Development and validation of deep learning models for qualitative classification of benign and malignant enlarged cervical lymph nodes based on ultrasound images

Introduction

Background

Lymph nodes are vital immune organs distributed throughout the body. Cervical lymph nodes (CLNs) constitute approximately one-third of all lymph nodes in the human body (1). CLN enlargement is a common clinical manifestation resulting from lesions within the neck or other regions, characterized by changes in size and morphology. Pathologically, abnormal CLNs are classified as benign or malignant. Benign pathologies include reactive hyperplasia, tuberculous lymphadenitis, and granulomatous inflammation. Malignant pathologies encompass lymphoma (primary to lymphoid tissue) and metastasis from solid malignancies. Accurate qualitative diagnosis of CLNs provides critical information for clinical decision-making, directly guiding subsequent patient management, helping to avoid unnecessary biopsies, and enabling precise staging and prognostic assessment in patients with malignancy.

Rationale and knowledge gap

Histopathological diagnosis remains the gold standard for differentiating benign from malignant CLNs; however, it is an invasive diagnostic procedure. Ultrasonography, as a non-invasive, simple, cost-effective, and convenient imaging modality, especially high-frequency ultrasound, is widely used in clinical practice as the method of choice for detecting and diagnosis of CLNs. B-mode ultrasound demonstrates significantly higher sensitivity than other conventional imaging modalities for detecting malignant CLNs in patients with head and neck malignancies (2). Among conventional imaging modalities, computed tomography can clearly delineate the morphology, borders, and enhancement characteristics of enlarged lymph nodes. However, it has limited sensitivity to metabolic activity, hindering effective differentiation between inflammatory and neoplastic processes (3). Magnetic resonance imaging, while capable of comprehensive assessment of lymph node signal characteristics and surrounding tissue involvement through multiparametric imaging, fails to reliably detect micrometastases (4,5). Ultrasonography can assess lymph nodes based on morphology and vascular distribution for initial evaluation, but it exhibits low specificity and is operator-dependent (6,7). With the advancement of artificial intelligence, machine learning has rapidly progressed in the field of medical imaging, demonstrating substantial promise across key clinical dimensions, including tumor risk stratification, molecular subtype prediction, disease diagnosis, and prognostic assessment. Notable examples are: Dembrower et al.’s Inception-ResNet-v2 convolutional neural network (CNN) model trained for predicting breast cancer risk in women (8); Zhao et al.’s machine learning model utilizing ultrasound videos to predict thyroid cancer lymph node metastasis (9); and Gu et al.’s model based on ultrasound images for the early prediction of patient response to neoadjuvant chemotherapy (10). Compared with the subjective diagnosis of imaging physicians, deep learning (DL), as a specialized machine learning approach, can automatically learn discriminative features from images for clinical prediction, thereby offering enhanced diagnostic accuracy and reproducibility (11,12).

Objective

Therefore, this study was designed to develop DL models to perform qualitative diagnosis of benign and malignant CLNs. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2024-576/rc).

Methods

Patients and study design

This retrospective cohort study utilized consecutively collected ultrasound images and pathological data from patients who underwent ultrasound-guided biopsy of CLNs at West China Hospital from January 2020 to December 2023. Inclusion criteria included: (I) clear pathological diagnosis of targeted CLNs; and (II) ultrasound examination performed before biopsy. Exclusion criteria included: (I) incomplete or unclear pathological diagnosis; and (II) incomplete or poor quality of B-mode and Color Doppler Flow Imaging images. Ultimately, a total of 2,697 patients and 3,014 abnormal CLNs were included. Ultimately, all 3,014 CLNs were randomly assigned at the image level to prevent data leakage. Using a computer-generated randomization sequence, the images were randomly allocated in a 7:2:1 ratio to a primary cohort (training set: n=2,110; validation set: n=603) and the test cohort (n=301). The detailed patient selection process was illustrated in Figure 1. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Review Committee of West China Hospital, Sichuan University (No. 2020-1219), and individual consent for this retrospective analysis was waived.

Figure 1 Overview of included patients and cohort allocation. CLN, cervical lymph node.

