1. Introduction

Thyroid nodules are found with increasing frequency, likely due to the widespread use of imaging modalities including ultrasonography (US) and computed tomography (CT). Although more than 95% of the detected nodules are clinically insignificant benign lesions, thyroid nodules are clinically important as they may represent thyroid malignant tumor [1, 2]. Despite an increase in the prevalence of malignant thyroid nodules and invasive examinations, the mortality rate has remained unchanged from 1973; this observation is attributed to overdiagnosis of subclinical thyroid nodules [3, 4]. Therefore, management of incidental thyroid nodules on cross-sectional imaging has been discussed [5–7].

With the recent spread of dual-energy CT (DECT) or spectral CT scanner, dual-energy imaging is widely used in daily practice. DECT offers information on material decomposition, which is acquired from specific changes in the attenuation of each material at two different tube voltages [8–10]. The resulting images provide a spectral attenuation curve plotted within 40 to 140 keV, which allows for characterization of tissues [8–10]. Previously, iodine uptake in thyroid tissues has been known to be associated with thyroid peroxidase (TPO) and sodium iodine symporter (NIS) [11–13]. Li et al. have demonstrated that quantitative ex vivo evaluations using noncontrast DECT have demonstrated a significant difference between benign and malignant thyroid lesions, most likely because cell membranes in normal follicular epithelial cells contain iodine transporters [14]. Therefore, it can be predicted that malignant thyroid nodules would contain less iodine and have lower attenuation on monochromatic images than benign nodules.

With the development of machine learning and the increasing availability of data sets, a quantitative mathematical approach has gained an important role in image diagnosis. Texture analyses based on gray-level changes offer objective features for the quantitative characterization of tissues. Using US and MRI, Acharya et al. [15] and Hao et al. [16] demonstrated significant differences in the texture features of benign and malignant thyroid tumors. We hypothesized that additional quantitative information gained by using multienergy virtual monochromatic image (VMI) data sets could improve the predictive performance of texture analysis.

Hence, the purpose of this study was to investigate the value of quantitative DECT analysis for differentiating between benign and malignant thyroid nodules in vivo, using two approaches: (1) iodine decomposition measurement and slope of CT attenuation and (2) texture analysis of monochromatic images (40 keV, 60 keV, and 80 keV).

2. Materials and Methods

This retrospective study was approved by our institutional review board, which waived the need for informed consent from patients.

2.3. Measurement of Iodine Decomposition and CT Attenuation Slope

Scan data were reconstructed as 3 mm thick axial images without overlapping of section intervals. Iodine material decomposition images at 120 kVp and monochromatic images at a range of 40 to 190 keV were reconstructed. We then selected CT slices that demonstrated the maximum diameter of the solid component within nodules of greater than or equal to 10 mm based on the minimum standard value for the fine-needle aspiration in Thyroid Imaging Reporting and Data System by American College of Radiology [18]. Regions of interest (ROIs) were manually drawn for each slice, under consensus between two radiologists (HT and KS, who had 5 and 7 years of experience in head and neck radiology, respectively). ROIs were drawn to include as much of the lesions, whilst avoiding cysts and overt calcifications. The following quantitative parameters were measured: [1] iodine concentration (mg/mL) at 120 kVp and [2] CT attenuation curve slope, plotted from 40 keV to 190 keV (Figure 1).

2.4. Image Segmentation and Texture Analysis

Thyroid nodules on 40 keV, 60 keV, and 80 keV monochromatic images were manually delineated by an attending radiologist (∗blinded∗, who had 5 years of experience in head and neck radiology) using OsiriX Imaging Software (http://www.osirix-viewer.com). Segmentation was performed on the image set where the tumor was best seen (typically 40 keV VMIs, which has been shown to improve visibility of other head and neck tumors) [19, 20], in conjunction with VMIs at other energies if needed. The solid component was carefully defined to avoid obvious cystic tissue and gross calcification. When severe artifacts were seen within nodules, only artifact-free slices were used for texture analysis. The same ROIs were used to generate VMI data sets for each patient at all 3 energies, taking advantage of the natural coregistration of VMI data sets produced from the same DECT acquisition. A total of 41 features, including 12 histogram features, 5 gray-level co-occurrence matrix (GLCM) features, 11 gray-level run-length (GLRL) features, 4 gray-level gradient matrix (GLGM) features, and 9 Law’s features were extracted from each contour, using in-house developed MATLAB-based texture analysis software (MathWorks, Natick, Massachusetts). A detailed summary of all textual features in details measured is shown in Appendix 1, with mean values presented in Figure 2.

