Thyroid nodules are a common clinical finding, with a prevalence of approximately 4% to 7% in the general population, and are most often detected by ultrasonography.1 2 Although most thyroid nodules are benign, distinguishing malignant from benign nodules remains a clinical priority to avoid unnecessary procedures and ensure timely intervention.3 To standardise risk stratification, various Thyroid Imaging Reporting and Data Systems (TIRADS) have been developed,4 5 including the ACR-TIRADS (American College of Radiology),6 the K-TIRADS (Korean Society of Thyroid Radiology),7 and the European Thyroid Association.8 Recognising the need for a system tailored to the Chinese healthcare context, the Chinese Artificial Intelligence Alliance for Thyroid and Breast Ultrasound proposed the Chinese TIRADS (C-TIRADS) in 2021.2 However, existing TIRADS models primarily focus on sonographic characteristics and often overlook relevant clinical risk factors (eg, patient age, sex, and cervical lymph node [LN] involvement).9 In clinical practice, radiologists frequently incorporate such clinical information into their assessments, contributing to inconsistency and variability in TIRADS classification.

Papillary thyroid carcinoma accounts for approximately 80% to 90% of all thyroid cancers and is typically characterised by indolent behaviour.10 11 A substantial proportion of new cases involve papillary thyroid microcarcinoma, defined as tumours measuring less than 10 mm in diameter, which generally carry a favourable clinical prognosis.12 Increasing recognition of the indolent nature of papillary thyroid microcarcinoma has raised concerns regarding potential overdiagnosis and overtreatment. However, current risk stratification strategies that rely solely on imaging features may either overestimate or underestimate malignancy risk, depending on the patient’s broader clinical context. Approaches that incorporate clinical risk factors into TIRADS classification could address these limitations and enhance diagnostic accuracy, supporting more individualised patient management.

This study aimed to develop and externally validate a predictive model that integrates both imaging characteristics and clinical risk factors to refine the C-TIRADS classification system. To our knowledge, this is the first nomogram-based model to incorporate clinical risk factors into the C-TIRADS framework. The tool is designed to assist radiologists in improving diagnostic consistency and supporting more informed and individualised clinical decision making in the management of thyroid nodules.

This retrospective diagnostic study included patients with thyroid nodules who underwent surgical resection at two tertiary hospitals in China. The derivation cohort comprised patients treated at Sichuan Provincial People’s Hospital from January to December 2022, while the external validation cohort was drawn from Affiliated Hospital of North Sichuan Medical College during the same period. Inclusion criteria were: (1) thyroid nodules confirmed by postoperative pathology and (2) preoperative ultrasonography of the thyroid and cervical LNs with complete imaging and clinical records. Exclusion criteria were: (1) unclear pathological diagnosis; (2) incomplete clinical data; or (3) poor-quality ultrasound images.

Two junior radiologists, blinded to clinical and pathological information, independently classified all nodules according to the C-TIRADS criteria. Subsequently, two senior radiologists re-evaluated the cases and adjusted the classifications based on additional clinical risk factors, including patient demographics and cervical LN findings. Any modification from the initial C-TIRADS classification was defined as ‘classification optimisation’ (^*C-TIRADS), encompassing both upgrades and downgrades.

Structured data collection forms were used to record clinical and sonographic variables. The collected data included patient sex, age, nodule size, number of nodules, C-TIRADS classification, and the presence of abnormal cervical LNs on ultrasonography.

Sonographic features that directly determine the C-TIRADS score (such as solidity, echogenicity, aspect ratio, microcalcification, and margin irregularity) were not included independently in the multivariable analysis to avoid collinearity. Based on clinical relevance and univariate regression analysis, six predictors were selected for model development, namely, patient sex, age-group (≤40, 40-60, and >60 years),13 14 nodule size, number of nodules (single vs multiple), presence of abnormal cervical LNs, and original C-TIRADS classification.

A binary logistic regression model was developed using the derivation cohort from Sichuan Provincial People’s Hospital (n=909). For categorical variables with more than two levels, dummy variables were created. The C-TIRADS category 5 was used as the reference group as it represents the highest level of suspicion and the most definitive management pathway (surgical resection), making it an appropriate clinical baseline to estimate relative malignancy risk and the need for reclassification. Model performance in the derivation cohort was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), and calibration was assessed by comparing predicted probability (PP) with observed outcomes using calibration plots.

