Ultrasound-based deep learning radiomics with explainable machine learning for predicting benign and malignant thyroid nodules in lymphocytic thyroiditis: a multicenter study Authors: Ji Zhou, Wenwu Lu, Wei Wei, Ziyi Liu, Wang Zhou, Wei Peng, Wenbo Ding, Xin Wu, Chaoxue Zhang Published: 2026 ABSTRACT AbstractBackgroundThyroid nodules (TNs) with lymphocytic thyroiditis (LT) complicate ultrasound (US) assessment, and few prediction models have been validated in this population. The aim of this study is to develop and validate a US-based deep learning radiomics nomogram (DLRN) model for predicting malignancy in TNs coexisting with LT.MethodsThis multicenter hybrid study enrolled 640 patients with pathologically confirmed TNs and concomitant LT. The cohort comprised a training set from Anhui Medical University First Affiliated Hospital (Jixi Road Campus), Anhui Medical University First Affiliated Hospital (High-tech Campus), and Affiliated Hospital of Integration Chinese and Western Medicine with Nanjing University of Traditional Chinese Medicine (retrospective, n=411), an external test set from Yijishan Hospital of Wannan Medical College (retrospective, n=162), and a validation set (prospective, n=67). Deep learning (DL) and handcrafted radiomics features were extracted separately from US images. After feature fusion, the features were Z-score normalized and selected using the Mann-Whitney U test or Chi-squared test, with Fisher’s exact test used when appropriate, followed by the least absolute shrinkage and selection operator (LASSO) regression. A DL radiomics (DLR) model was developed based on 10 machine learning algorithms and subsequently incorporated with independent clinical factors through logistic regression to build a DLRN. Model performance was evaluated by the area under the receiver operating characteristic curve (AUC), calibration and decision curve, along with SHapley Additive Explanations (SHAP) analysis for interpretability.ResultsLogistic regression was selected as the optimal DLR algorithm, achieving AUCs of 0.90 [95% confidence interval (CI): 0.87–0.93] in the training set, 0.90 (95% CI: 0.84–0.94) in the external test set, and 0.82 (95% CI: 0.71–0.91) in the prospective validation set. The DLRN, refined by integrating independent clinical factors via logistic regression, demonstrated superior discrimination with AUCs of 0.93, 0.94, and 0.92 in the training, external test, and prospective validation sets, respectively. The DLRN model demonstrated robust clinical applicability, as evidenced by calibration and decision curve analyses. SHAP analysis enhanced DLRN interpretability by quantifying feature contributions to individual predictions and overall model output.ConclusionsThe DLRN model was designed for the preoperative identification of malignant TNs in LT, thereby facilitating personalized treatment.Keywords: Lymphocytic thyroiditis (LT), thyroid cancer, ultrasound (US), deep learning (DL), radiomics Introduction The prevalence of thyroid nodules (TNs) is high in the general population, with a subset of cases ultimately diagnosed as malignant (1). Lymphocytic thyroiditis (LT) is a common autoimmune disease characterized by lymphocytic infiltration of the thyroid, positivity for thyroid peroxidase antibody (TPOAb)/thyroglobulin antibody (TgAb), and often accompanied by diffuse thyroid enlargement or nodule formation. Epidemiological studies indicate that the concurrent incidence rate of LT and papillary thyroid carcinoma is approximately 23% (2-4). Some studies suggest that LT increases the risk of developing thyroid cancer (5,6). However, the presence of LT may interfere with the differential diagnosis of benign and malignant TNs during ultrasound (US) examination (5). Surgical resection remains the primary treatment for differentiated thyroid cancer (7). Therefore, early and precise identification of thyroid cancer in this population is crucial, as it can reduce unnecessary surgeries while ensuring timely intervention to improve prognosis. However, there is currently a lack of diagnostic models specifically optimized for TNs in LT patients. Although guidelines strongly recommend thyroid and cervical lymph node US evaluations for all cases of confirmed or suspected TNs (7), this approach has limitations: the diagnostic accuracy of conventional US for malignant TNs exhibits significant operator dependent variability, constrained by the examiner’s experience and subjective judgment, leading to notable intra-observer and interobserver discrepancies (8). Additionally, in the context of LT, nodules may exhibit atypical US features that overlap between benign and malignant lesions, including ambiguous morphological changes, margin characteristics, and echogenicity patterns (9). In recent years, radiomics has become a prominent research focus and garnered widespread attention due to its significantly enhanced diagnostic efficacy and prognostic assessment