2017 7th IEEE International Conference on System Engineering and Technology (ICSET 2017), 2 - 3 October 2017, Shah Alam,

Malaysia

Texture Analysis and Classification in

Ultrasound Medical Images for
Determining Echo Pattern Characteristics

Hanung Adi Nugroho1, *, Made Rahmawaty1, 2, Lina Choridah3, Reni Indrastuti3

Yuli Triyani1, 2, Igi Ardiyanto1 3)
Department of Radiology, Faculty of Medicine,
1)
Department of Electrical and Information Technology, Universitas Gadjah Mada,
Faculty of Engineering, Universitas Gadjah Mada, Yogyakarta, Indonesia
Yogyakarta, Indonesia
2)
Department of Electrical Engineering,
Politeknik Caltex Riau, Pekanbaru, Indonesia
*)
[email protected]

Abstract—Ultrasound is one of the imaging modalities The use of computer aided diagnosis (CADx) system can
commonly used for detecting mass abnormalities of nodule. The assist the radiologists in decision making for detecting the mass
observation of ultrasound images is conducted by the abnormalities of nodule [3, 4]. Echo pattern called
radiologists, which tend to be subjective. Therefore, the use of echogenicity is one of the breast imaging reporting and data
computer aided diagnosis (CADx) system based on image system (BIRADS) characteristics which can be used to assist
processing can assist the radiologists to give more objective radiologists in analysing ultrasound images for determining the
decision-making for detecting the mass abnormalities of nodule. mass abnormalities of nodules [5].
This study proposes an approach to identify echo pattern
characteristic of nodule by analysing some extracted texture This study proposes an approach to determine echo pattern
features. A total of 343 ultrasound images consisting of 191 solid characteristics for distinguishing the mass of cystic and solid
and 152 cystic nodules are used in this study. Three classifiers, nodules based on texture-based features followed by involving
namely Naïve Bayes, support vector machine (SVM) and multi- some classifiers. The structure of this paper is organised as
layer perceptron (MLP) classifier are involved to measure the follows. Section II describes the material and proposed. The
performance of proposed approach. Generally, MLP classifier results and evaluation are provided in Section III while the
achieves the best performance in classifying nodule with the conclusion is presented in Section IV.
accuracy of 93.00%, Kappa of 0.86 and AUC of 0.974. These
results show that the proposed approach successfully identifies II. MATERIAL AND APPROACH
echo pattern characteristic of cystic and solid nodules on the
ultrasound images. The five main steps conducted in this study are pre-
processing, segmentation, feature extraction, feature selection
Keywords—ultrasound image, echo pattern characteristic, and classification as depicted in Fig. 1. Firstly, original image
nodule, texture analysis, multi-layer perceptron. is cropped in order to obtain the region of interest (RoI) of
nodule. The RoI is then converted to grey scale followed by
I. INTRODUCTION adaptive median filtering and morphological operation to
According to the global cancer statistics (GLOBOCAN) eliminate the speckle noises, marker and labels. Afterwards,
data that is issued by International Agency for Research on the enhanced image is then segmented based on active contour
Cancer (IARC), cancer is one of the threatening diseases based without edge (ACWE) to acquire the nodule area.
on the number of cases and mortality in the worldwide. Texture-based features which consist of histogram-based,
Around 14.1 million of new cases and 8.2 million number of GLCM and GLRLM features are extracted from segmented
mortality recorded in 2012 [1]. Breast, thyroid, cervix, kidney nodule. Correlation-based feature selection (CFS) is conducted
and liver cancers are the most prevalent cancers and the leading for gaining relevant features to increase the performance of
cause of death for women in the world. classification. Some classifiers involved are multi-layer
Ultrasound is one of the imaging modalities commonly perceptron (MLP), support vector machine (SVM) and naïve
used to detect and classify the mass abnormalities. Compared Bayes (NB) classifier.
to other modalities, ultrasound has several benefits especially A. Image Database
no radiation, faster, cheaper and potentially to increase the
number of detected nodules. However, it highly depends on In this study, 343 ultrasound images which consist of 54
the ability and experience of the radiologists [1, 2]. breast, 57 kidney, 94 cervix, 106 liver and 32 thyroid images in
24 bit BMP format taken from the Laboratory of Parasitology,
Faculty of Medicine, Universitas Gadjah Mada are used. Two

