Prediction of suspicious thyroid nodule using artificial neural network based on radiofrequency ultrasound and conventional ultrasound: A preliminary study

Ultrasonics 99 (2019) 105951

Contents lists available at ScienceDirect

Ultrasonics journal homepage: www.elsevier.com/locate/ultras

Prediction of suspicious thyroid nodule using artificial neural network based on radiofrequency ultrasound and conventional ultrasound: A preliminary study

T

Chunrui Liua, Linzhou Xieb, Wentao Konga, Xiaoling Luc, Dong Zhangb, Min Wua, Lijuan Zhangd, ⁎ Bin Yangc, a

Department of Ultrasound, Nanjing Drum Tower Hospital, The Affiliated Hospital of Nanjing University Medical School, Nanjing 210008, China Key Laboratory of Modern Acoustics (MOE), Department of Physics, Collaborative Innovation Center of Advanced Microstructure, Nanjing University, Nanjing 210093, China c Department of Ultrasound, Jinling Hospital, Medical School of Nanjing University, Nanjing 210002, China d Department of Ultrasound, Nanjing Pukou Hospital, Nanjing 210031, China b

A R T I C LE I N FO

A B S T R A C T

Keywords: Thyroid ultrasound Radiofrequency ultrasound Texture analysis Gray Level Co-occurrence Matrix (GLCM) Artificial Neural Network (ANN) Thyroid Imaging Reporting and Data System (TI-RADS)

This study explored the use of backscattered radiofrequency ultrasound signals combined with artificial neural network (ANN) technology to differentiate benign and malignant thyroid nodules, in comparison with conventional ultrasound techniques. The proposed method uses the gray level co-occurrence matrix algorithm and principal component analysis to identify principal characteristics for use as inputs in the ANN. The dataset consisted of 131 ultrasound images, of which 59 were benign and 72 were malignant, as determined by subsequent surgeries. The nodules were divided randomly into training, validation, and testing groups. Receiver operating characteristic curves (ROC) were drawn to compare the diagnostic efficiency of the ANN when applied to radiofrequency and conventional ultrasound images. The sensitivity, specificity, and accuracy of the ANN in predicting malignancy from the radiofrequency ultrasound images were 100, 91.5, and 96.2%, respectively; from conventional ultrasound, the corresponding values were 94.4, 93.2, and 93.9%, respectively. The area under the receiver operating characteristic curve (AUC) was also higher for radiofrequency than conventional ultrasound (AUC = 0.945 vs. 0.917, 95% confidence interval = 0.901–0.998 vs. 0.854–0.979, using a P-value of 0.26). We then classified each nodule into new risk categories according to the output of each sample generated by the proposed method. The malignancy risks in the proposed Categories 3, 4, and 5 were 0, 18.8, and 94.5%, respectively, compared with 0, 55.1, and 88.2% using the American College of Radiology’s Thyroid Imaging Reporting and Data System. Thus, this preliminary study initially indicated that the proposed method of using radiofrequency ultrasound and the ANN was more accurate at predicting malignancy and stratifying thyroid nodules than conventional ultrasound methods, thus offering significant potential to reduce the number of unnecessary thyroid biopsies.

1. Introduction Although thyroid nodules are a common clinical problem, especially in women and the elderly, only approximately 8–16% of thyroid nodules harbor thyroid cancer [1], and physicians must be able to differentiate the few cancerous thyroid nodules from the vast majority of benign ones. Fine needle aspiration (FNA) is the reference standard for the clinical diagnosis of thyroid cancer. This test is invasive and may result in physical and psychological harm to the patients. Furthermore, the test yields an indeterminate, non-diagnostic, or suspicious

⁎

cytological diagnosis in 15–30% of cases [2]. Therefore, FNA is only recommended when nodules are suspected of being malignant, and better noninvasive tools are therefore required for the clinical management of thyroid nodules. In contrast, thyroid ultrasound is inexpensive and noninvasive, and is therefore the primary imaging modality for patients with known or suspected thyroid cancer, which may be incidentally detected on computed tomography (CT), magnetic resonance imaging (MRI), or thyroidal uptake on 18F-flurodeoxyglucose positron emission tomography (18FDG-PET) scan [2]. Various thyroid ultrasound techniques

Corresponding author. E-mail address: [email protected] (B. Yang).

https://doi.org/10.1016/j.ultras.2019.105951 Received 7 August 2018; Received in revised form 18 June 2019; Accepted 23 June 2019 Available online 24 June 2019 0041-624X/ © 2019 Elsevier B.V. All rights reserved.

