International Journal of Computer Trends and Technology (IJCTT) – volume 11 number 6 – May 2014

Comparative Analysis of Textural Features Derived

from GLCM for Ultrasound Liver Image
Classification
Aborisade, D.O.#1, Ojo, J. A.#2, Amole, A.O.#3, Durodola A.O.#4
#1,2,3
Department of Electronic and Electrical Engineering,
Ladoke Akintola University of Technology, Ogbomoso.
P.M.B. 4000, Ogbomoso, Oyo-State, Nigeria.
#4
Department of Family Medicine,
Bowen University Teaching Hospital, Ogbomoso,
Oyo State, Nigeria.

Abstract— Comparative analysis of nine textural feature images characterization is the extraction of set of features
measures derived from gray-level co-occurrence matrix from the region(s) of interest (ROI) among the normal and
obtained from the region(s) of interest (ROI) among the normal abnormal anatomical structures that appear in the patient’s
and abnormal anatomical structures that appear in the patient’s ultrasound images, for ultrasonic liver tissues classification.
ultrasound liver images is presented in this paper. Selection of The characterization of liver images in this work is based
the most robust discriminating features for classification on texture analysis techniques. There exist a considerable
experiment is performed through analysis of each feature number of texture analysis techniques. The most common are
classes’ separability power. The results analysis shows that first order statistics, grey level co-occurrence matrix and
cluster prominence, cluster shade, maximum probability, fractal geometry. In this paper, gray-level co-occurrence
and entropy have high classes’ separability power and matrix (GLCM) method is used to extract nine textural
were selected for the classification of liver ultrasound features used to categorize the ultrasonic liver images into
images into normal liver (NL), primary liver cell carcinoma normal liver (NL), primary liver cell carcinoma (PLCC) and
(PLCC) and hepatocellular carcinoma (HCC) at 0.4, 0.4, 0.2 and hepatocellular carcinoma (HCC).
0.6 sensitivity respectively.
However, for a real-time ultrasonic liver image
Keywords— Liver tissue, Feature extraction, Feature selection. classification problems large number of features are not
necessary and thus dimensionality reduction is performed. An
approach to dimensionality reduction is feature selection.
I. INTRODUCTION Without employing feature selection technique many of the
Applications of medical imaging have received a great extracted features could be either redundant or even irrelevant
attention in medical and healthcare sector. Imaging techniques to the classification task. In this work, selection of the most
for computer aided diagnosis of diseases are of various kinds robust discriminating features to better represent the target
such as magnetic resonance imaging (MRI), single positron concept is performed through analysis of each feature classes’
emission computerized tomography (SPECT), computerized separability power. The results analysis shows that cluster
tomography (CT) and ultrasound. However, ultrasound prominence, cluster shade, maximum probability, and entropy
imaging is widely used technique in the diagnosis of soft have high classes’ separability power and were selected for
tissues, due to its ability to visualize human tissue without the classification of liver ultrasound images.
deleterious effects [1]. The remaining of this paper is organized as follows:
Liver is a large organ in the body that cleans the blood Section II presents feature extraction scheme and overview of
and produces bile which helps the body to deal with the fats various image classification methods. The experimental
we eat. However liver tissue is prone to diseases such as cyst, results and conclusion was given in Section III while
alcoholic cirrhosis, and carcinoma [2]. Many researchers in conclusion was presented in Section IV.
different contexts have proved in their findings that early II. TEXTURE ANALYSIS AND CLASSIFICATION METHODS
detection and treatment of liver diseases is the only way to
reduce the mortality [3, 4]. ltrasound images play an important Our approach to feature extraction and selection in
role to detect anatomical and functional information of liver addition to the overview of some image classification methods
tissue for diagnosis [5, 6]. Ultrasound are generally complex are presented in this section.
in nature for physicians and radiologists to be examine by A. Feature Extraction
simple visual inspection based on their individual experiences
and knowledge, thus computer aided diagnosis system (CAD) Feature extraction is a crucial step for any pattern
is required for supporting the detection and characterization of recognition task especially for ultrasonic liver tissues
liver tissue from ultrasound images [7-9]. The aim of liver classification. Generally, ultrasound images present various