978-1-5386-0383-3/17/$31.00 ©2017 IEEE 23

 2017 7th IEEE International Conference on System Engineering and Technology (ICSET 2017), 2 - 3 October 2017, Shah Alam, Malaysia

types nodule contained in these images are 191 solid and 152 • Seven features of grey level run length matrices
cystic nodules. The radiologists manually marked the nodule (GLRLM), namely short run emphasis (SRE), long run
area which is then used as the ground truth image. emphasis (LRE), low grey-level run emphasis (LGRE),
high grey-level run emphasis (HGRE), gray level non-
uniformity (GLN), run length non-uniformity (RLN)
and run percentage (RP). GLRLM is superior in
mapping the pattern textures which have similar long-
pixels.
E. Feature Selection
The purpose of feature selection step is to obtain the most
significant subset of extracted features. In this research, CFS
subset evaluation proposed by Hall [11] is used to select the
relevant features. The worth of subset attributes is evaluated
by considering the ability of individual predictive in each
Fig. 1. Block diagram of the proposed approach feature along with the degree of redundancy among them. The
subset feature which has a high correlation to the class and low
B. Pre-processing correlation to the other features is selected.
Spackle noise is one of the barrier factors in interpreting F. Classification
diagnosing ultrasound image since it reduces the contrast The objective of classification step is to identify the
image. Therefore, pre-processing is conducted for de-noising characteristics of the selected features, patterns and structures
ultrasound image while preserving the important information. within an image in order to classify the mass of nodule images
Normalisation is also necessary to increase the contrast into solid and cystic classes. The performance of three
between nodule and background. Adaptive median filter has classifiers, namely NB, MLP and SVM are evaluated.
been used widely to reduce speckle noises which processes
each of pixel noise in different windows sizes to determine a 1) Naïve Bayes classifier
new value for specific pixel [6]. Filtered image is subsequently Naïve Bayes (NB) classifier is the simplest form of
conducted by erosion and dilation based morphological Bayesian network as shown in Fig. 2 and one of the most
operation. Erosion is a shrinking operation, whereas dilation is efficient and effective inductive algorithms for machine
expansion operation [7]. learning [12]. Assuming that an underlying probabilistic
model allows capturing uncertainty model with the principal
C. Segmentation way by determining the probability of the outcomes, it may
Segmentation aims to separate all possible nodule area in solve the diagnostic and predictive problems. Naïve Bayes is
the image. This is the key point in image analysis. The mathematically formulated in (1).
enhanced image is segmented by active contour without edge
(ACWE) for distinguishing between nodule and non-nodule
area. ACWE is a combination of active contour method and
minimisation of energy based-segmentation. It is able to detect
the objects based on the curve evaluation technique [8].
D. Feature extraction
Features extraction is one of the methods for facilitating the
process of image analysis and pattern recognition. Extracted Fig. 2. An example of simple form naïve Bayes [12]
features provide the information about composition and spatial
properties of fundamental image elements. Moreover, texture
௡
feature is able to describe the mass characteristics of solid and ‫݌‬ሺ‫ ܥ‬ൌ ൅ሻ ‫݌‬ሺ‫ݔ‬௜ ȁ‫ ܥ‬ൌ ൅ሻ
cystic nodules. In this research, 33 texture features are ݂௡௕ ሺ‫ܧ‬ሻ ൌ ෑ (1)
‫݌‬ሺ‫ ܥ‬ൌ െሻ ‫݌‬ሺ‫ݔ‬௜ ȁ‫ ܥ‬ൌ െሻ
extracted from segmented nodule in the RoI image consisting ௜ୀଵ
of:
Here, Naïve Bayes classifier is defined by ݂௡௕ ሺ‫ܧ‬ሻ while‫ܥ‬
• Six histogram-based features such as mean, skewness, represents the classification variable, ൅is the positive class,
energy, standard deviation, smoothness and entropy [9]. and Ȃ is negative class.
• Twenty features of grey level co-occurrence matrices 2) Multi-layer perceptron classifier
(GLCM) in which a total of five features are obtained in Multi-layer perceptron (MLP) classifier is a network which
four directions 0, 45, 90 and 135 degrees. They are contains successive layers, such as input layer, hidden layer
angular second moment (ASM), inverse different and output layer based on the use back propagation as a
moment (IDM), contrast, entropy and correlation. learning mechanism. Fig. 3 shows a structure network of MLP
GLCM is the second order texture feature by [13].
considering the relation of pixel adjacency which was
proposed by Haralick et al. [10].

