Ultrasonic Radiomics Utilised to Predict Lateral Neck Lymph-Node Metastasis in Papillary Thyroid Carcinoma

By Xi Cai, Sichen Chen, Yan Ding, Fengsheng Zhou, Yu Zhang

Affiliations

1. Department of Medical Ultrasound, Wuxi People’s Hospital of Nanjing Medical University, Wuxi, China

doi: 10.29271/jcpsp.2025.06.742

ABSTRACT
Objective: To investigate the clinical significance of preoperative prediction for detecting metastasis in lateral neck lymph nodes (LNLN) among patients diagnosed with papillary thyroid carcinoma (PTC).
Study Design: Descriptive study.
Place and Duration of the Study: Department of Medical Ultrasound, Wuxi People’s Hospital of Nanjing Medical University, Wuxi, China, from March 2021 to March 2023.
Methodology: Data from PTC patients with complete ultrasound and clinical records were collected, including 112 patients in the training group and 49 in the validation group. Based on pathological analysis, the individuals were categorised into the LNLN metastasis- positive group (50 cases) and the LNLN metastasis-negative group (111 cases). Logistic regression analysis showed that independent factors affected gender (OR = 3.167) as well as the maximum diameter of the tumour (OR = 1.177) and LNLN metastasis. Based on the above two independent influencing factors, the clinical model and ultrasound image feature model were constructed, respectively. A total of 6 non-zero coefficients of ultrasound radiomics features were screened by the LASSO regression dimensionality reduction to construct ultrasound radiomics models.
Results: Significant differences existed in Rad scores (RS) between the LNLN metastasis-positive group and the LNLN metastasis-negative group in the training set (p <0.05). The ROC curve indicated that the combined model exhibited a significantly higher area under the curve (AUC) compared to the ultrasound radiomics model. The calibration curve demonstrated high calibration for both the ultrasound radiomics model and the combined model, in good consistency with the actual results.
Conclusion: The clinical-ultrasonic feature-radiomics model holds significant clinical value in predicting LNLN metastasis in PTC patients.

Key Words: Radiomics, Lateral neck lymph node, Papillary thyroid carcinoma, Preoperative prediction, Logistic regression analysis.

INTRODUCTION

Papillary thyroid carcinoma (PTC), the predominant form of differentiated thyroid carcinoma, comprises approximately 80-85% of all cases^1,2 and tends to spread to the lymph nodes in the neck region. PTC with lateral neck lymph node (LNLN) metastasis is linked to a higher likelihood of local recurrence and reduced overall survival rate, although the prognosis is fine for most PTC cases.³ Surgical treatment is the radical approach for thyroid carcinoma (TC). Preventive central lymph node dissection (PCND) has been recommended for PTC patients to decrease the local recurrence rate and enhance disease- specific survival (DSS). However, controversy still exists regarding therapeutic LNLN dissection.⁴
 

Postoperative preventive 131I treatment is recommended to enhance disease-free and overall survival due to high-risk features and the potential for recurrence, albeit with unavoidable transient mild sexual dysfunction in both females and males.⁵ Therefore, enhanced preoperative precision in diagnosing LNLN metastasis is a crucial research focus. It can assist surgeons in determining appropriate surgical strategies pre- operatively, which improves patient outcomes.

High-resolution ultrasound is an important imaging modality for the preoperative assessment of CLN status. However, the sensitivity of ultrasound in predicting LNLN metastasis is 0.70, while its specificity is 0.84. The precision of the outcomes relies on the expertise and proficiency possessed by the ultrasonographer.⁶ Up to 40% of TC have pathological lymph node metastasis at diagnosis, which cannot be diagnosed by preoperative imaging.

