Multimodal model enhances qualitative diagnosis of hypervascular thyroid nodules: integrating radiomics and deep learning features based on B-mode and PDI images

Introduction

In recent years, there has been a marked increase in the incidence of thyroid cancer, and 72.6% of cases occurring in Asia (1-3). Hypervascular thyroid nodules refer to rich blood supply nodules identified using blood flow imaging techniques such as color Doppler flow imaging (CDFI) or power Doppler imaging (PDI). Benign hypervascular nodules predominantly consist of follicular adenomas and nodular goiters, whereas malignant cases mainly encompass follicular carcinoma, medullary carcinoma, or follicular variant of papillary thyroid carcinoma (4-6). Some benign hypervascular thyroid nodules exhibit borderline behavior and require close follow-up, like oncocytic adenoma (7). Malignant hypervascular nodules are often associated with aggressive clinical behavior and less favorable prognoses (6,8,9). For example, follicular carcinoma poses a high risk of early metastasis (9), while medullary carcinoma is commonly linked to hereditary syndromes (8), both necessitating timely diagnosis and intervention.

Accurately differentiating between benign and malignant hypervascular nodules remains a significant clinical challenge. In clinical practice, distinguishing between benign and malignant follicular tumors typically depends on histological evidence such as vascular invasion or capsular infiltration, features that cannot be directly visualized using conventional imaging modalities (10). The imaging characteristics of malignant nodules frequently overlap with those of benign lesions, complicating differentiation (10,11). Alexander et al. demonstrated that ultrasound features, such as hypoechoic, microcalcifications, and infiltrative margins, are associated with a high suspicion of malignancy (12). However, the ultrasound features associated with hypervascular nodules, like follicular carcinoma, were inconsistent among different studies, and the positive predictive values of these ultrasound features are low (ranging from 55.6% to 61.2%) (10,11). Traditional blood flow imaging techniques are capable of visualizing vascular signals. PDI is sensitive in detecting slow flow within small vessels, enabling it to better visualize blood supply and delineate intranodular vascular architecture within thyroid nodules without being affected by the angle of ultrasound beam incidence (13). Current evidence indicates that vascularization correlates with the malignant potential of nodules (14). Moon et al. suggested that combining grayscale ultrasound features with PDI patterns improves diagnostic accuracy for differentiating benign from malignant thyroid nodules (13). Although traditional blood flow imaging techniques offer insights into blood flow patterns, they fail to objectively characterize the spatial complexity of vascular distribution, which contains critical diagnostic information beyond visual assessment. Therefore, developing a non-invasive, objective diagnostic tool to accurately predict nodule characteristics is essential for optimizing clinical decisions and reducing patient risks.

Artificial intelligence (AI) models present a highly promising solution for addressing the limitations of traditional diagnostic methods. AI models show improved diagnostic accuracy and help reduce differences in ultrasonographic feature analysis of thyroid nodules among observers, leading to more consistent results (15). Several AI and machine learning techniques have been developed to assist in the classification of thyroid nodules and the early detection of cancers. These include enhancements to the American College of Radiology Thyroid Imaging Reporting and Data System (TIRADS), which can also be applied manually (16,17). Another important advantage of AI systems is their ability to yield more systematized results, which can reduce inter-observer variability and contribute to the standardization of results obtained through the application of various TIRADS classification systems (18,19). As far as we know, however, the AI models developed for differentiate hypervascular thyroid nodules are limited.

Therefore, this study aims to develop machine learning models based on B-mode and PDI images to achieve accurate qualitative diagnosis of hypervascular thyroid nodules. The approach seeks to minimize reliance on invasive procedures and enhance diagnostic consistency across healthcare settings, ultimately supporting precision-based management of thyroid nodules. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-183/rc).

Methods

Study design and participants

This study included a retrospective cohort of patients with thyroid nodules diagnosed by surgical pathology at West China Hospital from August 2022 to December 2023.