24
 2017 7th IEEE International Conference on System Engineering and Technology (ICSET 2017), 2 - 3 October 2017, Shah Alam, Malaysia

3) Support vector machine classifier As shown in Table I, there is a difference average value of
Support vector machine (SVM) classifier is a supervised selected features between solid and cystic masses. In solid
learning algorithm which is used for binary classification mass, four selected features, i.e. contrast at 450, deviation,
problem that looking for an optimal hyperplane to separate energy and LRGE have greater average values than that of
class sample. The kernel function is used to map the data into cystic mass. Whilst the SRE feature in solid mass has smaller
the higher dimensional. Furthermore, the optimal hyperplane in average value than that of cystic mass.
the high dimensional data is distinguished by feature space
[14]. TABLE I. AVERAGE VALUES OF SELECTED FEATURES
Masses
Features
Cystic Solid
Contrast at 450 52.21 212.29
Deviation 12.06 31.52
Energy 0.31 0.42
SRE 0.50 0.39
LGRE 76.00 132.91

C. Performance Evaluation of Classifiers

The dataset is divided into 10-folds in which nine folds are
used as training dataset and one fold is used as testing dataset.
Fig. 3. MLP network structure [13]
For all classifiers, determination the right parameters is the
point to obtain good classification results. In order to measure
the performance of three used classifiers, some parameters
III. RESULTS AND EVALUATION such as accuracy, sensitivity, specificity, Kappa statistic, area
under curve (AUC) and computation time are involved as
A. Pre-processing and segmentation presented in Table II. Furthermore, receiver operating
The RoI of nodule in ultrasound image is determined by characteristic (ROC) curve is also considered to support the
radiologists. In general, RoI image still contains artefacts analysis performance of classifiers as shown in Fig. 5 and Fig.
(marker and label) and also the spackle noises obtained from 6. The closer the curve to the top and left-hand borders, the
coherent interference of backscattered echoes. Therefore, better the classifier performance.
filtering process without causing any blur effect is required
such as adaptive median filter. Moreover, the morphological TABLE II. THE EVALUATION RESULTS OF CLASSIFIERS PERFORMANCE
operation is proposed to soften the filtered images. ACWE 33 full features 5 selected features
segmentation method is subsequently applied to separate the Parameters
NB MLP SVM NB MLP SVM
nodule from its background. The selection of ACWE is based Accuracy (%) 90.67 92.71 93.00 92.13 93.00 93.00
on the result of a comparative study by Triyani et al. [4]. They Sensitivity (%) 90.54 90.97 94.44 88.34 91.55 93.83
stated that ACWE was the best method for segmenting nodule Specificity (%) 90.77 94.15 91.96 95.56 94.18 92.39
on breast ultrasound image. The results of pre-processing and Kappa statistic 0.811 0.853 0.858 0.842 0.859 0.858
Area under curve
segmentation step are presented in Fig. 4. 0.959 0.978 0.926 0.969 0.974 0.927
(AUC)
Computation time(s) 0.02 5.57 0.14 0.001 0.39 0.02

As shown in Table II, the use of five selected features

(a) (b) successfully improves the performance of all classifiers
indicated by higher accuracy achieved and less computation
time needed to build the models. It shows that these features
are the relevant features in identifying echo pattern
characteristics of the nodules. All of the classifiers generate
(c) (d) Kappa values more than 0.8 indicating that these classifiers
Fig. 4. The images result of (a) RoI (b) adaptive median filter have a high agreement in classifying the characteristics of
(c) morphological operation and (d) ACWE nodule ultrasound images [15].
SVM classifier gains the highest sensitivity at 93.83%
B. Feature Extraction and Feature Selection showing that SVM classifier has a good performance in
This study analyses 33 extracted texture features from first recognising cystic nodules. While Naive Bayes classifier has
and second statistical orders. CFS subset evaluation is then the best performance in recognising solid nodule with the
applied to find the relevant features in order to increase the specificity of 95.56%. Generally, the evaluation results show
performance of classification, reduce the computational time that MLP classifier outperforms Naïve Bayes and SVM
and simplify the process. Finally, five relevant features are classifiers by considering the highest values achieved in terms
selected from the 33 full features as summarised in Table I. accuracy of 93.00%, Kappa of 0.86, AUC of 0.974 and also
based on the ROC curve. However, it needs more time of
computation.