Radiomics, as a post-processing technology in medical imaging, converts images into quantitative data. It identifies features that are imperceptible to the naked eye and reflects tumour heterogeneity, assisting in disease diagnosis, tumour recurrence, and prognosis assessment.^7,8 Radiomics has a certain value in pre- dicting LNLN metastasis in PTC.^3,9 However, they have not deeply investigated the efficacy of single grey-scale ultrasonic radio-mics in assessing LNLN metastasis in PTC. Therefore, this study investigated the clinical utility of a nomogram incorporating ultrasonic radiomics and clinical risk factors for the preoperative prediction of LNLN metastasis in patients with PTC.

METHODOLOGY

This study was performed on patients diagnosed with PTC and treated at the Affiliated Hospital of Nanjing Medical University, Wuxi People's Hospital, China. The surgical procedures were conducted from March 2021 to March 2023. The inclusion criteria required patients to have complete ultrasound imaging data where no markers interfered with image analysis, undergo postoperative follow-up for over one year, and possess complete clinical data that have been confirmed by pathology. As for the exclusion criteria, patients with multifocal PTC, a history of thyroid surgery, or those who could not be excluded for LNLN metastasis were not included in the study. The LNLN group included LN in regions II-V according to the guidelines for the diagnosis and treatment of thyroid nodules and differentiated thyroid cancer (second edition). Fifty patients with PTC confirmed by postoperative pathology, exhibiting either LNLN metastasis or LNLN recurrence more than one year after surgery, were classified as LNLN metastasis-positive. The other 111 PTC patients were defined as the negative group. Participants were allocated into the training group (n = 112) and the validation group (n = 49) at a randomisation ratio of 7:3. Since the work referred to a retrospective analysis, the informed consent of the subjects was exempted.

The American GE LOGIQ E9 colour Doppler ultrasound diagnostic apparatus was used with the L9-5 probe at 6-9 MHz. Subjects were placed in a supine posture, ensuring complete visibility of the anterior neck region. The following signs were observed and recorded for tumours: Maximum diameter, shape, margin, aspect ratio, presence or absence of calcification, and relationship with the capsule. Table I compares ultrasonic characteristics between the validation and training groups.

Table I: Clinical data between validation and training groups.
 

Characteristics	Training group (n = 112)	Validation group (n = 49)	t/χ2	p-value
Age (years)	-	-	-	-
Mean ± SD	40.90 ± 12.24	41.00 ± 14.62	0.044	0.965
Gender	-	-	-	-
Male	40	17	0.016	0.901
Female	72	32	-	-
Maximum tumour diameter (mm)	13.70 ± 6.34	12.92 ± 7.74	0.191	0.849
Shape	-	-	-	-
Irregular	39	20	0.528	0.468
Regular	73	29	-	-
Margin	-	-	-	-
Irregular	85	38	0.052	0.820
Smooth	27	11	-	-
Aspect ratio >1	-	-	-	-
Present	72	28	0.739	0.390
Absent	40	21	-	-
Calcification	-	-	-	-
Present	84	39	0.399	0.528
Absent	28	10	-	-
Capsular invasion status	-	-	-	-
Present	42	23	1.261	0.261
Absent	70	26	-	-
Note: p is calculated from the χ2 test for qualitative variables or by the t-test for continuous variables.

Table II: Clinical data and ultrasound image features between the LNLN metastasis-positive group and the LNLN metastasis-negative group in the training set.

Characteristics	LNLN metastasis (+)	LNLN metastasis (-)	p-value
Age (years)	40.29 ± 14.04	41.18 ± 11.41	0.718
Mean ± SD	-	-	-
Gender	-	-	-
Male	19	21	0.006
Female	-	-	-
Maximum tumour diameter (mm)	17.94 ± 8.07	11.77 ± 4.19	<0.001
Shape	-	-	-
Irregular	13	26	0.728
Regular	22	51	-
Margin	-	-	-
Irregular	27	58	0.835
Smooth	8	19	-
Aspect ratio >1	-	-	-
Present	21	51	0.523
Absent	14	26	-
Calcification	-	-	-
Present	27	57	0.724
Absent	8	20	-
Capsular invasion status	-	-	-
Present	8	34	0.031
Absent	27	43	-
p is calculated from the χ2 test for qualitative variables or by the t-test for continuous variable.