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Review Committee of West China Hospital, Sichuan University (ethical approval number: 2021-171) and informed consent was obtained from all individual participants. Clinical trial registration number: ChiCTR2100049742.

Patients included in this study granted written permission for anonymized data use for research purposes at the time of biopsy or surgery. The inclusion criteria were as follows: (I) pathologic confirmation of thyroid nodules by surgery, (II) availability of PDI for evaluation, and (III) hypervascularity (defined as Adler grade ≥2). The exclusion criteria were the following: (I) multiple thyroid lesions, (II) history of thyroid surgery, and (III) incomplete clinical records or ultrasound images. Eligible patients were divided into a training cohort (n=220) and a test cohort (n=95), with no overlap between the two groups. Ultimately, a total of 315 patients were included in the analysis (Figure 1).

Figure 1 Patient selection flow diagram.

Data collection and image acquisition

The clinical profiles and pathological diagnosis of patients were collected from electronic medical records and ultrasound report. Ultrasound was performed within 3 days before surgery. The protocol for image acquisition is as follows: Patients lie supine with their necks exposed and shoulders elevated for mild neck extension. A high-frequency liner probe (L3-12A, Samsung RS80, South Korea) is used to adjust the gain, depth, and focus position. After a comprehensive scan, static images with nodules clearly centered are saved. For PDI images, reduce pulse repetition frequency to detect low-velocity blood flow and adjust color gain to eliminate background noise. Blood flow distribution is assessed on the nodule’s long-axis section to determine the Adler grade (0–3), and images with the most abundant blood flow are saved. During the procedure, excessive probe pressure should be avoided to prevent attenuation of blood flow signals.

Tumor segmentation, feature extraction and feature selection

The overall AI workflow was shown in Figure 2. One B-mode image and one PDI image of the tumors’ major axis planes were selected for model development. For B-mode images, thyroid nodules were manually segmented. Procedures mentioned above was performed by a senior radiologist with more than 5 years of practice experience blinded to pathological results. For PDI images, we used a red-channel thresholding method in the red, green, blue (RGB) color space to extract blood flow regions. Pixels with high red values and relatively low green and blue values were identified as blood, as blood typically appears with strong red color in the images.

Figure 2 Overall workflow of model construction and evaluation. AUC, area under the curve; ROC, receiver operating characteristic; SHAP, SHapley Additive exPlanations; TABPFN, Tabular Prior-data Fitted Network; XGBoost, extreme gradient boosting.

Based on segmented images, 1,910 radiomics feature extraction were implemented by using PyRadiomics library (version 3.0.1) (20,21). All ultrasound images were normalized and the grey level values were scaled by a factor of 100. Image discretization was performed with the bandwidths of 5 and the voxel array shift of 300 to prevent negative values. The final feature set integrates base features [91], mathematical transformation features [364], wavelet features [364], and LoG-filtered features [273], comprehensively characterizing nodule morphology and vascular properties.

A pre-trained ResNet model is employed to extract abstract deep learning features (Figure S1). Input images are converted from single-channel to three-channel format, resized to 224×224 pixels, and normalized using ImageNet mean and standard deviation. Features are extracted from the global average pooling layer of ResNet, yielding 2048 vectors. Principal component analysis is applied to reduce redundancy, retaining principal components that account for ≥95% cumulative variance, resulting in 1,000 features. This process captures high-dimensional patterns such as boundary complexity and internal heterogeneity, complementing radiomics features.

The 2,910 hybrid features (1,910 radiomics + 1,000 deep learning) undergo Z-score normalization to eliminate scale differences. Analysis of variance (ANOVA) F-tests are used to select features significantly associated with pathological diagnosis labels. The top 10 features from grayscale images and PDI images are selected (22,23).

