www.ietdl.

org
Published in IET Image Processing
Received on 28th February 2012
Revised on 7th October 2012
Accepted on 8th November 2012
doi: 10.1049/iet-ipr.2012.0073

ISSN 1751-9659

Wavelet statistical texture features-based

segmentation and classification of brain computed
tomography images
A. Padma Nanthagopal1, R. Sukanesh2
1
Research Scholar, Tiruchy Anna University, Tiruchy-625023, India
2
Department of Electronics and Communication Engineering, Thiagarajar College of Engineering, Madurai-625015, India
E-mail: [email protected]

Abstract: A computer software system is designed for segmentation and classification of benign and malignant tumour slices in
brain computed tomography images. In this study, the authors present a method to select both dominant run length and co-
occurrence texture features of wavelet approximation tumour region of each slice to be segmented by a support vector
machine (SVM). Two-dimensional discrete wavelet decomposition is performed on the tumour image to remove the noise.
The images considered for this study belong to 208 tumour slices. Seventeen features are extracted and six features are
selected using Student's t-test. This study constructed the SVM and probabilistic neural network (PNN) classifiers with the
selected features. The classification accuracy of both classifiers are evaluated using the k fold cross validation method. The
segmentation results are also compared with the experienced radiologist ground truth. Quantitative analysis between ground
truth and the segmented tumour is presented in terms of segmentation accuracy and segmentation error. The proposed system
provides some newly found texture features have an important contribution in classifying tumour slices efficiently and
accurately. The experimental results show that the proposed SVM classifier is able to achieve high segmentation and
classification accuracy effectiveness as measured by sensitivity and specificity.

1 Introduction have made segmentation more difficult. Therefore the

challenges for segmentation and classification of brain CT
Computed tomography (CT) uses special X-ray equipment to images have raised many different approaches. For
obtain many images from different angles and joins them example, Wei et al. [1], Lauric and Frisken [2] proposed to
together to produce multiple cross sectional images of the segment the soft tissues by means of pixel intensity-based
head. CT scanning provides more detailed information on segmentation method. Sharma et al. [3] proposed
head injuries, stroke, brain tumours and other brain co-occurrence texture features by means of bidirectional
diseases. In recent years, medical CT images have been associative memory-type artificial neural network to
extensively applied in clinical diagnosis. CT image can segment and classify the soft tissues in brain CT images.
assist physicians to detect and locate pathological changes Tong et al. [4] proposed to segment cerebrospinal fluid and
with more accuracy. CT images can be distinguished for brain matter by means of non-identical clustering
different tissues according to different grey levels. The techniques and decision tree classifier in two stages and the
images, if processed appropriately can offer a wealth of accuracy is less compared with our proposed method.
information, which is significant in assisting doctors in Padma and Sukanesh [5] proposed dominant run length
medical diagnosis. Many diagnostic imaging techniques can texture features by means of support vector machine (SVM)
be performed for the early detection of any abnormal classifier to segment and classify the brain CT images.
changes in tissues and organs such as CT imaging and Padma Nanthagopal and Sukanesh Rajamony [6] proposed
magnetic resonance imaging (MRI). Image segmentation is combined co-occurrence, grey level and new edge features
the process of partitioning a digital image into a set of by means of SVM classifier to segment the tumour from
pixels. An accurate and fast reproducible segmentation brain CT images. Rajendran and Madheswaran [7] proposed
technique is required in various applications. The results of co-occurrence texture features by means of Naive Bayesian
segmentation are obtained for classification and analysis classifier to classify benign and malignant tumour images.
purposes. The limitations for CT scanning of head images Ganesan and Radhakrishnan [8] proposed to segment the
are because of low soft tissue contrast (particularly the tumour by means of a genetic algorithm, Choplet et al. [9]
brain), which affect the edges and yield different objects proposed wavelet based co-occurrence texture features by
within the same range of intensity. All these limitations means of an SVM classifier to classify the magnetic

