic Neural
Network with GA Optimization," Fifth International Conference on
Intelligent Computation Technology and Automation (ICICTA), pp.7679, 12-14 Jan. 2012.
Sarma Mousmita and Kandarpa Kumar, "Segmentation of Assamese
phonemes using SOM," NCETACS, pp.121-125, March 2012J.
Breckling, Ed., The Analysis of Directional Time Series: Applications
to Wind Speed and Direction, ser. Lecture Notes in Statistics. Berlin,
Germany: Springer, 1989, vol. 61.
M. Zhang, K. E. Sakaieand and S. E. Jones, "Toward whole-brain
maps of neural connections: Logical framework and fast
implementation," IEEE MMBIA, pp.193-197, 9-10 Jan. 2012.
P. T. K Shri, and N. Sriraam, "EEG based detection of alcoholics using
spectral entropy with neural network classifiers," ICoBE, pp.89-93, 2728 Feb. 2012.

[25]
[26]

[27]
[28]
[29]
[30]
[31]

[32]

S. S. Kumar, R. S. Moniand J. Rajeesh, "Liver tumor diagnosis by

gray level and contourlet coefficients texture analysis," ICCEET,
pp.557-562, 21-22 March 2012.
T. Halgaswaththa, A. S. Atukorale, M. Jayawardena and J.
Weerasena,"Neural
network
based
phylogenetic
analysis,"
International Conference on Biomedical Engineering (ICoBE), pp.155160, 27-28 Feb. 2012.
C. M. Bishop, Neural Networks for Pattern Recognition, New York,
Oxford Univ. Press, 1995.
A. Padma and R.Sukanesh, A Wavelet Based Automatic
Segmentation of Brain Tumor in CT Images Using Optimal Statistical
Texture Features, IJIP, Vol. 5, No. 5, 2011.
Howard Demuth and Mark Beale, Neural Network Toolbox Users
Guide, 2000.
Mohd Fauzi Othman, Mohd Ariffanan and Mohd Basri, Probabilistic
Neural Network for Brain Tumor Classification, 2nd International
Conference on Intelligent Systems, Modelling and Simulation, 2011.
Kailash D. Kharat, Pradyumna P. Kulkarni and M. B. Nagori, Brain
Tumor Classification Using Neural Network Based Methods,
International Journal of Computer Science and Informatics, pp 22315292, Vol. 1, No. 4, 2012.
G. M. Josin and P. F. Liddle, Neural Network Analysis of the Pattern
of Functional Connectivity between Cerebral Areas in Schizophrenia.,
Biological Cybernetics, Vol. 84, No. pp. 117-122, Feb 2001.