Internal content classification of ultrasound thyroid nodules based on textural features

Ultrasound (US) is one of the best imaging modalities on thyroid identification. The suspicious thyroid is indicated in the existence of palp able nodules whose solid or cystic composition. Solid nodules have high possibility to be malignant than cystic. An effort to detect and classify the internal content of thyroid nodule has become challenge problem in radiology area. Operator dependence of ultrasound imaging makes it complicated due to missing interpretation among radiologists. Objective Computer Aided Diagnosis (CAD) was designed to solve it which works on texture analysis of histogram statistic, gray level co-occurrence matrice (GLCM) and gray level run length matrices (GLRLM). The fine-needle aspiration cytology (FNAC) is not needed because the textural pattern is s ignificantly different between solid and cystic nodules. Multi-layer perceptron (MLP) was adopted to do classification process for 72 US thyroid images yield an accuracy of 90.28%, the sensitivity of 87.80%, specificity of 93.55% and precision of 94.74%.


Introduction
Thyroid nodules are solid or liquid lumps in the thyroid gland.The incidences of nodule are the common problem in society and increasing with age [1].About 4-7% of adults possess palpable nodules, biopsy, and ultrasound imaging detects these well [2].Thyroid nodules are more common in elderly patients, women and people with iodine deficiency [3].More 90% of thyroid nodules are benign and do not require special treat ment.However, 5% nodules are malignant, require early detection and comprehensive treatment [4].
Several studies have explained identification of ultrasound thyroid nodules whether benign or malignant [5,6,7].An important parameter to identify thyroid malignancies is classifying the internal content of nodule [8,9].Most malignant nodules are solid on ultrasound imaging with sensitivity 69%-75% [5].However, ult rasounds are dependent operator modalities, leads to high subjective interpretation among radiologists whose vary on experience and expertise [1].Fine-needle aspiration cytology (FNAC) is a further procedure to ensure nodules composition.Noted, not all of the nodules should undergo this procedure, FNAC is probably unnecessary if the nodule dominantly cystic [5].An unwelcome thing if the surgery is undergoing on suspected malignant nodule, but it turns out the results of surgery shown a benign nodule.Therefore, we proposed an objective CAD to overcome these matter.Textural features embedded on CAD to recognize a solid and cystic pattern.
We have prepared 72 ultrasound image taken fro m department radiology of Sardjito hospital.All of the textural values are classified by multi-layer perceptron.The performance of the system is validated by confusion table consist of accuracy, sensitivity, specificity, precision and Negative Predictive Value (NPV).This study aims to classify thyroid nodules into solid and cystic as core parameters on thyroid cancer.Hopefully, the results can be used as objective consideration in making an accurate diagnosis.

Materials and Methods
The steps of this study are illustrated on the flowchart as shown in Fig. 1.The theoretical foundations of textural algorith ms and perceptron and a briefly rev iew of ult rasound imaging and pre-processing were presented here.Features of extraction were carried out to obtain measurable textural values on varies images of thyroid nodules based on the statistical histogram, GLCM , and GLRLM .M LP was adopted to classify all of the features and validated by confusion matrix.

Pre-processing
Pre-processing is mean an effort to reduce marking area in which nodule exist by the median filter.Marking area lead to improper extracted textural features.The main idea of the med ian filter is to run through each pixel along an image and replace each intensity with the median of neighboring values.The mask pattern of neighbors is called window which slides pixel by pixel over an image.The median is defined by middle value after all the entries in the window are sorted computationally.

