You are on page 1of 4

2016 International Conference on Information Technology Systems and Innovation (ICITSI)

Bandung – Bali, October 24 – 27, 2016


ISBN: 978-1-5090-2449-0

Classification of Breast Ultrasound Images based on


Posterior Feature

Tianur 1,a, Hanung Adi Nugroho 1,b, Muzni Sahar1,c, Reni Indrastuti2,a, Lina Choridah2,b
Igi Ardiyanto1,d Department of Radiology, Faculty of Medicine
Department of Electrical Engineering and Information Universitas Gadjah Mada
Technology, Faculty of Engineering Yogyakarta, Indonesia
Universitas Gadjah Mada Email : 2,areni.dr@gmail.com, 2,blinailmiah@gmail.com
Yogyakarta, Indonesia
Email : 1,atianur@mail.ugm.ac.id, 1,badinugroho@ugm.ac.id,
1,c
muzni@ mail.ugm.ac.id, 1,digi@ugm.ac.id

Abstract—Ultrasonography (USG) check-up is a common way for computer aided diagnosis (CAD) system on breast cancer
breast cancer screening, but the result is highly subjective on the lesions detection is necessary. There are many features that can
operator. Therefore, a system capable to objectively diagnose be used to aid breast cancer lesions diagnosis. This paper aims
breast cancer is necessary. One of the features of breast cancer is to classify type of lesions on breast ultrasound images based on
posterior acoustic patterns. It categorized into four classes which
are enhancement, shadowing, combined pattern, and no posterior
posterior acoustic feature.
acoustic feature. This paper proposes a scheme by extracting area Many lesions have posterior acoustic enhancement or no
suspected to have posterior acoustic features and background posterior acoustic [8] [9] [10] [11]. Therefore, posterior
features. The dataset consists of 98 breast USG images which are
acoustic enhancement can be marked as an important feature
classified into 69 posterior acoustic enhancement cases and 29 no
posterior acoustic cases. Firstly, a pre-processing of breast USG to determine breast lesions characteristics [10]. On some cases,
images is conducted to eliminate speckle noise, marker, and label. lesions with enhancement acoustic posterior have similar
Secondly, segmentation is using region growing method, and characteristic with lesions no posterior acoustic. This matter
followed by extracting posterior area and its background. Feature often causes misinterpretation on lesions and different
extraction is conducted on both of areas using histogram method. diagnosis among radiologists.
Finally, classification is using Multilayer Perceptron (MLP).
Performance of the proposed method successfully achieves This paper is organizing as follows: Section II represents
accuracy of 87.79%, sensitivity of 92.75% and specificity of the proposed theoretical basis for region of interest, adaptive
82.75% using six histogram features. It shows that this method is median filter, region growing, histogram and MLP. Material
succesful in classifying the breast USG images. Therefore, it has and methods are described in section II. Experimental results
potential to be implemented in an automated breast computer and discussion are described in section IV, and finally section
aided diagnosis (CAD) system. V is the conclusions.
Keywords—USG images, adaptive median filter, region II. THEORETICAL BASIS
growing, histogram, MLP
A. Region of Interest (RoI)
I. INTRODUCTION
ROI is a process of extracting an area by cropping USG
Breast cancer is a cancer commonly suffered by women in image at specified coordinate in order to extract area containing
the world. In 2013, breast cancer is a prevalent disease in breast cancer lesions from USG image. The images are cropped
Indonesia which about 0.8% with Daerah Istimewa Yogyakarta by initial coordinate and end coordinate. That coordinates form
as the highest prevalence about 2.4% [1]. a square in which every pixel within that area will be saved as
USG is one of the most effective and economic screening new image. An illustration is depicted in Fig. 1.
tools to observe breast cancer lesions. There are several 0,0 0,1 0,2 0,3 0,4
parameters that indicate breast cancer lesions characteristics
1,0 1,1 1,2 1,3 1,4 1,1 1,2 1,3
which are shape, orientation, margin, echo pattern and posterior
acoustic feature [2] [3] [4] [5]. Generally, commonly used 2,0 2,1 2,2 2,3 2,4 2,1 2,2 2,3

characteristics are shape, margin, and echo pattern. Ultrasound 3,0 3,1 3,2 3,3 3,4 3,1 3,2 3,3
(US) images contain much noise which made object on blur 4,0 4,1 4,2 4,3 4,4 New Image
effect image. Various intensity cause decrease in Signal Noise Original Image
Ratio (SNR) [6] [7]. One of the drawbacks in USG image is Fig. 1. Cropping Operation
high dependence on its operator. Therefore, a development

978-1-5090-2449-0/16/$31.00 ©2016 IEEE


In Fig. 1, the size of original image is 5x5 pixels. A new 2) Standard Deviation is a feature indicating contrast
image with 3x3 pixel is formed by cropping from initial value. Standard deviation is formulated in (2).
coordinate (1,1) and ended with coordinate (3,3). This new
image has same pixel value with the original image at range
(1,1) to (3,3). (2)