Single-modality and combined model construction

To develop machine learning models, we employed five modeling strategies based on B-mode and PDI images in training set. Five machine learning models—random forest (RF), logistic regression (LR), support vector machine (SVM), eXtreme Gradient Boosting (XGBoost) (24,25), and Tabular Prior-data Fitted Network (TABPFN) (26)—were trained, with hyperparameter tuning and cross-validation to optimize performance. After training, the optimal model was identified through systematical performance evaluation. Combined model was constructed by fusing features from B-mode and PDI images. A joint feature set was created by integrating key features selected from both modalities, and the best model was retained for training with consistent hyperparameters to ensure comparability.

Performance assessment and statistical analysis

Model performance was comprehensively evaluated through confusion matrix, receiver operating characteristic (ROC) curve, and SHapley Additive exPlanations (SHAP) feature analysis. The confusion matrix calculates accuracy, precision, recall, and F1-score to intuitively reflect classification results. The ROC curve plots the relationship between true positive rate and false positive rate, with area under the curve (AUC) value quantifying overall discriminative ability. SHAP identifies the key discriminative features. These methods collectively assess classification accuracy, threshold robustness, and interpretability.

Categorical data were analyzed with the Chi-squared test, and continuous variables were analyzed with Mann-Whitney U test. Statistical tests were conducted using SPSS version 25.0 (IBM Corp., Armonk, NY, USA) and Python version 3.8.5 (Python Software Foundation, Wilmington, DE, USA). All tests were two-sided, and a P value <0.05 was considered to be statistically significant.

Results

Patient characteristics

A total of 315 patients with thyroid nodules were included, divided into a training cohort (n=220) and a test cohort (n=95). Baseline characteristics showed no significant differences between cohorts (Table 1). Pathological diagnoses confirmed 69.8% (220/315) malignancies, predominantly classical papillary carcinoma (56.5%, 178/315), follicular variant of papillary carcinoma (8.6%, 27/315), and medullary carcinoma (3.2%, 10/315). Benign cases (30.2%, n=95) included follicular adenoma (14.6%, n=46), nodular goiter (7.9%, n=25), and neoplasms of uncertain potential (5.1%, n=16).

Data for training set (n=220), test set (n=95), and overall set (n=315) are displayed as median (interquartile range, interquartile range) for continuous variables and count (percent) for categorical variables. Age is represented in years, and tumor size are presented by major axis of nodules (mm).

Feature selection

Through ANOVA F-test, radiomic features were proven to be more important than deep learning features in differencing malignant and benign hypervascular thyroid nodules. For B-mode images, strongly associated features included logarithm_firstorder_Energy, original_firstorder_Energy, wavelet-L_firstorder _TotalEnergy, wavelet-L_firstorder_Energy, squareroot_firstorder_TotalEnergy, squareroot_firstorder_Energy, original_firstorder_TotalEnergy, square_glszm_ SizeZoneNonUniformity, logarithm_firstorder_TotalEnergy, square_glrlm_RunLengthNonUniformity. For PDI images, strongly associated features included squareroot_firstorder_Minimum, log-sigma-1-0-mm-3D_glszm_RunLengthNonUniformity, log-sigma-1-0-mm -3D_glszm_SizeZone NonUniformity, diagnostics_Mask-original_VolumeNum, log-sigma-3-0-mm-3D_ngtdm_Busyness, logarithm_glszm_GrayLevelNonUniformity, wavelet-H_glszm_SizeZoneNonUniformity, logarithm_glszm_GrayLevelNonUniformity, wavelet-H_glszm_ RunLengthNonUniformity and wavelet-H _glcm_Correlation. These statistically significant features were subsequently employed to develop machine learning classification models.

Performance of single-modality models

The diagnostic performance of machine learning models for differentiating hypervascular thyroid nodules is summarized in Table 2 and Figure 3A. In the B-mode analysis, TABPFN and LR shared the highest accuracy (0.86), with TABPFN additionally achieving precision of 0.87, recall of 0.94, F1-score of 0.90, and AUC of 0.88. SVM performed robustly with an accuracy of 0.84, precision of 0.85, recall of 0.94, F1-score of 0.89, and the highest AUC (0.89) among B-mode models. XGBoost and RF exhibited similar accuracy (0.81–0.80) and AUC values (0.83–0.85), though XGBoost demonstrated a slightly higher recall (0.88 vs. 0.85). LR and SVM shared identical recall (0.94) and nearly equivalent F1-scores (0.90 vs. 0.89), but LR had a marginally higher accuracy (0.86 vs. 0.84). Overall, SVM excelled in AUC for two modalities.