IET Image Process., 2013, Vol. 7, Iss. 1, pp. 25–32 25

doi: 10.1049/iet-ipr.2012.0073 & The Institution of Engineering and Technology 2013
 17519667, 2013, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-ipr.2012.0073 by Readcube (Labtiva Inc.), Wiley Online Library on [22/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
resonance images and the classification rate of SVM is less malignant slices of eight patients are acquired from GE
compared with our proposed method. Kharrat et al. [10] light speed 64 scanner. The resolutions of all the images are
proposed co-occurrence texture features by means of an standard and the images are normalised by the standard
SVM classifier to classify normal, benign and malignant reconstruction algorithm called Z filtering, since the slice
tumour images, but in these [7, 9, 10] three methods, profile, image noise and image artefacts are controlled by
segmentation could not be performed and the accuracy is this technique. All images are performed on 512 × 512
less compared with our proposed method. Padma and reconstruction matrix, 2–2.5 mm slice thickness, 110 Kv,
Sukanesh [11] proposed wavelet co-occurrence texture 63 mA and 1 s scan time in Siemens scanner and 130 Kv,
features by means of bidirectional associative memory type 83 mA and 0.5 s scanning time for GE light speed scanner.
artificial neural network to segment the brain tumour in CT Every single pixel of the two-dimensional (2D) slice is
images. Padma and Sukanesh [12] proposed co-occurrence described by the value known as density measured in
texture features by means of a probabilistic neural network Hounsfield unit (HU). As a result of image acquisition, a
(PNN) classifier to segment the brain CT images. The series of 512 × 512 16 bit grey scale images with 256
above literature survey showed that all the above methods different grey levels (0–255), with Hounsfield values
are considered co-occurrence texture features only and some ranging from −1000 to 1000 HU are produced.
of the methods are proposed to obtain classification only
and some of the methods are proposed to obtain 2.2 Segmentation of tumour using SVM classifier
segmentation only, therefore feature selection could not be
conducted in all the published literature. In our research, we choose the SVM [13] classifier to
In this study, we perform a combination of dominant run segment the shape of tumour information. SVM performs
length and the co-occurrence matrix method to improve robust non-linear classification with a kernel trick. SVM is
diagnostic accuracy in classifying benign from malignant independent of the dimensionality of feature space and the
tumour slices. The purpose of this study is to extract results obtained are very accurate. It combines linear
combined texture information from the segmented tumour algorithms with linear or non-linear kernel functions that
region. The segmentation of the tumour region can be make it a powerful tool in the machine learning community
performed by using an SVM classifier. We construct the with applications such as data mining and medical
SVM and PNN classifiers with the selected features using imaging applications. To apply SVM into non-linear data
Student’s t-test and evaluate the performance of the SVM distributions, the data can be implicitly transformed to a
and PNN classifiers using a round robin method. The high dimensional feature space where a linear separation
results of this research are helpful in classifying benign and might become possible. In this study, we choose a linear
malignant tumours and improving the diagnosis of tumour function.
slices. To build the SVM segmentation model, feature vectors of
tumour and non-tumour area are extracted. Five rectangles
2 Materials and methods with five points to cover the tumour and non-tumour area
are selected. For each selected rectangle, four vertices and
The proposed system is divided into five phases (i) image one centre point of the rectangle are concerned. For each
acquisition, (ii) segmentation of tumour region, (iii) image point, we use two properties of position and intensity of
preprocessing, (iv) feature extraction and feature selection selected point to form the feature vector or training vector.
and (v) classification and evaluation. The proposed ti = (xi, yi, Ii (xi, yi)) i = 1 to n represent the number of
methodology is applied to real datasets representing brain training vectors.
CT images with a dimension of 512 × 512 and all images
are in DICOM format. The proposed algorithm is (xi, yi) and Ii (xi, yi) represent the position and intensity of
implemented in Matlab 7.2 platform and run on 3.0 GHz, the selected points.
512 MB RAM Pentium IV systems in Microsoft Windows We totally have 25 feature sets from the tumour and the
operating systems. non-tumour area. These 50 feature vectors or feature sets
are given as input to the SVM classifier to segment the
2.1 Image acquisition tumour shape. After simple morphological operations, the
segmentation result is as shown in Fig. 1. Five training
The performance of the method has been tested on a real samples are obtained from each rectangle area. Hence, we
dataset collected from M/s Devaki MRI and CT scans, have a 25 × 3 matrix of tumour area and 25 × 3 matrix of
Madurai; Aarthi MRI and CT scans, Madurai, India from non-tumour area.
two different high-end multi-slice CT scanners such as
Siemens Somatom Sensation 64, GE light speed 64. The 2.3 Preprocessing the segmented tumour image
rotation time (RT) of Siemens was 0.5 ms, but the RT of based on 2D discrete wavelet decomposition
GE light speed was 0.8 ms. The dataset consists of
volume CT data of 16 patients (8-benign, 8-malignant). Preprocessing of the tumour image is an important step
Number of slices varies across the patient’s 10–13 benign towards extracting texture features from the tumour region.
slices and 10–13 malignant slices and 2 mm slice thickness A two-level discrete wavelet decomposition of tumour
in Siemens Somatom Sensation 64 and 2.5 mm slice image is performed, which results in four sub-bands. In a
thickness in GE light speed 64. In total, there are 2D discrete wavelet decomposition [14], the tumour image
208 slices belonging to two main categories: 104-benign, is represented by one approximation and three detailed
104-malignant and the size of the tumour is equal to or images representing the low- and high-frequency contents
larger than 1.5 cm. Five to six lesions are extracted from image, respectively. The approximation can be furthered to
each patient and the 104 benign slices of eight patients are produce one approximation and three detailed images at the
acquired from Siemens somatom scanner and the 104 next level of decomposition, and so on until the required