We emphasise that the primary outcome variable for model training was the pathological diagnosis (binary: malignant vs benign). The C-TIRADS optimisation, defined as upgrading or downgrading the original category based on PP thresholds, was a post-model clinical decision rule applied to the model output, not the outcome used for model development.

Internal validation was performed using bootstrap resampling with 1000 samples to obtain bias-corrected estimates of model performance and 95% confidence intervals (95% CIs). A fixed random seed was set to ensure reproducibility. The bias-corrected C-statistic was 0.728, compared with the original apparent performance of 0.730 (a difference of 0.002), confirming the model’s stable discriminative ability (online supplementary Table 1).

The final model was applied to the external cohort from Affiliated Hospital of North Sichuan Medical College (n=750) to evaluate its generalisability. Model discrimination was evaluated by calculating the AUC in the validation set, and calibration was assessed using calibration curves.

A nomogram was developed based on the final multivariable regression model to provide a visual tool for clinical application. Each predictor was assigned a score, and the total score corresponded to the PP of C-TIRADS classification optimisation.

Decision curve analysis and clinical impact curves were used to evaluate the clinical utility of the nomogram by quantifying the net benefit across a range of threshold probabilities. Specifically, the nomogram generates a PP indicating whether a nodule’s original C-TIRADS classification should be modified after integrating clinical information. For clinical decision making, we pre-specified probability cut-offs: PP ≥60% (upgrade), PP <30% (downgrade), and PP ≥30% but <60% (unchanged). Based on these thresholds, the model’s recommendations were translated into optimised C-TIRADS categories, which were then compared with radiologists’ optimisation decisions and surgical pathology findings, as appropriate. These thresholds are reported in the Results section and were applied consistently across all performance tables

To ensure consistent ROC analysis, all AUCs were calculated using continuous PPs rather than ordinal risk categories. For the original C-TIRADS system, the five-level ordinal classification was transformed into a continuous malignancy probability score using proportional-odds (ordinal logistic) regression. This standard statistical method was employed to model the ordered nature of the C-TIRADS categories and to derive a continuous probability of malignancy for each category, enabling fair comparison in ROC analysis against other models. For the optimised ^*C-TIRADS system, PPs were directly obtained from the final multivariable logistic regression model. The ROC curves and corresponding AUCs were constructed using these continuous predictions.

Statistical analyses and data visualisation were performed using SPSS (Windows version 26.0; IBM Corp, Armonk [NY], United States) and RStudio (version 2022). Categorical variables were reported as number of cases or percentages, with group comparisons conducted using Chi squared test or Fisher’s exact test, as appropriate. Multivariable logistic regression analysis was conducted to identify independent predictors. Model discrimination was evaluated using ROC curves, while calibration curves were used to assess model accuracy. Clinical decision and impact curves were established to assess practical clinical utility. A two-tailed P value of <0.05 was considered statistically significant.

All models were trained to predict pathological malignancy. The optimised ^*C-TIRADS classifications presented here were derived by applying predefined probability thresholds to the model’s malignancy predictions.

A total of 1659 patients with thyroid nodules were included in the study, comprising 909 patients in the derivation cohort and 750 in the external validation cohort. In the derivation cohort, 71.8% of patients were women, and the majority (90.8%) had nodules measuring ≤30 mm. Approximately 81.7% of patients showed no abnormal cervical LNs on ultrasonography. The rate of C-TIRADS optimisation was 60.6%. In the external validation cohort, similar distributions were observed, with a higher proportion of nodules >30 mm (Table 1).

Univariate binary regression analysis revealed that several variables were either significantly associated (P<0.05) or showed a trend towards association (0.05 < P < 0.1) with C-TIRADS optimisation. These variables included patient sex, age, nodule size (10-30 mm), number of nodules, solid composition, blurred margins, aspect ratio >1, abnormal cervical LNs, and C-TIRADS category (Table 2 and online supplementary Table 2).

A multivariable binary logistic regression model was developed to identify independent predictors associated with C-TIRADS optimisation. Six predictors were independently associated with the outcome. The key predictors of C-TIRADS optimisation were male sex, age 40 to 60 years, thyroid nodule size (per 1-mm increase), multiple thyroid nodules, presence of abnormal cervical LNs, and original C-TIRADS 4A category (online supplementary Table 3). A nomogram model was constructed based on these six independent predictors (Fig 1).