Figure 3 ROC curves of models in test set. (A) ROC curve of the B-mode based model; (B) ROC curve of the PDI based model. AUC, area under the curve; PDI, Power Doppler imaging; ROC, receiver operating characteristic; TABPFN, Tabular Prior-data Fitted Network; XGBoost, extreme gradient boosting.

Among the PDI-mode models, the SVM achieved the highest AUC of 0.86, with an accuracy of 0.82, precision of 0.81, recall of 0.97, and F1-score of 0.88). The TABPFN model demonstrated the highest recall rate (0.98) and tied with LR for the top accuracy (0.86), while also achieving a precision of 0.84, F1-score of 0.91, and AUC of 0.84. LR exhibited perfect recall (1.00) and the highest F1-score (0.91) but had a relatively lower AUC of 0.82. XGBoost and RF showed comparable accuracy (0.81 vs. 0.80) and F1-scores (0.87), though XGBoost had a marginally higher AUC (0.83 vs. 0.82). Precision values across all PDI models ranged from 0.79 to 0.84 (Table 2 and Figure 3B). The SVM model was selected as the optimal model for constructing the fused model due to its strong AUC performance in both B-mode (0.89) and PDI (0.86).

Performance of SVM models

For the test sets, the B-mode SVM (Table 3 and Figure 4A,4B) achieved an accuracy of 0.84, precision of 0.85, recall of 0.94, F1-score of 0.89, and AUC of 0.89, while the PDI SVM model (Table 3 and Figure 4C,4D) demonstrated a test set AUC of 0.86, accuracy of 0.82, precision of 0.81, recall of 0.97, and F1-score of 0.88. Feature importance analysis (Figure 4E,4F) revealed key discriminative features for the B-mode model, such as logarithm_firstorder_Energy, original_firstorder_Energy and wavelet-L_firstorder_TotalEnergy, while the PDI model relied on features like squareroot_firstorder_Minimum, logarithm_glrlm_GrayLevelNonUniformity, and diagnostics_Mask-original_VolumeNum, which significantly contributed to its performance. Although these radiomics features are engineered descriptors, they often reflect underlying tumor biology. For instance, TotalEnergy indicates tissue compactness or cellular density, while Minimum signal intensity may suggest necrotic regions or intratumoral heterogeneity. In addition, wavelet-based features capture multi-scale texture variations that are typically associated with tumor structural complexity.

Figure 4 Performance evaluation of single-modality SVM models in train and test. (A) Confusion matrix of B-mode SVM; (B) ROC curve of B-mode SVM; (C) confusion matrix of PDI SVM; (D) ROC curve of PDI SVM; (E) SHAP analysis of B-mode SVM; (F) SHAP analysis of PDI SVM. Color bar on the right of confusion matrix represents the magnitude of the counts in each cell using color intensity from light blue (low count) to dark blue (high count). AUC, area under the curve; PDI, Power Doppler imaging; ROC, receiver operating characteristic; SHAP, SHapley Additive exPlanations; SVM, support vector machine.

