Development and validation of a predictive model for thyroid nodule malignancy risk based on intralesional and perilesional ultrasound radiomic features

Introduction

Thyroid carcinoma (TC) ranks among the most prevalent malignancies worldwide, with its incidence showing a sustained upward trend (1). Differentiated TC constitutes the overwhelming majority of TC cases and exhibits an overall survival rate as high as 92%, reflecting a relatively favorable prognosis (1,2). Ultrasound, as the preferred imaging modality for discriminating between benign and malignant thyroid nodules (TNs), plays an irreplaceable role in clinical practice (3,4). With continuous advancements in ultrasound technology, image resolution has improved substantially, enabling precise observation of lesion characteristics such as location, size, shape, calcification patterns, and internal echogenicity. Nevertheless, the diagnostic accuracy of ultrasound for TNs remains approximately 74–82%, as a considerable proportion of nodules lack pathognomonic malignant radiological features and therefore evade definitive detection (5). Chinese Thyroid Imaging Reporting and Data System (C-TIRADS) 4a/b nodules, in particular, present a formidable diagnostic challenge due to their broad malignancy risk range (5–80%), which introduces substantial diagnostic uncertainty and complicates clinical decision-making (6). To enhance diagnostic precision, ultrasound-guided fine-needle aspiration biopsy (US-FNAB) and molecular diagnostic techniques have been widely incorporated into TN screening protocols (2). Although these approaches have improved diagnostic accuracy to some extent, US-FNAB, as an invasive procedure, has several limitations. First, not all C-TIRADS 4a/b nodules warrant aspiration. Second, biopsy outcomes are susceptible to variability related to operator experience and sampling site selection. Third, approximately 30% of fine-needle aspiration specimens yield indeterminate results (7). Consequently, the development of a novel, accurate, and non-invasive risk stratification tool has become an urgent clinical imperative.

The advent of radiomics technology offers an innovative solution to this diagnostic challenge. Radiomics enables high-throughput extraction of high-dimensional quantitative features from medical images, transforming traditionally qualitative visual assessments into digital data suitable for in-depth analysis and revealing latent biological information embedded within images (8,9). Artificial intelligence-based radiomics further extends this capability by leveraging machine learning (ML) algorithms to automate feature extraction, enable intelligent feature selection, and construct high-precision predictive models. Compared with conventional subjective qualitative assessment, radiomics offers enhanced objectivity, superior reproducibility, and greater informational dimensionality (10). Multiple studies have confirmed that ultrasound-based radiomics models demonstrate strong diagnostic efficacy in distinguishing benign from malignant TNs (11,12). However, existing research has predominantly focused on imaging features within the lesional parenchyma, largely overlooking the diagnostic information contained in the perilesional region. The perilesional area refers to the tissue architecture within a defined proximity to the lesion margin and encompasses critical microenvironmental components such as adjacent normal parenchyma, vascular networks, and fibrous connective tissue. Tumor growth, infiltration, and invasion induce a cascade of pathophysiological changes in surrounding tissue, including angiogenesis, inflammatory cell infiltration, fibrotic proliferation, and extracellular matrix remodeling—reflecting microenvironmental restructuring (13). These changes not only manifest the biological behavior of the lesion but may also convey substantial diagnostic information. Based on this rationale, the present study aimed to construct a multimodal integrated predictive model that combines intralesional and perilesional radiomic features with clinical characteristics and to systematically evaluate its performance in differentiating benign and malignant C-TIRADS 4a/b TNs. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0143/rc).

Methods

Study subjects

This retrospective study enrolled patients with TNs who underwent thyroid ultrasound and US-FNAB at Jiaxing Maternal and Child Health Care Hospital between August 2023 and September 2025. All participants received a definitive pathological diagnosis based on US-FNAB results. After applying the inclusion and exclusion criteria, 147 eligible patients were included. Using a random sampling methodology, patients were randomly allocated in a 7:3 ratio to the training cohort (n=102) and validation cohort (n=45). The training cohort was used for model development, and the validation cohort was used to evaluate model generalizability. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by our Institutional Review Board (approval No. 2026-Y-017). As this study employed a retrospective analysis design, the requirement for individual informed consent was waived. The study workflow is illustrated in Figure S1.