value (10-13). Unlike traditional medical imaging, radiomics utilizes artificial intelligence to convert medical imaging data into quantitative information beyond the scope of conventional visual interpretation, thereby improving the accuracy of diagnosis and prediction (10,14). Previous studies have demonstrated that US radiomics shows promising application prospects in various medical fields, such as thyroid cancer and breast cancer, yielding clinically significant results (15,16). Deep learning (DL), a subfield of artificial intelligence, employs artificial neural networks to extract patterns and make predictions from large datasets (17,18). DL radiomics (DLR), which integrates handcrafted radiomic features with convolutional neural network (CNN) derived DL features, can significantly improve the accuracy of tumor prediction and prognostic assessment (19-22). However, research utilizing DL and radiomics methods for the preoperative prediction of benign versus malignant TNs in the context of LT is still limited. Furthermore, previous studies have often focused solely on the region of interest (ROI), overlooking the diagnostic value of the peritumoral region. Compared to purely intra-tumoral radiomics, incorporating information from the peri-tumoral zone can significantly improve the predictive performance of models (23,24). Therefore, quantifying features both within and surrounding nodules through DLR may significantly enhance the discriminatory power for differentiating benign from malignant nodules. This study aims to leverage the diagnostic value of DL and radiomic features derived from both the interior and periphery of TNs. By integrating US imaging characteristics with clinical data, we construct a comprehensive prediction model for TNs in the context of LT. This approach seeks to provide a more precise and personalized diagnostic tool for this specific patient population. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0539/rc). Methods Patients This study employs a hybrid study design with both retrospective and prospective components. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital of Anhui Medical University (No. PJ2024-12-72). All participating hospitals were informed of and agreed to this study. Written informed consent was obtained from all participants in the prospective dataset, while a waiver of informed consent was granted for the retrospective parts. The retrospective dataset enrolled 573 patients who underwent thyroid surgery across four medical centers between January 2021 and December 2024. A total of 253 cases from Anhui Medical University First Affiliated Hospital (Jixi Road Campus) and Anhui Medical University First Affiliated Hospital (High-tech Campus), 158 cases from Affiliated Hospital of Integration Chinese and Western Medicine with Nanjing University of Traditional Chinese Medicine, and 162 cases from Yijishan Hospital of Wannan Medical College. Inclusion criteria required: (I) histopathologically confirmed LT with coexisting TN(s); (II) complete preoperative thyroid US images and clinical records. Exclusion criteria comprised: (I) suboptimal image quality or incomplete clinicopathological data; (II) with other neoplastic diseases; (III) prior preoperative interventions including radiofrequency or microwave ablation, radiotherapy, or chemotherapy. The prospective cohort recruited 67 patients from Anhui Medical University First Affiliated Hospital (Jixi Road Campus) and Anhui Medical University First Affiliated Hospital (High-tech Campus) according to the following criteria: Inclusion criteria: (I) voluntary participation with signed informed consent; (II) presence of sonographically suspicious TN(s); (III) completion of a fine-needle aspiration (FNA) biopsy guided by US, followed by molecular testing; (IV) scheduled for thyroid surgery within two weeks after US. Exclusion criteria: (I) previous thyroid surgery, ablation, or radiotherapy; (II) concurrent other malignancies; (III) postoperative pathology’s confirmed absence of LT. The patient recruitment flow was documented in Figure 1. The flowchart of patient recruitment and enrollment. "Our three centers" indicate the Anhui Medical University First Affiliated Hospital (Jixi Road Campus), Anhui Medical University First Affiliated Hospital (High-tech Campus), and the Affiliated Hospital of Integration Chinese and Western Medicine with Nanjing University of Traditional Chinese Medicine. "Center 4" indicates the Yijishan Hospital of Wannan Medical College. LT, lymphatic thyroiditis; US, ultrasound. US imaging acquisition and clinicopathological analysis In this retrospective study, US images were retrieved from the Picture Archiving and Communication System (PACS). For patients who underwent multiple preoperative thyroid US, only the most recent preoperative US examination before surgery was included. Two radiologists independently assessed image quality in the PACS