25
 2017 7th IEEE International Conference on System Engineering and Technology (ICSET 2017), 2 - 3 October 2017, Shah Alam, Malaysia

ACKNOWLEDGMENT
The authors would like to thank the Intelligent Systems
research group members in the Department of Electrical
Engineering and Information Technology, UGM, for sharing
and great discussion. The authors also gratefully appreciate the
Directorate General of Higher Education, Ministry of
Research, Technology and Higher Education, Republic of
Indonesia for financially supported by the Research Grant
“Penelitian Tim Pasca Sarjana” Universitas Gadjah Mada, No.
2546/UN1.P.III/DIT-LIT/LT/2017.
REFERENCES
[1] J. M. Baran and J. G. Webster, "Design of low-cost portable ultrasound
Fig. 5. ROC curves of the 33 full features systems," in 2009 Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (EMBC), 2009, pp. 792-
795.
[2] H. A. Nugroho, Y. Triyani, M. Rahmawaty, I. Ardiyanto, and L.
Choridah, "Performance analysis of filtering techniques for speckle
reduction on breast ultrasound images," in 2016 International
Electronics Symposium (IES), 2016, pp. 450-454.
[3] D.-R. Chen, Y.-L. Huang, and S.-H. Lin, "Computer-aided diagnosis
with textural features for breast lesions in sonograms," Computerized
Medical Imaging and Graphics, vol. 35, pp. 220-226, 2011.
[4] Y. Triyani, H. A. Nugroho, M. Rahmawaty, I. Ardiyanto, and L.
Choridah, "Performance analysis of image segmentation for breast
ultrasound images," in 2016 8th International Conference on
Information Technology and Electrical Engineering (ICITEE), 2016, pp.
1-6.
[5] L. Levy, M. Suissa, J. F. Chiche, G. Teman, and B. Martin, "BIRADS
ultrasonography," European Journal of Radiology, vol. 61, pp. 202-211,
Fig. 6. ROC curves of 5 selected features 2007/02/01/ 2007.
[6] H. K. N. Yusufiyah, H. A. Nugroho, T. B. Adji, and A. Nugroho,
"Feature extraction for classifying lesion's shape of breast ultrasound
images," in 2015 2nd International Conference on Information
IV. CONCLUSION Technology, Computer, and Electrical Engineering (ICITACEE), 2015,
pp. 102-106.
An approach to identify echo pattern characteristic of [7] A. K. Jain, Fundamentals of digital image processing: Prentice-Hall,
nodule ultrasound image based on texture analysis has been Inc., 1989.
proposed. Three classifiers, namely Naïve Bayes (NB), [8] T. Chan and L. Vese, "An active contour model without edges," Scale-
support vector machine (SVM) and multi-layer perceptron Space Theories in Computer Vision, pp. 141-151, 1999.
(MLP) are used to measure the performance of proposed [9] H. A. Nugroho, S. A. Akbar, and E. E. H. Murhandarwati, "Feature
extraction and classification for detection malaria parasites in thin blood
approach in classifying solid and cystic nodules. By using five smear," in 2015 2nd International Conference on Information
selected features, these classifiers are able to achieve higher Technology, Computer, and Electrical Engineering (ICITACEE), 2015,
accuracy and less time of computation than that of the 39 full pp. 197-201.
features. [10] R. M. Haralick and K. Shanmugam, "Textural features for image
classification," IEEE Transactions on systems, man, and cybernetics,
The evaluation results show that the proposed approach vol. SMC-3, pp. 610-621, 1973.
successfully to identify echogenicity characteristic of cystic [11] M. A. Hall, "Correlation-based feature selection for machine learning,"
and solid nodules. The best performance in recognising the The University of Waikato, 1999.
[12] H. Zhang, "The optimality of naive Bayes," AA, vol. 1, p. 3, 2004.
cystic nodule is achieved by SVM classifier with the sensitivity [13] I. Belakhdar, W. Kaaniche, R. Djmel, and B. Ouni, "A comparison
of 93.83%. Whilst the best performance in recognising the between ANN and SVM classifier for drowsiness detection based on
solid nodule is gained by NB classifier with the specificity of single EEG channel," in 2016 2nd International Conference on
95.56%. In overall, MLP gains the best performance with the Advanced Technologies for Signal and Image Processing (ATSIP), 2016,
accuracy of 93.00%, Kappa of 0.86, AUC of 0.974 and also the pp. 443-446.
curve of ROC almost arrived the upper left corner of the plot. [14] H.-D. Cheng, J. Shan, W. Ju, Y. Guo, and L. Zhang, "Automated breast
cancer detection and classification using ultrasound images: A survey,"
These results indicate that the proposed approach has a Pattern recognition, vol. 43, pp. 299-317, 2010.
potential to be implemented in the development of a CADx [15] A. J. Viera and J. M. Garrett, "Understanding interobserver agreement:
system for assisting the radiologists in determining the mass the kappa statistic," Fam Med, vol. 37, pp. 360-363, 2005.
abnormalities of ultrasound nodules.

26