br Laws Texture Energy analyzes

Laws’ Texture Energy analyzes the energy texture on the image by convolving the input image into nine filtered energy images using 5 by 5 laws’ convolutional masks. These masks are produced by multiplying two laws’ vectors (Laws, 1980). The vectors used in this study are listed in Table 1, while the process of multiplication is illustrated in Fig. 7.
According to the literature, even though there are 16 convolu-tional masks produced, there are only nine of them present the tex-ture energy as listed in Table 2.
2.4. Classification
Support vector machine (SVM) was adopted to classify shape, margin and orientation characteristic (external characteristic). Each of these three characteristics has two classes, so SVM has bet-ter accuracy result (Savelonas et al., 2008; Liu et al., 2017).
Fig. 10. GUI of CAD thyroid cancer system.
Please cite this article as: H. A. Nugroho, Zulfanahri, E. L. Frannita et al., Computer aided diagnosis for thyroid cancer system based on internal and external characteristics, Journal of King Saud University – Computer and Information Scienceshttps://doi.org/10.1016/j.jksuci.2019.01.007
H.A. Nugroho et al. / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx 7
Whereas multilayer perceptron (MLP) has better performance when it is used to classify more than two classes. In this case, MLP is used to classify content and echogenicity characteristics which are identified as internal characteristic (Nugroho et al., 2017b).
The concept of SVM is to find the best hyper-plane that can sep-arated the POM 1 being classify. The best hyper-plane is formed when the margin between each class border reach the optimum value (Nugroho and Witarto, 2003), as illustrated in Fig. 8. Dataset of classes is denoted as fx1; ; xng and class label from xi is presented as yi 2 fþ1; 1g then the border of class a and b are formulated in (13).
xi:w þ b1 for yi ¼ 1 ð13Þ Subsequenty, optimization constraint is defined in (14).
Lagrangian formula is used to solve the optimization with addi-tional constraints ai 0 as Lagrange coefficient and LD is the Lagrangian dual problem, so the best hyper-plane is formulated in (15).
Fig. 11. Result of pre-processing, (a) process of ROI on input image, (b) RoI image, (c) adaptive median filter result, (d) SRBF result.
Please cite this article as: H. A. Nugroho, Zulfanahri, E. L. Frannita et al., Computer aided diagnosis for thyroid cancer system based on internal and external characteristics, Journal of King Saud University – Computer and Information Scienceshttps://doi.org/10.1016/j.jksuci.2019.01.007
8 H.A. Nugroho et al. / Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
n
subject to
The concept of MLP is based on two main steps in creating the model through learning phase. First one is feed forward phase to calculate the error between the actual and the target with initial weight on each neuron. The second one is backward phase to prop-agate backward to renew the weight. The process is called as back error propagation (BEP) (Duda et al., 2001).
Assume that xi is the input data or extracted feature value, oj is the hidden layer, yk is the output class, wij is the weight between input and hidden layer, while wjk is the weight between hidden and output layer with one hidden layer. The architecture of MLP can then be designed as in Fig. 9.
2.5. Classification & diagnosis rule
There are several important rules that need to be described first in designing thyroid cancer system – CAD. Classification rule defines which characteristic needs to be classified first before others. For example in external characteristics, margin has to be
classified first to determine the shape and orientation characteris-tic. When a nodule has smooth margin, motor units means the shape of the nodule should be ‘round to oval’ only, but when it has irregular margin then the shape can be either ‘round to oval’ or irregular.