workstation, excluding cases with blurred images or missing key information. All included US images were scored using Chinese Thyroid Imaging Reporting and Data System (C-TIRADS) guidelines. For nodules with differing classifications, the two physicians jointly determined the final assessment. Clinicopathological data—including age, sex, nodule size, molecular testing results and FNA cytopathology—were collected through the electronic medical record system. Final benign/malignant determination of TNs was based on postoperative histopathological diagnosis as the gold standard. In the prospective study, we implemented a standardized acquisition protocol for systematic examination of enrolled patients. FNA was performed by a radiologist with over 10 years of interventional US experience, with concurrent documentation of nodule characteristics and archival of raw images. Post-procedural specimens underwent appropriate processing for cytopathological diagnosis and molecular testing for BRAF, RAS, or TERT gene mutations. The detailed image acquisition protocol and preprocessing procedures are described in Appendix 1. Image segmentation and feature extraction All nodules were included in the analysis and identified by unique nodule identifiers (IDs). The vast majority of patients had a single nodule, and only a small proportion had multiple nodules. For each nodule, one representative image at its maximum cross-sectional plane was selected for analysis. A radiologist with over 5 years of thyroid US experience manually delineated the ROI along the nodule margin using ITK-SNAP software, blinded to pathological results and clinical information. Subsequently, PyRadiomics was employed to extract radiomic features from the ROI. The ROI was expanded by 3 mm outward and cropped to capture both intra-nodular and peri-nodular imaging features. A VGG19 architecture was implemented to extract DL features from the cropped images. To evaluate feature stability, 50 randomly selected US images were independently segmented by a second equally qualified radiologist. Interobserver reproducibility was assessed using the intraclass correlation coefficient (ICC), with features demonstrating ICC >0.75 considered acceptably reproducible. The complete feature extraction workflow is illustrated in Figure 2. Appendix 2 provides a comprehensive description of the complete feature extraction process. Flowchart of comprehensive nomogram modeling for the prediction of benign or malignant status in TNs with LT. AUC, area under the curve; DCA, decision curve analysis; DLRN, deep learning radiomics nomogram; LASSO, least absolute shrinkage and selection operator; LT, lymphocytic thyroiditis; ROI, region of interest; TN, thyroid nodule; US, ultrasound. Feature selection and score construction The extracted radiomics features and DL features were subjected to early feature fusion, followed by standardization using Z-score normalization. Statistical screening was performed using the Mann-Whitney U test for continuous features and the χ2 test (or Fisher’s exact test when expected frequencies <5) for categorical features, with features showing no significant differences between groups (P>0.05) being excluded. Subsequently, the least absolute shrinkage and selection operator (LASSO) regression was implemented utilizing five-fold cross-validation with L1 regularization, which progressively compressed coefficients while simultaneously mitigating multicollinearity and preventing overfitting. The selected features were used to build diagnostic models through 10 ML algorithms: logistic regression, decision tree, support vector machine (SVM), k-nearest neighbors (KNN), linear discriminant analysis (LDA), Bagging, AdaBoost, random forest, multilayer perceptron (MLP), and naïve Bayes. Ten machine learning algorithms were evaluated and compared for model selection, aiming to establish the DLR model and generate the DLR score. Model performance was evaluated on both external test and prospective validation set. The complete feature selection process is detailed in the Appendix 3 and Figure S1. Development and evaluation of a DLRN model Univariate logistic regression initially screened clinical and US variables associated with benign/malignant differentiation in the training set (P<0.05). Significant predictors were subsequently analyzed via multivariate logistic regression to identify independent risk factors for malignancy (P<0.05). The DLR signatures yield a continuous DLR score, which was then combined with C-TIRADS category and gene mutation status in a logistic regression model. The output of this regression constitutes the final DLRN model, presented as a nomogram for individualized risk assessment. The DLRN model is intended for use in patients who have undergone FNA with indeterminate cytology (Bethesda III–IV) and for whom molecular testing results are available. For comparison, a clinical model was constructed using