ISSN: 2231-5381 http://www.ijcttjournal.org Page 239

 International Journal of Computer Trends and Technology (IJCTT) – volume 11 number 6 – May 2014

granular structures as texture and the analysis of ultrasound where;  i and  j are the GLCM mean of the first and
image is analogous to the problem in texture analysis.
However textural features are those characteristics such as second components ,  i and  j are the GLCM
smoothness, fitness and coarseness of certain pattern variances of the first and second components.
associated with the image. There exist a considerable number (d) Cluster shade and cluster prominence characterises the
of texture feature analysis techniques. In this work gray-level tendency of clustering of the pixels in the region of
co-occurrence matrix (GLCM) is used for the extraction of interest.
textural features.
Gray level co-occurrence matrix (GLCM) [10, 11] is a
second-order statistics methods, which is based on (local)
Cluster Shade    (i  j     ) p(i, j)
i j i j
3
(5)
information about gray levels in pair of pixels. The matrix 4
Cluster Pr om.    (i  j     ) p (i, j ) i j (6)
defined over the image with distribution of co-occurring i j
values of given offset. Let Q be an operator that defines the
position of two pixels are relative (offset), and an image f , (e) Entropy is a statistical measure of randomness that can be
with L possible intensity levels. Let G be a matrix with used to characterizes the texture of an image
element g define number of times that pair of pixel with
ij
intensities Z i and Z j occur in f with specified position in
Entropy    p(i, j) log p(i, j)
i j
(7)

Q (1  i, j  L ) . Mathematically we have a co-occurrence (f) Dissimilarity is a measure of distance between pairs of

matrix C which defined over an n  m image I , offset objects (pixels) in the region of interest.
(x, y ) .

n m 1, if ( p, q )  i and I ( p  x, q  y )  j

Dissimilarity   i j
i  j p (i , j ) (8)
Cx, y (i, j)     (1)
0, otherwise
p 1 q 1
(g) Maximum probability measures the maximum likelihood
of producing the pixels of interest.
Haralick [12] described 14 statistic feature measures that
can be calculated from the co-occurrence matrix with the
Max. Pr obability  max . p(i, j) for all (i, j ) (9)
intent to describe the texture of the images. In this work we
extracted nine of those, homogeneity, correlation,
dissimilarity, contrast, clustering shade, clustering prominence, (h) Energy provides the sum of squared elements in the
entropy, energy and maximum probability from the region(s) GLCM. It has values between 0 and 1.
of interest (ROI) of the liver image.

(a) Contrast is a measure of intensity contrast between a pixel

Energy 
i, j 
P( i, j) 2 (10)

and its neighbour over the whole image. For a “constant”

Selection of a set of appropriate input feature variables is
image (no variation) contrast is zero.
an important issue in the building of a classifier [13]. The
2
purpose of feature variable selection is to find the smallest set
Cont. 
i j
i  j p (i, j ) (2) of features that can result in satisfactory predictive
performance. Because of the curse of dimensionality [14], it is
often necessary and beneficial to limit the number of input
(b) Local homogeneity measures the closeness of the features in a classifier in order to have a good predictive and
distribution of elements in the GLCM to the GLCM less computationally intensive model.
diagonal. For a diagonal GLCM, homogeneity is 1. Numerous feature selection methods have been developed
in the pattern recognition literature [14], [15]. In this work the
1
Homog .  
i j
1 i  j
2
p(i, j ) (3) best discriminating feature is selected by analyzing the
classes’ separability power of each feature.
(c) Correlation is a measure of how correlated a pixel is to its B. Image Classification Methods
neighbour over the whole image. It is 1 or -1 for a
Briefly reviewed of some image classification techniques
perfectly positively or negatively correlated image and
are given in the following subsection.
infinity for a constant image.
1) k-Nearest Neighbor [16]: k- Nearest Neighbor
(i   i )( j   j ) p (i, j ) classifier is based on learning by analogy, that is by
Correl.   i j  i j
(4)
comparing a given test sample with training samples which

