J Vis (2013) 16:1928

DOI 10.1007/s12650-012-0153-y

R E G UL A R P A P E R

A. Padma Nanthagopal R. Sukanesh Rajamony

Classification of benign and malignant brain tumor

CT images using wavelet texture parameters
and neural network classifier
Received: 6 September 2011 / Revised: 26 May 2012 / Accepted: 22 October 2012 / Published online: 7 December 2012
 The Visualization Society of Japan 2012

Abstract Computational methods are useful for medical diagnosis because they provide additional information that cannot be obtained by simple visual interpretation. As a result, an enormous amount of computer
vision research effort has been targeted at achieving automated medical image analysis. In this paper, we
present the combination of wavelet statistical texture features (WST) obtained from two-level discrete
wavelet-transformed (DWT) images and wavelet co-occurrence texture features (WCT) obtained from twolevel DWT detail images for the classification of abnormal brain tissues into benign, malignant tumor of CT
images. Our proposed system consists of four phases: (1) segmentation of region of interest, (2) discrete
wavelet decomposition, (3) feature extraction and feature selection, and (4) classification and evaluation.
The support vector machine is employed to segment the shape of tumor information. A combination of both
WST and WCT texture features is extracted from tumor region of two-level discrete wavelet-transformed
images. Genetic algorithm (GA) is used to select the optimal texture features from the set of extracted
features. The probabilistic neural network classifier (PNN) is built to classify the abnormal brain tissues into
benign, malignant tumor images and evaluate the performance of classifier by comparing the classification
results of the PNN classifier with linear vector quantization (LVQ) neural network classifier, back propagation neural network (BPN) classifier. The results of PNN, LVQ, BPN classifiers for the texture analysis
methods are evaluated using statistical analysis and receiver operating characteristic analysis. From the
experimental results, it is inferred that the best classification performance is achieved by PNN than LVQ and
BPN classifiers. The system has been tested with real data of 80 benign, malignant CT brain tumor images
and has achieved satisfactory results.
Keywords Computed tomography (CT)  Discrete wavelet transform (DWT)  Receiver operating
characteristic (ROC) analysis  Support vector machine (SVM)  Probabilistic neural network (PNN)

1 Introduction
Computed tomography (CT), sometimes called CAT scan, uses special X-ray equipment to obtain many
images from different angles and join them together to produce multiple cross-sectional images of the head.
CT scanning provides more detailed information on head injuries, stroke, brain tumors, and other brain
diseases. In recent years, medical CT images have been applied in clinical diagnosis widely. That can assist
physicians to detect and locate pathological changes with more accuracy. CT images can be distinguished
A. Padma Nanthagopal (&)
Trichy Anna University, Trichy 625 023, India
E-mail: giri.padma2000@gmail.com
R. Sukanesh Rajamony
Electronics and Communication Engineering, Thiagarajar College of Engineering, Madurai 625 015, India