The combined model, built by integrating the B-mode and PDI SVM models, showed exceptional performance on the training set with an accuracy of 0.96 and AUC of 1.00 (Table 3 and Figure 5A,5B). However, its test set performance declined, achieving an accuracy of 0.78 and AUC of 0.89. Compared to single-modality models, the fused model outperformed them on the training set (p<0.001), but its test set AUC was only marginally higher than those of B-mode (0.89) and PDI (0.86). While the fused model demonstrated superior test set recall (0.95) and F1-score (0.86) compared to some single-modality models, its lower accuracy (0.78) and precision (0.78) suggest potential overfitting or generalization limitations (Table 3 and Figure 5). To mitigate potential overfitting observed in the fused model, we implemented dropout layers (dropout rate =0.3) and L2 weight regularization (λ=1e−4) during model training to enhance generalization and constrain complexity. Feature importance analysis (Figure 5C) revealed key discriminative features included wavelet-L_firstorder_Energy_tumor, wavelet-L_firstorder_TotalEnergy_tumor and logarithm_firstorder_Energy_tumor. Additionally, we conducted a comparative experiment using SVM models trained on the top 10 features selected from radiomics, deep learning, and combined feature sets (Figure S2 and Table S1), to further evaluate the discriminative power of each modality and their combination.

Figure 5 Performance evaluation of combined models. (A) Confusion matrix; (B) ROC curve; (C) SHAP analysis. Color bar on the right of confusion matrix represents the magnitude of the counts in each cell using color intensity from light blue (low count) to dark blue (high count). AUC, area under the curve; ROC, receiver operating characteristic; SHAP, SHapley Additive exPlanations; SVM, support vector machine.

Discussion

This research focused on improving the diagnosis of hypervascular thyroid nodules by developing machine learning models that leverage both B-mode and PDI features. Critical distinguishing characteristics were meticulously analyzed from segmented tumor regions and integrated into hybrid models. Among the tested models, the SVM emerged as the most effective single-modality model, achieving high accuracy and robust AUC values in test sets. A combined model that fused both modalities exhibited nearly perfect performance during training, although its generalizability was slightly diminished during testing. The surpassing performance of the combined model demonstrated its capacity to capture complementary diagnostic patterns. These findings underscore the clinical reliability of SVM and emphasize the significance of integrating multimodal imaging features to enhance diagnostic precision. In summary, such models exhibit substantial potential as tools to augment traditional diagnostic workflows.

AI systems can identify subtle diagnostic patterns concealed within large datasets, offering clinicians valuable, objective decision support and enhancing diagnostic reliability, particularly for complex cases. Several studies investigated diagnostic performance of AI models in differentiating hypervascular thyroid nodules. Yu et al. constructed integrative model incorporating sonographic radiomic signatures, conventional ultrasound features, and clinical parameters and demonstrated an AUC of 0.844 in differentiating follicular adenoma from carcinoma (10). Yoon et al. constructed a nomogram using ultrasound features and cytopathology results to predict the malignancy of thyroid nodules diagnosed as undetermined follicular lesions and achieved AUCs of 0.769–0.817 (27). On another study, Shin et al. achieved an accuracy of only 0.741 and 0.69 using artificial neural network and SVM, respectively, based preoperative ultrasonography in differentiating follicular adenoma from carcinoma (28). Zhang et al. developed a deep learning model to differentiate pathologically proven atypical and typical medullary thyroid carcinoma from follicular thyroid adenoma and achieved better performance than junior radiologists (29). Compared with previously reported studies, the combined model integrating B-mode and PDI features demonstrated near-perfect training performance (AUC: 1.0). However, its test performance (AUC: 0.89) indicated challenges in generalizability, potentially attributable to feature redundancy or a lack of sufficient external validation.

The combined model’s superior recall and F1-score compared to single-modality models underscored the complementary value of multimodal imaging. B-mode provides morphological details such as boundary irregularity and calcification patterns (27,29), while PDI quantifies vascular spatial complexity, including disordered “penetrating” flow in malignancies versus organized “peripheral” patterns in benign nodules (13). Multimodal ultrasound provides comprehensive characteristic information of thyroid nodules from diverse perspectives, thereby enhancing diagnostic accuracy (30). Synergistic integrating with multimodal ultrasound, AI has opened up new possibilities for enhancing diagnostic efficacy. Several studies have explored the development of multimodal ultrasound-based diagnostic models leveraging deep learning techniques, yielding promising preliminary results (31,32). This synergy aligns with clinical needs for comprehensive diagnostic tools that reduce reliance on biopsies.