Table III: Estimates of the independent variables and related parameters of the clinical-ultrasound characteristics-imaging histology logistic regression equations for the training group.

Variables	β	SE	Wald χ2	p-value	OR (95% CI)
Maximum tumour diameter	0.111	0.049	5.505	0.025	1.117 (1.014-1.231)
Gender	1.077	0.512	4.425	0.035	2.935 (1.076-8.004)
RS	3.985	1.170	11.598	<0.001	53.784 (5.428-532.943)
Constant	-5.262	1.103	22.769	<0.001	0.005
Note: β represents the regression coefficient; OR represents the odds ratio; CI represents the confidence interval; RS represents the Rad scores. Value of p is determined by the logistic regression model.

Figure 1: Radiomics model and combined model establishment. (A) Correlation heatmap of ultrasonic radiomics features. (B) Nomogram for predicting LNLN metastasis in PTC using the combined model.

The grey-scale ultrasound images of the largest diameter section of the tumour in the training group were imported into open-source software 3D-slicer (version 5.4.0) in JPG format. A sonographer with over five years of experience manually delineated the regions of interest (ROI) along the contours of the lesions, ensuring complete coverage of the entire lesion area. Another physician in the superficial ultrasound group with more than 10 years of experience confirmed the ROI of lesions.

A consensus was reached through discussion in case of disagreement, and the ROI was re-delineated by the sonographer. The radiomics features were extracted based on the ROI using the Radiomics package in 3D Slicer. The feature preprocessing and feature screening were then performed in the following order. First, maximum and minimum values were normalised. All features of each dimension for all samples were linearly normalised to (0, 1). Next, LASSO regression was used to screen out the ultrasound radiomics feature correlation heatmap. Penalty coefficients were adjusted using 5-fold cross- validation to select features and their corresponding coefficients predicting PTC LNLN metastasis from the training group. A weighted linear combination was used to construct the radiomics label, and RS was assigned to each patient.

Selected clinical data, ultrasonic sonogram features, and RS were combined. The researchers utilised multivariate logistic regression to develop their integrated model, with a nomogram created for visualising the model.

The statistical analysis was conducted using SPSS 27.0 and R software (version 4.3.2). The Kolmogorov-Smirnov test was used to determine whether the distribution was normal. Continuous variables following normal distribution were presented as x ± s, while measurement data not following normal distribution were expressed as M (P₂₅, P₇₅). Independent samples t-test or Mann-Whitney U test was used for intergroup comparison. Qualitative data were presented in terms of percentages, and intergroup comparisons were conducted using chi-square tests. Software R was utilised to implement the LASSO regression model, employing the Glmnet r package. Multivariate logistic regression was used to construct models of clinics, ultrasonic feature, radiomics, and their combined model. The work evaluated the performance of each model by analysing its receiver operating characteristic (ROC) curve. The statistical significance was determined when p was <0.05.

RESULTS

The validation and training groups did not differ significantly concerning age, gender, tumour maximum diameter, shape, margin, aspect ratio, calcification, or capsular invasion status (Table I).

Among the 112 patients in the training set, there were 35 cases in the LNLN metastasis-positive group and 77 cases in the LNLN metastasis-negative group. There was a noticeable disparity in gender distribution between the two groups (p <0.05). Multivariate logistic regression analysis showed that gender was an independent influencing factor for predicting LNLN metastasis in PTC patients (OR = 3.167; p = 0.007). A clinical model was constructed: Logit(p) = -2.405 + 1.153* gender based on the independent influencing factors.

Figure 2: Model performance comparison. (A) ROC curve of models in the training set; (B) ROC curve of the models in the validation set; (C) Calibration curve of the ultrasonic radiomics model; (D) Calibration curve of the combined model.