B. Adaptive Median Filter


Adaptive median filter is a popular widely used method 3) Skewness is the feature that measures asymmetry
compared to standard median method. In this paper, adaptive toward intensity rate. A negative value indicates contrast bias
median filter is used to diminish or eliminate disturbance such to left whilst positive value indicates contrast bias to right.
a speckle noise which commonly found in USG image. In Skewness formula is shown in (3).
USG image there are also other disturbance such as label and
marker which made by radiologist’s. Adaptive median filter (3)
could eliminate that label and marker. Despite of the effort,
heavily weighted label and marker will still remain. Adaptive 4) Energy is a feature to measure distribution of pixel
median filter will process each pixel with different window intensity toward range of grayscale. Energy is represented in
size. Maximum value, minimum value, and median have to be (4).
acquired for each window. The filter determines new value
based from that pixels [6] [12]. (4)
C. Region Growing
5) Entropy is a feature indicating image complexity.
Region growing is pixel grouping method into bigger area
Entropy formula is shown in (5).
or region by adding every neighboring pixel that has similar
value as illustrated in Fig. 2. This method leads to segment
lesion area to separate it from the background. log (5)

6) Smoothness is a feature measuring subtleness of the an


seed image. High smoothness value indicates the image has subtle
intensity. Smoothhness formula is shown in (6).

 1
1
(6)
1
E. MLP (Multilayer Perceptron)
MLP is a neural network consisting of neurons. The
neurons consist of several layers, which are input layer, hidden
Fig. 2. Region growing layer and output layer. Input layer receives external data input
and forward it to hidden layer. The hidden layers consist of one
This segmentation method needs several points (seed) or more layers. All input neurons are connected to every
which will grow to all direction until desired iteration. This neuron in hidden layer. There is no limit for hidden layers
paper initializes a square region to accelerate segmentation count and neurons count for each of its layer. A typical
process. multilayer perceptron is shown in Fig. 3.
D. Histogram
Histogram is a diagram to display frequency for every Input Layer Hidden Layer Output Layer
intensity value in the image pixels. it is also a simple method V11 Z1
W21
W22
to extract texture. Rough image has high contrast with wide X1 V12
V13 W21 Y1
intensity range, meanwhile subtle image has low intensity with V21
V22
Z2 W22

narrow range [13]. The features used in this paper are: X2 V23 W31
Z3 W32

1) Mean is the object of contrast rate. The variable i b1


Y2

represents grayscale value of an image, p(i) represents i b2


appearance probability and L represents the upper bound of
maximum grayness value. Mean is mathematically formulated Fig. 3. Multilayer perceptron
in (1).
The input value for each node in hidden layer (Zi_in) is
described as:
. (1)
_ 1 (7)
with Xi is the data input value, while Vij and b1j are weight
value and bias of hidden layers. USG Image

Preprocessing
1
_ _
(8) RoI
1
Zi describes the sigmoid activation function of hidden layer.
This activation function is the output value of a neuron. Adaptive Median Filter
The values of each node in output layer (Yi_in) are
described as: Segmentation and Capture area

_ 2 (9) Histogram Feature Extraction

Yi describe the sigmoid activation function of output layer.


Multilayer Perseptron

Classification
1 (MLP)
_ _
(10)
1
Then, output values from output layers are compared with Posterior Acoustic Enhancement,
target value (T) as described in (11).
No posterior acoustic
(11)
E is an error value used to fix or update weight and bias Fig. 5. Block diagram system
value.
Fig. 5 shows the process of proposed methods to classify
III. MATERIAL AND METHODS posterior feature into posterior accoustic enhancement and no
In this paper, 98 breast cancer images are obtained with posterior acoustic.
permission from Departemen Radiologi Sardjito Yogyakarta. IV. RESULTS AND DISCUSSION
The images consist of 69 lesions with posterior acoustic
enhancement and 29 no posterior acoustic. Examples of In this research, there are four stages, such as data
posterior acoustic enhancement and no posterior acoustic aquisition, prepocessing, feature extraction and classification.
enhancement are shown in Fig. 4. A. Preprocessing
The initial process is region of interest (RoI) to determine
the area of breast lesions in USG images. It is conducted
manually by an operator. RoI area is filtered with adaptive
median filter to reduce speckle noise and marker.

a. USG images b. RoI

a. Posterior acoustic enhancement b. No acoustic posterior feature

Fig. 4. Characteristic of posterior feature.

Posterior acoustic enhancement is the bright area due to


ultrasound waves are forwarded below the lesion, which made
it brighter than the surrounding area. No posterior acoustic
lesions is the lesions that does not have bright shade below.
c. Adaptive median filter
Characteristic of posterior feature is used to classify.
Fig. 6. Result of image preprocessing.