Ultrasonics 99 (2019) 105951

C. Liu, et al.

the performance of the proposed method for predicting the risk of thyroid carcinoma.

have varying sensitivities and specificities for the diagnosis of malignant thyroid nodules. The American College of Radiology (ACR) established the Thyroid Imaging Reporting and Data System (TI-RADS) to evaluate the risk for malignancy based on the number of suspicious ultrasound features. However, these evaluations are strongly influenced by the subjective judgment of the radiologist with regard to image feature recognition [3,4]. Previous studies have reported new ultrasonic techniques, such as strain elastography, acoustic radiation force impulse imaging, and contrast-enhanced ultrasound; however, these methods remain controversial because their efficacy has not yet been confirmed [5–7]. Elastography may be affected by several factors, including respiration; precompression; and nodule location, size, calcification, and inflammation. This method may yield misleading results in benign nodules with coarse calcifications or cystic or mixed cystic-solid lesions [8]. Contrast-enhanced ultrasound is a promising noninvasive method for the differential diagnosis of benign and malignant thyroid nodules. However, overlapping data indicating both benign and malignant nodules in this method limits the interpretation of tumor microvascularity. Furthermore, nodule size affects contrast-enhanced ultrasound examination and interpretation. For example, nodules smaller than 10 mm do not show vascularization, whereas larger nodules appear hypervascular [9]. Therefore, an objective, noninvasive, and comparatively accurate method of differentiating benign and malignant thyroid nodules remains a high priority. Radiofrequency signals contain a wealth of information and structural detail that are usually lost in the conventional ultrasound images. The radiofrequency ultrasound local estimator is a newly developed tissue characterization tool that extracts original radiofrequency signals from tissues, allowing the actual organization and mechanical properties of the investigated tissue to be distinguished. Preliminary studies of radiofrequency ultrasound have been promising, and the method has been shown have broader prospective applications in identifying prostate and breast cancers and grading fatty liver [10–13]. To date, few studies on radiofrequency ultrasound’s thyroid cancer detection performance have been reported. This paper’s research group previously demonstrated that a nonlinear approach to identifying pathological changes in thyroid nodules based on statistical analyses of radiofrequency ultrasound signals can differentiate malignant from benign thyroid nodules [14]. To the best of our knowledge, nowhere else in the literature has radiofrequency ultrasound image recognition been performed by texture analysis. However, the raw radiofrequency signal, which is currently processed to generate a B-mode image that is optimal for human perception, is likely to contain additional information that is imperceptible to humans but may yet be useful to determine the malignancy of a lesion. This information may be obtained with the application of deep learning, and indeed, artificial neural networks (ANNs) are often used to solve image recognition problems by parallel processing methods. ANNs have several advantages: they do not need to start with a hypothesis or a priori identification of potential key variables; they requires less formal statistical training; they can implicitly detect complex nonlinear relationships between dependent and independent variables, as well as all possible interactions between predictor variables; and there are multiple training algorithms available [15]. Some previous studies have used ANNs for pattern recognition in thyroid nodules [16–20] and thyroiditis [21]. However, these works have been limited to classifying thyroid nodules as only either malignant or benign, without further differentiation into other relevant categories. Therefore, we developed a new method to predict suspicious thyroid nodules using radiofrequency ultrasound and ANN. First, radiofrequency data were acquired to reconstruct radiofrequency ultrasound images using MATLAB software. Then, regions of interest (ROIs) were outlined by a radiologist and textural features of the ROIs were analyzed using the gray-level co-occurrence matrix (GLCM) algorithm and principal component analysis (PCA) to obtain characteristic values, which were then used to train the ANN. Finally, we tested

2. Materials and methods 2.1. Patients This study was approved by the Ethics Committee of the Drum Tower Hospital, Medical School of Nanjing University. From July 2017 to December 2017, 131 patients (77 women and 54 men, mean age of 43 years, age range of 18–77 years) with thyroid nodules were examined preoperatively with their informed consent, using conventional ultrasound and radiofrequency ultrasound. All patients in this study then underwent total thyroidectomy or hemi-thyroidectomy. 2.2. Ultrasound data acquisition A Vinno 70-color Doppler ultrasound system and an X6-16L broadband probe with probe frequencies of 1–25 MHz (Suzhou, China) was used for all examinations. The patients were placed in the supine position. The examinations were performed by one practiced investigator experienced in thyroid ultrasound. Thyroid nodules were first located by two-dimensional ultrasound imaging, and their composition, echogenicity, shape, margin, and echogenic foci were assessed. The size of a lesion was defined as the maximum diameter of multiple cross sections. Color Doppler flow imaging was used to observe blood flow in the lesion. Diagnostic results were scored on the ACR TI-RADS scale [3,4], where TI-RADS 1 indicates a normal thyroid gland; TI-RADS 2 indicates benign conditions (0% malignancy); TI-RADS 3 indicates probably benign nodules (< 5% malignancy); TI-RADS 4 indicates suspicious nodules (5–80% malignancy rate); TI-RADS 5 indicates probably malignant nodules (malignancy > 80%); and TI-RADS 6 indicates biopsy-proven malignant nodules. Then, the ultrasound device was switched to the radiofrequency functional status; the mechanical index (MI), probe depth, and probe frequency were fixed at 1.2, 38 mm, and 10 MHz, respectively, with the thermal index of soft tissue (TIS) at 0.8 and a total gain range between 95 and 110. The time gain compensation (TGC) was kept constant at the near and far fields, and the focus was located at the center of the image. The settings were saved to ensure consistent test conditions. 2.3. Histopathologic analysis Through routine paraffin sectioning and hematoxylin-eosin staining, the histopathologic results were judged by experienced pathologists. Nodular goiter, thyroid adenoma, and thyroiditis conditions were classified as benign. Papillary thyroid carcinoma, follicular thyroid carcinoma and so on classed as malignant. 2.4. Data analysis An overview of the proposed methodology is shown in Fig. 1. 2.4.1. Radiofrequency analysis 2.4.1.1. ROI selection. Each group of ultrasonic radiofrequency signal data consisted of a total of 312 sound beams, and each sound beam contained 2856 sampling points. First, we reconstructed the original radiofrequency signal data into radiofrequency ultrasound images using MATLAB software (MathWorks, Natick, MA). Fig. 2a and c show a malignant thyroid nodule in the left lobe of a 47-year-old female patient and a benign nodule in the left lobe of a 52year-old male patient. Fig. 2b and d show the corresponding ROIs of nodules selected manually by an experienced sonographer. In order to analyze the partial textural features of each radiofrequency ultrasound image, each image was divided into nine equal parts as shown in Fig. 3. 2