26 IET Image Process., 2013, Vol. 7, Iss. 1, pp. 25–32

& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-ipr.2012.0073
 17519667, 2013, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-ipr.2012.0073 by Readcube (Labtiva Inc.), Wiley Online Library on [22/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org

Fig. 1 Segmentation process using SVM classifier

level is reached. The rows of approximation coefficients are structural abnormalities in different tissues. As the tissues
convolved with both a low-pass filter and a high-pass filter present in the brain are difficult to classify using the shape
and the results are column down sampled. The columns of or the grey-level intensities, texture feature extraction
both down sampled results are convolved with both a is found to be very important for further classification. The
low-pass and high-pass filter and the results are row down 17 spatial features are extracted from the two-level
sampled. The resulting four matrices are the next-level wavelet approximation tumour image of each slice by using
approximation and detailed coefficients. The wavelet 2D both dominant grey-level run length and grey-level
decomposition process is shown in Fig. 2. A1 and A2 co-occurrence matrix method.
represent the wavelet approximations at first and second The dominant grey-level run length matrix φ(d, θ) [15] is as
levels, respectively, and are a low-frequency part of the follows
images. H1, V1, D1, H2, V2 and D2 represent the details
of horizontal, vertical and diagonal directions at first and 

f(d, u) = P i, j|d, u , 0 , i ≤ Ng , 0 , j ≤ Rmax
second levels, respectively, and are a high-frequency part of
the images. After 2D wavelet decomposition at second level (1)
is performed on the tumour image, the approximation at the
second level is obtained to replace the original image to be where Ng is the maximum grey level and Rmax is the
used for texture analysis. Approximation at the second level maximum run length. The element P(i, j|θ) specified the
is more homogeneous than original tumour image after estimated number of runs where a given image contains a
removing high-frequency detail information. This will make run length j for a grey level i in the direction of angle θ.
the texture features extracted based on dominant run length Four dominant grey-level run length matrices corresponding
and co-occurrence matrix method more significant. to θ = 0°, 45°, 90° and 135° are computed and the
following four dominant run length texture features [16]
2.4 Feature extraction and feature selection such as short-run low-grey-level emphasis, short-run
high-grey-level emphasis, long-run low-grey-level emphasis
2.4.1 Feature generation: Texture analysis is a (LLGE) and long-run high-grey-level emphasis (LHGE) are
quantitative method that can be used to quantify and detect extracted from the two-level wavelet approximation tumour
image of each slice’s dominant grey-level run length matrix
and take the average of all the features extracted from four
dominant grey level run length matrices.
The grey-level co-occurrence matrix ¢ (d, θ) [17] as
follows


f(d, u) = P i, j|d, u , 0 , i ≤ Ng , 0 , j ≤ Ng (2)

where Ng is the maximum grey level. The function P(i, j/d, θ)

is the probability matrix of two pixels, which are located
within an inter-sample distance d and direction θ have a
grey level i and grey level j. Four grey-level co-occurrence
matrices corresponding to θ = 0°, 45°, 90° and 135°
directions are computed with distance d(= 1, 2) and the
following 13 textural-based Haralick features [18] are
extracted from the two-level wavelet approximation tumour
image of each slice’s grey-level co-occurrence matrix and
take the average of all the features extracted from four
Fig. 2 Two-level discrete wavelet decomposition grey-level co-occurrence matrices.