Acquisition and interpretation of ultrasound images

Ultrasound examination and biopsy of CLNs were performed by experienced sonographers. Ultrasound images were acquired using high-frequency linear array probes, including the Philips IU22 (probe L12-5, L18-4; Philips, Amsterdam, Netherlands), Hitachi Vison Preirus (probe L13-3; Hitachi Medical Corp., Tokyo, Japan), Mindray Resona7 (probe L14-5, L14-3; Mindray, Shenzhen, China), and Siemens ACUSON OXANA 2 (probe L14-5; Siemens Healthineers, Erlangen, Germany). Standard image acquisition involved obtaining the maximum diameter plane (longitudinal view) and its vertical plane (transverse view), ensuring the lesion was centered within the image. For DL analysis, one representative standard B-mode image from the longitudinal view was selected per lesion.

Two senior radiologists, blinded to the pathological diagnosis, independently reviewed the ultrasound images. They evaluated and documented the following ultrasound features of CLNs using a standardized form: size (long axis, short axis, long-to-short axis ratio), cortical echogenicity (anechoic, hypoechoic, isoechoic, hyperechoic), cortical homogeneity (homogeneous, inhomogeneous), corticomedullary demarcation (clear, unclear), shape (oval, round, irregular), margin (circumscribed, uncircumscribed), cystic degeneration (absent, present) and calcification (present, absent).

Pathological diagnosis

The pathologic diagnosis of CLNs was obtained from the Department of Pathology at West China Hospital. Given the diverse diagnoses of enlarged CLNs, the samples were classified into benign and malignant groups. Benign CLNs encompassed diagnostic outcomes such as normal lymph nodes, reactive hyperplasia, the presence of normal lymph node cells, the absence of abnormal lymph node cells, lymphoproliferative disorders, and inflammatory lesions of lymph nodes. Malignant CLNs primarily consisted of diagnostic findings indicative of metastatic malignant tumors, the presence of atypical hyperplastic cells, and lymphoma.

CLNs segmentation

The region of interest segmentation was conducted through manual delineation along the margins of the CLNs to minimize potential interference from surrounding structures such as adjacent lymph nodes, blood vessels, and muscles on the lymph node images. In cases where multiple lymph nodes were present within a single ultrasound image, only the punctured lymph node was designated as the target for segmentation (Figure 2).

Figure 2 Segmentation examples of CLN regions of interest. The green circles indicate boundary of CLNs. CLN, cervical lymph node.

Development of DL models

Following the segmentation process, the images were resized to a dimension of 224×224×3 through isotropic scaling. The padding for any excess areas was performed using zero-value pixels. Subsequently, the resized images underwent Gaussian normalization, resulting in normalized images with means and standard deviations (SDs) of (0.485, 0.456, 0.406) and (0.229, 0.224, 0.225), respectively. Each element corresponded to the mean and SD of the respective image channel.

CNN and Vision Transformer were used to conducted DL models. Finally, a variety of CNN models, including the ResNet series (ResNet50, ResNet101, and pretrained ResNet50), VGG16, DenseNet121, and EfficientNet-B0, as well as Vision Transformer models such as VisionTransformer_B_16, SwinTransformer, and Max_VisionTransformer (13-18), were developed and validated for CLNs qualitative diagnosis.

Statistical analysis

Statistical analysis was performed using SPSS (version 26.0) and Python (version 3.8.5) software. Receiver operating characteristic curves were generated, and model performance was evaluated by calculating accuracy, precision, recall, F1 scores. Generalization ability and model fit were assessed using loss curves. Statistical performance differences between models were compared using the DeLong test. Continuous variables were presented as mean ± SD, and compared using Student’s t-test. Categorical variables were expressed as frequencies and percentages, with between-group comparisons performed using Chi-squared test. All tests were two-sided, and P value <0.05 was considered as statistically significant.