3. Results

3.2. Texture Analysis for Differentiating from Benign and Malignant Thyroid Nodules

Appendix 2 summarizes measurements of 41 texture features at 40, 60, and 80 keV using DECT. Significant differences between benign and malignant thyroid nodules were observed between the histogram mean for 80 keV (P = 0.019), histogram median at 40 and 80 keV (P = 0.034 and 0.001, respectively), GLCM contrast for 80 keV (P = 0.031), GLGM skewness at 40 and 80 keV (P = 0.030, respectively), and mean gradients (MGR) and gradient variance (VGR) in GLGM at 40, 60, and 80 keV (P = 0.028–0.044). The AUC, sensitivity, and specificity of selected texture features are shown in Table 2. Based on ROC analysis, the most effective discrimination between benign and malignant nodules was provided by the histogram mean and median at 80 keV, with AUCs of 0.77 for both, sensitivities of 57.1 and 71%, and specificities of 85 and 75%, respectively. Figure 3 (malignant case) and Figure 4 (benign case) present a representative case in which selected texture features can differentiate between benign and malignant thyroid nodules at high keV on DECT, whereas material decomposition data were unable to discriminate between them.

Table 2. Diagnostic performance of selected texture features to differentiate benign thyroid nodules from malignant nodules.

Texture parameter	Benign lesions (n = 14)		Malignant lesions (n = 20)		P value	Cut-off	Sensitivity (%)	Specificity (%)	AUC
	Mean	SD	Mean	SD
Mean in histogram
40 keV	1124.8	39.6	1091.3	52.3	0.052	N/A	N/A	N/A	0.682
60 keV	1091.9	26.2	1059.8	94.5	0.163	N/A	N/A	N/A	0.664
80 keV	1085.7	18.5	1071.9	9.4	0.019 ∗	<1080.9	57.1	85.0	0.771¶
Median in histogram
40 keV	1124.8	39.2	1099.7	27.1	0.034 ∗	<1112.0	64.3	85.0	0.686
60 keV	1092.8	26.0	1072.6	39.4	0.103	N/A	N/A	N/A	0.679
80 keV	1086.7	18.6	1071.9	9.5	0.014 ∗	<1080.9	71.0	75.0	0.771¶
Contrast in GLCM
40 keV	23.6	18.9	18.0	12.5	0.309	N/A	N/A	N/A	0.629
60 keV	21.3	20.1	17.2	27.1	0.635	N/A	N/A	N/A	0.629
80 keV	18.5	15.3	8.4	4.3	0.031 ∗	<15.12	50.0	85.0	0.693
Skewness in GLGM
40 keV	36.3	12.3	28.6	7.6	0.030 ∗	<37.00	50.0	85.0	0.696
60 keV	36.0	12.2	28.8	7.4	0.064	N/A	N/A	N/A	0.671
80 keV	36.4	12.0	28.8	7.4	0.030 ∗	<29.49	78.6	65.0	0.696
MGR in GLGM
40 keV	5.792	3.559	9.340	5.805	0.035 ∗	>3.301	95.0	35.7	0.689
60 keV	5.041	2.880	7.993	4.479	0.038 ∗	>3.692	85.0	50.0	0.693
80 keV	4.850	2.696	7.615	4.199	0.038 ∗	>3.639	85.0	50.0	0.704
VGR in GLGM
40 keV	19974.2	10745.3	30478.1	16424.1	0.044 ∗	>15036	85.0	42.4	0.689
60 keV	18824.2	10530.9	28949.7	15235.9	0.039 ∗	>12098	95.0	35.7	0.696
80 keV	17957.5	10154.8	28800.9	15431.2	0.028 ∗	>14288	85.0	50.0	0.721

• GLCM = gray-level co-occurrence matrix; GLGM = gray-level gradient matrix; MGR = mean gradients; VGR = a variance of gradients; AUC = area under receiver operating characteristic curve;  ^∗significant differences are defined as P < 0.05; ^¶highest AUC among 41 texture features.

4. Discussion

Six texture features on three monochromatic images met the desired performance standard for distinguishing between benign and malignant thyroid nodules in vivo. Notably, DECT material decomposition and CT attenuation slopes did not show significant differences. These results suggest that this approach to texture analysis using extracted monochromatic images may be a valuable imaging tool for diagnosing thyroid tumors using noncontrast CT. Ultrasound (US) has been used to assess thyroid nodules, and US-guided biopsies are a standard procedure in the establishment of the diagnosis of benign and malignant thyroid nodules. However, US is dependent on the operator’s technique and does not clearly visualize the mediastinal lesions. Furthermore, thyroid nodules are often unintentionally identified by CT. The texture analysis is a postprocessing technique that assesses the tissue characteristics of the tumors without exposing the patient to additional radiation. If a thyroid nodule is discovered on DECT in daily clinical practice, we believe that texture analysis using extracted monochromatic images can provide additional information for further examination, including US.