IET Image Process., 2013, Vol. 7, Iss. 1, pp. 25–32 27

doi: 10.1049/iet-ipr.2012.0073 & The Institution of Engineering and Technology 2013
 17519667, 2013, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-ipr.2012.0073 by Readcube (Labtiva Inc.), Wiley Online Library on [22/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
2.4.2 Feature selection: Feature selection is the process where xi is the ith training pattern from each class, n is the
of reducing the dimension of feature vector. For feature dimension of the input vectors, m is the number of training
selection, every feature is observed. Significant features are patterns in each class and σ is a smoothing parameter
selected by the calculation of mean values for every feature corresponding to standard deviation of Gaussian distribution.
in benign tumour class and malignant tumour class. The PNN architecture is composed of many
Student’s t-test [19] is employed as a method to separate interconnected processing units or neurons organised in
both the classes. Two-sample Student’s t-test considered successive layers: input layer, pattern layer, summation
each feature independently. It is assumed that both classes layer, decision layer or output layer. The input layer unit
of data values are distributed normally and had similar does not perform any computation and simply distributes
variances. Test statistics are calculated as follows the input to the neurons in the pattern layer. On receiving
a pattern x from the input layer, the neuron Xij of the
Xb − Xm pattern layer computes its output. The output of the pattern
t =  (3)
Varb Varm layer is represented by
+
nb nm
− x−xij 2
where, xb and xm represent mean values from benign and Yij (X ) = e s2 i, j = 1, . . . , n, x = x1 , x2 , . . . , xn
malignant classes. varb and varm represent variances of (6)
benign and malignant classes. nb and nm consist of numbers
of instance in each class. This t value followed Student n represents the number of training sets and σ is the
t-test with (nb + nm− 2) degrees of freedom. On the basis of smoothing parameter. The summation layer neurons
test statistics and degrees of freedom, the significance compute pattern x being classified into Gij by summing
P-value [20] is calculated. Only the optimal features are and averaging the output of all neurons that belong to the
selected based on the condition P < 0.001 (i.e.) if the same class.
P-value of the feature is less than 0.001, then the The output of the summation layer is represented by
corresponding feature is selected and used for classification.   2 

−x−xij  / s2
n
1 i
2.5 Classifier Gij (X ) = e
ni K=1
Classification is the process where a given test sample is
i, j = 1, . . . , n, x = x1 , x2 , . . . , xn (7)
assigned a class on the basis of knowledge gained by the
classifier during training.
where ni denotes the total number of training sets or patterns
2.5.1 Support vector machine: SVM [21] is a in class Gij. If the a priori probabilities for each class are the
supervised learning method and is used for one-class and same, and the losses associated with making an incorrect
n-class classification problems. Its basic idea is to transform decision for each class are the same, the decision layer unit
a non-linear dividing problem into a linear one by some kind classifies pattern x in accordance with Bayes’s decision rule
of kernel function. To apply SVM into non-linear data based on the output of all the summation layer neurons
distributions, the data can be implicitly transformed to a  
high-dimensional feature space where a separation might Oi (x) = max Gij (x) , i, j = 1, . . . , n (8)
become possible. SVM has the following advantages: 1) It is
designed to use limited samples to obtain optimum solution
for practical problem; 2) It can ensure to find the global where Oi(x) denotes the estimated class of pattern x and n is
rather than local optimum solution. 3) The generalisation the total number of classes in the training sets or pattern.
ability of SVM is very good with relatively less The input feature set is classified as benign class, if the
computational complexity. One major problem is that the total input to the decision unit is positive and the input
computation is complex; the idea of kernel function solves feature set is classified as malignant class, if the total input
this difficulty. By a properly selected kernel function, we can to the decision unit is negative.
reduce the time of training without lost of any precision. The
problem is a non-linear classification one; therefore we use a 2.5.3 Classification performance evaluation: There
non-linear SVM that uses a Gaussian kernel function are many statistical methods for evaluating and estimating
the performance of the classifier. Some of the best-known

  2 methods are round robin (10-fold cross-validation) and
K X , Xi = exp −s  x − xi  (4)
jackknife methods. The goal should be to obtain
classification accuracy of the system. A common set-up to
2.5.2 Probabilistic neural network classifier: PNNs measure the classification performance is by dividing the
[22] can be used for classification problems based on total number of data into a training set, a validation set and
Bayesian classification and classical estimators for a test set. The classifier is trained with the training set,
probability density function. It uses the exponential whereas the validation set is used to decide when to stop
activation function instead of the sigmoidal activation training. The test data are only used after training by means
function. Consider the two class problem, namely benign of which an independent classification performance can be
class and malignant class. The PNN uses the following measured. The accuracy of the classifier is evaluated based
estimator for the probability density function given by on the error rate. This error rate [23] can be described by
the terms true and false positive and true and false negative

