Development and internal validation of a multivariable MRI-based radiomics prediction model for preoperative cervical lymph node metastasis in thyroid cancer

Introduction

Thyroid cancer (TC) is one of the most common malignant tumors, and its incidence rate is on the rise globally. Papillary TC (PTC) accounts for 84% of TC cases, making it the most common thyroid malignancy (1-3). Patients with PTC can expect a generally good prognosis and relatively long survival (4). Although PTC is considered an indolent tumor, the presence of cervical lymph node metastasis (LNM) can elevate the clinical risk (5). Neck LNM is a high risk factor for local recurrence (6), and is the focus of surgical treatment for TC (7). According to the 2015 American Thyroid Association (ATA) guidelines, prophylactic lymph node dissection (LND) does not significantly improve long-term survival (5) and is associated with a high risk of complications, such as permanent hypoparathyroidism and recurrent laryngeal nerve damage (8). Consequently, the accurate preoperative evaluation of the cervical lymph node status in patients with TC is critical to the formulation of individualized surgical plans and determining patient prognosis.

Ultrasonography (US) is the most commonly used method for the preoperative assessment of cervical lymph node involvement but is limited by factors such as instrument resolution and operator experience. Moreover, US cannot sufficiently evaluate the parapharyngeal and cervical lymph nodes (9), with the sensitivity of US diagnosis being reported to be 63% (10). Computed tomography (CT) can be used to evaluate LNM, but it is not routinely recommended (5).

With the development of computer science and imaging postprocessing, radiomics has emerged as a promising diagnostic technology. Radiomics involves the high-throughput extraction of a wide range of quantitative features and can transform medical images into available high-dimensional data. These data can then be quantitatively analyzed, and predictive models can be established through artificial intelligence methods to support clinical decision-making (11,12). The diagnostic value of radiomics for LNM, peritumoral invasion, and prognosis has been established in several studies (13-15). Moreover, radiomics has been demonstrated to be superior to US and CT in the preoperative prediction of LNM in patients with PTC (16,17) Magnetic resonance imaging (MRI) offers superior soft‑tissue contrast and multiparametric capabilities, yet its radiomic potential for LNM prediction has been rarely explored.

This study aimed to develop a radiomics model based on T2-weighted imaging (T2WI) and contrast-enhanced T1-weighted imaging (T1C+) and to evaluate its ability to preoperatively diagnose cervical lymph node status in patients with TC. We present this article in accordance with the TRIPOD reporting checklist (18) (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0259/rc).

Methods

Study design and participants

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This prospective study was approved by the Ethics Committee of The First Affiliated Hospital of Guangxi Medical University (approval No. 2021[KY‑E‑034]). All patients provided written informed consent before the study.

Initially, 135 patients who were pathologically diagnosed with TC after surgery between April 2017 and January 2019 were included. Of these patients, 97 were ultimately enrolled; the flowchart of participant inclusion is shown in Figure 1. The inclusion criteria were as follows: (I) pathologically confirmed TC; (II) previous cervical LND; (III) preoperative MRI; (IV) no previous history of thyroid surgery, biopsy, or head and neck cancer; and (V) no history of neck radiotherapy before MRI. Meanwhile, the exclusion criteria were: (I) no neck LND and (II) poor MRI imaging quality.

Figure 1 Flowchart of patient enrollment. MR, magnetic resonance; ROI, region of interest; TC, thyroid cancer.

Sample size

A formal sample size calculation was not performed a priori; the sample was determined by the number of eligible consecutive cases during the study period. With 61 LNM-positive events and 10 final features retained in the model, the events-per-variable (EPV) ratio was 6.1, which is below the conventional threshold of 10–20 for logistic regression.

Outcome assessment

The reference standard for cervical LNM was histopathological examination of all resected lymph nodes from the surgical specimens. Pathological assessment was performed by two experienced pathologists who were blinded to the MRI findings and radiomics predictions. LNM was defined as the presence of any metastatic focus in at least one cervical lymph node.