solely clinical features. Besides, to evaluate the incremental value of the DLRN model over C-TIRADS alone, we compared the performance of C-TIRADS alone versus the full DLRN model across all datasets. The predictive accuracy of the DLR model, clinical model, and DLRN model were assessed using receiver operating characteristic (ROC) curve analysis, with quantitative metrics including area under the ROC curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value. Statistical differences between the three models’ ROC curves were evaluated using DeLong’s test. Calibration was examined through the Hosmer-Lemeshow test and calibration curves to quantify agreement between predicted probabilities and actual outcomes. Decision curve analysis (DCA) visualized the net benefit of the model across varying high-risk probability thresholds and evaluated clinical utility. To enhance interpretability, SHapley Additive Explanations (SHAP) analysis quantified individual feature contributions to model predictions and identified DL and radiomics features influencing benign/malignant classification. Additionally, to assess within-patient clustering effects, we performed a sensitivity analysis using cluster-robust standard errors clustered at the patient level in the DLRN model on the training set. Statistical analysis Data analysis and modeling were performed using Python (version 3.8) with the PyTorch (version 1.12.0) DL framework and scikit-learn (version 1.2.1) for feature selection and model evaluation. Baseline statistical analyses were conducted using MedCalc (version 20.2) and SPSS (version 26.0). Continuous variables were compared using the Mann-Whitney U test, and categorical variables were compared using the χ2 test (or Fisher’s exact test when expected frequencies were <5). All tests were two-sided, and P<0.05 was considered statistically significant. To assess the potential impact of within-patient clustering, we performed a sensitivity analysis using cluster-robust standard errors clustered at the patient level in the DLRN model on the training set (Table S1). Results Clinicopathological characteristics This study ultimately enrolled 640 patients with LT and suspicious TNs. The training set was constructed from retrospective cases from Anhui Medical University First Affiliated Hospital (Jixi Road Campus), Anhui Medical University First Affiliated Hospital (High-tech Campus), and Affiliated Hospital of Integration Chinese and Western Medicine with Nanjing University of Traditional Chinese Medicine (n=411), while retrospective data from Yijishan Hospital of Wannan Medical College formed the external test set (n=162). Prospective data constituted the validation set (n=67). Thyroid cancer cases accounted for 80.78% (332/411), 75.93% (123/162), and 79.10% (53/67) of the training, test, and validation sets, respectively. Table 1 summarizes the clinicopathological characteristics across all three datasets. Table S2 shows the comparison of baseline characteristics between benign and malignant groups in the training set. Table 1Clinicopathological characteristics in the training, test and validation datasetsCharacteristicsTraining set (n=411)External test set (n=162)Prospective validation set (n=67)Age (years)43.94 (11.07)46.40 (11.96)45.01 (12.38)Sex Male8.27 (34/411)13.50 (22/162)8.96 (6/67) Female91.73 (377/411)86.42 (140/162)91.04 (61/67)Tumor size (mm)9.78 (6.89)11.94 (9.42)9.31 (4.63)Tumor location Left51.82 (213/411)45.06 (73/162)29.85 (20/67) Right45.99 (189/411)52.47 (85/162)65.67 (44/67) Isthmus2.19 (9/411)2.47 (4/162)4.48 (3/67)Gene mutation Positive52.55 (216/411)53.09 (83/162)59.70 (40/67) Negative47.45 (195/411)46.91 (76/162)40.30 (27/67)C-TIRADS 31.46 (6/411)14.20 (23/162)0 (0/67) 4A9.49 (39/411)34.57 (56/162)34.33 (23/67) 4B43.31 (178/411)28.40 (46/162)34.33 (23/67) 4C45.26 (186/411)18.52 (30/162)28.36 (19/67) 50.49 (2/411)3.09 (5/162)2.99 (2/67)Data are presented as mean (SD) or % (n/N). C-TIRADS, Chinese Thyroid Imaging Reporting and Data System; SD, standard deviation. Clinicopathological characteristics in the training, test and validation datasets Data are presented as mean (SD) or % (n/N). C-TIRADS, Chinese Thyroid Imaging Reporting and Data System; SD, standard deviation. Univariate logistic regression analysis of clinical characteristics revealed significant associations between nodule status and nodule size, genetic mutations, and C-TIRADS classification in the training set (Table 2). Notably, nodule size was excluded from the clinical model due to loss of statistical significance in multivariate logistic regression (Table 3). Subsequently, based on these identified clinical risk factors, a clinical model for nodule status assessment was developed using multivariate logistic regression. The clinical model achieved AUC values of 0.84 [95% confidence interval (CI): 0.80–0.87], 0.81 (95% CI: 0.74–0.87), and 0.82 (95% CI: 0.71–0.90) in the training, test, and