Ultrasonics 99 (2019) 105951

C. Liu, et al.

Fig. 1. Flowchart of the proposed method, which combines radiofrequency ultrasound and an artificial neural network (ANN) to detect potentially malignant thyroid nodules. RF = radiofrequency, 2D = 2-dimensional, ROI = region of interest, TI-RADS = Thyroid Imaging Reporting and Data System.

levels of two pixels separated by a certain distance can be correlative. Thus, with a given offset, the GLCM was used in this study to describe the textural features of the ultrasound images. The features of the GLCM can be described by the following scalar quantities.

Fig. 3a and b illustrate the ROI divisions in Fig. 2b and d; Fig. 3c and d are the corresponding histopathologic images, where Fig. 3c is a papillary thyroid carcinoma (X40), and Fig. 3d is a nodular goiter (X40). 2.4.1.2. Textural feature analysis. The gray-level co-occurrence matrix (GLCM) algorithm is a texture analysis method that extract texture information from images by identifying similar gray-level pixels that are distributed continuously in a space [22,23]. In addition, the gray

(1) The angular second moment (ASM) describes the distribution of gray level degree, where P (i,j) is the normalized gray-level co-

Fig. 2. Radiofrequency ultrasound images reconstructed from the original radiofrequency data using MATLAB software: (a) malignant thyroid nodule, (b) selected ROI for the malignant thyroid nodule, (c) benign thyroid nodule, and (d) selected ROI for the benign thyroid nodule. 3

Ultrasonics 99 (2019) 105951

C. Liu, et al.

Fig. 3. ROI division and histopathologic images: (a) division of ROI in Fig. 2b, (b) division of ROI in Fig. 2d, (c) histology for Fig. 2a (X40), and (d) histology for Fig. 2c (X40). k

occurrence matrix.

ASM =

k

μx =

k

∑i =1 ∑ j=1 P (i, j)2

i=1 j=1

k

CON =

μy =

k

∑ ∑ j·P (i, j) i=1 j=1

∑i =1 ∑ j=1 (i − j)2P (i, j)

k

(2)

sx2 =

k

∑ ∑ P (i, j) × (i − μx )2 i=1 j=1

k

k

∑i =1 ∑ j=1

P (i, j ) 1 + (i − j )2

k

s y2 =

k

ENT = − ∑

i=1

(4)

X=

k

[ ∑i = 1 ∑ j = 1 (ij ) P (i, j )] − μx μ y sx s y

(9)

2.4.2. Characteristic values analysis After solving the GLCM of the image, the characteristic values of the nine parts had to be analyzed, which required the execution of extensive calculations. Therefore, PCA, which is a widely used dimension reduction method in image analysis, was used here to identify the most important characteristic values. The given sample matrix was:

(5) Correlation (COR) indicates the consistency of the image texture over a certain distance. k

k

∑ ∑ P (i, j) × (j − μy )2 i=1 j=1

k

∑ j=1 P (i, j)log(P (i, j))

(8)

(3)

(4) Entropy (ENT) indicates the randomness and irregularity of all pixel intensity.

COR =

(7)

k

(3) The inverse different moment (IDM) indicates the change of the gray level in a certain part.

IDM =

(6)

(1)

(2) The contrast (CON) describes the image resolution. k

k

∑ ∑ i·P (i, j)

x ⋯ x1n ⎞ ⎛ 11 ⋮ ⋱ ⋮ ⎟ ⎜ x x ⋯ mn ⎠ ⎝ m1

(10)

(5) Zero-mean normalization was used to normalize the matrix. The normalized matrix became:

In Eq. (5), 4

Ultrasonics 99 (2019) 105951

C. Liu, et al.

Cmn =

c ⋯ c1n ⎛ 11 ⎞ ⋮ ⋱ ⋮ = (p1 , p2 , ⋯, pn ) ⎜ ⎟ c ⋯ c m 1 mn ⎝ ⎠

Table 1 Characteristics of patients and thyroid nodules.