T
 as follows
1 1 m
x − xi x − xi
f (X ) = exp (5)
( 2P) s
n/2 n m
i=1
2 s2 Sensitivity = TP/(TP + FN )∗100 (9)

28 IET Image Process., 2013, Vol. 7, Iss. 1, pp. 25–32

& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-ipr.2012.0073
 17519667, 2013, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-ipr.2012.0073 by Readcube (Labtiva Inc.), Wiley Online Library on [22/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
Specificity = TN /(TN + FP)∗100 (10)

Accuracy = (TP + TN )/(TP + TN + FP + FN ) (11)

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 classified as
positive. Sensitivity is the ability of the method to identify
malignant cases. Specificity is the ability of the method to
identify benign cases. Accuracy is the proportion of
correctly diagnosed cases from the total number of cases.
To make the classification results comparable and for
exhaustive data analysis, the classification performance [24]
is evaluated by the round robin method, in each step one
dataset is left out and the classifier is trained using the rest
and the classifier is applied to the left out dataset. This
procedure is repeated such that each dataset is left out once.
For example, in the n-sample images, the round robin
method trains the classifier with n − 1 samples and then
uses the one remaining sample as a test sample.
Classification is repeated until all n samples have been used
once as a test sample.

3 Results and discussion

The real CT brain images are collected from eight benign
tumour patients and eight malignant tumour patients and Fig. 5 Segmented results of sample images
each patient has approximately 13 slices. The slices that are
clearly focusing on the brain are selected from the available
slices of a CT brain tumour image. We considered 208 step is to determine the relevance of each selected feature to
slices in this study as described in ‘Materials and methods’. the process of classifying benign and malignant slices.
They are in DICOM format. Fig. 3 shows the results of During the evaluation process by using Student’s t-test, the
input real CT image and the segmented tumour image best textural features selected are LLGE, LHGE, energy,
based on SVM classifier. Fig. 4 shows the wavelet entropy, variance and inverse difference moment (IDM).
approximation and detail in horizontal, vertical and The feature selection results are consistent with the
diagonal directions at second level of wavelet decomposition. knowledge of radiologists. For example, the feature IDM
The segmented results of two slices from two different that measures the homogeneity of a slice and the feature
patients with malignant tumour image and three slices from LLGE that captures the homogeneous nature of the texture
three different patients with benign tumour image, the feature. This is consistent with the radiologists; the presence
corresponding ground truth images are as shown in Fig. 5. of homogeneity suggests that an abnormal slice is benign.
From the wavelet approximation tumour image of each Next, feature variance measures the heterogeneity of a slice
slice, we extract the four dominant run length texture and the feature LHGE captures the heterogeneous nature of
features and 13 co-occurrence texture features. Since the the texture features. This is also consistent with the
dimensionality of feature space is comparable with the radiologists; the presence of heterogeneity suggests that an
number of samples, feature selection is carried out using abnormal slice is malignant. The average mean feature
Student’s t-test as mentioned in feature selection. The next value and the average 95% confidence interval (CI) of
mean value across all the 104 benign and 104 malignant
slices are represented in Table 1. These six features are
given as inputs to the SVM and PNN classifiers, and the
classification accuracy is evaluated by means of the round
robin method.

Table 1 Mean and 95% CI of mean of features selected using

Fig. 3 Tumour segmentation using SVM classifier Student’s t-test

Feature Benign slices (96) Malignant slices (96)

name
Mean 95% CI of mean Mean 95% CI of mean

entropy 0.5678 0.5284–0.6148 0.5748 0.5210–0.6184

energy 0.6527 0.6434–0.6903 0.6823 0.6217–0.7143
variance 0.9586 0.9163–0.9875 0.9578 0.9154–0.9867
IDM 0.6345 0.6243–0.6737 0.6576 0.6153–0.6638
LLGE 0.6327 0.6240–0.6758 0.6583 0.6240–0.6778
Fig. 4 Wavelet approximation and details of segmented tumour LHGE 0.9987 0.9426–0.9876 0.9578 0.9257–0.9866
images