The model demonstrated good discrimination, with an AUC of 0.730 (95% CI=0.697-0.762) in the derivation cohort (online supplementary Fig a). Internal validation using 1000 bootstrap samples yielded a bias-corrected C-statistic of 0.728, indicating stable model performance (online supplementary Table 1). Calibration curves showed good agreement between PPs and observed outcomes (online supplementary Fig b).

Diagnostic thresholds were evaluated to stratify risk. A PP of ≥60% or <30% was considered indicative of a high likelihood of classification change: a PP of ≥60% suggested upgrading, while a PP of <30% suggested downgrading; PPs between 30% and 60% indicated that the classification was likely to remain unchanged. A detailed summary of sensitivity, specificity, and overall accuracy across these thresholds is presented in online supplementary Table 4.

When applied to the external cohort, the model achieved an AUC of 0.865 (95% CI=0.839-0.891) [online supplementary Fig c], demonstrating excellent generalisability. Calibration plots again confirmed close agreement between predicted and observed probabilities (online supplementary Fig d). At the 60% probability threshold, sensitivity was 85.0%, specificity was 69.0%, and overall accuracy was 79.7% in the external validation cohort. Diagnostic performance metrics across various risk thresholds of the final prediction model were analysed in the external validation population (online supplementary Table 5).

Decision curve analysis (Fig 2a) demonstrated that the nomogram model provided greater net clinical benefit across a wide range of threshold probabilities compared with treating all or no patients. The clinical impact curve (Fig 2b) showed that the number of true positives closely approximated the predicted number across relevant thresholds. The observed distribution of histopathological outcomes was as follows: in the derivation cohort, 769 nodules (84.6%) were confirmed malignant and 140 (15.4%) were benign; in the validation cohort, 434 nodules (57.9%) were malignant and 316 (42.1%) were benign.

Comparison of diagnostic efficacy between the original C-TIRADS and optimised C-TIRADS classifications demonstrated superior performance of the optimised model in both the derivation and validation cohorts (Fig 2c and d, respectively). The optimised classification achieved higher AUC values for differentiating benign from malignant nodules (AUC=0.97 vs 0.94 in the derivation cohort; AUC=0.97 vs 0.95 in the external validation cohort). The predictive model tended to improve C-TIRADS classification by upgrading category 4A nodules to category 4B or 4C, reflecting enhanced clinical utility (Table 3 and Fig 2).

A 55-year-old man underwent ultrasound examination, which revealed a solid hypoechoic thyroid nodule in the right lobe measuring approximately 7.1 × 6.4 mm² (Fig 3a). Simultaneously, abnormal LNs were detected on the ipsilateral side of the neck, characterised by indistinct corticomedullary differentiation and suspected microcalcifications (Fig 3b). According to the conventional C-TIRADS system, the nodule was initially classified as category 4B. However, application of the nomogram model yielded a cumulative score of 155 points, corresponding to a malignancy risk of >90%. Based on this result, the TIRADS category was optimised and upgraded to category 5 (Fig 3c). Subsequent histopathological examination confirmed the diagnosis of papillary thyroid microcarcinoma with cervical LN metastasis.

This study retrospectively analysed the sonographic characteristics and clinical risk factors of 1659 thyroid nodules from two large tertiary hospitals in western China, with the aim of optimising the C-TIRADS classification. A predictive model integrating clinical parameters and imaging features was developed and externally validated, demonstrating high diagnostic performance (AUC=0.865 in external validation) and clinical benefit, as evidenced by decision curve analysis.

Despite the widespread adoption of various TIRADS frameworks globally,2 4 5 6 7 8 fundamental methodological limitations persist. Current models, such as ACR-TIRADS,6 primarily focus on ultrasound features and rely heavily on consensus-driven rather than statistically validated risk stratification systems.6 15 Although TIRADS demonstrates robust sensitivity in clinical settings, its specificity remains relatively limited.16 Interobserver variability is another key concern—radiologists’ subjective interpretation of ultrasound features can result in inconsistent classification outcomes.17 To address these limitations, various strategies have been proposed, including the integration of artificial intelligence techniques to reduce observer subjectivity.18 19 20 Artificial intelligence has shown promise in matching or even surpassing the specificity achieved by radiologists; however, their clinical implementation remains constrained by challenges in interpretability and low acceptance in routine practice.