MRI acquisition

MRI scans were performed for all patients 1–2 weeks before surgery using a MAGNETOM Verio 3.0-T device (Siemens Healthineers, Erlangen, Germany) equipped with an 8-channel orthogonal coil. Patients were positioned supine, head-first. The chin was elevated, the head was supported with a padded head cushion, and the hands were placed at the sides. The head was positioned within the coil so that the eyebrows were at the center of the coil. The MRI scanning included a T2WI fat-suppressed T2-weighted imaging sequence [time to repeat/time to echo (TR/TE) =2,000 ms/88 ms, layer thickness =4 mm, continuous scanning, field of view (FOV) =200 mm, number of excitations (NEX) =1] and T1WI sequence (TR/TE =210 ms/2.6 ms, layer thickness =4 mm, FOV =200 mm, NEX =2). The patients were trained in breathing techniques prior to the MRI examination, and they were required to maintain the same level of breath throughout the procedure. Example images are shown in Figure 2.

Figure 2 Example of ROI delineation. (A,B) T1C+ and T2WI images of the same patient, respectively. (C,D) Delineated ROIs (yellow outline). (E) Three-dimensional image of the ROI. ROI, region of interest; T1C+, T1-weighted heterogeneous contrast-enhanced imaging; T2WI, T2-weighted imaging.

Tumor segmentation and feature extraction

MR images were used for tumor segmentation. First, T2WI and T1C+ images were input into the research workstation. Second, the lesions were manually segmented independently by a radiologist with 5 years of experience and a senior radiologist with 15 years of experience; for any discrepant results, the Deputy Director of Radiology was consulted to determine the final result. Semiautomatic delineation of the region of interest (ROI) was conducted along the edge of the lesion, and it was ensured that the lesion was completely visualized in the coronal and sagittal position. Areas of necrosis or calcification in the lesion were avoided. An example of ROI delineation is shown in Figure 2. Inter-observer reproducibility was assessed on a random subset of 30 cases; all extracted features demonstrated good consistency (intraclass correlation coefficient ≥0.89). In cases of multifocal TC, the largest lesion was selected for further analysis. Features from the T2WI and T1C+ images were extracted via Artificial Intelligence ToolKit version 3.0.0 (GE HealthCare, Chicago, IL, USA). A total of 1,691 features were extracted and classified into six groups: (I) gray-level histogram features, (II) shape features, (III) texture features, (IV) gray-level cooccurrence matrix features, (V) gray-level run length matrix features, and (VI) gray-level size zone matrix features.

Model construction

Patients were randomly assigned to training and validation set at a ratio of 7:3. In order to model the radiomics features in the training set and avoid overfitting, the Boruta algorithm was used for feature screening. The Boruta algorithm is a packaging algorithm based on random forest, which iteratively searches for all possible subsets of related properties and performs top-down screening of related features by comparing the importance of the original properties with the importance of the random implementation (19,20). After the sample size of the training set was accounted for, the 10 most important features were finally retained. Subsequently, the final model was constructed with these 10 features via a random forest algorithm and 10-fold cross-validation.

Model validation and statistical analysis

In the radiomics features model analysis, receiver operating characteristic (ROC) curve analysis was conducted to evaluate the performance of the model, with the area under the curve (AUC), accuracy, specificity, sensitivity, and accuracy metrics being recorded. The predictive performance of the combined model was rigorously assessed in the internal validation cohort, followed by an evaluation of its potential clinical utility. The default probability threshold of 0.5 was used to define LNM status (i.e., patients with a predicted probability ≥0.5 were classified as LNM-positive). In the statistical analysis of clinical characteristics, continuous variables were tested with the Mann-Whitney test, while categorical variables were tested with the Pearson chi-squared test. SPSS 26.0 (IBM, Armonk, NY, USA) was used to conduct these analyses. For all tests, P<0.05 was considered statistically significant. All statistical tests were two-sided.

Results

Patient characteristics

This study included 97 patients with TC (61 lymph node-positive and 36 lymph node-negative). Tables 1,2 present the baseline data of patients. Age, sex, histological type, and tumor location did not differ significantly between lymph node-positive and -negative groups (P>0.05).