Adaptive median filter is very effective to reduce speckle


noise, marker, and label without make blurr effect. This results
are also proven by other researchers [7]. It is shown in Fig. 6.
Then the image is segmented using region growing method. classify the breast USG images and can be implemented in
It separate lesion from background. Next step is to define breast computer aided diagnosis (CAD) system.
posterior area and area beside of posterior area as shown in Fig.
7. ACKNOWLEDGMENT
The authors would like to thank to Department of
Radiology, RSUP dr. Sardjito and colleagues in Intelligent
Systems Research Group at the Department of Electrical
b. Posterior area Engineering and Information Technology, Universitas Gadjah
Mada.
REFERENCES
[1] WHO, "Latest word cancer statistics Global cancer burden rises to 14.1
million new cases in 2012," 2013.
[2] I. Kemenkes R I, " Situasi penyakit kanker," 2015.
a. Capturing process c. Beside of posterior area [3] A. S. Hong, E. L. Rosen, M. S. Soo, and J. A. Baker, “BI-RADS for
sonography: positive and negative predictive values of sonographic
Fig. 7. Segmentation result features,” Am. J. Roentgenol., vol. 184, no. 4, pp. 1260–1265, 2005.
[4] W.-C. Shen, R.-F. Chang, W. K. Moon, Y.-H. Chou, and C.-S. Huang,
B. Feature Extraction “Breast Ultrasound Computer-Aided Diagnosis Using BI-RADS
Features,” Acad. Radiol., vol. 14, no. 8, pp. 928–939, Aug. 2007.
Feature extraction is a method to extract feature from USG [5] G. Bouzghar, B. J. Levenback, L. R. Sultan, S. S. Venkatesh, A.
image in order to recognize characteristic of the image. In this Cwanger, E. F. Conant, and C. M. Sehgal, “Bayesian Probability of
paper, histogram is used to extract texture of the posterior area Malignancy With BI-RADS Sonographic Features,” J. Ultrasound
and the surrounding area. Feature extraction produces six Med., vol. 33, no. 4, pp. 641–648, Apr. 2014.
histogram feature such as mean, standard deviation, skewness, [6] L. Sellami, O. Sassi, K. Chtourou, and A. Hmida, “Breast Cancer
Ultrasound Images SequenceExploration Using BI-RADS Features
energy, entropy,and smothness for aech area. Extraction: Towards an AdvancedClinical Aided Tool for Precise
C. Classification Lesion Characterization,” IEEE Trans. NanoBioscience, vol. PP, no. 99,
p. 7, Oct. 2015.
Next step is classifying the breast cancer lesion into [7] D. A. Khusna, H. A. Nugroho, and I. Soesanti, “Performance analysis of
posterior acoustic enhancement or no posterior acoustic. This edge and detailed preserved speckle noise reduction filters for breast
process used MLP method with k-fold cross validation. The 98 ultrasound images,” in 2015 2nd International Conference on
Information Technology, Computer, and Electrical Engineering
data consist of 69 lesions with posterior acoustic enhancement (ICITACEE), 2015, pp. 76–80.
and 29 no posterior acoustic. Because the limitation of the data, [8] F. N. Radzi and N. Yahya, “Evaluation of digital speckle filters for
it is divided into several data groups. One of the data groups is ultrasound images,” in Control System, Computing and Engineering
testing data, while the other is classified as training data. It is (ICCSCE), 2014 IEEE International Conference on, 2014, pp. 337–342.
conducted recursively for all of data groups. Therefore, each [9] J. I. Arribas, J. Cid-Sueiro, and C. Alberola-Lopez, Estimation of
group is classified as testing data and training data. posterior probabilities with neural networks: an application to
microcalcification detection in breast cancer diagnosis. Piscataway, NJ,
Performance of the proposed method successfully achieves USA: John Wiley & Sons Inc.–IEEE Press, 2007.
accuracy of 87.79%, sensitivity of 92.75% which is a [10] J. Cui, B. Sahiner, H.-P. Chan, C. Paramagul, A. Nees, L. M. Hadjiiski,
possibility rate of posterior acoustic enhancement classified as and Y.-T. Wu, “Characterization of posterior acoustic features of breast
posterior acoustic enhancement, and specificity of 82.75% masses on ultrasound images using artificial neural network,” 2008, p.
which is a possibility rate of no posterior acoustic feature 691521.
[11] S. P. Weinstein, E. F. Conant, C. Mies, G. Acs, S. Lee, and C. Sehgal,
classified as no posterior acoustic feature. “Posterior acoustic shadowing in benign breast lesions sonographic-
V. CONCLUSION pathologic correlation,” J. Ultrasound Med., vol. 23, no. 1, pp. 73–83,
2004.
This paper proposes a scheme to classify of breast cancer [12] K. E. Maturen, A. P. Wasnik, J. E. Bailey, E. G. Higgins, and J. M.
lessions based on posterior feature. The proposed method Rubin, “Posterior acoustic enhancement in hepatocellular carcinoma,” J.
Ultrasound Med., vol. 30, no. 4, pp. 495–499, 2011.
consists of four stages such as data acquisition, preprocessing,
[13] H. D. Cheng, J. Shan, W. Ju, Y. Guo, and L. Zhang, “Automated breast
feature extraction and classification. It succesfully classifies cancer detection and classification using ultrasound images: A survey,”
breast lessions into two classes, namely posterior acoustic Pattern Recognit., vol. 43, no. 1, pp. 299–317, Jan. 2010.
enhancement and no posterior acoustic. Performance of the
proposed method successfully achieves accuracy of 87.79%,
sensitivity 92.75% and specificity of 82.75% using six
histogram features. It shows that this method has a potential to

You might also like