(11)

Benign

Malignant

Patients (n) % Female TSH (mU/L) Age Nodule size (mm)

73.1 1.5 ± 1.1 45.3 ± 11.2 20.1 ± 9.3

45.9 1.7 ± 0.9 47.9 ± 7.4 12.9 ± 5.9

The covariance matrix can be obtained from eigenvectors, and was here used as a transformation matrix to reduce the characteristic values.

Nodule location Left lobe Right lobe

34 25

40 32

⎛ cov (p1 , p1 ) ⋯ cov (p1 , pn ) ⎞ ⋮ ⋱ ⋮ Σ=⎜ ⎟ ⎜ cov (p , p ) ⋯ cov (p , p ) ⎟ n 1 n n ⎠ ⎝

ACR TI-RADS 3 4 5

31 22 6

0 27 45

Where

c ⎛ 1i ⎞ c2i ⎜ ⎟, i ∈ {1, 2, ⋯, n} pi = ⎜⋮⎟ ⎝ cmi ⎠

(12)

(13)

The eigenvalues of the covariance matrix were ranked:

ACR TI-RADS = American College of Radiology Thyroid Imaging Reporting and Data System; TSH = thyroid stimulating hormone

(14)

λ1 ≥ λ2 ≥ ⋯≥λn The k-th principal component contribution rate ak was:

λk ak = λ1 + λ2 +⋯+λn

Table 2 Ultrasonographic features of the thyroid nodules.

(15)

The top k accumulative contribution rate Σp was:

Σp =

k ∑i = 1

λi

λ1 + λ2 +⋯+λn

(16)

2.4.3. Pattern recognition In this study, MATLAB’s neural net pattern recognition tool (NPRTOOL) was used to distinguish benign thyroid nodules from malignant ones with script and parameters established for this purpose. The Levenberg–Marquardt algorithm was used to train the ANN and mean squared error (MSE) was chosen as the performance function. After training, the feed forward ANN shown in Fig. 4 was created, with three layers, 21 input features, 10 neurons in the hidden layer, and two output classes. The network included a sigmoid transfer function in the hidden layer, and a SoftMax transfer function in the output layer. W was the weight matrix and b was an offset vector. The number of neurons in the hidden layer was calculated by the empirical formula:

m=

n+l +α

Benign

Malignant

Composition

Mixed cystic and solid Solid or almost completely solid

6 53

1 71

Echogenicity

Hyperechoic or isoechoic Hypoechoic Very hypoechoic

12 47 0

2 60 10

Shape

Wider-than-tall Taller-than-wide

59 0

53 19

Margin

Smooth Lobulated or irregular margin Extensive extrathyroidal extension

55 3 1

3 49 20

Echogenic foci

None Macrocalcifications Peripheral (rim) calcifications Punctate echogenic foci

41 6 5 7

31 4 0 37

significant difference. To verify our hypothesis that thyroid radiofrequency ultrasound can provide more tissue characteristic information than conventional ultrasound imaging techniques, texture and ANN analyses were also performed on B mode ultrasound images. The ROI selections were determined by the same ultrasound physician.

(17)

where m is the number of neurons in the hidden layer, n is the number of inputs in the input layer, l is the number of outputs in the output layer, and α is a constant between one and 10.

3. Results 3.1. Ultrasonographic features and histological results

2.4.4. Statistical analysis Quantitative data were expressed as means with standard deviations. The diagnostic efficacy of radiofrequency ultrasound and ultrasound ANN was assessed based on the receiver operating characteristic curve (ROC) and calculations of the sensitivity, specificity, and accuracy with regard to predicting the malignancy of thyroid nodules. For all tests, a p-value < 0.05 was considered to indicate a statistically

Among the 131 thyroid nodules that were surgically removed, 59 were benign (46 nodular goiter and 13 thyroid adenoma) and 72 were malignant (all papillary thyroid carcinomas). The characteristics of the patients and thyroid nodules are in Table 1. The Ultrasonographic features of the thyroid nodules are presented in Table 2.

Fig. 4. Feed forward ANN used in the proposed method. 5

Ultrasonics 99 (2019) 105951

C. Liu, et al.

3.3. ANN training and test results

Table 3 21 principal components in radiofrequency ultrasound imaging. Principal component

Contribution rate

Accumulative contribution rate

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

34.35741 16.46211 7.745968 6.187108 5.716695 3.672446 3.454407 2.630857 2.110509 1.971752 1.76026 1.507122 1.386126 1.06412 0.971929 0.917456 0.801567 0.768166 0.63404 0.556221 0.529509

34.35741 50.81952 58.56549 64.7526 70.46929 74.14174 77.59615 80.227 82.33751 84.30926 86.06952 87.57665 88.96277 90.02689 90.99882 91.91628 92.71784 93.48601 94.12005 94.67627 95.20578