20

A. Padma Nanthagopal, R. Sukanesh Rajamony

for different tissues according to their different gray levels. The images if processed appropriately can offer
a wealth of information which is significant to assist doctors in medical diagnosis. A lot of research efforts
have been directed toward the field of medical image analysis with the aim to assist in diagnosis and clinical
studies (e.g., Duncan and Ayache 2000). Pathologies are clearly identified using automated CAD system
(e.g., Tourassi 1999). It also helps the radiologist in analyzing the digital images to bring out the possible
outcomes of the diseases. The medical images are obtained from different imaging systems such as magnetic
resonance imaging (MRI), CT imaging and ultra sound imaging. The CT has been found to be the most
reliable method for early detection of tumors because this modality is the mostly used in radio therapy
planning for two main reasons. The first reason is that scanner images contain anatomical information which
offers the possibility to plan the direction and the entry points of radio therapy rays which have to target only
the tumor region and to avoid other organs. The second reason is that CT scan images are obtained using
rays, which is same principle as radio therapy. This is very important because the intensity of radio therapy
rays has been computed from the scanned image. Advantages of using CT include good detection of
calcification, hemorrhage and bony detail plus lower cost, short imaging times and widespread availability.
The situations include patient who are too large for MRI scanner, claustrophobic patients, patients with
metallic or electrical implant and patients unable to remain motionless for the duration of the examination
due to age, pain or medical condition. For these reasons, this study aims to explore methods for classifying
the abnormal brain tissues into benign, malignant tumors in CT images. Image classification is the process of
assigning a set of pixels with similar properties. Accurate, fast and reproducible classification techniques are
required in various applications. Brain tumor is any mass that results from abnormal growths of cells in the
brain. It may affect any person in almost in any stage. Brain tumors can have a variety of shapes and sizes; it
can appear any location and in different image intensities. Brain tumors can be benign, malignant. Lowgrade gliomas and meningiomas which are benign tumors represent the most common type of tumor.
Glioblastoma multiform is a malignant tumor and represents the most common brain neoplasm. Benign
brain tumor has a homogeneous structure and does not contain cancer cells. They can be either simply be
monitored radio logically or surgically eradicated and they seldom grow back. Malignant brain tumors have
a heterogeneous structure and contain cancer cells. They can be treated by radiotherapy, chemotherapy or a
combination of both.
Many techniques have been reported for the classification of brain tumors in MR images most notably,
support vector machine classifier (SVM) (e.g., Shebang and Hiremath 2009), neural network classifiers (e.g.,
Reddick et al. 2009) and knowledge-based techniques (e.g., Clark et al. 1998), the SVM classifier for the
diagnosis of breast cancer in ultrasonic images (e.g., Chang et al. 2003a, b), method for automated segmentation and classification of brain MR images in which an SVM classifier (e.g., Chang et al. 2003a, b)
was used to classify normal and abnormal images with statistical features, SVM classifier was used for the
classification of MRI image and also shows that classification rate of SVM was higher when compared to
neural networks self-organizing maps (e.g., Chaplet et al. 2006), EM algorithm to detect the abnormalities
but this method (e.g., Gering et al. 2002) requires high computational effort. The knowledge-based techniques are also proposed to make more efficient segmentation and classification results, but these techniques
required more intensive training (Fig. 1).

Fig. 1 Segmentation process of SVM classifier

Classification of benign and malignant brain tumor

21

In medical image analysis, the determination of tissue type (normal or pathological) and classification of
tissue pathology are performed by using texture analysis. To solve the texture classification problem, many
approaches have been developed such as multi-channels methods (e.g., Schad et al. 1993), multi-resolution
analysis (e.g., Freeborough and Fox 1998), optimal Gabor filters (e.g., Dunn et al. 1995) for texture
segmentation, tree structured wavelet transform for texture classification (e.g., Chang and Kuo 1993). Gabor
filters are poor due to their lack of orthogonality that results in redundant features at different scales while
wavelet transform is capable of representing texture at more suitable scale, by varying the spatial resolution,
and there is also a wide range of choices for the wavelet function. Most classification techniques methods
offer intensity-based statistical features. However, in our approach, we adopt wavelet-based statistical
texture features to classify the abnormal brain tissues.
2 System overview
The proposed system is divided into four phases: (1) segmentation of region of interest (ROI), (2) image
decomposition using DWT, (3) feature extraction and feature selection and (4) classification and evaluation.
The SVM is used to segment the shape of tumor information, and the two-dimensional DWT at 2nd level is
performed on the ROI of the tumor image to remove the noise. The combination of WST features obtained
from two-level discrete wavelet-transformed images and WCT features obtained from two-level discrete
wavelet-transformed detail images is presented. The extracted texture features are optimized by genetic
algorithm (GA) (e.g., Brill et al. 1992). The optimal texture features are fed to PNN (e.g., Yang 2011), LVQ
(e.g., Tang et al. 2007), BPN (e.g., Fausett 1994) classifiers to classify abnormal brain tissues into benign,
malignant tumors.
2.1 Segmentation of ROI
In our research, we choose the SVM (e.g., El-Naqa et al. 2002) classifier to segment the shape of tumor
information. SVM performs the robust nonlinear classification with kernel trick. SVM is independent of the
dimensionality of the feature space and that the results obtained are very accurate. It outperforms other
classifiers even with small numbers of available training samples. SVM is a supervised learning method and
is used for one-class and n-class classification problems. It combines linear algorithms with linear or
nonlinear kernel functions that make it a powerful tool in the machine learning community with applications
such as data mining and medical imaging applications. To apply SVM into nonlinear data distributions, the
data can be implicitly transformed to a high-dimensional feature space where a linear separation might
become possible. In this study, we chose a linear function.
In order to build the segmentation model, feature vectors of tumor and nontumor area are extracted. Five
rectangles inside the tumor area and nontumor area are selected to train the SVM model. For each selected
rectangle, four vertices and one center point of the rectangle are concerned. For each point of rectangle, we
use two features of position and intensity of selected point to form the feature vector.
1. Position = (xi,yi), i = 1, , n
2. Intensity = Ii (xi, yi), i = 1, , n, n represents the number of training vectors.
We select five rectangles from tumor and nontumor area. So, we select totally 25 feature sets from tumor
and 25 feature sets from nontumor area, respectively. These 50 feature vectors or feature set are given as
input to the SVM classifier to segment the tumor shape.
2.2 Discrete wavelet decomposition using DWT
Daubechies wavelet filter of order two is used and found to yield good results in classification of the brain
tumor CT images. A two-level DWT of region of interest (ROI) is performed which results in four subbands. In 2DDWT (e.g., Van et al. 1999), the image is represented by one approximation and three detail
images representing the low- and high-frequency contents image, respectively. The approximation can be
further to produce one approximation and three detail images at the next level of decomposition, and
wavelet decomposition process is shown in Fig. 2. Figure 3 shows the wavelet decomposed tumor images at
2nd level. A1 and A2 represent the wavelet approximations at 1st and 2nd level, respectively, and are low-