Inclusion and exclusion criteria

Inclusion criteria were as follows: TNs classified as C-TIRADS 4a or 4b on ultrasound; nodules subjected to US-FNAB with a definitive pathological diagnosis; high-quality ultrasound images with complete imaging and clinical data; and, in cases of multiple nodules, inclusion of the largest lesion.

Exclusion criteria were prior thyroid surgery or radiofrequency ablation, suboptimal ultrasound image quality, concurrent malignancies at other anatomical sites, coexisting hyperthyroidism, pregnancy, or incomplete clinical data.

US-FNAB

Patients were positioned supine, and under real-time ultrasound guidance, the aspiration needle was advanced into the target nodule. Specimens were immediately smeared onto glass slides, fixed, and stained with hematoxylin and eosin for cytomorphological examination. US-FNAB results were classified according to the 2017 Bethesda System for Reporting Thyroid Cytopathology, with Bethesda categories V and VI defined as malignant (14).

Clinical and ultrasound characteristics

Clinical and ultrasound characteristics were extracted from the electronic medical record system, including sex, age, echotexture, maximum diameter, aspect ratio, shape, margin, calcification, vascularity, and C-TIRADS classification. One horizontal and one longitudinal ultrasound image capturing the maximum lesion diameter were selected for each nodule and stored in JPG format.

Delineation of regions of interest (ROIs)

All ultrasound images underwent grayscale normalization to reduce the impact of contrast variability on texture analysis. Two physicians with more than 5 years of experience in thyroid ultrasonography independently delineated intralesional ROI using ITK-SNAP version 3.8.0 under blinded conditions (http://www.itksnap.org). In cases of substantial discrepancies between the two physicians, a senior physician with over 15 years of experience adjudicated the final ROI. Perilesional ROI were generated by a uniform 2-mm outward expansion from the intralesional ROI boundary. This expansion distance was selected based on the rationale that it adequately captures the tumor-associated microenvironment since excessive expansion might introduce irrelevant normal tissue, thereby diminishing the diagnostic utility of perilesional features. All segmented ROIs were stored in neuroimaging informatics technology initiative (NIFTI) format, as shown in Figures S2,S3.

Radiomic feature extraction and selection

Following ROI delineation, radiomic features were extracted from intra- and perilesional ROIs using the PyRadiomics module (version 3.0.1) in Python 3.9 (15). Feature categories included shape, first-order statistics, and second-order texture features derived from the gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, gray-level dependence matrix, and neighboring gray tone difference matrix. Feature reproducibility was assessed using the intraclass correlation coefficient (ICC), with ICC values exceeding 0.75 indicating acceptable reliability.

After radiomic feature extraction, a large volume of high-dimensional features was obtained, inevitably containing redundant elements. A five-step feature optimization strategy was implemented for selection and dimensionality reduction: (I) removal of outliers; (II) Z-score standardization of all radiomic features; (III) Spearman correlation analysis to eliminate features with inter-feature correlation coefficients exceeding 0.9, thereby reducing multicollinearity; (IV) recursive feature elimination to exclude features with minimal predictive contribution; and (V) least absolute shrinkage and selection operator (LASSO) regression for further refinement, with 10-fold cross-validation to identify the optimal regularization parameter (lambda). Radiomic features with non-zero coefficients at the optimal lambda were retained for model construction.

Construction and validation of predictive models

Clinical characteristics selection

Univariate analysis was conducted to identify clinical and ultrasound variables associated with malignant TNs, using P<0.05 as the significance threshold. Variables identified as significant were entered into multivariate logistic regression (LR) analysis to determine independent predictors.