131 samples were chosen randomly for the neural net training. They were divided randomly into three groups: 91 in training groups, 20 in validation groups, and 20 in testing groups (Fig. 4). Tables 5 and 6 show the diagnostic efficiencies of ANN applied to the different groups of both radiofrequency and conventional ultrasound images. The sensitivity, specificity, and accuracy of the ANN in predicting malignant thyroid nodules from radiofrequency ultrasound images were 100, 91.5, and 96.2%, respectively; from conventional ultrasound images, the corresponding values were 94.4, 93.2, and 93.9%, respectively. The overall accuracy of the ANN applied to conventional ultrasound was 93.9%, which is 2.3% lower than the accuracy of 96.2% when applied to radiofrequency ultrasound.

3.4. Area under the receiver operating characteristic curve To verify our hypothesis that thyroid radiofrequency ultrasound can provide more tissue characteristic information than conventional ultrasound imaging techniques, Fig. 5 shows the area under the receiver operating characteristic curve (AUC) for both radiofrequency (AUC = 0.945, 95% confidence interval: 0.901–0.998) and conventional ultrasound (AUC = 0.917, 95% confidence interval: 0.854–0.979). Although the analysis showed higher values for radiofrequency ultrasound, there was no statistically significant difference, with p = 0.26.

Table 4 21 principal components in conventional ultrasound imaging. Principal component

Contribution rate

Accumulative contribution rate

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

22.68251515 15.21078459 9.085176354 6.919928621 6.775661798 4.462871702 3.958133432 3.12436334 2.846899206 2.546828149 2.024027704 1.974354076 1.855036074 1.485172688 1.306226706 1.206554501 1.035420479 0.97202431 0.86771033 0.755258938 0.704751094

22.68252 37.8933 46.97848 53.8984 60.67407 65.13694 69.09507 72.21943 75.06633 77.61316 79.63719 81.61154 83.46658 84.95175 86.25798 87.46453 88.49995 89.47198 90.33969 91.09495 91.7997

3.5. Classify the new risk categories According to the 2017 ACR TI-RADS classification [3,4], TI-RADS 1–6 correspond to different degrees of malignant risk from low to high. Based on the ANN output of this study, we attempted to reclassify the thyroid nodules for comparison with the 2017 ACR TI-RADS categories. The output of the ANN is a vector: output (a1, a2), a1 + a2 = 1.5, a1 ∈ [0.5, 1]. If a1 → 0.5, the input appears as a benign case. If a1 → 1, the input appears as a malignant case. We therefore considered that we could classify the thyroid nodules according to the ANN output. For 0 ≤a1 − 0.5 ≤ 10−9 , the thyroid nodules were classified as Category 3, in which the risk of malignancy was expected to be 0–5%. For 10−9 < a1 − 0.5 ≤ 0.5 − 10−9 , the thyroid nodules were classified as Category 4, in which the risk of malignancy was expected to be 5–80%. For 0.5 − 10−9 < a1 − 0.5 ≤ 0.5 the thyroid nodules were classified as Category 5, in which the risk of malignancy was expected to be 80–100%. The malignancy risks in the new Categories 3, 4, and 5 were 0, 18.8, and 94.5%, respectively (Table 7), in comparison with the ACR TIRADS values, where the malignancy risks were 0, 55.1, and 88.2% respectively (Table 1).

3.2. Textural feature results The vector acquired from GLCM had 180 dimensionalities, and PCA was used to reduce the dimensionalities of the vector. The 21 principal components obtained from radiofrequency and conventional ultrasound are shown in Tables 3 and 4, respectively.

Table 5 Diagnostic efficiencies of ANN applied to radiofrequency ultrasound images. Training group

Benign Malignant Total Sensitivity Specificity Accuracy

Validation group

Test group

All in ANN

Benign

Malignant

Benign

Malignant

Benign

Malignant

Benign

Malignant

40 3 43

0 48 48 100% 93.0% 96.7%

8 0 8

0 12 12 100% 100% 100%

6 2 8

0 12 12 100% 75.0% 90%

54 5 59

0 72 72 100% 91.5% 96.2%

6

Ultrasonics 99 (2019) 105951

C. Liu, et al.

Table 6 Diagnostic efficiencies of ANN applied to conventional ultrasound images. Training group