22

A. Padma Nanthagopal, R. Sukanesh Rajamony

Fig. 2 Two-level discrete wavelet decomposition

Fig. 3 Discrete wavelet decomposed tumor images at 2nd level

frequency part of the images. H1, V1, D1, H2, V2, D2 represent the details of horizontal, vertical and
diagonal directions at 1st and 2nd level, respectively, and are high-frequency part of the images.
Among the high-frequency sub-bands, the one whose histogram presents the maximum variance is the
sub-band that represents the clearest appearance of the changes between the different textures. The combination of WST features extracted from 2nd level of both low- and high-frequency sub-bands and WCT
features extracted from 2nd level of high-frequency sub-bands is useful to classify abnormal brain tissues
into benign, malignant tumors.
2.3 Feature extraction and feature selection
Texture analysis is a quantitative method that can be used to quantify and detect structural abnormalities in
different tissues. As the tissues present in brain are difficult to classify using shape or intensity level of
information, the texture feature extraction is founded to be very important for further classification. The
purpose of feature extraction is to reduce original data set by measuring certain features that distinguish one
region of interest from another. The analysis and characterization of textures present in the medical images
can be done by using the combination of WST features obtained from two-level discrete wavelet-transformed (both approximation and detail) images and WCT features obtained from two-level discrete wavelettransformed detail images.