Textural Features
Jain [17] defines texture as a relat ionship neighbor pixels intensity which is repeated forming a basic pattern.The basic pattern is mean texture elements or texels.But somet imes there is a complicated repetition pattern and poorly understood by human vision.Hence, the sum of methods is required to describe related pattern into measurable values as an objective information namely texture features .
The simplest way, texture features are extracted by the statistical histogram.Here, h istogram means as the discrete bar chart illustrating probabilities of each pixel intensity appears in the whole image.A great value states that the pixels on an intensity that often arise.In the gray-scale image as used in this study, the number of gray levels (L) as much as 256 level fro m 0 to 255.A value of 0 for dark b lack imagery and a maximu m of 255 to a maximu m wh ite image.The dominance of pixels with levels approaching 255 will tend to make the image appear bright white.Histogram of various textures can be seen in Fig. 6.Fig. 6.Histogram various gray-scale images [18] A histogram is useful to observe the spread of the intensity value that can be used as a basis to represent the image of the texture pattern.The values which are derived fro m histogram also named as first order statistic features.The following values are explained below.(1) In this case, i is the level of gray in the image of f and p(i) represents the probability of occurrence of i and L declare the highest value of gray levels.The above formu la will produce an average brightness of objects .
2) Mode [17] Mode values stated gray level that appears most of all pixels.
3) Variance 2 can be called variance or second order normalized mo ment because p(i) is a function of probability.M is the Mean value in the Eq.(1).(5)

4) Standard Deviation
Gray Level Co-occurrence Matrices (GLCM) was published by Haralick et al. [19] with 14 characteristic value that represents the texture pattern.GLCM was known as the 2 nd order of texture pattern calcu lation.Texture measurements on first order statistical of the histogram are not paid attention to the relationship of neighborhood pixels, while the 2 nd order statistic, the relationship between the couple two original image pixel is calculated [18].
For examp le, f(x,y) is an image with a size of Nx and Ny which has a pixel with a possibility of up to L level and  ⃗⃗ is a vector spatial direction.  (, ) is defined as the number of pixels with  ∈ 1, . .,  effect at the offset  ⃗⃗ to pixels with values  ∈ 1, . ., , which can be expressed by the following formula [20] : where the offset  ⃗⃗ an angle and pixel distance.Fig.7 [18] shows the four directions of GLCM.Fig. 8 [16] gives an examp le GLCM matrix for angle 0 o with 1-pixel d istance.Notation (0.0) states the relationship of two horizontal rows of pixels with pixel value 0 o followed by pixel value 0 on the right.Arranged matrix fro m this calculation was called GLCM framework matrix.Haralick et al. [19] formu lated 14 texture features that can be extracted fro m GLCM matrix.The notations that will be used are:  p(i,j) is a   ( ,  ) matrix which has been normalized by the size of the row i and column j.  px(i) a new matrix is the su m of all elements in the matrix row p(i,j) or can be writen as where Ng equal with L that is the number of quantization gray levels, that is 256.
1) Angular Second Moment (ASM) ASM is often also referred to as 2 nd order energy or uniformity.

6) Sum Average
7) Sum Entropy Just in case probability may be zero, and the log (0) is not defined, then the value of "log" was added with the arbitrary constant small value (p + e) where e constant is small positive in the calculation of entropy.

8) Sum
Variance where f19 is the greatest value entry element of p(i,j) matrix.
Additional features of GLCM are also derived fro m Clausi [12].

1)
Inverse Difference Inverse Difference Normalized Gray Level Run Length Matrices (GLRLM ) is a method to represent the texture published by Galloway [16].Gray Level is a level value of image intensity at successive either vertically, horizontally or d iagonally.The Run Length is the number of pixels occupied by the value of the intensity level.Fig. 9. GLRLM matrix [16] GLRLM features are represented by the notation p(i,j) refer to matrix M × N consist of row i intensity gray levels and column j of run length in succession in the direct ion of 0 0 (horisontal), 45 0 /135 0 (diagonal) or 90 0 (vert ical).Fig. 9 g ives an examp le o f the format ion of GLRLM matrix of 4x4 image with gray levels 0 to 3 (4 bit) at an angle 0 0 and 45 0 .
Xiaoou Tang [23] defines four new matrix o f GLRLM values to simp lify the texture wh ich are proposed by Galloway [16].

1) Gray Level Run Length Pixel Number Matrix
Based on GLRLM matrix, many textures that can be extracted.There are five texture features which is proposed by Galloway [16].