There were notable variances observed in the tumour's maximum diameter and its invasion status of the capsule among the two groups (p <0.05). Through multivariate logistic regression analysis, the maximum tumour diameter exerted an autonomous influence on predicting LNLN metastasis in PTC patients (OR = 1.177; p <0.001). An ultrasonic feature model was constructed based on the independent influencing factor: Logit(p) = -2.327 + 0.163* maximum lesion diameter (Table II).

A total of 837 raw features were extracted in the training set (Figure 1A). Six features were obtained to construct the radiomics model after feature preprocessing and selection: RS = 228.26 +0.063*original_glszm_GreyLevelNonUniformity-1.100*wavelet.LLH_firstorder_Minimum-233.352*wavelet.LHH_glrlm_HighGreyLevelRunEmphasis+0.854*wavelet.HLH_glszm_LargeAreaLowGreyLevelEmphasis+0.131×wavelet.LLL_glszm_GreyLevelNonUniformity+1.648*wavelet.LLL_glszm_SizeZoneNonUniformityNormalised. Rad score (RS) for each lesion was calculated based on the formula. The RS for the positive and negative LNLN metastasis groups were 0.51 ± 0.25 and 0.22 ± 0.19 points in the training group, respectively. The validation group exhibited statistically significant differences between the positive and negative LNLN metastasis groups, with RS scores of 0.68 ± 0.28 and 0.44 ± 0.23 points, respectively (p <0.05).

The purpose of this analysis was to leverage the high- throughput features extracted by ultrasound radiomics, which can objectively reflect tumour heterogeneity and classify lymph node metastasis in PTC. Since the ROIs drawn in the work were all derived from the maximum diameter section of the grey-scale ultrasound image, there were some feature loss. Therefore, a composite model was developed by integrating radiomics features from ultrasound images with clinical data, encompassing two independent influencing factors (gender and maximum tumour diameter).

PTC LNLN metastasis is the dependent variable (1 and 0 denote the presence/absence of LNLN metastasis, respectively). Clinical data (gender), ultrasonic features (maximum lesion diameter), and RS are included to establish the combined model: Logit(p) = -5.262 + (0.111 * maximum lesion diameter) + (1.077 * gender) + (3.985 * RS). Table III shows the training group's clinical-ultrasonic feature- radiomics logistic regression analysis. Figure 1B presents the nomogram. The cumulative scores of the model are derived from summing up the respective scores assigned to each influencing factor based on the nomogram model. The predicted probability is obtained in terms of the total scores.
 

The AUC for diagnosing PTC LNLN metastasis was 0.635 for the clinical model, 0.757 for the ultrasonic feature model, 0.843 for the combined model, and 0.824 for the radiomics model in the training group. The AUC for diagnosing PTC LNLN metastasis was 0.538, 0.741, 0.747, and 0.778 in the validation group, respectively. Both the training and validation groups showed that the AUC of the combined model outperformed that of a standalone model (Figure 2A, B). The accuracy rates for predicting PTC LNLN metastasis were 83.9 and 71.4% in the training group and the validation group, respectively, with the sensitivities of 62.9 and 100.0% and the specificities of 93.5 and 47.1%. The calibration curves indicated that the radiomics and combined models exhibited superior calibration performance, indicating better consistency with actual results (Figure 2C, D).

DISCUSSION

Accurate preoperative assessment of LNLN metastasis in patients with PTC is crucial for guiding surgical strategies. Ultrasound is widely used for preoperative evaluation of LNLNs in thyroid cancer. However, the consistency of evaluation results among different physicians varies significantly, and their diagnostic efficacy heavily depends on the subjective experience and technical skills of the physicians.³ Previous study focused on the ultrasound features of PTC tumours, exploring the correlation between combinations of suspicious ultrasound features and LNLN metastasis.¹⁰ However, these suspicious ultrasound features based on physicians' judgements are prone to inter-observer variability. Recently, radiomics has attracted increasing attention from researchers with the advancement of medical imaging analysis methods. Radiomics can objectively analyse quantitative features and capture additional information to reflect tumour heterogeneity.^11-14 The application value of radiomics in predicting LNLN metastasis in PTC patients has not yet been fully explored with limited related research.