IET Image Process., 2013, Vol. 7, Iss. 1, pp. 25–32 29

doi: 10.1049/iet-ipr.2012.0073 & The Institution of Engineering and Technology 2013
 17519667, 2013, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-ipr.2012.0073 by Readcube (Labtiva Inc.), Wiley Online Library on [22/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
Table 2 Segmentation accuracy and error for 16 patients with 208 slices
Patients (8 + 8) Benign (104) slices Malignant (104) slices

No of slices Seg accuracy Seg error, % No. of slices Seg accuracy Seg error, %

1 13 97.834 1.66 13 98.765 1.06

2 13 99.098 1.87 13 97.969 1.40
3 13 97.998 1.17 13 98.453 0.88
4 13 98.898 1.20 13 97.9564 0.89
5 13 98.8732 1.14 13 99.8390 1.56
6 13 99.780 0.87 13 98.8762 1.12
7 13 98.984 0.75 13 99.876 1.23
8 13 98.898 0.26 13 98.673 1.30
average accuracy 98.7954 1.115 98.67 1.18

3.1 Segmentation performance evaluation artificial neural network, achieving a segmentation accuracy
of 94.4%. Padma and Sukanesh [5] proposed optimal
We perform the quantitative results in terms of performance dominant grey level run length texture features by means of
measures such as segmentation accuracy and segmentation an SVM classifier, achieving an accuracy of 95.2%. Padma
error, which are computed for each benign and malignant Nanthagopal and Sukanesh Rajamony [6] proposed
tumour slices. The output of the segmented tumour is combined co-occurrence, grey-level and new edge features
compared with the ground truth (target). The ground truth is by means of an SVM classifier, achieving an accuracy of
obtained from the boundary drawings of the radiologist. 96.4%.
The quantitative results in terms of performance measures
such as segmentation accuracy and segmentation error for 3.2 Classification performance of the SVM, PNN
real data of 104 benign slices of eight patients, real data of classifiers
104 malignant slices of eight patients are obtained using
(12), (13) and are tabulated in Table 2. The features selected in feature selection are used to train the
SVM, PNN classifiers by 10-fold cross-validation method
Segmentation accuracy = (no. of pixels matched/ (round robin) and the accuracy of the classifiers is evaluated
and tabulated in Table 3. From Table 3, the accuracy of
total no. of tumor pixels in ground truth)∗100 (12)
SVM classifier has better classification accuracy than PNN
classifier.
Segmentation error = (no. of misclassified pixels/
no. of pixels in segmented tumor image)∗100 (13) 3.3 Comparison of the accuracy with other
reported results
The segmentation accuracy is calculated as the direct ratio of Computer software system is designed for the classification of
the number of tumour pixels common for ground truth and the benign and malignant tumour slices. For example Padma and
proposed method output to the total ground truth tumour Sukanesh [12] considered co-occurrence texture features to
pixels. The segmentation error is calculated as the number segment the tumour from brain CT images by means of
of misclassified pixels to the total number of pixels in bidirectional associative memory type artificial neural
the segmented tumour region. The average segmentation network, achieving an accuracy of 93.75%. Sharma et al.
accuracy of 104 benign slices is 98.7954%, 104 malignant [3] proposed co-occurrence texture features by means of
slices is 98.67% and the average segmentation error of 104 bidirectional associative memory type artificial neural
benign slices is 1.115% and 104 malignant slices is 1.18%, network, achieving an accuracy of 95.4%. Padma and
respectively. Sukanesh [5] proposed optimal dominant grey-level run
From the segmentation results of real sample CT slices and length texture features by means of an SVM classifier,
the results of segmentation accuracy, this methodology is achieving an accuracy of 95.8%. Padma Nanthagopal and
effective and the segmentation accuracy is high compared Sukanesh Rajamony [6] proposed combined co-occurrence,
with the following existing methods. The existing methods grey level and new edge features by means of an SVM
are: Sharma et al. [3] proposed co-occurrence texture classifier, achieving an accuracy of 96.25%. Kharrat et al.
features by means of bidirectional associative memory type [10] considered wavelet co-occurrence texture features to
classify the normal, benign and malignant tumour images
Table 3 Classification performances of the SVM, PNN
by means of an SVM-based classifier, achieving an
classifiers with 208 slices accuracy of 96.29%. Padma and Sukanesh [11] proposed
the wavelet co-occurrence texture features by using a PNN
Classification parameter SVM PNN classifier, achieving an accuracy of 97%. In this proposed
method, we found that combined wavelet-based dominant
TP 102 101
TN 102 101
run length and co-occurrence texture features were
FP 2 3 represented in SVM-based feature space to classify benign
FN 2 3 and malignant tumour slices efficiently, and these newly
sensitivity in % 98.07% 96.19% combined features should be included in the feature set and
specificity in % 98.07% 96.19% achieving the highest classification accuracy of 98.07%
classification
accuracy in % 98.07% 97.11% using 10-fold cross-validation method when compared with
other conventional texture analysis methods.