ISSN: 2231-5381 http://www.ijcttjournal.org Page 240

 International Journal of Computer Trends and Technology (IJCTT) – volume 11 number 6 – May 2014

are similar to it. The training samples are described by equivalent to performing structural risk minimization to
n attributes. Each sample represents a point in achieve good generalization [23, 24]. Finding the optimal
n  dimensional space. In this way, all of the training samples hyper-plane implies solving a constrained optimization
are stored in an n  dimensional pattern space. When given an problem using quadratic programming.
unknown sample, a k-nearest neighbor (k-NN) classifier
searches the pattern space for the k training samples which are The dimensionality of the feature space is determined by
closest to the unknown sample. These k training samples are the number of support vectors extracted from the training data.
the k-nearest neighbors of the unknown sample [16, 17]. The SVM can locate all the support vectors, which exclusively
“Closeness” is defined in terms of a distance metric, such as determine the decision boundaries. To estimate the
Euclidean distance. The Euclidean distance between two misclassification rate (risk), the so called leave-one-out
points or samples X 1  ( x11 , x12 ,  , x1n ) and procedure is used. It removes one of training samples,
performs training using the remaining training samples, and
X 2  ( x 21 , x 22 ,  , x 2 n ) is computed as; tests the removed sample with the newly derived hyper-plane.
It repeats this process for all of the samples, and the total
n number of errors becomes the estimation of the risk [25].
dist.( X 1 , X 2 )    (x 1i  x 2i ) 2 (11)
i 1

The basic steps of the k-NN algorithm are; III. RESULTS AND DISCUSSION
 To compute the distances between the new sample The ultrasound image datasets used in the experiment are
and all previous samples, have already been provided by St.Gregory’s specialist clinic and ultrasound
classified into clusters; diagnostic service, Yemetu, Ibadan. The data samples which
 To sort the distances in increasing order and select was scanned with HP Deskjet F2400 at both vertical and
the k samples with the smallest distance values; horizontal resolution of 200dpi with a bit depth of 24 were
 To apply the voting principle. A new sample will be acquired in off-line mode from a Shimedzu 350XL ultrasound
added (classified) to the largest cluster out of k machine. Ninety samples of ultrasound liver images are used
selected samples [18]. in the experiments. Out of these samples 18 are normal, 42
primary liver cell carcinoma (PLCC) and 30 hepatocellular
2) Bayes Classifier [19-21]: The Bayes classifier is carcinoma (HCC) images. First all images are registered into
applied to investigate the feasibility of classifying texture the database through intensity based image registration
image, since from the statistical viewpoint, it represents the method. Nine textural features derived from gray level co-
optimum measure of performance. The Bayesian decision rule occurrence matrix are extracted from the region of interests
classifies an observation to the class that has the highest a (ROIs) among the normal and abnormal ultrasound images.
posteriori probability among the classes. One of the ways to The between classes distance of the extracted features was
represent a pattern classifier is in terms of a set of computed to selects the best discriminant features for the the
discriminant functions g i ( X ), i  1,  , K where K denote classification of liver ultrasound images for every three class
cases shown in Fig. 1 to 3.
total number of classes. The classifier is to assign a feature
vector X to class  i if g i ( X )  g j ( X ) for all j  i . Let us The input data set is divided into two equal halves for
assume that the distribution of feature vectors X within the ith training and the testing as presented in Table 1. The result of
class P ( X  i ) is a multivariate normal distribution with the experiment is shown in Table 2. Based on the available
data and the experiments conducted, it was found from the
mean vector  i and covariance matrix C i and the a priori result presented in Table 2 and Figure 4 that, cluster
probabilities are equal for all classes. Under such an prominence, cluster shade, maximum probability and entropy
assumption, the discriminant functions can be defined as; have high classes’ separability power than other features and
1 are then selected as the best discriminant features for the the
g i ( X )   ( X   i ) T C i1 ( X   i )
2 classification of liver ultrasound images for the three class
(12) cases at 0.4, 0.4, 0.2 and 0.6 sensitivity respectively.
1
 log Ci  log P(i )
2
TABLE 1