Radiomic prediction models

Using the selected radiomic features, radiomics scores (rad-scores) were calculated for each patient as weighted linear combinations of individual feature coefficients. LR, decision tree (DT), and light gradient boosting machine (LightGBM) were employed to construct radiomic prediction models based on horizontal intralesional, longitudinal intralesional, and combined horizontal and longitudinal intra- and perilesional features within the training cohort, with performance evaluated in the validation cohort. These algorithms have distinct properties, and their performance was evaluated using default parameters. LR, as a linear model, offers strong interpretability and stability (16); DT requires minimal data preprocessing and provides intuitive interpretability but is prone to instability and overfitting (17); LightGBM, an ensemble method based on gradient boosting, enables efficient training with low computational resource consumption (18). Model performance was comprehensively assessed using receiver operating characteristic (ROC) curves, area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The model with the highest AUC and balanced clinical metrics was selected as the optimal radiomic model.

Clinical-radiomics combined model

The independent clinical predictors were integrated with the rad-scores from the optimal radiomic model to construct a clinical-radiomics combined model. ROC curves were generated in both training and validation cohorts to evaluate predictive performance. Calibration curves were plotted to assess agreement between predicted and actual probabilities, with the Hosmer-Lemeshow (HL) test used to evaluate calibration (P>0.05 indicating adequate calibration). Decision curve analysis (DCA) was performed to quantify the net benefit across a range of threshold probabilities, thereby evaluating the clinical utility of the models in decision-making.

Interpretability analysis

To enhance model interpretability, the Shapley additive explanations (SHAP) framework was applied to analyze feature importance in the optimal predictive model (19). SHAP, grounded in Shapley value theory from cooperative game theory, quantifies each feature’s contribution to predictions at both the individual and aggregate levels by computing its marginal effect on model output.

Statistical analysis

Statistical analyses were performed using SPSS version 29.0 (IBM, New York, USA) and Python version 3.9 (http://www.python.org). For continuous factors with normal distribution, the independent sample t‑test was used, whereas, for continuous factors without normal distribution, the Mann-Whitney U test was used. Categorical variables were expressed as frequency (percentage) [n (%)] and compared using the χ² test or Fisher’s exact test, as appropriate.

Results

Baseline characteristics

This study ultimately enrolled 147 patients with TNs, aged 22–83 years. The malignancy rate was 51.1%. No significant differences were observed in baseline characteristics between the training and validation cohorts (all P>0.05), indicating satisfactory balance between groups (Table S1). Univariate analysis revealed that age, echotexture, calcification, aspect ratio, and shape were significantly associated with malignant nodules in the training cohort (P<0.05), whereas gender, calcification, aspect ratio, and shape were significantly associated with malignancy in the validation cohort (P<0.05) (Table S2). Multivariate LR analysis identified microcalcification, aspect ratio >1, and irregular shape as independent predictors of malignant TNs (Table S3).

Selecting radiomics features

In the consistency assessment, ICC values for ROI features independently delineated by the two physicians all exceeded 0.75, indicating satisfactory reproducibility and stability of manual segmentation. Radiomic features were extracted from three ROIs: 467 features from the horizontal intralesional region, 467 from the longitudinal intralesional region, and 1,868 from the integrated horizontal and longitudinal intra- and perilesional regions. Through a multi-step feature-selection process, LASSO regression (Figure 1) retained the most representative features: 12 from the horizontal intralesional region, five from the longitudinal intralesional region, and 20 from the integrated horizontal and longitudinal intra- and perilesional regions (Figure 2). Pearson correlation analysis demonstrated that the absolute correlation coefficients for the vast majority of feature pairs were <0.7, indicating low multicollinearity among retained features and an effectively reduced risk of model overfitting (Figure 3).

Figure 1 LASSO regression for ultrasound radiomics features screening. (A-C) Characteristic coefficient changing curve with lambda. (D-F) Lambda selection graph. LASSO, least absolute shrinkage and selection operator.

Figure 2 The optimal radiomics features selected by LASSO regression. LASSO, least absolute shrinkage and selection operator.

Figure 3 Heatmap of correlations among the optimal radiomics features.

Models performance evaluation

Rad-scores were constructed based on the selected features. Three algorithms—LR, DT, and LightGBM—were systematically applied to compare predictive performance across different ROI-based models (Table S4).