Classification of benign and malignant brain tumor

23

Algorithm for feature extraction is as follows

Obtain the sub-image blocks, starting from the top left corner.
Decompose sub-image blocks using two-level 2D DWT.
Derive spatial gray-level dependence matrix (SGLDM) or gray-level co-occurrence matrices (e.g.,
Haddon and Boyce 1993) for two-level detail sub-bands of DWT with 1 for distance and 0, 45, 90 and
135 for h and averaged.
From these co-occurrence matrices, the following nine Haralick texture features (e.g., Haralick et al.
1973) called WCT features are extracted.
The WST features are extracted from two-level discrete wavelet-transformed (DWT) images.
Combination of the both WST and WCT features is used for classification.
Then, the feature values are normalized by subtracting minimum value and dividing by maximum value
minus minimum value. Maximum and minimum values are calculated based on the training data set. In the
data set, if the feature value is less than the minimum value, it is set to minimum value. If the feature value is
greater than the maximum value, it is set to maximum value. Normalized feature values are then optimized
by feature selection algorithms. Table 1 shows the WST and WCT features extracted using SGLDM
method.
2.3.1 Feature selection
Feature selection is the process of choosing subset of features relevant to particular application and improves
classification by searching for the best feature subset, from the fixed set of original features according to a
given feature evaluation criterion (i.e., classification accuracy). Optimized feature selection reduces data
dimensionalities and computational time and increases the classification accuracy. The feature selection
problem involves the selection of a subset of features from a total number of features, based on a given
optimization criterion. Y denotes the subset of selected features, and Z denotes the set of remaining features.
So, S = Y U Z at any time. J(Y) denotes a function evaluating the performance of Y. J depends on the
particular application. Here, J(Y) denotes the classification performance of classifying the abnormal brain
tissues into benign, malignant tumor from CT images using the set of features in X.
2.3.2 Genetic algorithm
We consider the standard GA to begin by randomly creating its initial population. Solutions are combined
via a crossover operator to produce offspring, thus expanding the current population of solutions. The
individuals in the population are then evaluated via a fitness function, and the less fit individuals are
eliminated to return the population to its original size. The process of crossover, evaluation and selection is
repeated for a predetermined number of generations or until a satisfactory solution has been found. A
mutation operator is generally applied to each generation in order to increase variation. In the feature
selection formulation of the genetic algorithm, individuals are composed of chromosomes: a 1 in the bit
position indicates that feature should be selected; 0 indicates this feature should not be selected. As an

Table 1 WST and WCT features extracted using SGLDM method

S. no
WST features
1
2
3
WCT features
1
2
3
4
5
6
7
8
9

Second order WST and WCT features

Mean (MN)
Standard deviation (SD)
Energy (ENER)
Entropy (ENT, measure the disorder of an image)
Energy (ENE, measure the textural uniformity)
Contrast (CON, measure the local contrast in an image)
Sum average (SA, measure the average of the gray level within an image)
Variance (VAR, measure the heterogeneity of an image)
Correlation (COR, measure a correlation of pixel pairs on gray levels)
Max probability (MP, determine the most prominent pixel pair in an image)
Inverse difference moment (IDM, measure the homogeneity of an image)
Cluster tendency (CT, measure the grouping of pixels that have similar gray values)

24

A. Padma Nanthagopal, R. Sukanesh Rajamony

example chromosome 00101000 means the 3rd and 5th features are selected. That is the chromosome
represents Y = {3, 5} and Z = {1, 2, 4, 6, 7, 8}. Fitness function for given chromosome X is defined as
FitnessY J Y  penaltyY

where Y is the corresponding feature subset and penalty(Y) = w 9 (|Y| - d) with a penalty coefficient w.
The size value d is taken as a constraint, and a penalty is imposed on chromosomes breaking this constraint.
The chromosome selection for the next generation is done on the basis of fitness. The fitness value decides
whether the chromosome is good or bad in a population. The selection mechanism should ensure that fitter
chromosomes have a higher probability survival. So, the design adopts the rank-based roulette-wheel
selection scheme. If the mutated chromosome is superior to both parents, it replaces the similar parent. If it
is in between the two parents, it replaces the inferior parent; otherwise, the most inferior chromosome in the
population is replaced. The selected optimal feature set based on the test data set is used to train the PNN
classifier to classify the abnormal brain tissues into benign, malignant tumors in brain CT images. Table 2
shows the best features selected using GA during the execution.
The WCT features energy (ENE), entropy (ENT), variance (VAR), inverse difference moment (IDM)
and WST feature energy are present in most of the feature vectors or feature set selected by GA. The
remaining features are least significant. The classification accuracy of 97.5 % is obtained with five features
using GA. Therefore, these minimum numbers of features are possible to classify the tumor images.
2.4 PNN classifier
Classification is the process where a given test sample is assigned a class on the basis of knowledge gained
by the classifier during training. Probabilistic neural networks (PNNs) are a widely used classification
methodology, because of its simplicity; robustness to noise, fast training speed, fast online speed, no local
minima issues, training samples can be added or removed without extensive retraining. Their main task is
the classification of unknown feature vectors into predefined classes, where the probability density function
(PDF) of each class is estimated by kernel functions. PNNs are supervised neural network models, closely
related to the Bayes classification rule and Parzen nonparametric probability density function estimation
theory. Their training procedure consists of a single pass over all training patterns, thereby rendering PNNs
faster to train, compared to the feed forward neural networks (FNNs).
The PNN architecture is composed of many interconnected processing units or neurons organized in
successive layers: input layer, pattern layer, summation layer, decision layer or output layer. The input layer
unit does not perform any computation and simply distributes the input to the neurons in the pattern layer.
On receiving a pattern x from the input layer, the neuron xij of the pattern layer computes its output. The
output of pattern layer is represented by
kxxij k