Model performance

Three predictive models were evaluated via ROC analysis, with the results being presented in Figures 3,4. In the training set, the AUC, accuracy, specificity, and sensitivity of the T2WI model were 0.914, 0.714, 0.800, and 0.667, respectively; those of the T1C+ model were 0.959, 0.607, 0.900, and 0.444, respectively; and those of the combined model were 0.963, 0.927, 0.917, and 0.933, respectively. The diagnostic performance from validation was comparable: the AUC, accuracy, specificity, and sensitivity of the T2WI model were 0.853, 0.714, 0.800, and 0.667, respectively; those of the T1C+ model were 0.881, 0.607, 0.900, and 0.444, respectively; and those of the combined model were 0.900, 0.786, 0.500, and 0.944, respectively. The AUC of combined model was slightly higher than those of the stand-alone models, indicating its superiority.

Figure 3 ROC curves of the three RF models in the training set. AUC, area under the curve; CE, contrast-enhanced, referring to T1C+ model; RF, random forest; ROC, receiver operating characteristic.

Figure 4 ROC curves of the three RF models in the validation set. AUC, area under the curve; CE, contrast-enhanced, referring to T1C+ model; RF, random forest; ROC, receiver operating characteristic.

Discussion

This study aimed to develop and internally validate an MRI radiomicsbased prediction model for preoperative cervical LNM in patients with TC. The combined T2WI + T1C+ model achieved an AUC of 0.900 in the validation set, which is comparable or slightly superior to prior CT-based models (AUCs 0.71–0.77) (17,21) and US-elastography radiomics models (AUC 0.832) (22). Comparison between training and validation performance. The combined model showed a decline in AUC from 0.963 (training) to 0.900 (validation), and specificity dropped markedly from 0.917 to 0.500. This discrepancy suggests notable model optimism, likely attributable to the limited sample size and the low EPV of 6.1. This underscores that our findings are preliminary and must be interpreted with caution. MRI, as a powerful imaging modality, can visualize the internal structures of objects at the micro and macro scale and provide 2D and 3D imaging in a noninvasive manner (23). Therefore, MRI radiomics features may serve as a reliable means to preoperatively diagnosing LNM. This study has several strengths. It is one of the few studies to explore MRI‑based radiomics for LNM prediction in TC.

In this study, MRI radiomics features were examined in patients with TC as a means to assessing lymph node status. A complete randomization group design was used to minimize bias. The Boruta algorithm was used to select the features. This algorithm selects all features associated with the dependent variable, rather than those specific to the model, in order to minimize the model cost. This further provides a comprehensive understanding of the influencing factors of the dependent variable, leading to more suitable and efficient feature selection. Therefore, we used the Boruta algorithm for feature selection in the first stage, retaining the 10 most optimal features. In subsequent stages, the final model was constructed through a random forest algorithm. An important purpose of radiomics is to construct predictive models of treatment response through use of phenotypic features of tumor obtained from medical images. This capability is essential for personalized medicine, in which treatments are tailored to the characteristics of individual patients and their tumors (23). To our knowledge, there are only a handful of radiomic studies that have examined the determination of LNM status in patients with TC. One study examined the use of CT imaging features in predicting central LNM (CLNM) in patients with PTC, reporting a validation set AUC of 0.77 [95% confidence interval (CI): 0.55–0.99] (24). In a multicenter study, radiomics features extracted from CT images were significantly correlated with CLNM, with ROC curve values 0.747 (95% CI: 0.706–0.782), 0.710 (95% CI: 0.634–0.786), and 0.764 (95% CI: 0.654–0.875) in the training, internal validation, and external validation sets, respectively (17). Another study demonstrated that radiomics based on preoperative shear wave elastography images and the corresponding B-mode ultrasound images have the potential to predict LNM, with the AUC in the validation cohort being 0.832 (95% CI: 0.749–0.916) (22). In our study, we evaluated the ability of MRI radiomics to determine cervical lymph node status. We found that the models consisting of radiomics features from T2WI and T1C+ images performed well diagnostically, providing greater accuracy than those reported in the three aforementioned studies. The model with the combined features of the two image types produced a better diagnostic performance (AUC =0.9).