The dominance of radiomics features over deep learning-derived ones in feature selection contrasts with trends in other cancer studies, where deep learning often outperforms handcrafted features. Ma et al. (33) developed a diagnostic model for thyroid nodule classification by leveraging deep learning features, which exhibited superior diagnostic performance compared to conventional radiomics models (AUC: 0.948–0.961). Ardakani et al. (34) integrated radiomics features with deep learning features extracted from ultrasound images to construct a hybrid model for differentiating benign from malignant thyroid nodules. Their results demonstrated that the hybrid model achieved better diagnostic performance than either radiomics-only or deep learning-only approaches. This discrepancy may reflect domain-specific nuances in thyroid imaging, where textural and morphological descriptors capture subtle variations in nodule heterogeneity more effectively than abstract deep learning patterns.

SVM’s consistency across both modalities highlights its suitability for high-dimensional and is critical consideration for clinical deployment. The high recall rates of SVM models minimize false negatives—a vital attribute given the aggressive nature of malignant hypervascular thyroid nodule. Research on SVM models for differentiating benign and malignant thyroid nodules has demonstrated significant potential. By leveraging their inherent capability to handle high-dimensional data, SVM models employ meticulously designed kernel functions to project data into optimized feature spaces, thereby improving classification accuracy—a critical advantage when addressing the complexities of thyroid nodule datasets (35,36). Nevertheless, the diagnostic performance of SVM models remains highly dependent on effective feature selection and high-quality data. Features that are poorly engineered or issues with data quality can substantially diminish model effectiveness. Moreover, the limited interpretability of SVM models, stemming from their less transparent decision-making process, has hindered their broader clinical adoption, despite their technical merits (37).

Several limitations warrant attention. First, the retrospective, single-center design and modest sample size may limit generalizability, particularly given the predominance of papillary carcinoma in the cohort. Prospective validation across different institutions, especially different operators and machines, as well as diverse populations, including rarer subtypes like medullary carcinoma, is essential to confirm robustness. In this study, all ROI segmentation tasks were performed by a single physician, introducing potential bias. Future implementation of artificial intelligence tools or models for ROI segmentation offers a promising approach to address this limitation. Second, reliance on 2D images and single-plane analysis (major axis) may overlook 3D spatial information, which could enhance diagnostic accuracy. Future studies should explore volumetric features or multiplanar fusion to better represent lesion complexity. Third, the performance gap in the combined model between training and testing signals potential overfitting, emphasizing the need for stricter regularization or external cohorts to validate feature stability.

Conclusions

In conclusion, the combined models leveraging B-mode and PDI images demonstrate excellent performance in differentiating benign from malignant hypervascular thyroid nodules. Radiomic features were extracted and SVM models were employed to construct the optimal integrated framework. These tools improve malignancy detection accuracy, support personalized patient management, and advance precision-based clinical workflows for thyroid lesion evaluation.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-183/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-183/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-183/prf