Yij x e

r2

i; j 1 . . . n;

x x1 ; x2 . . .xn

n represents number of training sets, r is the smoothing parameter. The summation layer neurons compute
the pattern x being classified into Gi by summing and averaging the output of all neurons that belong to the
same class.
The output of summation layer is represented by
Table 2 Best features selected by GA
S. no.
1
2
3
4
5
6
7
8
9
10

Feature set
IDM, ENT, ENE, VAR, CON, MN
IDM, CON, ENE, MP, VAR, ENER
ENT, IDM, VAR, IDM, CT, ENER
IDM, ENT, CT, ENE, CON, SD
CON, IDM, VAR, ENT, ENE, ENER
ENT, SA, IDM, ENE, VAR, MN
VAR, ENT, ENE, SA, IDM, ENER
ENT, CON, ENE, VAR, IDM, ENER
IDM, ENT, CT, CON, VAR, MN
ENE, ENT, MP, CON, COR, SD

Classification accuracy (%)

96
96
97.5
97
97.5
97
97.5
97.5
96
97

Classification of benign and malignant brain tumor

Gi x 1=ni

Xni
K1

25

kxxij k
r2

i; j 1 . . . n;

x x1 ; x2 . . .xn

where ni denotes the total number of training sets or patterns in class Gi. If the priori probabilities for each
class are the same, and the losses associated with making an incorrect decision for each class are the same,
the decision layer unit classifies the pattern x in accordance with Bays decision rule based on the output of
all the summation layer neurons:
Oi X maxGi X ;

i 1. . .n

where Oi(x) denotes the estimated class of the pattern x, and n is the total number of classes in the training
sets or patterns.
3 Results and discussion
Our proposed method is implemented on real human brain CT data set. Figure 4a represents benign tumor CT
image, and Fig. 4b represents the malignant tumor CT image. The input data set consists of 80 images: 40
images are malignant, 40 images are benign. The input data set is randomly partitioned into training and test
sets which are classified using PNN, LVQ, BPN classifiers. This section describes the combined waveletbased texture analysis method of classifying the brain abnormal tissues into benign, malignant tumors in CT
images. A comparative study of the classification accuracy is performed for both wavelet-based texture
analysis method and SGLDM method. The combination of WST features and WCT features extracted as
given in the feature extraction algorithm, and the optimal texture feature set is selected by GA based on the
classification performance of PNN, LVQ, BPN classifiers. From the feature selection, it is found that the
optimal feature set which gives the good classification performance are WST features like energy and the
WCT features like energy, entropy, variance and inverse difference moment. These four optimal WCT
features and one WST feature form the feature vectors or feature set. These feature vectors are given as input
to the PNN, LVQ, BPN classifiers to classify the abnormal brain tissues into benign, malignant tumors. The
classifiers are tested by using leave one out cross-validation method. The leave one out cross-validation can
be used as a method to estimate the classifier performance in unbiased manner. Here, each step, one data set is
left out and the classifier is trained using the rest and the classifier is applied to the left out data set. This
procedure is repeated such that each data set is left out once. The results are also analyzed by using receiver
operating characteristic (ROC) analysis (e.g., Fawcett 2006). ROC analysis is used to evaluate the performance of the classifiers. ROC curve is a graphical representation of sensitivity versus specificity as a
threshold parameter is varied. By calculating area under the ROC curve (AUC), we can measure the class
discrimination capability of the each classifier. The larger the area (the higher AUC value) means higher the
classification performance. Accuracy of the classifiers is evaluated based on the error rate. This error rate can
be described by the terms true and false positive and true and false negative as follows.
Sensitivity TP = TP FN