The combined model based on preoperative ultrasound radiomics and clinical features achieved an AUC of 0.843 (0.764–0.922) in the training set and 0.778 (0.643–0.914) in the validation set in the current study, both higher than those of the clinical and radiomics models alone. The nomogram visualisation demonstrated high predictive value for assessing LNLN metastasis status in PTC patients. Compared with Liu et al.’s study the diagnostic performance of the work was slightly lower.¹⁵ This discrepancy was attributed to the fact that the radiomics prediction model was constructed using first-order, texture, and wavelet-based features. Based on the training set data, a total of 837 original quantitative features were extracted. Six radiomics features were ultimately obtained after feature selection, including one first-order feature and five wavelet-based features. These features could quantify the texture-grey-level intensity, grey-level uniformity, and distribution in ultrasound images, reflecting microscopic characteristics such as echo intensity, internal structure, and heterogeneity of the lesions. PTC primary tumours with LNLN metastasis exhibited more pronounced texture coarseness and greater tumour heterogeneity. In contrast, Liu et al. incorporated grey-scale ultrasound and elastography features, providing richer data and enabling the extraction of more radiomics features.¹⁵ AUC (0.778) of the combined model in the validation set was lower than that reported in Lui et al.’s study,¹⁵ but the sensitivity reached 100%. This difference was due to the relatively smaller dataset in the present study, where the integration of radiomics features significantly improved predictive sensitivity.

Zhang et al. established an MRI-based clinical predictive model that demonstrated good discriminative performance for preoperative prediction of LNLN metastasis.¹⁶ However, MRI is not sensitive to calcifications, and respiratory and swallowing movements can significantly affect image quality. Moreover, MRI is expensive and not suitable for routine examination. Current studies suggest that the application of ultrasonic radiomics to evaluate the likelihood of lymph node metastasis in patients diagnosed with PTC is feasible. Tong et al. combined ultrasonic radiomics scores, ultrasonic reports, and CT reports, creating a nomogram for the preoperative prediction of LNLN metastasis.¹⁷ It shows good discriminative ability, calibration, and consistency with actual results. However, repeated examinations may lead to resource wastage. Inconsistencies between ultrasonic and CT reports increase inconvenience for patients and delay in diagnosis and treatment. Therefore, there is an urgent clinical need for a mature AI model based on ultrasound thyroid sonograms to address the variability in operator expertise and preoperative prediction.

Ultrasonic radiomics was applied to predict PTC LNLN metastasis preoperatively in this study. Six radiomics features were screened out. The RS score was obtained by summing up the products of the six features and their respective weight. The results showed that the RS in the positive LNLN metastasis group of PTC patients in the validation group and the training group was higher compared to the negative group. Quantitative radiomics features had a good classification effect on LNLN metastasis.

The radiomics model in this study had an AUC of 0.824 and 0.747 for predicting PTC LNLN metastasis in the training and validation groups, respectively. Clinical data and ultrasonic features showed that the diagnosis of PTC LNLN metastasis was influenced by gender and the maximum lesion diameter, which were found to be unrelated factors. The results were similar to the conclusions of Liu et al. and Park et al.^18,19 Further construction of a nomogram model combining RS clinical data and ultrasound features. The results show that the nomogram model had a higher AUC for predicting PTC LNLN metastasis in the training group than the radiomics model (0.843 vs. 0.824). This was validated in the validation group (0.778 vs. 0.747). The calibration curves indicated that the prediction curves of both the radiomics and combined models closely approximated the ideal ones. Consequently, the predictions of the two models had good consistency with actual results. The clinical-ultrasonic feature-radiomics model provides a reliable method for preoperative prediction of PTC LNLN metastasis.