Several important limitations must be acknowledged. First, the sample size is small (n=97), and the EPV of 6.1 is below the recommended minimum of 10–20, increasing the risk of overfitting. Although we employed Boruta selection and cross‑validation to mitigate this, the results should be considered hypothesis‑generating rather than definitive. Second, this was a single-centre study with internal validation only (random split); external validation in a larger, independent cohort is essential before clinical translation. Third, we did not perform decision curve analysis (DCA) or formal calibration testing (e.g., Hosmer-Lemeshow) due to the limited sample size, which we acknowledge prevents quantification of net clinical benefit and calibration accuracy. Fourth, manual tumor segmentation is time‑consuming and may limit routine applicability; automated segmentation methods are desirable. Fifth, we did not segment the cervical lymph nodes themselves due to their small size and poor MRI visualization; future work with higher‑resolution MRI could address this. Sixth, we did not include a clinical‑only model for comparison, which would have better illustrated the added value of radiomics. Despite these limitations, our study provides preliminary evidence that MRI radiomics may assist in preoperative risk stratification. Future multicenter studies with larger samples should incorporate DCA, calibration assessment, and automated segmentation to improve clinical applicability and generalizability.

Conclusions

T2WI and T1C+ imaging radiomics features were used to establish a model for preoperative prediction of cervical LNM in patients with TC. While the model shows promising discriminative ability, its low specificity precludes standalone clinical use at present. With further refinement, incorporation of clinical factors, and external validation in larger cohorts, MRI radiomics may eventually contribute to individualized surgical planning in TC.

Acknowledgments

The authors are deeply grateful to all the patients and the research staff for their contributions to this project.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0259/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0259/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0259/prf

Funding: This study received support from the Guangxi Scientific Research and Technology Development Project (No. 1598011-4). The funder had no role in study design, data collection, analysis, or manuscript preparation.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0259/coif). All authors report that this study was supported by the Guangxi Scientific Research and Technology Development Project (No. 1598011-4). The authors have no other 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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital of Guangxi Medical University (approval No. 2021[KY‑E‑034]), and written informed consent was obtained from all participants.