Specificity TN = FP TN

Accuracy TP TN= TP TN FN FP

where TN is the number of benign cases truly classified as negative, TP is the number of malignant cases
truly classified as positive, FN, malignant cases falsely classified as negative and FP, benign cases falsely

Fig. 4 a, b Benign, malignant tumor image (the red rectangle represents the ROIs of tumor area)

26

A. Padma Nanthagopal, R. Sukanesh Rajamony

classified as positive. Sensitivity measures the ability of the method to identify malignant cases. Specificity
measures the ability of the method to identify benign cases. Accuracy is the proportion of correctly
diagnosed cases from the total number of cases.
Neural network classifiers performance was analyzed based on the training and testing data sets. Both
training and testing sets are different. For designing the classifier, the training set contains 40 benign and 40
malignant images, and testing set contains 40 benign and 40 malignant images are considered. The PNN
structure designed has 16 neurons in the input units of input layer for giving 16 extracted features as inputs,
80 hidden neurons in the pattern units as training set contains 80 feature vectors and 2 neurons in the
summation unit for the 2 classes of benign, malignant tumor images. The LVQ network structure designed
has the input layer of LVQ network that has 16 input neurons for giving the extracted features, and the
output layer has 2 neurons as the number of classes to be identified as benign, malignant tumor images with
learning rate as 0.01, the number of hidden neurons as 80 and the number of epochs as 150 for the available
data. The BPN structure has designed with two hidden-layered architecture. Single hidden-layered architecture is not enough for converge the input patterns. This may be due to complexity of input feature vectors.
Hence, two hidden-layered neural network architecture is used. The first hidden layer consists of 33 neurons
(2 9 16 ? 1 = 33), that is, 2 9 number of nodes in the input layer ?1, because 16 features are given as
input to the classifier (four WCT features such as energy, entropy, variance, inverse difference moment are
extracted from two-level detail sub-bands. So 3 9 4 = 12, one WST feature of two-level wavelet-transformed images 1 9 4 = 4, so 12 ? 4 = 16 features) and the second layer consists of 16 neurons. For the
best performance of BPN, the proper number of nodes in the hidden layer is selected through trial-and-error
method based on number of epochs needed to train the network. It is observed that the network performed
well with 33 nodes. The learning rate for input and hidden layer is 0.4, moment = 0.2 and the error allowed
is 0.01. Binary sigmoid function is used. The initial weights are randomly selected from [-0.5, 0.5]. After
successful training with 450 epochs, the satisfactory results are obtained.
Table 3 shows the classification performances of PNN, LVQ, BPN classifiers with 80 images. The
accuracy of PNN, LVQ, BPN classifiers is 97.5, 95, 88.75 % for the same training and testing data sets.
Table 4 shows the classification performances of the classifiers using different cross-validation methods.
The accuracy of the PNN classifier in three-, five-, tenfold cross-validation methods is 96.6, 97.1 and
97.5 %, respectively, for the same training and testing data sets. Similarly, the accuracy of LVQ, BPN
classifiers in three-, five-, tenfold cross-validation methods is 94, 94.6, 95 and 87.6, 88.3, 88.75 %,
respectively. The accuracy of the PNN classifier in tenfold cross-validation method is high when compared
with the PNN classifier for 3, 5 cross-validation methods and also the other classifiers.
To justify the choice of wavelet domain for the same data set without applying wavelet transform, in the
gray-level domain, the performance of the PNN, LVQ, BPN classifiers is shown in Table 5. The accuracy of
the PNN classifier in wavelet domain is 97.5 %, and the gray-level domain is 95 %. The accuracy of the
PNN classifier in the wavelet domain is high when compared with the PNN classifier in the gray-level
domain. Table 6 shows the classification performances of our proposed technique and SGLDM method. The
classification accuracy of our proposed method is 97.5 % high when compared with the SGLDM method.