Benign Malignant Total Sensitivity Specificity Accuracy

Validation group

Test group

All in ANN

Benign

Malignant

Benign

Malignant

Benign

Malignant

Benign

Malignant

39 2 41

2 48 50 96.0% 95.1% 95.6%

9 1 10

1 9 10 90.% 90.0% 90.0%

7 1 8

1 11 12 91.7% 87.5% 90.0%

55 4 59

4 68 72 94.4% 93.2% 93.9%

4. Discussion 4.1. ANN The accuracies of the proposed ANN at predicting malignancy in the training, validation, and test groups were 96.7, 100, and 90%, respectively, using the radiofrequency ultrasound images; and 95.6, 90, and 90.0%, respectively, using the conventional ultrasound images. The ANN could rapidly recognize linear patterns; non-linear patterns with threshold impacts; and categorical, step-wise linear, or even contingency effects. Different algorithms such as the Levenberg–Marquardt algorithm, cost-sensitive random forest, and genetic algorithms are widely used in various ANNs, with no uniform standard for choosing a network structure. In any ANN, with increasing numbers of samples, the accuracy will also increase. Therefore, although the number of samples in this study was large enough, the classifications of the thyroid nodules would be more reliable and the classifying efficiency further improved with additional samples [15,16,19,24,25]. We can calculate the number of neurons in different layers based on Eq. (17). In order to identify all of the malignant samples, we prioritized ANN training to discern as many malignant thyroid nodules as possible, instead of optimizing performance for both malignant and benign samples. Our experiments showed that the network with 10 neurons performed the best at the given task (Table 8). Although the ANN will fail with insufficient neurons, excessive neurons in the hidden layer increases the computation cost and can result in over-fitting. The complexity of the Levenberg–Marquardt algorithm in the ANN training is O ((M3 /6) ), where M is the dimension of the weight matrix W in Fig. 4. Thus, the computation time will be extremely long for large values of M, and therefore the PCA algorithm was used to reduce the computation time for the ANN training. For comparison, the ANN required approximately 18.16 and 390.74 s to train 100 times with and without the PCA algorithm. Therefore, the PCA algorithm clearly improved the efficiency of ANN training, which would be especially relevant for large data sets.

4.2. New risk categories In contrast to the diagnostic methods currently available, the output of the proposed ANN can be classified into the newly proposed risk categories. Most previous works have used texture-based features in combination with a support vector machine (SVM) classifier in order to accomplish the tasks of nodule identification and classification based on their malignancy risk and characterization of the type of malignancy [16]. TI-RADS was established to categorize thyroid nodules and stratify their malignancy risk. The sonographic TI-RADS lexicon includes sonographic descriptors for composition, echogenicity, shape, margin, and echogenic foci, and the number and complexity of these sonographic factors have a considerable impact on inter-observer concordance and diagnostic performance [3]. In this study, we classified the output of different radiofrequency ultrasound and ANN samples into different categories according to malignancy risk. The new

Fig. 5. Analysis of the area under the receiver operating characteristic curve (AUC) for: (a) radiofrequency ultrasound and (b) conventional ultrasound.

7

Ultrasonics 99 (2019) 105951

C. Liu, et al.

suspicious thyroid nodules using radiofrequency ultrasound images combined with an ANN, in comparison with the same ANN applied to conventional ultrasound images. The sensitivity, specificity, and accuracy of the ANN in predicting the malignancy of thyroid nodules were 100, 91.5, and 96.2%, respectively, when applied to radiofrequency ultrasound images; when applied to conventional ultrasound images, the corresponding values were 94.4, 93.2, and 93.9%, respectively. The AUC for radiofrequency ultrasound was higher than for conventional ultrasound. In addition, each nodule was classified into new malignancy risk categories according to the ANN output. Of course, although these preliminary results suggested the proposed method could help sonographers to identify risky thyroid nodules and reduce the number of unnecessary thyroid biopsies, more data will be collected and analyzed in our future study to further confirm the feasibility and accuracy of the proposed method.

Table 7 Malignancy risks of the newly proposed categories based on radiofrequency ultrasound and ANN results.

Category 3 Category 4 Category 5

Number of samples

Number of malignant samples

Risk of malignancy

42 16 73

0 3 69

0 18.8 94.5%

Table 8 Dependence of ANN performance on the number of neurons. Number of neurons

5

6

7

8

9

10

11

12

13

14

PPV (%) Accuracy (%)

96.2 91.6

98.2 95.4

94.7 93

96.5 95.4

94.5 95.4

100 96.2

100 96.9

96.5 95.4

96.6 96.9

98.2 96.2

Acknowledgement

PPV = positive predictive value.

This work was partially supported by the Jiangsu Province Key Research & Development Plan (No. BE2018703), the QingLan Project, and the Fundamental Research Funds for the Central Universities (No. 14380120). This study was conducted with the informed consent of each study participant.

Categories 3, 4, and 5 were characterized by malignancy risks of 0, 18.8, and 94.5%, respectively, compared with 0, 55.1, and 88.2% for TI-RADS. Thus, the new categories can better distinguish the malignant nodules, especially for the indeterminate TI-RADS 4 category. Although the ANN output value of any given nodule can vary, to a certain extent, there is a one-to-one correspondence between the ANN output value and the incidence of malignancy. The new categories allow for a selection of suspicious nodules to be submitted to FNA, thereby avoiding unnecessary thyroid biopsies.