Integrating clinical risk factors may enhance risk stratification for thyroid nodules, as suggested by a growing body of evidence.21 In alignment with this, our study incorporated clinical variables including patient age, sex, number of nodules, and cervical LN status into the predictive model, thereby more accurately reflecting routine clinical diagnostic workflows. While previous studies22 23 24 suggested that male patients with thyroid nodules, particularly those with indeterminate fine-needle aspiration cytology undergoing molecular testing, exhibit a higher malignancy risk,25 our study did not identify a significant difference in thyroid cancer incidence between sexes. This discrepancy may be attributable to methodology differences, as molecular testing was not performed in our cohort and all diagnoses were confirmed through postoperative histopathology. The absence of statistical significance for male sex may reflect population-specific characteristics, such as regional variation in risk factor distribution or age composition.26 These methodological and demographic differences may have attenuated the observed sex-related effect. Nonetheless, male patients in our study were assigned higher risk scores, suggesting an association with malignancy risk, despite the lack of statistical significance.

Compared with previous models that primarily focused on intrinsic ultrasound features of thyroid nodules,27 28 29 our nomogram offers a more comprehensive assessment. Although the individual contributions of factors such as sex and age were relatively modest, they reflected subtle clinical patterns often considered by radiologists during decision making. The C-TIRADS optimisation approach demonstrated clear advantages, particularly in reducing unnecessary invasive procedures without compromising diagnostic accuracy, achieving an AUC of 0.972. Furthermore, the new model indicated that a risk threshold of ≥60% favoured the recommendation for C-TIRADS optimisation, whereas a threshold of <30% favoured exclusion. The integration of complex imaging data with clinical information represents a core competency for radiologists.30 With appropriate standardised training and communication frameworks in place, radiologists are well positioned to leverage quantitative metrics generated by the new model into routine diagnostic workflows. This advancement holds promise for improving diagnostic consistency and accuracy in clinical practice.

This study has several limitations that should be acknowledged. First, the optimisation of the TIRADS classification was influenced by radiologists’ subjective judgement, which may have contributed to interobserver variability. Second, although data collection was conducted by trained junior radiologists, observer variation and the subjective nature of ultrasound interpretation may have affected the model’s performance.31 Third, internal validation using bootstrap resampling may have overestimated model performance due to potential overfitting; therefore, external validation was essential to confirm generalisability. Fourth, owing to the retrospective design, only a limited set of clinical parameters (eg, sex, age, and cervical LN status) was included. Other relevant factors such as body mass index, environmental exposures, nodule location, family history of thyroid cancer, and radiation exposure history,32 33 were not assessed. Finally, the study cohort exclusively comprised cases confirmed by surgical pathology, resulting in a relatively low proportion of benign lesions, which may have introduced selection bias. The exclusion of patients diagnosed solely by fine-needle aspiration was intentional but may have affected the generalisability of the findings.

To address the limitations of the present study, future research should aim to standardise the application of TIRADS by adopting unified classification frameworks and implementing regular training programmes to enhance interobserver consistency. Prospective multicentre studies involving broader and more diverse populations are warranted, incorporating a wider range of clinical risk factors to improve predictive accuracy. In particular, data regarding family history, radiation exposure, and other relevant variables across centres would support more comprehensive risk assessment and enhance the generalisability of prediction models. In addition, including patients with fine-needle aspiration–confirmed benign nodules may help achieve a more balanced representation of benign and malignant cases. The development and application of nomogram-based structured training programmes for radiologists could also be explored to further improve diagnostic consistency and clinical utility. While the widespread adoption of a revised classification system will require time, we hope that the findings of this study may contribute to that transition.

We developed and externally validated a nomogram-based predictive model that integrates imaging features and clinical risk factors to optimise C-TIRADS classification for thyroid nodules. The model demonstrated good discrimination and calibration across internal and external cohorts, offering a practical tool to assist radiologists in refining diagnostic assessments and improving clinical decision making. Future research incorporating additional clinical variables and prospective validation is warranted to further strengthen the model’s applicability across diverse clinical settings.

Concept or design: Y Liang, Y Zou, P He, Q Chen.
Acquisition of data: Y Liang, Y Zou, Z Zou, B Ren.
Analysis or interpretation of data: Y Liang, S Peng, Y Zou.
Drafting of the manuscript: Y Liang, Y Zou, HM Yuan, Z Zou.
Critical revision of the manuscript for important intellectual content: P He, Y Zou.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

The authors have disclosed no conflicts of interest.