Results

Patient characteristics

A total of 3,014 CLNs from 2,697 patients were included in this study. These comprised 1,489 (49.4%) benign cases and 1,525 (50.6%) malignant cases. Among the benign cases, lymphocytic infiltration was observed in 1,097 cases (36.4%), reactive hyperplasia in 250 cases (8.3%), granulomatous inflammation in 90 cases (3%), and tuberculous lymphadenopathy in 21 cases (0.7%). Within the malignant group, metastatic CLNs accounted for the majority (1,325/1,525, 86.9%). Malignancy prevalence was significantly higher in males (54.2%, 561/1,036) than in females (48.7%, 964/1,978) (P<0.05). The overall cohort age ranged from 7 to 87 (43.81±14.77) years. Benign cases spanned 7–84 (43.97±14.08) years, while malignant cases ranged from 9 to 87 (43.66±15.42) years. No statistically significant difference in age was observed between benign and malignant groups (P>0.05).

Ultrasound features associated with the diagnosis of CLNs

Univariate analysis demonstrated that all ultrasonographic features evaluated in this study showed significant associations with CLNs diagnosis (Table 1). Malignant CLNs exhibited significantly larger dimensions than benign CLNs in both long-axis (17.61±9.68 vs. 16.41±7.97 mm) and short-axis measurements (9.15±5.65 vs. 6.65±3.47 mm, both P<0.001). Conversely, the long-to-short axis ratio was significantly smaller in malignant CLNs (2.09±0.83) compared to benign CLNs (2.65±1.11, P<0.001). The majority of malignant CLNs exhibited hyperechoic cortex (48.59%) relative to the echogenicity of surrounding cervical soft tissue, whereas most benign CLNs displayed hypoechoic cortex (56.95%). Compared to benign CLNs, malignant CLNs demonstrated significantly higher rates of inhomogeneous cortex (64.59% vs. 7.99%), unclear corticomedullary demarcation (97.11% vs. 60.51%), and ill-defined margins (75.15% vs. 62.06%) (all P<0.001). Among irregularly shaped CLNs (n=465), 71.61% (333/465) were malignant. Cystic degeneration and calcification were hallmark features of malignancy, with 87.98% (227/258) and 88.25% (751/851) of such cases confirmed as malignant, respectively.

Development and validation of DL models

For the classification of benign and malignant CLNs, eight advanced DL models were developed. Utilizing the validation set from the primary cohort, the generalization ability and performance of these models were evaluated through accuracy, recall, and loss curves (Figure 3). The results indicated that EfficientNet-B0 outperforms other models, achieving the highest accuracy of 75% on the validation set (Figure 4).

Figure 3 Generalization ability and performance of DL models on the validation set. (A) Recall curves of DL models for the malignant CLNs group. (B) Recall curves of DL models for the benign CLNs group. (C) Accuracy curves of DL models. (D) Loss curves of DL models. CLN, cervical lymph node; DL, deep learning.

Figure 4 Performance of EfficientNet-B0 model. Recall curves and accuracy curve of EfficientNet-B0 model on the validation set.

Diagnostic performance of DL models and ultrasonography

Based on the test cohort, the diagnostic performance of DL models and ultrasonography was evaluated and compared (Table 2, Figure 5). All DL models demonstrated satisfactory performance in qualitative diagnosis of CLNs, achieving area under the curve (AUC) values ranging from 0.56 to 0.81, accuracy from 0.62 to 0.73, sensitivity from 0.53 to 0.79, and specificity from 0.43 to 0.74. As indicated in Table 2, the VGG16 model exhibited the best performance with an AUC of 0.81 [95% confidence interval (CI): 0.77–0.86], accuracy of 0.73, sensitivity of 0.71, and specificity of 0.74. In comparison to ultrasonography, the VGG16, ResNet101, and ResNet50 models showed significantly superior predictive performance (P<0.05).

Figure 5 ROC curves of DL models and ultrasound. (A) ROC curves of ResNet series models. (B) ROC curve of DenseNet121 model. (C) ROC curve of VGG16 model. (D) ROC curve of EfficientNet-B0 model. (E) ROC curve of Max_VisionTransformer model. (F) ROC curve of SwinTransformer model. (G) ROC curve of VisionTransformer_B_16 model. (H) ROC curve of ultrasound. AUC, area under the curve; CI, confidence interval; DL, deep learning; ROC, receiver operating characteristic.