Table 3 Classification results of PNN, LVQ, and BPN classifiers with 80 images
Classifier
PNN
LVQ
BPN

No. of images

Training

160
160
160

B
40
40
40

Testing
M
40
40
40

B
39
38
35

M
39
38
36

Images misclassified

Accuracy (%)

2
4
9

97.5
95
88.75

B benign image, M malignant image

Table 4 Classification accuracy of 3, 5, 10 cross-validation methods with 80 images
Cross-validation methods

Classification accuracy

Threefold
Fivefold
Tenfold

PNN (%)
96.6
97.1
97.5

LVQ (%)
94
94.6
95

BPN (%)
87.6
88.3
88.75

Classification of benign and malignant brain tumor

27

Table 5 Classification performances of PNN, LVQ, BPN classifiers with 80 images

Parameter used

Wavelet domain (%)

TP
TN
FP
FN
Sensitivity in %
Specificity in %
Accuracy in %

PNN
39
39
2
0
100
95.1
97.5

LVQ
38
38
3
1
97.4
92.6
95

Gray-level domain (%)

BPN
36
35
5
4
90
87.5
88.75

PNN
39
37
3
1
97.5
92.5
95

LVQ
37
36
4
3
92.5
90
91.25

BPN
35
34
6
5
7.5
85
86.25

Table 6 Classification accuracy of the proposed technique with 80 images

S. no.
1
2
3
4
5
6

Technique
WT ? SGLDM ? GA ? PNN
SGLDM ? GA ? PNN
WT ? SGLDM ? GA ? LVQ
SGLDM ? GA ? LVQ
WT ? SGLDM ? GA ? BPN
WT ? GA ? BPN

Classification accuracy (%)

97.5
95
95
91.25
88.75
86.25

Feature selection is carried out using GA. There are 9 WCT features from detail sub-bands and 3 WST
features from each sub-band. So, totally 9 9 3 = 27 WCT features and 3 9 4 = 12 WST features. Hence,
27 ? 12 = 39 features are extracted. The next step is to determine the relevance of each selected feature to
the process of classifying abnormal brain tissues into benign and malignant tumors. During the evaluation
process by using GA, some features may be selected many times as the number of generation increases. If
the feature was selected more times, that feature was given as more important in the feature selection. The
number of times the features selected was energy, entropy, variance and inverse difference moment of WCT
features and energy of WST feature.
The feature selection results were also consistent with the knowledge of radiologists. For example, the
feature inverse difference moment measures the homogeneity of an image, whereas the feature variance
measures heterogeneity of an image. This is consistent with radiologists the presence of homogeneity
suggests that an abnormal image is benign and the presence of heterogeneity suggests that an abnormal
image is malignant. The parameter set for the GA algorithm is as follows: population size is 30; crossover
probability is 1.0; mutation rate is 0.1; penalty coefficient is 0.5; and stopping condition is 100 generations.
In our application for evaluating classification accuracy, tenfold cross-validation method is done on the
data set collected from 80 images (40 benign, 40 malignant). The images are divided into ten sets. The nine
sets are used for training, and remaining set is used for testing. In the next iteration (210), nine sets are used
for training, and remaining set is used for testing. This process is repeated for ten times. The classification
accuracy is calculated by taking the average of all the correct classifications in ten iterations. Figure 5a, b

Fig. 5 a, b ROC analysis curves of classifiers in wavelet, gray-level domain

28

A. Padma Nanthagopal, R. Sukanesh Rajamony

shows the ROC analysis curve obtained for the classifiers in wavelet domain, gray-level domain. Results
show that if the number of sample images increased, we get the good classification accuracy in tenfold
cross-validation method.
4 Conclusion
As a conclusion, we have presented a method for combined wavelet-based texture feature extraction method
and selecting the optimal texture features using GA and evaluated the PNN, LVQ, BPN classifiers to classify
the abnormal brain tumor into benign and malignant. This method effectively works well for classification
of the abnormal brain tissues into benign and malignant tumor with high sensitivity, specificity and accuracy. The combined wavelet-based texture analysis method proposed in this work using the SVM classifier
enables proper tumor segmentation, and the PNN classifier enables tumor classification thereby saving time
and reducing the complexity involved. Results show that the combined wavelet-based statistical features
were yielding better results compared to the statistical features extracted directly from the image without
applying wavelet transform. This justifies the choice of using wavelet transform. The developed classification system is expected to provide valuable and accurate classification process for the physicians. This
approach has potential for further development because of this simplicity that will motivate to classify the
other types of abnormalities to the radiologists.
Acknowledgments The authors are grateful to Dr. S. Alagappan Chief Consultant and Radiologist, Devaki Scan Centre,
Madurai for providing CT images and validation.