This manuscript was initially posted as a preprint entitled ‘Development and validation of a clinical prediction model to aid radiologists optimize thyroid C-TIRADS classification’ on Research Square (DOI: 10.21203/rs.3.rs-3831900/v1). After peer feedback and extensive revisions undertaken collaboratively by the author team, the current version has substantially evolved and markedly differs from the preprint version.

This research was supported by Sichuan Science and Technology Program (Ref Nos.:2025ZNSFSC1751, 2026YFHZ0039), the University-Industry Collaborative Education Program (Ref No.: 250505236300920), the University-level Project of North Sichuan Medical College (Ref Nos.: CXSY24-06, CBY22-QNA48), and the Hospital-level Projects of the Affiliated Hospital of North Sichuan Medical College, China (Ref Nos.: 210930, 2023-2GC013, 2025LC010). The funders had no role in the study design, data collection/analysis/interpretation, or manuscript preparation.

This research was approved by the Ethics Committee of Sichuan Provincial People’s Hospital (Ref No.: ER20210347) and the Ethics Committee of Affiliated Hospital of North Sichuan Medical College, China (Ref No.: 2021ER436-1). The requirement for informed patient consent was waived by both Committees due to the retrospective nature of the research.

The supplementary material was provided by the authors, and some information may not have been peer reviewed. Accepted supplementary material will be published as submitted by the authors, without any editing or formatting. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by the Hong Kong Academy of Medicine and the Hong Kong Medical Association. The Hong Kong Academy of Medicine and the Hong Kong Medical Association disclaim all liability and responsibility arising from any reliance placed on the content.

1. 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

2. Zhou J, Song Y, Zhan W, et al. Thyroid imaging reporting and data system (TIRADS) for ultrasound features of nodules: multicentric retrospective study in China. Endocrine 2021;72:157-70. Crossref

3. Trimboli P. Complexity in the interpretation and application of multiple guidelines for thyroid nodules: the need for coordinated recommendations for “small” lesions. Rev Endocr Metab Disord 2025;26:223-7. Crossref

4. Park JY, Lee HJ, Jang HW, et al. A proposal for a thyroid imaging reporting and data system for ultrasound features of thyroid carcinoma. Thyroid 2009;19:1257-64. Crossref

5. Horvath E, Majlis S, Rossi R, et al. An ultrasonogram reporting system for thyroid nodules stratifying cancer risk for clinical management. J Clin Endocrinol Metab 2009;94:1748-51. Crossref

6. Tessler FN, Middleton WD, Grant EG, et al. ACR Thyroid Imaging, Reporting and Data System (TI-RADS): white paper of the ACR TI-RADS Committee. J Am Coll Radiol 2017;14:587-95. Crossref

7. Shin JH, Baek JH, Chung J, et al. Ultrasonography diagnosis and imaging-based management of thyroid nodules: revised Korean Society of Thyroid Radiology consensus statement and recommendations. Korean J Radiol 2016;17:370-95. Crossref

8. Russ G, Bonnema SJ, Erdogan MF, Durante C, Ngu R, Leenhardt L. European Thyroid Association guidelines for ultrasound malignancy risk stratification of thyroid nodules in adults: the EU-TIRADS. Eur Thyroid J 2017;6:225-37. Crossref

9. Chen Z, Wang JJ, Du JB, et al. Development and validation of a dynamic nomogram for predicting central lymph node metastasis in papillary thyroid carcinoma patients based on clinical and ultrasound features. Quant Imaging Med Surg 2025;15:1555-70. Crossref

10. Boucai L, Zafereo M, Cabanillas ME. Thyroid cancer: a review. JAMA 2024;331:425-35. Crossref

11. Zhang J, Xu S. High aggressiveness of papillary thyroid cancer: from clinical evidence to regulatory cellular networks. Cell Death Discov 2024;10:378. Crossref

12. Ma T, Semsarian CR, Barratt A, et al. Rethinking low-risk papillary thyroid cancers <1 cm (papillary microcarcinomas): an evidence review for recalibrating diagnostic thresholds and/or alternative labels. Thyroid 2021;31:1626-38. Crossref

13. Kwong N, Medici M, Angell TE, et al. The influence of patient age on thyroid nodule formation, multinodularity, and thyroid cancer risk. J Clin Endocrinol Metab 2015;100:4434-40. Crossref