1) Short Run Emphasis (SRE)
Based on matrix which is proposed by Xiaoou Tang [23] above, the value of SRE in Eq. ( 48) can be simplified into : where nr is the sum total of all value elements GLRLM or it can be written as 2) Long Run Emphasis (LRE) where np s the size of the value matrix GLRLM resolution or number of p ixels contained in the mat rix GLRLM.The value of np can be written as Chu et al [24] added 2 texture features that can be derived from GLRLM matrix as a follow.

Multi Layer Perceptron
Multilayer Perceptron (M LP) is an arrangement of a neural network with programming paradig m that is inspired by the performance of the microstructure of the human brain [26].Artificial neural networks are used extensively in the disciplines of art ificial intelligence for pattern recognition and classification [27].M LP architecture can be seen in Fig. 10.This network consists of a collect ion of neurons (nodes) in a hidden layer and on any ramifications contained weight value will always change as the process of learning.Input layer contains all textural features of a statistical h istogram, GLCM , and GLRLM.Fig. 10.Arcitecture of MLP [28] The amount of hidden layer and neuron can be adapted depend on the complexity of textural classification problems.Each value entered in neurons will p roduce output values by a sigmoid activation function as shown in Eq. ( 62).
Good knowledge is generated by new weights at each branch network at the end of the learning process.Back Error Propagation (BEP) is a supervised learning algorithm used in architecture MLP [26].BEP allows the adjustment of the price weights hidden neurons by means of creeping and turning error output.There are two stages in the learning process of the BEP is the direction forward and backward direction.In the forward direction, the signal propagated fro m the input layer to the hidden layer 1, the hidden layer 2, to the output layer.Then calculated the output of each neuron in each layer with activation function.In the backward direct ion, the error factor p ropagated from the output layer to the hidden layer two hidden layers 1, to the input layer.Then the weight of the branching connection in each layer is calculated and updated.The process is carried out continuously every copies and epoch.A copy is one training p rocess involving of textural features, while one epoch is one training process involving all data.Illustration of the learning process in the forward and backward direction can be seen in Fig. 11.
For each neuron in a layer (eg, yj) is calcu lated sum of the input value with the following formula: As for the output neurons in a layer is calculated activation function in Eq. ( 62), if a =1, the first derivative of this is : The value of yj and yj' are stored for calculation in backward direction.This process is done for the entire layer in the forward direction.The weight wi at every branch is updated on the backward direction following this equation: or where η is learning rate with the value between 0 to 1.And ti is training target which is used for learning reference of the system.Each neuron (eg zi) in backward direction layer are applied in calculat ion of error factor d. Then neurons in the output layer are applied: to neurons in other layers then : (70) and Adjustment of weight factor wij is done by follo wing Eq.(68) to Eq. (71).According to Fig. 10, every weight that is interconnected with neuron zi on backward direction is calculated using the following formula: where zi is the result of the previous calcu lation of the forward direction.The value di is then propagated return for the next layer to the input layer o f the M LP netwo rk.The new weights were adjusted by Eq. (67).Its process continues to next input copies until all epoch is comp leted.The new weights will appear at every train ing then at the end of the process resulting small error.Thus the MLP network is ready for use in testing process with thyroid nodule dataset.

Performance Indices
Validation methods are required to perform quality of features and classification.Performance indexs are determined by the following formulas:  T able 1. Performance index of confusion matrix

CAD System
All o f the steps in Fig. 1 are co mpiled in the executable graphical user interface (CAD) developed on MA TLAB.Its application supported radiologist to identify thyroid nodules composition objectively.CAD was built to decrease dependency of the ultrasound machine and missinterpretation diagnosis among radiolog ist as shown in Fig. 12.

Median Filtering
Pre-processing stage in this study is a screening process to remove the marking area on the ultrasound image with the med ian filter algorith m.Marking area makes the results of feature extract ion less precise.Samp le pre -processing results from the ROI image is displayed in Fig. 13.