References
Brill FZ, Brown DE, Martin WN (1992) Fast genetic selection of features for neural network classifiers. IEEE Trans Neural
Netw 3:324328
Chang T, Kuo C (1993) Texture analysis and classification with tree structured wavelet transform. IEEE Trans Image Process
2:429441
Chang RF, Wu WJ, Moon WK, Chou YH, Chen DR (2003a) Support vector machines for diagnosis of breast tumors on US
images. Acad Radiol 10:189197
Chang RF, Wu WJ, Moon WK, Chou YH, Chen DR (2003b) Improvement in breast tumor discrimination by support vector
machines and speckle-emphasis texture analysis. Ultrasound Med Biol 29:679686
Chaplet S, Patnaik LM, Jaganathan NR (2006) Classification of magnetic resonance brain images using wavelets as input to
support vector machine and neural network. Biomed Signal Process Control 1:8692
Clark MC, Hall LO, Goldgof DB, Velthuzien R, Murtagh FR, Silbiger MS (1998) Automatic tumor segmentation using
knowledge based techniques. IEEE Trans Med Imaging 17:187192
Duncan JS, Ayache N (2000) Medical image analysis: progress over two decade and challenges ahead. IEEE Trans Pattern
Anal Mach Intell 22:85106
Dunn C, Higgins WE (1995) Optimal Gabor filters for texture segmentation. IEEE Trans Image Process 4:947964
El-Naqa I, Yang Y, Wernick MN, Galatsanos NP, Nishikawa RM (2002) A support vector machine approach for detection of
micro calcifications. IEEE Trans Med Imaging 21:15521563
Fausett L (1994) Fundamentals of neural networks: architectures algorithms and applications. Prentice Hall International,
Englewood Cliffs, pp 335343
Fawcett T (2006) An introduction to ROC analysis. Pattern Recognit Lett 27:861864
Freeborough PA, Fox NC (1998) MR image texture analysis applied to the diagnosis and tracking of Alzheimers disease.
IEEE Trans Med Imaging 17:475479
Gering DT, Eric W, Grimson L, Kikins R (2002) Recognizing deviations from normalcy for brain tumor segmentation. Lecture
Notes Comput Sci 2488:388395
Haddon JF, Boyce JF (1993) Co-occurrence matrices for image analysis. IEE Electron Commun Eng J 5:7183
Haralick RM, Shanmugam K, Dinstein I (1973) Texture features for image classification. IEEE Trans Syst Man Cybern
3:610621
Reddick WE, Glass JO, Cook EN, Elkin TD, Deaton RJ (2009) Automated segmentation and classification of multispectral
magnetic resonance images of brain using artificial neural networks. IEEE Trans Med Imaging 16:911918
Schad LR, Bluml S, Zuna I (1993) MR tissue characterization of intracranial tumors by means of texture analysis. Magn Reson
Imaging 11:889896
Shebang DC, Hiremath PS (2009) Support vector machine (SVM) classifier for brain tumor detection. International conference
on advances in computing, communication and control, pp 444448
Tang QH, Liu BH, Chen YQ, Zhou XH, Ding JS (2007) Application of LVQ neural network combined with the genetic
algorithm in acoustic seafloor classification. Chin J Geophys 50:291298
Tourassi GD (1999) Journey towards computer aided diagnosisrole of image texture analysis. Radiology 213:317320
Van de Wouver G, Scheunders P, Van Dyck D (1999) Statistical texture characterization from discrete wavelet representation.
IEEE Trans Image Process 8:592598
Yang X (2011) Finger print smear detection based on sub band Feature representation. EURASIP J Adv Signal Process
34:325339