Horizontal intralesional model: The DT model achieved an AUC of 0.930 [95% confidence interval (CI): 0.886–0.967] in the training cohort, which declined to 0.767 in the validation cohort, indicating overfitting. The LR model demonstrated an AUC of 0.851 (95% CI: 0.775–0.915) in the training cohort and 0.808 (95% CI: 0.649–0.920) in the validation cohort. Accuracy, sensitivity, and specificity were consistently maintained at 0.78, indicating superior generalization performance (Figure 4A,4B).

Figure 4 Comparison of ROC curves of different models in the training and validation cohorts. AUC, area under the curve; DT, decision tree; LightGBM, light gradient boosting machine; LR, logistic regression; ROC, receiver operating characteristic.

Longitudinal intralesional model: the DT model yielded AUCs of 0.863 (95% CI: 0.792–0.925) and 0.813 (95% CI: 0.677–0.921) in the training and validation cohorts, respectively. This stable performance surpassed that of the LR model, highlighting the added value of longitudinal feature extraction (Figure 4C,4D).

Integrated horizontal and longitudinal intra- and perilesional model: In the training cohort, both LR and LightGBM models achieved AUCs exceeding 0.950, substantially outperforming single-region models. In the validation cohort, the LR model demonstrated optimal performance, with an AUC of 0.822 (95% CI: 0.670–0.940), accuracy of 0.800, sensitivity of 0.870, and specificity of 0.727. Although the improvement in validation AUC was relatively modest, the strong training performance and stable validation results support the feasibility and effectiveness of the multi-region feature integration strategy (Figure 4E,4F).

To further enhance predictive performance, independent clinical predictors were integrated with the combined horizontal and longitudinal intra- and perilesional rad-scores to construct a clinical-radiomic combined model. The multimodal fusion performance of LR, DT, and LightGBM algorithms was systematically compared (Table S4). In the training cohort, all three algorithms demonstrated exceptional discriminatory ability. The LightGBM model achieved the highest AUC of 0.974 (95% CI: 0.946–0.994), followed by LR (AUC: 0.959; 95% CI: 0.920–0.986) and DT (AUC: 0.940; 95% CI: 0.889–0.980). Sensitivity and specificity exceeded 0.86 for all models, indicating that feature fusion substantially improved predictive performance. In the validation cohort, the DT model exhibited the strongest generalization capacity, with an AUC of 0.882 (95% CI: 0.770–0.964). The LR model maintained stable performance, with an AUC of 0.866 (95% CI: 0.728–0.970), accuracy of 0.844, sensitivity of 0.913, and specificity of 0.773. The LightGBM model showed a decrease in AUC to 0.864, suggesting potential overfitting in more complex fusion scenarios. Overall, the LR algorithm demonstrated the most balanced and robust performance in the clinical-radiomics combined model (Figure 5A,5B). A systematic comparison of five LR-based predictive models was conducted (Table S5). Across both cohorts, AUC values increased progressively, with the clinical-radiomics combined model achieving the best performance and strongest generalization ability (Figure 6A,6B). Radiomic features significantly improved the limited sensitivity of the clinical-only model, providing more reliable support for clinical decision-making. HL test P values for the clinical-radiomic combined model exceeded 0.05 in both cohorts, indicating excellent calibration and concordance between predicted risk and actual observations. Calibration curves showed close agreement between predicted and observed probabilities, without appreciable deviation (Figure 6C). DCA revealed that across a wide threshold probability range (0.1–0.9), the clinical-radiomic combined model consistently provided greater net benefit than alternative models and both treat-all and treat-none strategies, indicating superior clinical utility (Figure 6D,6E).

Figure 5 ROC curve performance comparison of the clinical-radiomics combined model between the training and validation cohorts. AUC, area under the curve; DT, decision tree; LightGBM, light gradient boosting machine; LR, logistic regression; ROC, receiver operating characteristic.

Figure 6 Development and validation of the clinical-radiomics combined model. (A,B) ROC curves of different models based on logistic regression algorithm in the training (A) and validation cohorts (B). (C) Calibration curves of the clinical-radiomics combined model in the training and validation cohorts. (D,E) DCA of different models in the training (D) and validation cohorts (E). Model 1: clinical model. Model 2: horizontal intralesional model. Model 3: longitudinal intralesional model. Model 4: integrated horizontal and longitudinal intra- and perilesional model. Model 5: clinical-radiomics combined model. AUC, area under the curve; DCA, decision curve analysis; ROC, receiver operating characteristic.