Discussion

Enlarged CLNs serve as significant clinical indicators for various pathologies. Accurate early diagnosis of lymphadenopathy etiology can minimize diagnostic delays and reduce unnecessary invasive procedures. In this study, ultrasonography demonstrated high sensitivity but low specificity in diagnosing CLNs. The ultrasound features were found to be significantly associated with the malignancy of CLNs, suggesting that DL models based on ultrasound images have the potential to achieve satisfactory performance. Performance evaluation confirmed the superior diagnostic capability of DL models compared to ultrasonography (P<0.05), indicating their potential to enhance diagnostic precision and guide clinical decision-making.

In our study, univariate analysis yielded that the size (long axis, short axis, long-to-short axis ratio), cortical echogenicity, cortical homogeneity, corticomedullary demarcation (clear, unclear), shape, margin, cystic degeneration, and calcification are significantly and independently associated with the diagnosis of enlarged CLNs. The ultrasound features relative to malignancy of CLNs remain controversial according to the literature review. It was previously reported that features such as roundness, irregular margins, hyperechoic, absence of lymphatic gates, macroscopic necrosis, calcification, and peripheral or mixed vascular flow were often associated with malignant lymph nodes (2). Gupta et al. (19) suggested that malignant lymph nodes were often associated with round, homogeneous echogenicity, well-defined borders with abundant peripheral blood flow. In the study of Turgut et al. (20), the ultrasound features that were significantly associated with malignant lymph nodes were a decreased ratio of long-to-short axis ratio, abnormal or absent lymphatic hilum, and microcalcifications. Due to the overlap in imaging manifestations of benign and malignant lymph node lesions, which makes the characteristics of malignant lymph nodes summarized in different studies differ, coupled with the complex anatomy of the neck and the small size of some metastatic lymph nodes (21-23), which are easy to be missed on examination. Therefore, the implement of intelligence tools could reduce the operator-dependence of ultrasound and improve accurate qualitive diagnosis of lymph nodes.

Recent advances in artificial intelligence models for ultrasound image analysis have demonstrated promising performance in characterizing enlarged CLNs, achieving sensitivities of 74–91%, specificities of 72–87%, and AUC values of 0.76–0.92 (24-27). However, these models are predominantly trained on limited single-center datasets, constraining their generalizability. Furthermore, most studies focus on discriminating specific pathological subtypes, such as metastatic carcinoma, lymphoma, or tuberculous lymphadenitis (28-30), with comprehensive investigations encompassing the full spectrum of pathological types for holistic benign-malignant risk stratification remaining scarce.

Many studies have shown that DL models can be used to predict benign and malignant tumors and lymph node metastasis of malignant tumors, but most of the models are based on images of tumor lesions, and fewer studies are based on lymph node images. In this study, based on the ultrasound images of CLNs, we analyzed the image characteristics of the CLNs in terms of their benignity and malignancy, and constructed a prediction model capable of making a qualitative diagnosis of lymph node lesions in the neck based on the region of interest of the CLNs images. The models showed good diagnostic performance in both primary and test cohort, achieving the AUC of 0.81. The efficacy of the DL model in our study was lower than some of published articles, which could be related to the inclusion of more cases of tuberculosis as well as granulomatous inflammation in the present study, as there is much overlap between tuberculosis and metastatic lymph nodes in ultrasound images (31).

There are some limitations to our study. First, this is a retrospective single-center cohort study. Generalizability can be improved if data from additional institutions and long-term follow-up are added. Second, the study in this paper is only based on the B-mode ultrasound images of the target object for the construction of the DL model, and fails to incorporate more ultrasound images, such as ultrasonography, elastography, and so on. Thirdly, all the included cases are lymph nodes of a definitive nature by puncture biopsy, and do not include the more inclined to be benign with the follow-up observation of the lymph nodes, which may have a selective bias.