14. Pizzato M, Li M, Vignat J, et al. The epidemiological landscape of thyroid cancer worldwide: GLOBOCAN estimates for incidence and mortality rates in 2020. Lancet Diabetes Endocrinol 2022;10:264-72. Crossref

15. Tessler FN, Middleton WD, Grant EG, Hoang JK. Re: ACR Thyroid Imaging, Reporting and Data System (TI-RADS): white paper of the ACR TI-RADS Committee. J Am Coll Radiol 2018;15(3 Pt A):381-2. Crossref

16. Angelopoulos N, Goulis DG, Chrisogonidis I, et al. Diagnostic performance of European and American College of Radiology Thyroid Imaging Reporting and Data System classification systems in thyroid nodules over 20 mm in diameter. Endocr Pract 2025;31:72-9. Crossref

17. Jin Z, Pei S, Shen H, et al. Comparative study of C-TIRADS, ACR-TIRADS, and EU-TIRADS for diagnosis and management of thyroid nodules. Acad Radiol 2023;30:2181-91. Crossref

18. Wildman-Tobriner B, Buda M, Hoang JK, et al. Using artificial intelligence to revise ACR TI-RADS risk stratification of thyroid nodules: diagnostic accuracy and utility. Radiology 2019;292:112-9. Crossref

19. Wu SH, Li MD, Tong WJ, et al. Adaptive dual-task deep learning for automated thyroid cancer triaging at screening US. Radiol Artif Intell 2025;7:e240271. Crossref

20. Trimboli P, Colombo A, Gamarra E, Ruinelli L, Leoncini A. Performance of computer scientists in the assessment of thyroid nodules using TIRADS lexicons. J Endocrinol Invest 2025;48:877-83. Crossref

21. Kobaly K, Kim CS, Mandel SJ. Contemporary management of thyroid nodules. Annu Rev Med 2022;73:517-28. Crossref

22. Xu L, Li G, Wei Q, El-Naggar AK, Sturgis EM. Family history of cancer and risk of sporadic differentiated thyroid carcinoma. Cancer 2012;118:1228-35. Crossref

23. Iglesias ML, Schmidt A, Ghuzlan AA, et al. Radiation exposure and thyroid cancer: a review. Arch Endocrinol Metab 2017;61:180-7. Crossref

24. Saenko V, Mitsutake N. Radiation-related thyroid cancer. Endocr Rev 2024;45:1-29. Crossref

25. Figge JJ, Gooding WE, Steward DL, et al. Do ultrasound patterns and clinical parameters inform the probability of thyroid cancer predicted by molecular testing in nodules with indeterminate cytology? Thyroid 2021;31:1673-82. Crossref

26. Li X, Xing M, Tu P, et al. Urinary iodine levels and thyroid disorder prevalence in the adult population of China: a large-scale population-based cross-sectional study. Sci Rep 2025;15:14273. Crossref

27. Xiao J, Xiao Q, Cong W, et al. Discriminating malignancy in thyroid nodules: the nomogram versus the Kwak and ACR TI-RADS. Otolaryngol Head Neck Surg 2020;163:1156-65. Crossref

28. Xin Y, Liu F, Shi Y, Yan X, Liu L, Zhu J. A scoring system for assessing the risk of malignant partially cystic thyroid nodules based on ultrasound features. Front Oncol 2021;11:731779. Crossref

29. Zhou T, Hu T, Ni Z, et al. Comparative analysis of machine learning-based ultrasound radiomics in predicting malignancy of partially cystic thyroid nodules. Endocrine 2024;83:118-26. Crossref

30. Bluethgen C, Van Veen D, Zakka C, et al. Best practices for large language models in radiology. Radiology 2025;315:e240528. Crossref

31. He Z, Li Y, Zeng W, et al. Can a computer-aided mass diagnosis model based on perceptive features learned from quantitative mammography radiology reports improve junior radiologists’ diagnosis performance? An observer study. Front Oncol 2021;11:773389. Crossref

32. Kim Y, Roh J, Song DE, et al. Risk factors for posttreatment recurrence in patients with intermediate-risk papillary thyroid carcinoma. Am J Surg 2020;220:642-7. Crossref

33. Zhao J, Wen J, Wang S, Yao J, Liao L, Dong J. Association between adipokines and thyroid carcinoma: a meta-analysis of case-control studies. BMC Cancer 2020;20:788. Crossref