Model interpretability analysis

The SHAP framework was applied to interpret the clinical-radiomic combined model. Global feature importance analysis showed that the integrated horizontal and longitudinal intra- and perilesional rad-scores were the most influential predictors, markedly exceeding traditional clinical variables and highlighting the dominant contribution of radiomic features to malignancy prediction (Figure 7A). SHAP force plots provided individualized explanations for representative malignant and benign cases, visually illustrating each feature’s contribution through color gradients and vector magnitude (Figure 7B,7C). This transparent, case-level interpretability enhances clinician understanding and confidence in model outputs, thereby strengthening support for evidence-based clinical decision-making.

Figure 7 Model interpretation by SHAP. (A) The overall contribution of each feature to the model prediction classification (class 0: non-response, class 1: response), in order of importance. (B) SHAP force plot of patients with malignant thyroid nodules. (C) SHAP force plot of patients with benign thyroid nodules. SHAP, Shapley additive explanations.

Discussion

With the widespread adoption of ultrasound and increased health awareness, the detection rate of TNs has shown a sustained increase (20). The C-TIRADS grading system, in particular, has provided standardized criteria for clinical diagnosis (21,22). However, when nodules are classified as grade 4a or higher, patients often opt for immediate fine-needle aspiration or surgical intervention, resulting in a considerable number of benign nodules undergoing unnecessary invasive procedures. To address this clinical challenge, radiomics technology has demonstrated promising potential in differentiating benign from malignant TNs in recent years (23,24). For example, Liang et al. and Chen et al. developed predictive models based on ultrasound radiomic features that outperformed junior clinicians in diagnostic accuracy (25,26). Wang et al. constructed a multimodal ultrasound imaging diagnostic model that achieved an AUC of 0.829, demonstrating strong diagnostic performance (27). Nevertheless, prior studies have predominantly focused on imaging features of the lesion itself, largely overlooking the biologically significant information contained within the perilesional region (28).