30 IET Image Process., 2013, Vol. 7, Iss. 1, pp. 25–32

& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-ipr.2012.0073
 17519667, 2013, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-ipr.2012.0073 by Readcube (Labtiva Inc.), Wiley Online Library on [22/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
whenever there is change in the slice dataset and this
method can be applied to brain CT images only. The work
can be extended to other types of imaging such as liver CT
imaging, MRI imaging and ultrasound imaging as a future
work.

5 Acknowledgment
The authors are grateful to Dr. S. Alagappan Chief Consultant
and Radiologist, Devaki Scan Centre, Madurai for providing
CT images and validation.

6 References
1 Wei, K., He, B., Zhang, T., Shen, X.: ‘A novel method for segmentation
of CT head images’. Int. Conf. on Bio informatics and Biomedical
Fig. 6 Computation time to extract the features from each slice Engineering, Wuhan, July 2007, pp. 717–720
2 Lauric, A., Frisken, S.: ‘Soft segmentation of CT brain data’. Technical
report TR-2007-3, Tuffs University, M.A.
3 Sharma, N., Ray, A.K., Sharma, S., Shukla, K.K., Pradhan, S.,
3.4 Comparison of the computation time with Aggarwal, L.M.: ‘Segmentation and classification of medical images
other reported results using texture primitive features: application of BAM-type artificial
neural network’, J. Med. Phys., 2008, 33, (3), pp. 119–126
4 Tong, H.L., Mohammad, F., Ahmad, F., Ryoichi, K.: ‘Segmentation of
A computer software system is designed for classification CT Brain Images using unsupervised clustering’s’, J. Vis., 2009, 12, (2),
of benign and malignant tumour slices. The computation pp. 31–138
time to extract the texture features from each single image 5 Padma, A., Sukanesh, R.: ‘Automatic classification and segmentation of
is evaluated with the existing texture analysis methods brain tumor in CT images using optimal dominant gray level run
and the results of comparisons are represented using a bar length texture features’, Int. J. Adv. Comput. Sci. Appl., 2011, 2, (10),
pp. 53–59
graph as shown in Fig. 6. The existing methods such as 6 Padma Nanthagopal, A., Sukanesh Rajamony, R.: ‘A region-based
co-occurrence texture features proposed by Sharma et al. segmentation of tumour from brain CT images using Nonlinear
[3] and Padma and Sukanesh [12], optimal dominant Support Vector Machine classifier’, J. Med. Eng. Technol., 2012, 36,
grey-level run length texture features proposed by Padma (5), pp. 271–277
7 Rajendran, P., Madheswaran, M.: ‘An improved image mining
et al. [5], combined co-occurrence, grey-level and new edge technique for brain tumor classification using efficient classifier’,
features proposed by Padma Nanthagopal and Sukanesh Int. J. Comput. Sci. Netw. Secur., 2009, 6, (3), pp. 107–116
Rajamony [6], wavelet co-occurrence texture features 8 Ganesan, R., Radhakrishnan, R.: ‘Segmentation of computed
proposed by Kharrat et al. [10] and Padma and Sukanesh tomography brain images using genetic algorithm’, Inte. J. Soft
[11]. From the results, it is observed that the SVM with Comput., 2009, 4, (4), pp. 157–161
9 Choplet, S., Patnaik, L.M., Jaganathan, N.R.: ‘Classification of magnetic
wavelet-based combined dominant grey-level run length resonance brain Images using wavelets as input to support vector
method plus grey level co-occurrence matrix method is less machine and neural network’, Biomed. Signal Process. Control.,
computation time equal to 2.3 s when compared with other 2006, 1, (1), pp. 86–92
conventional texture analysis methods. 10 Kharrat, A., Gasmi, K., Ben Messaoud, M., Nacera, B., Abid, M.: ‘An
Hybrid approach for Automatic classification of Brain MRI using
genetic algorithm and support vector machine’, Leonardo J. Sci.,
2010, 1, (17), pp. 71–82
4 Conclusion 11 Padma, A., Sukanesh, R.: ‘A wavelet based automatic segmentation of
brain tumor in CT images using optimal statistical texture features’,
In this paper, we presented a method for combined dominant Int. J. Image Process., 2011, 5, (5), pp. 552–563