However, there are still certain limitations. The work is a retrospective analysis carried out at a single center, potentially leading to biased case selection and lacking an external validation dataset. Further multi-center research should prove the generalisability and robustness of nomograms. The ROI delineation was manual and selected at the maximum level of the tumor, which could not reflect the entire tumour. An automatic, reliable, and efficient tumour segmentation method (such as 3-D RoI-Aware U-Net).²⁰ Moreover patients with PTC were excluded from the work, which may lead to confusion regarding LN metastasis of PTC lesions.

CONCLUSION

Ultrasonic radiomics could non-invasively predict PTC LNLN metastasis preoperatively. Nomograms’ preoperative predictions were enhanced by integrating ultrasonic radiomics into clinical practice. Consequently, diagnostic accuracy was improved to develop individualised neck dissection plans for patients and guide postoperative radiotherapy decisions.

FUNDING:
The work was supported by the Top Talent Support Program-me for Young and Middle-Aged People of the Wuxi Health Committee (Grant No: HB2023001).

ETHICAL  APPROVAL:
The work was approved by the Ethics Committee and Review Board of Wuxi People’s Hospital of Nanjing Medical University (No: KY21005), and was conducted ethically following the World Medical Association Declaration of Helsinki.

PATIENTS’  CONSENT:
Due to retrospective nature of the study, the requirement for patients’ informed consent was waived according to the decision of Ethics Committee.

COMPETING  INTEREST:
The authors declared no conflict of interest.

AUTHORS’  CONTRIBUTION:
XC, SC: Conceptualisation, methodology, investigation, writing, and revision. Both authors contributed equally to this work.
YD: Project administration and review.
FZ: Resourcing and data curation.
YZ: Visualisation and supervision.
All authors approved the final version of the manuscript to be published.