Classification and Evaluation
The classification process begins with the preparation of MLP classifier architecture for later conducted training and testing process on the system of identification of thyroid nodules.Determination of the nu mber of hidden layer neurons and MLP architecture that will be applied to the system should also be tested with Weka machine learn ing.Thus thyroid nodule identification system used will be more accurate and reliable in a variety of d ifferent data and keep fit at a time whenever used.Fro m the test of neurons number and layers, it can be obtained MLP arch itecture that consists of one hidden layer which co mposes of 100 neurons.The correlation of hidden layer and neurons number to accuracy of the classification results is shown in Fig. 14 and Fig. 15.As in bar chart, Fig. 14 shows that more and more layers are added to the MLP network system, the accuracy of the test results has narrowed.Accuracy stood at 60% and did not change again begin to layer 4 and so on.Single layer produce the highest accuracy of 90% co mpared with the nu mber of hidden layer to another.Fig. 15 shows that accuracy is not changed even though the number o f neurons increased.Start neurons 40 to 100 unchanged at all.So that the MLP architecture applied in this classification system is a single layer with 100 number of neurons .Fro m the conducted experiment results, the full combination features of the statistical histogram, GLCM and GLRLM has given best result than using each groups separately.Fig. 16 shows the results of each performance.

Accuracy Sensitivity Specificity Precision NPV
The second confusion describes the performance of 10fold cross validation.Each of data has possibility to become the training and testing set.The number of fold determine how much sample data and how many time t rain ing.As shown in Table 3, c ross-validation presented better accuracy than training-testing method.The number of misclassifying data in cross-validation was higher than training-testing method, i.e., 2 FP, and 5 FN.Sensitiv ity also decreased from 89.5% to 87.80%.The better performance could be obtained by some thyroid ultrasound images addition.

Conclusion
The goal of CAD thyroid nodule based on textural features to classify cystic and solid category has been going well.Textural patterns of thyroid nodules on ultrasound image are the main visual characteristic for the radiologist in identification.The calculation of statistical histogram and measurable values of GLCM also GLRLM can describe the textural pattern of cystic or solid nodules accurately.It increases the objectivity of doctors in diagnos ing thyroid nodules and reduces dependence on the ultrasound operator.In future work, it is suggested to use segmentation technique for localizing nodule area specifically.The work of M LP is influenced by some the dataset, as much of it then we get a robust network.Utilities another classifier are also needed to compare the quality of features .

Fig. 2 .
Fig. 2. General Electric P3 Voluson Doppler ultrasound machine Region of interest (ROI) localization was applied to get nodule area.Localized nodul area provide low co mputational cost and giving accurate features.ROI localizat ion is shown

Fig. 11 .
Fig. 11.Forward and backward BEP where TP (True Positive) is the number of positive data on the target that is classified positive on the system.TN (True Negative) is the number of negative data on the target that is classified negatively on the system.FP (False Positive) representation of the amount of negative data on the target that is classified positive systems and FN (False Negative) is the amount of positive data on the target system are classified negatively.above is grouped in confusion matrix table.It shows the relationship between the actual class or target with predicted class.Actual class is mean radiologist diagnosis and clinical pathology report fro m the hospital.Predicted class is output classification fro m CAD.By looking at the value of Accuracy, Sensitivity, Specificity, Precision and Negative Predict ive Value (NPV), the performance of the classification on the classification of thyroid nodules can be obtained.The arrangement of confusion matrix can be seen in Table. 1.

Fig. 12 .
Fig. 12. CAD of thyroid nodule classification based on textural features

Fig. 16 .
Fig. 16.Comparation of textural features performance Two confusions matrix are p resented to evaluate textural features extraction and classification.The first confusion matrix is used to validate MLP learn ing outcomes with testing images.All 72 thyroid images were d ivided into 36 training set and 36 testing set.Its depends on the number of the training set, as much of training set then the CAD become smarter.Table 2 illustrates the result which there are five incorrect classificat ion consists of 2 FP and 3 FN.It means that two cysts are predicted as solids and three solids as cystic.T able 2. T he peformance of testing set