Funding: This study was supported by the Natural Science Foundation of Sichuan Province under Grant (No. 2023NSFSC1722) and the Natural Science Foundation of Sichuan Province under Grant (No. 2024YFFK0226).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-183/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Review Committee of West China Hospital, Sichuan University (ethical approval number: 2021-171) and informed consent was obtained from all individual 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. Ferlay J, Ervik M, Lam F, et al. Global Cancer Observatory: Cancer Today. Lyon, France: International Agency for Research on Cancer; 2024.
2. Kitahara CM, Schneider AB. Epidemiology of Thyroid Cancer. Cancer Epidemiol Biomarkers Prev 2022;31:1284-97. [Crossref] [PubMed]
3. Lebbink CA, Links TP, Czarniecka A, et al. 2022 European Thyroid Association Guidelines for the management of pediatric thyroid nodules and differentiated thyroid carcinoma. Eur Thyroid J 2022;11:e220146. [Crossref] [PubMed]
4. Zhang F, Mei F, Chen W, et al. Role of Ultrasound and Ultrasound-Based Prediction Model in Differentiating Follicular Thyroid Carcinoma From Follicular Thyroid Adenoma. J Ultrasound Med 2024;43:1389-99. [Crossref] [PubMed]
5. Liu X, Zhu Q, Liu J, et al. Ultrasound characteristics of medullary thyroid carcinoma. Chinese Journal Medical Ultrasound 2021;18:266-71. (Electronic Edition).
6. Halada S, Baran JA, Bauer AJ, et al. Clinicopathologic Characteristics of Pediatric Follicular Variant of Papillary Thyroid Carcinoma Subtypes: A Retrospective Cohort Study. Thyroid 2022;32:1353-61. [Crossref] [PubMed]
7. Asa SL, Mete O. Oncocytic Change in Thyroid Pathology. Front Endocrinol (Lausanne) 2021;12:678119. [Crossref] [PubMed]
8. Pelizzo MR, Mazza EI, Mian C, et al. Medullary thyroid carcinoma. Expert Rev Anticancer Ther 2023;23:943-57. [Crossref] [PubMed]
9. Yamazaki H, Sugino K, Katoh R, et al. Management of follicular thyroid carcinoma. Eur Thyroid J 2024;13:e240146. [Crossref] [PubMed]
10. Yu B, Li Y, Yu X, et al. Differentiate Thyroid Follicular Adenoma from Carcinoma with Combined Ultrasound Radiomics Features and Clinical Ultrasound Features. J Digit Imaging 2022;35:1362-72. [Crossref] [PubMed]
11. Yoon JH, Kim EK, Youk JH, et al. Better understanding in the differentiation of thyroid follicular adenoma, follicular carcinoma, and follicular variant of papillary carcinoma: a retrospective study. Int J Endocrinol 2014;2014:321595. [Crossref] [PubMed]
12. Alexander EK, Cooper D. The importance, and important limitations, of ultrasound imaging for evaluating thyroid nodules. JAMA Intern Med 2013;173:1796-7. [Crossref] [PubMed]
13. 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]
14. Li W, Ge Z, Cai S, et al. Diagnostic value of greyscale ultrasound combined with superb microvascular imaging in thyroid nodules: a systematic review and meta-analysis. Quant Imaging Med Surg 2025;15:440-54. [Crossref] [PubMed]
15. Sorrenti S, Dolcetti V, Radzina M, et al. Artificial Intelligence for Thyroid Nodule Characterization: Where Are We Standing? Cancers (Basel) 2022;14:3357. [Crossref] [PubMed]
16. Ulisse S, Baldini E, Lauro A, et al. Papillary Thyroid Cancer Prognosis: An Evolving Field. Cancers (Basel) 2021;13:5567. [Crossref] [PubMed]
17. Baldini E, Tuccilli C, Pironi D, et al. Expression and Clinical Utility of Transcription Factors Involved in Epithelial-Mesenchymal Transition during Thyroid Cancer Progression. J Clin Med 2021;10:4076. [Crossref] [PubMed]
18. Prieditis P, Radzina M, Mikijanska M, et al. Non-Marked Hypoechogenic Nodules: Multicenter Study on the Thyroid Malignancy Risk Stratification and Accuracy Based on TIRADS Systems Comparison. Medicina (Kaunas) 2022;58:257. [Crossref] [PubMed]