Conclusions

In conclusion, the ultrasound characteristics of CLNs were significantly and independently associated with malignancy. DL models utilizing ultrasound images demonstrated satisfactory performance in qualitative diagnosis of enlarged CLNs, with the VGG16 model achieving the highest AUC of 0.81 (95% CI: 0.77–0.86). These DL models utilizing ultrasound images demonstrated promising performance in the qualitative CLNs. This approach enhanced the diagnostic accuracy of preoperative ultrasound assessment, thereby allowing a subset of patients to avoid unnecessary biopsies and optimizing clinical decision-making.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2024-576/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2024-576/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2024-576/prf

Funding: The study was supported by the Achievement Conversion and Guidance Project of Chengdu Science and Technology Bureau (No. 2017-CY02-00027-GX).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2024-576/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Review Committee of West China Hospital, Sichuan University (No. 2020-1219), and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Iwanaga J, Lofton C, He P, et al. Lymphatic System of the Head and Neck. J Craniofac Surg 2021;32:1901-5. [Crossref] [PubMed]
2. Ryu KH, Lee KH, Ryu J, et al. Cervical Lymph Node Imaging Reporting and Data System for Ultrasound of Cervical Lymphadenopathy: A Pilot Study. AJR Am J Roentgenol 2016;206:1286-91. [Crossref] [PubMed]
3. Cho JK, Ow TJ, Lee AY, et al. Preoperative (18)F-FDG-PET/CT vs Contrast-Enhanced CT to Identify Regional Nodal Metastasis among Patients with Head and Neck Squamous Cell Carcinoma. Otolaryngol Head Neck Surg 2017;157:439-47. [Crossref] [PubMed]
4. Siegel MJ, Acharyya S, Hoffer FA, et al. Whole-body MR imaging for staging of malignant tumors in pediatric patients: results of the American College of Radiology Imaging Network 6660 Trial. Radiology 2013;266:599-609. [Crossref] [PubMed]
5. Liu Y, Wang R, Ding Y, et al. A predictive nomogram improved diagnostic accuracy and interobserver agreement of perirectal lymph nodes metastases in rectal cancer. Oncotarget 2016;7:14755-64. [Crossref] [PubMed]
6. Acu L, Oktar SÖ, Acu R, et al. Value of Ultrasound Elastography in the Differential Diagnosis of Cervical Lymph Nodes: A Comparative Study With B-mode and Color Doppler Sonography. J Ultrasound Med 2016;35:2491-9. [Crossref] [PubMed]
7. Chang JM, Leung JWT, Moy L, et al. Axillary Nodal Evaluation in Breast Cancer: State of the Art. Radiology 2020;295:500-15. [Crossref] [PubMed]
8. Dembrower K, Liu Y, Azizpour H, et al. Comparison of a Deep Learning Risk Score and Standard Mammographic Density Score for Breast Cancer Risk Prediction. Radiology 2020;294:265-72. [Crossref] [PubMed]
9. Zhao HN, Yin H, Li MH, et al. Contrast-enhanced ultrasound image sequences based on radiomics analysis for diagnosis of metastatic cervical lymph nodes from thyroid cancer. Gland Surg 2024;13:1437-47. [Crossref] [PubMed]
10. Gu J, Tong T, He C, et al. Deep learning radiomics of ultrasonography can predict response to neoadjuvant chemotherapy in breast cancer at an early stage of treatment: a prospective study. Eur Radiol 2022;32:2099-109. [Crossref] [PubMed]
11. Peng S, Liu Y, Lv W, et al. Deep learning-based artificial intelligence model to assist thyroid nodule diagnosis and management: a multicentre diagnostic study. Lancet Digit Health 2021;3:e250-9. [Crossref] [PubMed]
12. Zheng X, Yao Z, Huang Y, et al. Deep learning radiomics can predict axillary lymph node status in early-stage breast cancer. Nat Commun 2020;11:1236. [Crossref] [PubMed]
13. Lee JG, Jun S, Cho YW, et al. Deep Learning in Medical Imaging: General Overview. Korean J Radiol 2017;18:570-84. [Crossref] [PubMed]