After the introduction of high-resolution CT scanners, previous studies have demonstrated that imaging features including punctuate calcification, nodule size, and irregularity could help detect malignant thyroid nodules [22–24]. DECT has spread widely at a variety of institutions and has been used in daily practice. Li et al. demonstrated that there was a potential to diagnose ex vivo thyroid nodules using DECT by detecting the amount of iodine and changes in HU attenuation slopes in tissues [14]. Previous studies have demonstrated that the expressions of TPO mRNA and NIS in malignant thyroid nodules were less than in benign lesions [11–13]. In the process of producing thyroid hormones, TPO incorporates inorganic iodine into organic iodine in the follicular epithelial cells. NIS expression on the basolateral membrane of the thyroid follicular cells leads to iodine accumulation. Malignant thyroid nodules have comparatively less iodine, due to the decreased expression of TPO mRNA and NIS, than benign thyroid nodules. However, Ringel et al. also demonstrated that there was no significant difference in the NIS expression between malignant and benign thyroid nodules [12]. Additionally, we found no significant differences between benign and malignant thyroid nodules using material decomposition and attenuation slopes for monochromatic images obtained in vivo. Although it is difficult to determine the exact reason for this difference, we propose several possible causes. First, other tissues within the scan range, including fat, muscle, and bone, would affect the attenuation of X-rays at different energies and make it difficult to detect low iodine concentrations. Li et al. demonstrated that the ex vivo iodine concentrations within goiters, follicular adenomas, and malignant groups were 0.42 ± 0.37 mg/mL, 0.51 ± 0.42 mg/mL, and 0.41 ± 0.2 mg/mL, respectively [14]. In addition, Ringel et al. also demonstrated that there was no significant difference in NIS expression between malignant and benign thyroid nodules [12]. These differences might be difficult to detect in the presence of more complex tissue components. Second, the dual source scanner used in this study may have more significant problems with cross-scatter than the fast kilovoltage-switching dual-energy scanner used in the previous study. Cross-scatter results in noise due to the collision of two X-rays emitted from different sources at a perpendicular angle to each other [25, 26].

Measures of texture analysis such as smoothness and regularity in pixels represent both distributions and relationships in gray-level patterns [27]. Texture analysis has been applied to the field of oncology to analyze predictors, treatment responses, and histologic classification [28–32]. Heterogeneity in the malignant tissues is recognized to reflect cellularity, necrosis, and angiogenesis, which are associated with disorders of attenuation [33]. Previous studies demonstrated that texture analysis of images obtained from US and MRI is useful for distinguishing between benign thyroid nodules and malignancies with less-uniform texture parameters, classifiers, and applications [15, 16, 34–36]. Since thyroid nodules are often detected by CT scan before performing further US, with texture analysis using DECT, unnecessary further examinations of the thyroid lesions can be potentially prevented.

Al Ajmi et al. previously described first-order statistics for texture features that could be used to classify benign parotid tumors and predict the effectiveness of induction chemotherapy in patients with head and neck carcinoma, respectively [32, 37]. In this study, the histogram mean and median at 80 keV were higher for benign than malignant lesions. Furthermore, these factors resulted in the best AUCs for differentiating between the two. GLGM MGR and VGR, which are indicative of gray-level distortions, were lower for benign lesions. These findings would suggest that, in 80 keV images, benign thyroid lesions have less heterogeneity than malignant thyroid lesions. At higher keV, tumor and soft tissue attenuations are inferior to those of lower keV on monochromatic images; however, image noise decreases at higher keV [20, 38]. CT imaging of the thyroid is expected to be more susceptible to beam hardening artifacts and image noise, due to the proximity of the clavicles and shoulders. Therefore, a decreased standard deviation at higher keV might have a small influence on mean texture features between benign and malignant thyroid nodules, compared with imaging at lower keV.

Despite these findings, some limitations to the present study need to be considered. In our retrospective study, our investigations of texture analysis on monochromatic images were performed in a small number of patients with a high prevalence of malignant thyroid nodules rather than in patients with incidental malignant thyroid nodules. This may lead to overfitting results. Second, this is a preliminary first demonstration regarding the potential use of DECT texture analysis based on a relatively small number of patients with thyroid nodule; certainly, future investigation both using a larger number of patients and comparing DECT with US is required. Third, benign thyroid nodules were followed for at least for 1 year clinically; therefore, their development into malignant lesions may not have been detected within the study time frame. Lastly, obvious cystic changes in lesions were excluded from quantitative measurements because it was thought that cystic tissue might affect the accurate analysis of solid components. Therefore, a potentially confounding effect of interreader error would be a source of bias. These problems could be solved in future studies with a larger sample size.

5. Conclusions

In conclusion, this study investigated the imaging characteristics of benign and malignant thyroid nodules in vivo using quantitative DECT. If a thyroid nodule is discovered on DECT in daily clinical practice, texture analysis of monochromatic images acquired at high keV may be a more promising diagnostic approach for thyroid nodules than DECT material decomposition and CT attenuation slopes.

Disclosure

K.O. is an employee of Siemens Healthcare K.K. Authors who are not employees of Siemens Healthcare K.K. monitored and had control of inclusion of any data and information submitted for the publication.

Conflicts of Interest

All authors declare that they have no conflicts of interest.