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. Fang F, Gong Y, Liao L, et al. Value of Contrast-Enhanced Ultrasound for Evaluation of Cervical Lymph Node Metastasis in Papillary Thyroid Carcinoma. Front Endocrinol (Lausanne) 2022;13:812475. [Crossref] [PubMed]
2. Boucai L, Zafereo M, Cabanillas ME. Thyroid Cancer: A Review. JAMA 2024;331:425-35. [Crossref] [PubMed]
3. Chen X, Chen X, Xie W, et al. BRAF-activated ARSI suppressed EREG-mediated ferroptosis to promote BRAF(V600E) (mutant) papillary thyroid carcinoma progression and sorafenib resistance. Int J Biol Sci 2025;21:128-42. [Crossref] [PubMed]
4. Liu W, Zhang D, Jiang H, et al. Prediction model of cervical lymph node metastasis based on clinicopathological characteristics of papillary thyroid carcinoma: a dual-center retrospective study. Front Endocrinol (Lausanne) 2023;14:1233929. [Crossref] [PubMed]
5. Ouyang H, Li X, Xia F, et al. Clinical significance of level VII lymph node metastasis in papillary thyroid cancer. Sci Rep 2025;15:41827. [Crossref] [PubMed]
6. Wang Y, Li X, Gang Q, et al. Pathomics and single-cell analysis of papillary thyroid carcinoma reveal the pro-metastatic influence of cancer-associated fibroblasts. BMC Cancer 2024;24:710. [Crossref] [PubMed]
7. Puccini M, Ambrosini CE, Rossi L, et al. Sentinel node mapping in thyroid cancer: an overview. Front Med (Lausanne) 2023;10:1163151. [Crossref] [PubMed]
8. Haugen BR, Alexander EK, Bible KC, et al. 2015 American Thyroid Association Management Guidelines for Adult Patients with Thyroid Nodules and Differentiated Thyroid Cancer: The American Thyroid Association Guidelines Task Force on Thyroid Nodules and Differentiated Thyroid Cancer. Thyroid 2016;26:1-133. [Crossref] [PubMed]
9. Jin S, Bao W, Yang YT, et al. Establishing a prediction model for lateral neck lymph node metastasis in patients with papillary thyroid carcinoma. Sci Rep 2018;8:17355. [Crossref] [PubMed]
10. Hu Q, Zhang WJ, Liang L, et al. Establishing a Predictive Nomogram for Cervical Lymph Node Metastasis in Patients With Papillary Thyroid Carcinoma. Front Oncol 2021;11:766650. [Crossref] [PubMed]
11. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-6. [Crossref] [PubMed]
12. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
13. Zhang XY, Zhang D, Zhou W, et al. Predicting lymph node metastasis in papillary thyroid carcinoma: radiomics using two types of ultrasound elastography. Cancer Imaging 2025;25:13. [Crossref] [PubMed]
14. Zhang Y, Liu L, Han M, et al. Unraveling the power of radiomics: prediction and exploration of lymph node metastasis in stage T1/2 esophageal squamous cell carcinoma. NPJ Precis Oncol 2025;9:176. [Crossref] [PubMed]
15. Li G, Li L, Li Y, et al. An MRI radiomics approach to predict survival and tumour-infiltrating macrophages in gliomas. Brain 2022;145:1151-61. [Crossref] [PubMed]
16. Agyekum EA, Ren YZ, Wang X, et al. Evaluation of Cervical Lymph Node Metastasis in Papillary Thyroid Carcinoma Using Clinical-Ultrasound Radiomic Machine Learning-Based Model. Cancers (Basel) 2022;14:5266. [Crossref] [PubMed]
17. Li J, Wu X, Mao N, et al. Computed Tomography-Based Radiomics Model to Predict Central Cervical Lymph Node Metastases in Papillary Thyroid Carcinoma: A Multicenter Study. Front Endocrinol (Lausanne) 2021;12:741698. [Crossref] [PubMed]
18. Collins GS, Reitsma JB, Altman DG, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ 2015;350:g7594. [Crossref] [PubMed]
19. Cui Y, Yang W, Ren J, et al. Prognostic value of multiparametric MRI-based radiomics model: Potential role for chemotherapeutic benefits in locally advanced rectal cancer. Radiother Oncol 2021;154:161-9. [Crossref] [PubMed]
20. Dong F, Li Q, Xu D, et al. Differentiation between pilocytic astrocytoma and glioblastoma: a decision tree model using contrast-enhanced magnetic resonance imaging-derived quantitative radiomic features. Eur Radiol 2019;29:3968-75. [Crossref] [PubMed]
21. Peng Y, Zhang ZT, Wang TT, et al. Prediction of Central Lymph Node Metastasis in cN0 Papillary Thyroid Carcinoma by CT Radiomics. Acad Radiol 2023;30:1400-7. [Crossref] [PubMed]
22. Jiang M, Li C, Tang S, et al. Nomogram Based on Shear-Wave Elastography Radiomics Can Improve Preoperative Cervical Lymph Node Staging for Papillary Thyroid Carcinoma. Thyroid 2020;30:885-97. [Crossref] [PubMed]
23. Iacopetta D, Ceramella J, Baldino N, et al. Targeting Breast Cancer: An Overlook on Current Strategies. Int J Mol Sci 2023;24:3643. [Crossref] [PubMed]
24. Motiei M, Mansouri SS, Tamimi A, et al. The diagnostic accuracy of MRI radiomics in axillary lymph node metastasis prediction: a systematic review and meta-analysis. Int J Surg 2025;111:6412-26. [Crossref] [PubMed]

(English Language Editor: J. Gray)