14. Eminaga O, Eminaga N, Semjonow A, et al. Diagnostic Classification of Cystoscopic Images Using Deep Convolutional Neural Networks. JCO Clin Cancer Inform 2018;2:1-8. [Crossref] [PubMed]
15. Vij R, Arora S. Modified deep inductive transfer learning diagnostic systems for diabetic retinopathy severity levels classification. Biomed Signal Process Control 2025;99:106885.
16. Abd El-Ghany S, Mahmood M A, Abd El-Aziz AA. Adaptive Dynamic Learning Rate Optimization Technique for Colorectal Cancer Diagnosis Based on Histopathological Image Using EfficientNet-B0 Deep Learning Model. Electronics 2024;13:3126.
17. Choudhary R, Deepak A, Krishnasamy G, et al. An optimized bidirectional vision transformer based colorectal cancer detection using histopathological images. Biomed Signal Process Control 2025;102:107210.
18. Alruily M, Mahmoud AA, Allahem H, et al. Enhancing Breast Cancer Detection in Ultrasound Images: An Innovative Approach Using Progressive Fine-Tuning of Vision Transformer Models. Int J Intell Syst 2024;2024:6528752.
19. Gupta A, Rahman K, Shahid M, et al. Sonographic assessment of cervical lymphadenopathy: role of high-resolution and color Doppler imaging. Head Neck 2011;33:297-302. [Crossref] [PubMed]
20. Turgut E, Celenk C, Tanrivermis Sayit A, et al. Efficiency of B-mode Ultrasound and Strain Elastography in Differentiating Between Benign and Malignant Cervical Lymph Nodes. Ultrasound Q 2017;33:201-7. [Crossref] [PubMed]
21. Majumdar KS, Rao VUS, Prasad R, et al. Incidence of Micrometastasis and Isolated Tumour Cells in Clinicopathologically Node-Negative Head and Neck Squamous Cell Carcinoma. J Maxillofac Oral Surg 2020;19:131-5. [Crossref] [PubMed]
22. Yamazaki Y, Chiba I, Hirai A, et al. Clinical value of genetically diagnosed lymph node micrometastasis for patients with oral squamous cell carcinoma. Head Neck 2005;27:676-81. [Crossref] [PubMed]
23. Lee DW, Ji YB, Sung ES, et al. Roles of ultrasonography and computed tomography in the surgical management of cervical lymph node metastases in papillary thyroid carcinoma. Eur J Surg Oncol 2013;39:191-6. [Crossref] [PubMed]
24. Qian T, Zhou Y, Yao J, et al. Deep learning based analysis of dynamic video ultrasonography for predicting cervical lymph node metastasis in papillary thyroid carcinoma. Endocrine 2025;87:1060-9. [Crossref] [PubMed]
25. Valizadeh P, Jannatdoust P, Pahlevan-Fallahy MT, et al. Diagnostic accuracy of radiomics and artificial intelligence models in diagnosing lymph node metastasis in head and neck cancers: a systematic review and meta-analysis. Neuroradiology 2025;67:449-67. [Crossref] [PubMed]
26. Park J, Kim S, Lim JH, et al. Development of a multi-modal learning-based lymph node metastasis prediction model for lung cancer. Clin Imaging 2024;114:110254. [Crossref] [PubMed]
27. Zhang M, Lyu S, Yang L, et al. A nomogram based on ultrasound radiomics for predicting the invasiveness of cN0 single papillary thyroid microcarcinoma. Gland Surg 2023;12:1735-45. [Crossref] [PubMed]
28. Huang X, Gan X, Feng J, et al. Nomograms for predicting cervical central lymph node metastases and high-volume cervical central lymph node metastases in papillary thyroid carcinoma. Gland Surg 2025;14:421-35. [Crossref] [PubMed]
29. Cui QL, Yin SS, Fan ZH, et al. Diagnostic Value of Contrast-Enhanced Ultrasonography and Time-Intensity Curve in Differential Diagnosis of Cervical Metastatic and Tuberculous Lymph Nodes. J Ultrasound Med 2018;37:83-92. [Crossref] [PubMed]
30. Dai L, Zheng L, Li Y, et al. Analysis and prediction of contralateral central lymph node metastasis risk in unilateral papillary thyroid carcinoma with ipsilateral lateral cervical lymph node: a retrospective clinical study. Gland Surg 2025;14:380-90. [Crossref] [PubMed]
31. Li L, He L, Xiong M, et al. Diagnostic value of contrast-enhanced ultrasound combined with serum procalcitonin in tuberculous lymph nodes and metastatic lymph nodes. Clinics (Sao Paulo) 2025;80:100541. [Crossref] [PubMed]