Appendix A. Supplementary material Supplementary data to this article can be found online at https:// doi.org/10.1016/j.ultras.2019.105951. References

4.3. Benefits and limitations

[1] K.D. Burman, L. Wartofsky, Thyroid nodules, N. Engl. J. Med. 373 (2015) 2347–2356. [2] B.R. Haugen, E.K. Alexander, K.C. Bible, G. Doherty, S.J. Mandel, Y.E. Nikiforov, F. Pacini, G. Randolph, A. Sawka, M. Schlumberger, 2015 American thyroid association management guidelines for adult patients with thyroid nodules and differentiated thyroid cancer, Thyroid (2015). [3] E. Horvath, S. Majlis, R. Rossi, C. Franco, J.P. Niedmann, A. Castro, M. Dominguez, An ultrasonogram reporting system for thyroid nodules stratifying cancer risk for clinical management, J. Clin. Endocrinol. Metab. 94 (2009) 1748–1751. [4] F.N. Tessler, W.D. Middleton, E.G. Grant, J.K. Hoang, L.L. Berland, S.A. Teefey, J.J. Cronan, M.D. Beland, T.S. Desser, M.C. Frates, L.W. Hammers, U.M. Hamper, J.E. Langer, C.C. Reading, L.M. Scoutt, A.T. Stavros, ACR Thyroid Imaging, Reporting and Data System (TI-RADS): white paper of the ACR TI-RADS committee, J. Am. College Radiol.: JACR 14 (2017) 587–595. [5] A.E. Samir, M. Dhyani, A. Anvari, J. Prescott, E.F. Halpern, W.C. Faquin, A. Stephen, Shear-wave elastography for the preoperative risk stratification of follicular-patterned lesions of the thyroid: diagnostic accuracy and optimal measurement plane, Radiology 277 (2015) 565–573. [6] Q. Peng, C. Niu, M. Zhang, Q. Peng, S. Chen, Sonographic characteristics of papillary thyroid carcinoma with coexistent Hashimoto's thyroiditis: conventional ultrasound, acoustic radiation force impulse imaging and contrast-enhanced ultrasound, Ultrasound Med. Biol. 45 (2019) 471–480. [7] F. Gu, L. Han, X. Yang, H. Liu, X. Li, K. Guo, Z. Zhao, X. Zhou, W. Luo, Value of timeintensity curve analysis of contrast-enhanced ultrasound in the differential diagnosis of thyroid nodules, Eur. J. Radiol. 105 (2018) 182–187. [8] D. Cosgrove, R. Barr, J. Bojunga, V. Cantisani, M.C. Chammas, M. Dighe, S. Vinayak, J.-M. Xu, C.F. Dietrich, WFUMB guidelines and recommendations on the clinical use of ultrasound elastography: part 4. Thyroid, Ultrasound Med. Biol. 43 (2017) 4–26. [9] P.S. Sidhu, V. Cantisani, C.F. Dietrich, O.H. Gilja, A. Saftoiu, E. Bartels, M. Bertolotto, F. Calliada, D.A. Clevert, D. Cosgrove, A. Deganello, M. D'Onofrio, F.M. Drudi, S. Freeman, C. Harvey, C. Jenssen, E.M. Jung, A.S. Klauser, N. Lassau, M.F. Meloni, E. Leen, C. Nicolau, C. Nolsoe, F. Piscaglia, F. Prada, H. Prosch, M. Radzina, L. Savelli, H.P. Weskott, H. Wijkstra, The EFSUMB guidelines and recommendations for the clinical practice of Contrast-Enhanced Ultrasound (CEUS) in non-hepatic applications: update 2017 (long version), Ultraschall Der Med. 39 (2018) E2-E44. [10] F. Imani, P. Abolmaesumi, E. Gibson, A. Khojaste, M. Gaed, M. Moussa, J.A. Gomez, C. Romagnoli, M. Leveridge, S. Chang, D.R. Siemens, A. Fenster, A.D. Ward, P. Mousavi, Computer-aided prostate cancer detection using ultrasound RF time series: in vivo feasibility study, IEEE Trans. Med. Imag. 34 (2015) 2248–2257. [11] M. Moradi, P. Abolmaesumi, D.R. Siemens, E.E. Sauerbrei, A.H. Boag, P. Mousavi, Augmenting detection of prostate cancer in transrectal ultrasound images using SVM and RFR time series, IEEE Trans. Biomed. Eng. 56 (2009) 2214–2224. [12] J. Wang, C.S. Kang, X.F. Liu, T.T. Li, Y. Wang, T.H. Feng, Z.J. Li, J.P. Xue, K.L. Shi, Clinical value of radiofrequency ultrasonic local estimators in classifying breast