In the present study, microcalcification, aspect ratio >1, and irregular shape were identified as independent predictors of malignant TNs through multivariate LR analysis, showing substantial agreement with previous literature (29). Microcalcification is among the most pathognomonic ultrasound features of malignant TNs and is typically defined as punctate hyperechoic foci measuring less than 1 millimeter in diameter. These calcific deposits are generally considered to be associated with rapid tumor cell proliferation, necrosis, and stromal changes, reflecting pathological processes characteristic of malignancy (30,31). Benign TNs tend to conform to the anatomical architecture of normal thyroid tissue and encounter relatively low resistance to horizontal growth; consequently, their aspect ratio usually remains below 1. In contrast, malignant TNs often exhibit marked cellular atypia and infiltrative growth patterns, resulting in centripetal three-dimensional expansion and an increased likelihood of aspect ratios exceeding 1. Moreover, malignant TNs frequently display irregular margins, such as lobulated or spiculated contours (32,33). Although these conventional ultrasound features have diagnostic value, models based solely on them face notable limitations. In our study, the clinical model achieved a sensitivity of only 65.2% in the validation cohort, indicating that approximately one-third of malignant nodules were at risk of misclassification. Recognizing these limitations, this study integrated radiomics by employing three ML algorithms—LR, DT, and LightGBM—to construct horizontal intralesional, longitudinal intralesional, and integrated horizontal and longitudinal intra- and perilesional models using ultrasound images of TNs. This multi-regional ROI delineation strategy represents both a technical innovation and a biologically meaningful approach. The horizontal intralesional model captures imaging features in the horizontal plane, primarily reflecting spatial distribution, textural heterogeneity, and morphological characteristics in this cross-section. In the validation cohort, the LR-based horizontal model demonstrated robust stability, with an AUC of 0.808 (95% CI: 0.649–0.920), confirming satisfactory generalization ability. However, information from a single imaging plane may be insufficient to fully characterize the three-dimensional nature of lesions. Accordingly, we developed a longitudinal intralesional model focusing on the sagittal plane to better assess longitudinal growth patterns and depth of infiltration. This model also exhibited excellent stability, with the DT algorithm achieving an AUC of 0.813 (95% CI: 0.677–0.921), suggesting that longitudinal-plane features provide complementary biological information. TNs do not exist in isolation but instead interact dynamically with surrounding tissues. As malignant cells proliferate and invade, they induce inflammatory infiltration in adjacent regions and progressively disrupt normal thyroid parenchyma. The perilesional region is rich in microenvironmental components—including immune cells, extracellular matrix, and neovasculature—and dysfunction within this niche can promote tumor progression and invasion (34,35). Building on this biological rationale, the integrated horizontal and longitudinal intra- and perilesional model combined radiomic features derived from both lesional and perilesional regions. In the validation cohort, the LR-based integrated model achieved the best overall performance, with an AUC of 0.822 (95% CI: 0.670–0.940), accuracy of 80.0%, sensitivity of 87.0%, and specificity of 72.7%, thereby validating the effectiveness of multi-regional feature fusion. To further enhance predictive performance and clinical applicability, we integrated the identified independent clinical predictors with radiomic features from the integrated intra- and perilesional regions to develop a clinical-radiomic combined model. Among the three ML algorithms evaluated, LR demonstrated the greatest stability. In the validation cohort, the LR-based clinical-radiomic combined model achieved an AUC of 0.866 (95% CI: 0.728–0.970), with all performance metrics meeting standards for potential clinical application. Compared with the purely clinical model, this combined approach yielded substantial improvements across multiple dimensions, particularly in AUC, highlighting the synergistic value of integrating clinical and radiomic information. The clinical implementation of a predictive model depends not only on discriminative ability but also on calibration and net clinical benefit. The HL test indicated good calibration, supporting the reliability of predicted probabilities. DCA demonstrated that the clinical-radiomic combined model provided net benefit across a range of threshold probabilities, supporting its utility in individualized clinical decision-making. The SHAP framework enabled transparent interpretation of model predictions. Global feature importance analysis revealed that rad-scores derived from intralesional and perilesional regions contributed most substantially to the model’s output, exceeding the contribution of traditional clinical features. This finding underscores the added value of radiomics and confirms that radiomic features capture biologically relevant information not fully represented by conventional ultrasound assessment. Among clinical features, microcalcification, shape, and aspect ratio exhibited relatively modest SHAP values, indicating that individual morphological indicators have limited discriminatory power when considered in isolation.

This study has some limitations. First, as a single-center retrospective study with a relatively limited sample size and no external validation cohort, the generalizability of the model requires further confirmation through multicenter, large-scale prospective studies. Second, radiomic feature extraction depends heavily on accurate ROI delineation. In this study, ROIs were manually contoured, which may introduce subjectivity. Future work should explore automated or semi-automated segmentation methods to improve reproducibility. Third, this study relied exclusively on conventional two-dimensional ultrasound images and did not incorporate advanced imaging modalities such as contrast-enhanced ultrasound. Future research integrating multimodal imaging data may further improve diagnostic performance. Fourth, storing ultrasound images in lossy JPG format may distort and lose pixel-level high-frequency texture information essential for obtaining stable radiomic features and compromise data reproducibility, so future studies are advised to adopt original DICOM files to retain complete image data and guarantee reliable radiomic feature extraction.

Conclusions

This study successfully developed an integrated clinical-radiomic prediction model based on the LR algorithm, combining radiomic features from intralesional and perilesional regions with independent clinical predictors. The model demonstrated excellent diagnostic performance in distinguishing benign from malignant C-TIRADS 4a/b TNs. These findings confirm the important diagnostic value of perilesional information and provide clinicians with a convenient, non-invasive, and efficient auxiliary diagnostic tool.