3) Support Vector Machine [22]: SVMs are primarily DISTRIBUTION OF SAMPLES IN TRAINING AND TESTING SETS
two-class classifiers that have been shown to be an attractive
and more systematic approach to learn linear or non-linear Training Testing
decision boundaries [23, 24]. The classifier constructs an PLCC 21 21
optimal separating hyper-plane between the classes in the HCC 15 15
Normal 9 9
dataset by maximizing the distance of either class from the
hyper-plane using the Gaussian radial basis kernel. This is

ISSN: 2231-5381 http://www.ijcttjournal.org Page 241

 International Journal of Computer Trends and Technology (IJCTT) – volume 11 number 6 – May 2014

TABLE 2
PERFORMANCE EVALUATION FOR NORMAL, HCC AND PLCC
LIVER FOR K-NN CLASSIFIER WITH K = 7.

Features Con Dis Ent En Hom MP CS CP

Spec. 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
Sens. 0.0 0.0 0.6 0.0 0.0 0.2 0.4 0.4
FPR NaN NaN 1.0 NaN NaN 1.0 1.0 1.0
FNR 0.88 0.88 0.95 0.88 0.88 0.9 0.93 0.93

Fig. 2 Primary Liver Cell Carcinoma (PLCC) Liver

Fig. 1 Hepatocellular Carcinoma (HCC) Liver Fig. 3 Normal Liver

IV. CONCLUSION
In this paper, nine textural features derived from gray
level co-occurrence matrix were extracted from the region of
interests (ROIs) among the normal and abnormal ultrasound
liver images. Results analysis of between classes value of each
feature shows that cluster prominence, cluster shade,
maximum probability, and entropy are the best discriminating
features selected for the classification of liver ultrasound
images and for diagnosing liver diseases based on the
following diagnostic indices; False-Positive Rate, False-
Negative Rate, Specificity and Sensitivity .

ISSN: 2231-5381 http://www.ijcttjournal.org Page 242

 International Journal of Computer Trends and Technology (IJCTT) – volume 11 number 6 – May 2014

(e)

(a)

(f)

(b)

(g)

(c)

(h)

(i)
(d)
Fig.4 Analysis of saparabilty power of each feature classes’

ISSN: 2231-5381 http://www.ijcttjournal.org Page 243

 International Journal of Computer Trends and Technology (IJCTT) – volume 11 number 6 – May 2014