The main benefits of this study are summarized as follows. (1) The proposed method has no operator dependency; all of the analyses are performed by computer. (2) According to several evaluation criteria (accuracy, sensitivity, specificity, and AUC), the results are robust and reproducible because the proposed method is based on texture analysis and an ANN strategy. (3) The newly formulated categories can differentiate benign and malignant classes using a single value. However, this study had the following limitations. (1) The number of thyroid ultrasound images we used to train the model was not sufficient for a large-scale data analysis, and the testing group data also came from the same group of patients at the same hospital, limiting the generalization of the results reported here. (2) Because not all types of thyroid nodules were included in the evaluated data, more samples are needed to fully evaluate the ANN model. (3) The ROIs in the study were selected manually by a sonographer and therefore depended on sonographer’s experience with the echographic appearance of thyroid nodules, which may lead to subjective deviation that could affect the output of the ANN. (4) Considering the huge data sets produced in radiofrequency ultrasound imaging, the computation time must be further reduced for practical applications. An optimal feature selection approach will be investigated in future works to reduce the number of parameters, perhaps by an advanced classifier such as an SVM. (5) For conventional ultrasound images, the performance of ANN may depend on the quality of the images provided by the ultrasound device. However, in additional to conventional ultrasound images, radiofrequency ultrasound signals that is independent on post image processing procedures adopted by different ultrasound devices. These preliminary results indicated that the performance of ANN combined with radiofrequency ultrasound signals is better than that combined with conventional ultrasound images. 5. Conclusion The present work proposed a method for the prediction of 8

Ultrasonics 99 (2019) 105951

C. Liu, et al.

lesions, J. Ultras Med. 32 (2013) 83–92. [13] J. Wang, C. Kang, T. Feng, J. Xue, K. Shi, T. Li, X. Liu, Y. Wang, Effects of instrument settings on radiofrequency ultrasound local estimator images: a preliminary study in a gallbladder model, Mol. Med. Rep. 8 (2013) 995–998. [14] H. Xu, C. Liu, P. Yang, J. Tu, B. Yang, D. Zhang, A nonlinear approach to identify pathological change of thyroid nodules based on statistical analysis of ultrasound RF signals, Sci. Rep. 7 (2017) 16930. [15] L.-C. Zhu, Y.-L. Ye, W.-H. Luo, M. Su, H.-P. Wei, X.-B. Zhang, J. Wei, C.-L. Zou, A model to discriminate malignant from benign thyroid nodules using artificial neural network, Plos One 8 (2013). [16] J. Chi, E. Walia, P. Babyn, J. Wang, G. Groot, M. Eramian, Thyroid nodule classification in ultrasound images by fine-tuning deep convolutional neural network, J. Digit. Imag. 30 (2017) 477–486. [17] J. Ma, F. Wu, T. Jiang, Q. Zhao, D. Kong, Ultrasound image-based thyroid nodule automatic segmentation using convolutional neural networks, Comput. Meth. Prog. Bio. 12 (2017) 1895–1910. [18] J.L. Ma, F. Wu, J. Zhu, D. Xu, D.X. Kong, A pre-trained convolutional neural network based method for thyroid nodule diagnosis, Ultrasonics 73 (2017) 221–230. [19] J. Ma, F. Wu, T.a. Jiang, J. Zhu, D. Kong, Cascade convolutional neural networks for automatic detection of thyroid nodules in ultrasound images, Med. Phys. 44 (2017) 1678–1691. [20] B.L. Guo, F.S. Ouyang, L.Z. Ouyang, Z.W. Liu, S.J. Lin, W. Meng, X.Y. Huang,

[21]

[22]

[23]

[24]

[25]

9

H.X. Chen, S.M. Yang, Q.G. Hu, Development and validation of an ultrasound-based nomogram to improve the diagnostic accuracy for malignant thyroid nodules, Eur. Radiol. 29 (2019) 1518–1526. Y.G. Shin, J. Yoo, H.J. Kwon, J.H. Hong, H.S. Lee, J.H. Yoon, E.K. Kim, H.J. Moon, K. Han, J.Y. Kwak, Histogram and gray level co-occurrence matrix on gray-scale ultrasound images for diagnosing lymphocytic thyroiditis, Comput. Biol. Med. 75 (2016) 257–266. S.J. Chen, S.N. Yu, J.E. Tzeng, Y.T. Chen, K.Y. Chang, K.S. Cheng, F.T. Hsiao, C.K. Wei, Characterization of the major histopathological components of thyroid nodules using sonographic textural features for clinical diagnosis and management, Ultrasound Med. Biol. 35 (2009) 201–208. S.J. Chen, C.Y. Chang, K.Y. Chang, J.E. Tzeng, Y.T. Chen, C.W. Lin, W.C. Hsu, C.K. Wei, Classification of the thyroid nodules based on characteristic sonographic textural feature and correlated histopathology using hierarchical support vector machines, Ultrasound Med. Biol. 36 (2010) 2018–2026. Q. Yu, T. Jiang, A. Zhou, L. Zhang, C. Zhang, P. Xu, Computer-aided diagnosis of malignant or benign thyroid nodes based on ultrasound images, Eur. Arch. Otorhinolaryngol. 274 (2017) 2891–2897. K.J. Lim, C.S. Choi, D.Y. Yoon, S.K. Chang, K.K. Kim, H. Han, S.S. Kim, J. Lee, Y.H. Jeon, Computer-aided diagnosis for the differentiation of malignant from benign thyroid nodules on ultrasonography, Acad. Radiol. 15 (2008) 853–858.