12 Padma, A., Sukanesh, R.: ‘Texture feature based analysis of segmenting
run length and co-occurrence texture features from wavelet soft tissues from brain CT images using BAM-type artificial neural
approximation tumour image of each slice and evaluated network’, J. Inf. Eng. Appl., 2011, 1, (4), pp. 34–43
the performance of SVM and PNN classifiers in segmenting 13 El-Naqa, I., Yang, Y., Wernick, M.N., Galatsanos, N.P., Nishikawa, R.
and classifying benign and malignant tumour slices. The M.: ‘A support vector machine approach for detection of micro
results of this research would be highly helpful in calcifications’, IEEE Trans. Med. Imag., 2002, 21, (12), pp. 1552–1563
14 Van de Wouver, G., Scheunders, P., Van Dyck, D.: ‘Statistical texture
improving diagnosis of tumour slices. The combined characterization from discrete wavelet representation’, IEEE Trans.
dominant run length and co-occurrence texture features Image Process., 1999, 8, (11), pp. 592–598
proposed in this work using SVM and PNN classifiers 15 Tang, X.: ‘Texture information in run length matrices’, IEEE Trans.
enable proper tumour segmentation and classification Image Process., 1998, 7, (11), pp. 234–243
16 Tang, X.: ‘Dominant run length method for image classification’, Woods
thereby saving time and reducing the complexity involved. Hole Report – 97-07, Woods Hole Oceanographic Institution, 1997
From the experimental results, it is inferred that the best 17 Khuzi, M., Besar, R., Zaki WMD, W., Ahmad, N.N.: ‘Identification of
classification performance is achieved by the SVM than the masses in digital mammogram using gray level co-occurrence matrices’,
PNN classifier. Use of large databases is expected to Biomed. Imag. Interv. J., 2009, 5, (3), pp. 109–119
improve system performance and ensure repeatability of the 18 Haralick, R.M., Shanmugam, K., Dinstein, I.: ‘Texture features for
Image classification’, IEEE Trans. Syst. Man Cybern., 1973, 3, (6),
resulting performance. This approach has potential for pp. 610–621
further development because of this simplicity that will 19 Levner, I., Bulitko, V., Lin, G.: ‘Feature extraction for classification of
motivate us to classify the other types of abnormalities to proteomic mass spectra: a comparative study’, Springer-Verlag Berlin
radiologists. The developed segmentation and classification Heidelberg, Stud Fuzz 207, 2006, 207, pp. 607–624
20 Soper D.S.: ‘P-value calculator for a student t-test (Online Software)’,
system is expected to provide a valuable and accurate http://www.danielsoper.com/statcalc3, 2011
classification process for the physicians. The limitation of 21 Zhu, Y., Tan, Y., Hua, Y., Wang, M., Zhang, G., Zhang, J.: ‘Feature
this method is that it needs new training for the classifiers selection and performance evaluation of support vector machine

IET Image Process., 2013, Vol. 7, Iss. 1, pp. 25–32 31

doi: 10.1049/iet-ipr.2012.0073 & The Institution of Engineering and Technology 2013
 17519667, 2013, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-ipr.2012.0073 by Readcube (Labtiva Inc.), Wiley Online Library on [22/12/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
based classifier for differentiating benign and malignant pulmonary 23 Liao Y.-Y., Tsui, P.-H., Yeh, C.-K.: ‘Classification of benign and
nodules by computed tomography’, J. Digit. Imaging, 2010, 23, (1), malignant breast tumors by ultrasound B-scan and nakagami-based
pp. 51–65 images’, J. Med. Biol. Eng., 2009, 30, (5), pp. 307–312
22 Mao, K.Z., Tan, K.-C., Ser, W.: ‘Probabilistic neural-network structure 24 Kim, J.K., Park, H.W.: ‘Statistical textural features for detection of micro
determination for pattern classification’, IEEE Trans. Neural Netw., calcifications in digitized mammograms’, IEEE Trans. Med. Imaging,
2000, 11, (4), pp. 1009–1016 1999, 18, (3), pp. 231–238

32 IET Image Process., 2013, Vol. 7, Iss. 1, pp. 25–32

& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-ipr.2012.0073