Acknowledgments

We appreciate the linguistic assistance provided by TopEdit (www.topeditsci.com) during the preparation of this manuscript.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0143/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0143/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0143/prf

Funding: This work was supported by the Jiaxing Science and Technology Plan Project (No. 2024AD30123), the Medical and Health Science Program of Zhejiang Province (No. 2025HY1222) and Jiaxing Health Science and Technology Plan Project (No. JWKZ-25004).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-0143/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 our Institutional Review Board (approval No. 2026-Y-017). As this study employed a retrospective analysis design, the requirement for individual informed consent 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. Forma A, Kłodnicka K, Pająk W, et al. Thyroid Cancer: Epidemiology, Classification, Risk Factors, Diagnostic and Prognostic Markers, and Current Treatment Strategies. Int J Mol Sci 2025;26:5173. [Crossref] [PubMed]
2. Ringel MD, Sosa JA, Baloch Z, et al. 2025 American Thyroid Association Management Guidelines for Adult Patients with Differentiated Thyroid Cancer. Thyroid 2025;35:841-985. [Crossref] [PubMed]
3. Batawil N, Alkordy T. Ultrasonographic features associated with malignancy in cytologically indeterminate thyroid nodules. Eur J Surg Oncol 2014;40:182-6. [Crossref] [PubMed]
4. Kwak JY, Han KH, Yoon JH, et al. Thyroid imaging reporting and data system for US features of nodules: a step in establishing better stratification of cancer risk. Radiology 2011;260:892-9. [Crossref] [PubMed]
5. Pang T, Huang L, Deng Y, et al. Logistic regression analysis of conventional ultrasonography, strain elastosonography, and contrast-enhanced ultrasound characteristics for the differentiation of benign and malignant thyroid nodules. PLoS One 2017;12:e0188987. [Crossref] [PubMed]
6. Hu Y, Xu S, Zhan W. Diagnostic performance of C-TIRADS in malignancy risk stratification of thyroid nodules: A systematic review and meta-analysis. Front Endocrinol (Lausanne) 2022;13:938961. [Crossref] [PubMed]
7. Fu G, Chazen RS, MacMillan C, et al. Development of a Molecular Assay for Detection and Quantification of the BRAF Variation in Residual Tissue From Thyroid Nodule Fine-Needle Aspiration Biopsy Specimens. JAMA Netw Open 2021;4:e2127243. [Crossref] [PubMed]
8. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
9. 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]
10. Oberije C, Nalbantov G, Dekker A, et al. A prospective study comparing the predictions of doctors versus models for treatment outcome of lung cancer patients: a step toward individualized care and shared decision making. Radiother Oncol 2014;112:37-43. [Crossref] [PubMed]
11. Colakoglu B, Alis D, Yergin M. Diagnostic Value of Machine Learning-Based Quantitative Texture Analysis in Differentiating Benign and Malignant Thyroid Nodules. J Oncol 2019;2019:6328329. [Crossref] [PubMed]
12. Park VY, Lee E, Lee HS, et al. Combining radiomics with ultrasound-based risk stratification systems for thyroid nodules: an approach for improving performance. Eur Radiol 2021;31:2405-13. [Crossref] [PubMed]
13. Du JW, Li GQ, Li YS, et al. Identification of prognostic biomarkers related to the tumor microenvironment in thyroid carcinoma. Sci Rep 2021;11:16239. [Crossref] [PubMed]
14. Cibas ES, Ali SZ. The 2017 Bethesda System for Reporting Thyroid Cytopathology. Thyroid 2017;27:1341-6. [Crossref] [PubMed]
15. Aerts HJ, Velazquez ER, Leijenaar RT, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5:4006. [Crossref] [PubMed]
16. Wang QQ, Yu SC, Qi X, et al. Overview of logistic regression model analysis and application. Zhonghua Yu Fang Yi Xue Za Zhi 2019;53:955-60. [Crossref] [PubMed]