19. Gild ML, Chan M, Gajera J, et al. Risk stratification of indeterminate thyroid nodules using ultrasound and machine learning algorithms. Clin Endocrinol (Oxf) 2022;96:646-52. [Crossref] [PubMed]
20. van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
21. Liu Q, Jiang P, Jiang Y, et al. Prediction of Aneurysm Stability Using a Machine Learning Model Based on PyRadiomics-Derived Morphological Features. Stroke 2019;50:2314-21. [Crossref] [PubMed]
22. Jiang Z, Wang B, Han X, et al. Multimodality MRI-based radiomics approach to predict the posttreatment response of lung cancer brain metastases to gamma knife radiosurgery. Eur Radiol 2022;32:2266-76. [Crossref] [PubMed]
23. Jiang Z, Dong Y, Yang L, et al. CT-Based Hand-crafted Radiomic Signatures Can Predict PD-L1 Expression Levels in Non-small Cell Lung Cancer: a Two-Center Study. J Digit Imaging 2021;34:1073-85. [Crossref] [PubMed]
24. Chen T, He T, Benesty M, et al. Xgboost: extreme gradient boosting. R package version 04-2. 2015;1:1-4.
25. Ogunleye A, Wang QG. Enhanced XGBoost-Based Automatic Diagnosis System for Chronic Kidney Disease. Paper presented at: 14th IEEE International Conference on Control and Automation (ICCA); Jun 12-15, 2018; Anchorage, AK.
26. Hollmann N, Müller S, Purucker L, et al. Accurate predictions on small data with a tabular foundation model. Nature 2025;637:319-26. [Crossref] [PubMed]
27. Yoon JH, Lee HS, Kim EK, et al. A nomogram for predicting malignancy in thyroid nodules diagnosed as atypia of undetermined significance/follicular lesions of undetermined significance on fine needle aspiration. Surgery 2014;155:1006-13. [Crossref] [PubMed]
28. Shin I, Kim YJ, Han K, et al. Application of machine learning to ultrasound images to differentiate follicular neoplasms of the thyroid gland. Ultrasonography 2020;39:257-65. [Crossref] [PubMed]
29. Zhang R, Yi G, Pu S, et al. Deep learning based on ultrasound to differentiate pathologically proven atypical and typical medullary thyroid carcinoma from follicular thyroid adenoma. Eur J Radiol 2022;156:110547. [Crossref] [PubMed]
30. He H, Zhu J, Ye Z, et al. Using multimodal ultrasound including full-time-series contrast-enhanced ultrasound cines for identifying the nature of thyroid nodules. Front Oncol 2024;14:1340847. [Crossref] [PubMed]
31. Yi D, Fan L, Zhu J, et al. The diagnostic value of a nomogram based on multimodal ultrasonography for thyroid-nodule differentiation: A multicenter study. Front Oncol 2022;12:970758. [Crossref] [PubMed]
32. Wu S, Shu L, Tian Z, et al. Predictive Value of the Nomogram Model Based on Multimodal Ultrasound Features for Benign and Malignant Thyroid Nodules of C-TIRADS Category 4. Ultrason Imaging 2024;46:320-31. [Crossref] [PubMed]
33. Ma X, Xia L, Chen J, et al. Development and validation of a deep learning signature for predicting lymph node metastasis in lung adenocarcinoma: comparison with radiomics signature and clinical-semantic model. Eur Radiol 2023;33:1949-62. [Crossref] [PubMed]
34. Abbasian Ardakani A, Mohammadi A, Yeong CH, et al. Diagnosis of Thyroid Nodule Malignancy Using Peritumoral Region and Artificial Intelligence: Results of Hand-Crafted, Deep Radiomics Features and Radiologists' Assessment in Multicenter Cohorts. J Ultrasound Med 2025;44:1059-74. [Crossref] [PubMed]
35. Li L, Deng H, Chen W, et al. Comparison of the diagnostic effectiveness of ultrasound imaging coupled with three mathematical models for discriminating thyroid nodules. Acta Radiol 2024;65:441-8. [Crossref] [PubMed]
36. Zhu X, Li J, Li H, et al. Intranodular and perinodular ultrasound radiomics distinguishes benign and malignant thyroid nodules: a multicenter study. Gland Surg 2024;13:2359-71. [Crossref] [PubMed]
37. Liu M, Yao J, Yang J, et al. Bidirectional interaction directional variance attention model based on increased-transformer for thyroid nodule classification. Biomed Phys Eng Express 2024;11: [Crossref] [PubMed]