REFERENCES [25] V. C. Chen (2004),” Evaluation of Bayes, ICA, PCA and SVM Methods
for Classification”, RTO SET Symposium on Target Identification and
[1] Wen-Li Lee, Kai-Sheng Hsieh , Yung-Chang Chen, and Ying-Chng Recognition Using RF Systems, Oslo, Norway, Pp. 37.1-37.
Chen (2004), “A Study of Ultrasonic Liver Images Classification with
Artificial Neural Networks Based on Fractal Geometry and Multi-
resolution Analysis Biomedical Engineering” Applications, Basis &
Communications,16: 59-67
[2] Suganya, R. and S. Rajaram (2012), “Content Based Image Retrieval of
Ultrasound Liver Diseases Based on Hybrid Approach”, American
Journal of Applied Sciences 9 (6): 938-945, ISSN 1546-9239.
[3] Bendi Venkata Ramana, M.Surendra Prasad Babu, and N. B.
Venkateswarlu (2011), “A Critical Study of Selected Classification
Algorithms for Liver Disease Diagnosis” International Journal of
Database Management Systems (IJDMS), Vol.3, No.2, DOI:
10.5121/ijdms.2011.3207 101
[4] Meenalosini, S., J. Janet and E. Kannan (2012), “A Novel Approach in
Malignancy Detection of Computer Aided Diagnosis”, American
Journal of Applied Sciences 9 (7): 1020-1029.
[5] Xue, X., D.J. Lee, S. Antani and L.R. Long (2005), “Relevance
Feedback for Spine X-ray Retrieval”, in Proceedings of the 18th IEEE
Symposium on Computer Based Medical System, Jun. 23-24, IEEE
Xplore Press, pp: 197-202. DOI: 10.1109/CBMS.2005.94
[6] Suganya, R. and S. Rajaram (2012), “Content Based Image Retrieval of
Ultrasound Liver Diseases Based on Hybrid Approach”, American
Journal of Applied Sciences 9 (6): 938-945.
[7] Yeh, W.C., S.W. Haung and P.C. Li (2003), “Liver Fibrosis Grade
Classification with B-mode Ultrasound”, Ultrasound Med. Biol., 29:
1229-1235. PMID: 14553797
[8] Hong, H.M., Y.N. Sun and X.Z. Lin, (2002), Texture feature coding
method for classification of liver sonography. Comput. Med. Imag.
Graph., 6: 33-42.
[9] Aube, C., F. Oberi, N. Korali, M.A. Namour and D. Loisel (2002),
“Seminars in Ultrasound. CT MRI”, 23: 3-3.
[10] Meng Liang, ( 2012), “3D co-occurrence matrix based texture analysis
applied to cervical cancer screening”, Department of Information
Technology, UPPSALA UNIVERSITET.
[11] rafel C. Gonzalez, Richard E.Woods, ”Digital Image Processing.”,
Pearson Education, Inc. Upper Saddle River, new Jersey 07458. Third
Edition, pp.830-836.
[12] Robert M. Haralick, ”Statistical and structural approaches to texture.”,
Proc. IEEE, vol. 67, no. 5, pp. 786 - 804, 1979.
[13] Guoqiang Peter Zhang (2000), “Neural Networks for Classification: A
Survey”, IEEE Transactions on Systems, Man, and Cybernetics—Part
C: Applications and Reviews, Vol. 30, No. 4, pp. 451-462.
[14] P. A. Devijver and J. Kittler, Pattern Recognition: A Statistical
Approach. Englewood Cliffs, NJ: Prentice-Hall, 1982.
[15] K. Funahashi, Introduction to Statistical Pattern Recognition. New
York: Academic, 1990.
[16] Reza Entezari-Maleki, Arash Rezaei, and Behrouz Minaei-Bidgoli
“Comparison of Classification Methods Based on the Type of Attributes
and Sample Size” Department of Computer Engineering, Iran University
of Science & Technology (IUST), Tehran, Iran.
[17] J. Han, M. Kamber (2006), “Data Mining: Concepts and Techniques,”
Elsevier, Second Edition.
[18] M. Kantardzic (2003), “Data Mining: Concepts, Models, Methods, and
Algorithms,” John Wiley & Sons Publishing.
[19] R. O. Duda and P. E. Hart,(1973),”Pattern Classification and Scene
Analysis”, New York: Wiley.
[20] M. Nadler and E. P. Smith (1993),”Pattern Recognition Engineering”,
New York: Wiley.
[21] K. Fukunaga (1990), “Introduction to Statistical Pattern Recognition”,
Academic Press, San Diego.
[22] Aborisade, D.O and Adebayo I.G (2011), “Cork Stopper Classification
Using Feature Selection Method and SVM Based Classifier”,
International Journal of Signal Processing, Image Processing and Pattern
Recognition Vol. 4, No. 3.
[23] V. Vapnik (1995),” The Nature of Statistical Learning Theory”,
Springer, Berlin.
[24] C. Burges (1998), “Tutorial on support vector machines for pattern
recognition”, Data Mining Knowledge Discovery 2 (2) Pp. 955-974.

ISSN: 2231-5381 http://www.ijcttjournal.org Page 244