17. Rajaguru H. S R SC. Analysis of Decision Tree and K-Nearest Neighbor Algorithm in the Classification of Breast Cancer. Asian Pac J Cancer Prev 2019;20:3777-81. [Crossref] [PubMed]
18. Zhang J, Mucs D, Norinder U, et al. LightGBM: An Effective and Scalable Algorithm for Prediction of Chemical Toxicity-Application to the Tox21 and Mutagenicity Data Sets. J Chem Inf Model 2019;59:4150-8. [Crossref] [PubMed]
19. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems 2017;30.
20. Boers T, Braak SJ, Rikken NET, et al. Ultrasound imaging in thyroid nodule diagnosis, therapy, and follow-up: Current status and future trends. J Clin Ultrasound 2023;51:1087-100. [Crossref] [PubMed]
21. Torshizian A, Hashemi F, Khoshhal N, et al. Diagnostic Performance of ACR TI-RADS and ATA Guidelines in the Prediction of Thyroid Malignancy: A Prospective Single Tertiary Center Study and Literature Review. Diagnostics (Basel) 2023;13:2972. [Crossref] [PubMed]
22. Zhou J, Yin L, Wei X, et al. 2020 Chinese guidelines for ultrasound malignancy risk stratification of thyroid nodules: the C-TIRADS. Endocrine 2020;70:256-79. [Crossref] [PubMed]
23. Lee J, Song Y, Soh EY. Central lymph node metastasis is an important prognostic factor in patients with papillary thyroid microcarcinoma. J Korean Med Sci 2014;29:48-52. [Crossref] [PubMed]
24. Zhu YC, Jin PF, Bao J, et al. Thyroid ultrasound image classification using a convolutional neural network. Ann Transl Med 2021;9:1526. [Crossref] [PubMed]
25. Chen JH, Zhang YQ, Zhu TT, et al. Applying machine-learning models to differentiate benign and malignant thyroid nodules classified as C-TIRADS 4 based on 2D-ultrasound combined with five contrast-enhanced ultrasound key frames. Front Endocrinol (Lausanne) 2024;15:1299686. [Crossref] [PubMed]
26. Liang J, Huang X, Hu H, et al. Predicting Malignancy in Thyroid Nodules: Radiomics Score Versus 2017 American College of Radiology Thyroid Imaging, Reporting and Data System. Thyroid 2018;28:1024-33. [Crossref] [PubMed]
27. Wang SR, Zhu PS, Li J, et al. Study on diagnosing thyroid nodules of ACR TI-RADS 4-5 with multimodal ultrasound radiomics technology. J Clin Ultrasound 2024;52:274-83. [Crossref] [PubMed]
28. Cao Y, Zhong X, Diao W, et al. Radiomics in Differentiated Thyroid Cancer and Nodules: Explorations, Application, and Limitations. Cancers (Basel) 2021;13:2436. [Crossref] [PubMed]
29. Frates MC, Benson CB, Charboneau JW, et al. Management of thyroid nodules detected at US: Society of Radiologists in Ultrasound consensus conference statement. Radiology 2005;237:794-800. [Crossref] [PubMed]
30. Li X, Zhou W, Zhan W. Clinical and ultrasonographic features of medullary thyroid microcarcinomas compared with papillary thyroid microcarcinomas: a retrospective analysis. BMC Med Imaging 2020;20:49. [Crossref] [PubMed]
31. Yin L, Zhang W, Bai W, et al. Relationship Between Morphologic Characteristics of Ultrasonic Calcification in Thyroid Nodules and Thyroid Carcinoma. Ultrasound Med Biol 2020;46:20-5. [Crossref] [PubMed]
32. Kim EK, Park CS, Chung WY, et al. New sonographic criteria for recommending fine-needle aspiration biopsy of nonpalpable solid nodules of the thyroid. AJR Am J Roentgenol 2002;178:687-91. [Crossref] [PubMed]
33. Moon HJ, Kwak JY, Kim MJ, et al. Can vascularity at power Doppler US help predict thyroid malignancy? Radiology 2010;255:260-9. [Crossref] [PubMed]
34. Xia J, Shi Y, Chen X. New insights into the mechanisms of the extracellular matrix and its therapeutic potential in anaplastic thyroid carcinoma. Sci Rep 2024;14:20977. [Crossref] [PubMed]
35. Xu H, Ma J, Li N, et al. Comprehensive single-cell RNA analysis reveals intertumoral microenvironment heterogeneity and hub niche of carcinogenesis in thyroid cancer. NPJ Precis Oncol 2025;9:379. [Crossref] [PubMed]