REFERENCES

1. Filetti S, Durante C, Hartl D, Leboulleux S, Locati LD, New-bold K, et al. Thyroid cancer: ESMO clinical practice guidelines for diagnosis, treatment and follow-up. Ann Oncol 2019; 30(12):1856-83. doi: 10.1093/annonc/mdz400.
2. China anti-cancer association thyroid cancer professional committee. Integrated diagnosis and treatment guidelines for thyroid cancer in China (2022 Abbreviated Edition). Chin J Clin Oncol 2023; 50(7):325-30. doi: 10.12354/j.issn.1000- 8179.2023.20221410.
3. Dong L, Han X, Yu P, Zhang W, Wang C, Sun Q, et al. CT radiomics-based nomogram for predicting the lateral neck lymph node metastasis in papillary thyroid carcinoma: A prospective multicentre study. Acad Radiol 2023; 30(12): 3032-46. doi: 10.1016/j.acra.2023.03.039.
4. Brito JP, Hay ID, Morris JC. Low risk papillary thyroid cancer. BMJ 2014; 348:g3045. doi: 10.1136/bmj.g3045.
5. Chinese Society of Endocrinology, Thyroid and Metabolism Surgery Group of the Chinese Society of Surgery, China Anti- Cancer Association, Chinese Association of Head and Neck Oncology, Chinese Society of Nuclear Medicine, China Anti- Cancer Association, et al. Guidelines for the diagnosis and treatment of thyroid nodules and differentiated thyroid cancer (second edition). Chin J Endocrinol Metabol 2023; 39(03):181-226. doi: 10.3760/cma.j.cn311282-20221023- 00589.
6. Zhao H, Li H. Meta-analysis of ultrasound for cervical lymph nodes in papillary thyroid cancer: Diagnosis of central and lateral compartment nodal metastases. Eur J Radiol 2019; 112:14-21. doi: 10.1016/j.ejrad.2019.01.006.
7. Bera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol 2022; 19(2): 132-46. doi: 10.1038/s41571-021-00560-7.
8. Gillies RJ, Schabath MB. Radiomics improves cancer screening and early detection. Cancer Epidemiol Biomarkers Prev 2020; 29(12):2556-67. doi: 10.1158/1055-9965.EPI-20- 0075.
9. Zhang XY, Zhang D, Zhou W, Wang ZY, Zhang CX, Li J, et al. Predicting lymph node metastasis in papillary thyroid carcinoma: Radiomics using two types of ultrasound elastography. Cancer Imaging 2025; 25(1):13. doi: 10.1186/ s40644-025-00832-w.
10. Ding X, Liu Y, Zhao J, Wang R, Li C, Luo Q, et al. A novel wavelet-transform-based convolution classification network for cervical lymph node metastasis of papillary thyroid carcinoma in ultrasound images. Comput Med Imaging Graph 2023; 109:102298. doi: 10.1016/j.compmedimag.2023.102298.
11. Sun P, Wei Y, Chang C, Du J, Tong Y. Ultrasound-based nomogram for predicting the aggressiveness of papillary thyroid carcinoma in adolescents and young adults. Acad Radiol 2024; 31(2):523-5. doi: 10.1016/j.acra.2023.05.009.
12. Chen W, Lin G, Cheng F, Kong C, Li X, Zhong Y, et al. Development and validation of a dual-energy ct-based model for predicting the number of central lymph node metastases in clinically node-negative papillary thyroid carcinoma. Acad Radiol 2023; 31(1):142-56. doi: 10.1016/j.acra.2023.04.038.
13. Jiang L, Guo S, Zhao Y, Cheng Z, Zhong X, Zhou P. Predicting extrathyroidal extension in papillary thyroid carcinoma using a clinical-radiomics nomogram based on B-mode and contrast-enhanced ultrasound. Diagnostics (Basel) 2023; 13(10):1734. doi: 10.3390/diagnostics13101734.
14. Kim Y, Roh JL, Gong G, Cho KJ, Choi SH, Nam SY, et al. Risk factors for lateral neck recurrence of N0/N1a papillary thyroid cancer. Ann Surg Oncol 2017; 24(12):3609-16. doi: 10.1245/s10434-017-6057-2.
15. Liu T, Ge X, Yu J, Guo Y, Wang Y, Wang W, et al. Comparison of the application of B-mode and strain elastography ultrasound in the estimation of lymph node metastasis of papillary thyroid carcinoma based on a radiomics approach. Int J Comput Assist Radiol Surg 2018; 13(6):1617-27. doi: 10. 1007/s11548-018-1796-5.
16. Zhang H, Hu S, Wang X, He J, Liu W, Yu C, et al. Prediction of cervical lymph node metastasis using MRI radiomics approach in papillary thyroid carcinoma: A feasibility study. Technol Cancer Res Treat 2020; 19:1533033820969451. doi: 10.1177/1533033820969451. 
17. Tong Y, Li J, Huang Y, Zhou J, Liu T, Guo Y, et al. Ultrasound-based radiomic nomogram for predicting lateral cervical lymph node metastasis in papillary thyroid carcinoma. Acad Radiol 2020; 28(12):1675-84. doi: 10.1016/j. acra.2020.07.017.
18. Liu C, Xiao C, Chen J, Li X, Feng Z, Gao Q, et al. Risk factor analysis for predicting cervical lymph node metastasis in papillary thyroid carcinoma: A study of 966 patients. BMC Cancer 2019; 19(1):622. doi: 10.1186/s12885-019-5835-6.
19. Park VY, Han K, Kim HJ, Lee E, Youk JH, Kim EK, et al. Radiomics signature for prediction of lateral lymph node metastasis in conventional papillary thyroid carcinoma. PLoS One 2020; 15(1):e0227315. doi: 10.1371/journal.pone.022 7315.
20. Huang YJ, Dou Q, Wang ZX, Liu LZ, Jin Y, Li CF, et al. 3-D RoI-Aware U-Net for accurate and efficient colorectal tumor segmentation. IEEE Trans Cybern 2021; 51(11):5397-408. doi: 10.1109/TCYB.2020.2980145.