validation sets, respectively. Table 2Results of the univariate logistic regression analysis in the training setCharacteristicsUnivariate analysisOR (95% CI)P valueAge (years)0.99 (0.96–1.01)0.23Sex1.86 (0.64–5.45)0.26Tumor size (mm)0.95 (0.92–0.98)0.003*Tumor location0.70 (0.44–1.12)0.14Gene mutation7.21 (3.89–13.39)<0.001*C-TIRADS4.70 (3.11–7.09)<0.001*DLR-score1.91 (1.68–2.17)<0.001**, P<0.05. C-TIRADS, Chinese Thyroid Imaging Reporting and Data System; CI, confidence interval; DLR-score, deep learning radiomics score; OR, odds ratio. Results of the univariate logistic regression analysis in the training set *, P<0.05. C-TIRADS, Chinese Thyroid Imaging Reporting and Data System; CI, confidence interval; DLR-score, deep learning radiomics score; OR, odds ratio. Table 3Results of the multivariate logistic regression analysis in the training setCharacteristicsClinical model Comprehensive modelOR (95% CI)P valueOR (95% CI)P valueTumor size (mm)NA0.19NA0.31Gene mutation7.11 (3.63–13.93)<0.001*5.23 (2.27–12.09)<0.001*C-TIRADS4.57 (2.96–7.05)<0.001*3.86 (2.28–6.54)<0.001*DLR-scoreNANA1.92 (1.64–2.24)<0.001**, P<0.05. C-TIRADS, Chinese Thyroid Imaging Reporting and Data System; CI, confidence interval; DLR-score, deep learning radiomics score; NA, not available; OR, odds ratio. Results of the multivariate logistic regression analysis in the training set *, P<0.05. C-TIRADS, Chinese Thyroid Imaging Reporting and Data System; CI, confidence interval; DLR-score, deep learning radiomics score; NA, not available; OR, odds ratio. Establishment of DLR features This study extracted 464 radiomics features and 128 DL features from each patient’s US images. Following statistical analysis, 17 optimal features were selected for constructing the DLR score. Logistic regression was selected as the final modeling algorithm due to its stable and consistent predictive performance across different datasets. Its AUC values in the training set, external test set, and prospective validation set were 0.90 (95% CI: 0.87–0.93), 0.90 (95% CI: 0.84–0.94), and 0.82 (95% CI: 0.71–0.91), respectively (Figure 3A). As demonstrated in Figure 3B, the SHAP plot visualizes the direction and magnitude of contribution of DL and radiomics features on the predictions of the logistic regression model, enhancing model interpretability. Detailed statistical metrics for all 10 ML models are provided in Tables S3-S5. Performance comparison of machine learning algorithms and feature contribution analysis. (A) Performance comparison of ten machine learning algorithms on the training, external test, and prospective validation datasets. (A1) Training set, (A2) external test set, (A3) prospective validation set. (B) SHAP summary plot illustrating the contribution of each feature to the optimal DLR model output. AUC, area under the receiver operating characteristic curve; DLR, deep learning radiomics; KNN, k-nearest neighbors; LDA, linear discriminant analysis; ML, machine learning; MLP, multilayer perceptron; SHAP, SHapley Additive Explanations; SVM, support vector machine. Nomogram development and performance assessment Multivariate logistic regression identified gene mutation, nodule C-TIRADS classification, and the DLR score as independent risk predictors for malignant TNs in the context of LT (Table 3). Utilizing these factors, we developed the DLRN model and presented it as a nomogram (Figure 4A). Figure 4B calibration curve shows that the nomogram has good calibration, with Hosmer-Lemeshow test P values of 0.41, 0.82, and 0.93 in the training set, test set, and validation set, respectively. Table 4 presents the predictive performance of the DLRN model at the recommended optimal cutoff value for evaluating the nature of nodules in the context of LT across different datasets. ROC curves are shown in Figure 5A-5C. The comprehensive model exhibited excellent discriminative ability in the training set 0.93 (95% CI: 0.91–0.96), testing set 0.94 (95% CI: 0.90–0.97), and validation set 0.92 (95% CI: 0.83–0.97). Compared to the single DLR model, clinical model, and C-TIRADS, Delong’s test confirmed that DLRN model outperformed both the clinical model and the DLR model in all three groups (all P<0.05) (Tables S6,S7). DCA shows that across different datasets and various risk threshold probabilities, DLRN model provides a higher overall net benefit compared to both the clinical model and the DLR model (Figure 5D-5F). The optimal cutoff value for the Nomo-score to differentiate between benign and malignant nodules was determined by maximizing the Youden index. Figure 5G-5I illustrates the grouping of data using this optimal Nomo-score cutoff. Nomogram and calibration curves of the training set, external test set, and prospective validation set. (A) DLR nomogram for prediction of benign and malignant TNs in LT. (B) The calibration curves of the nomogram across three datasets. DLR, deep learning radiomics; LT, lymphocytic thyroiditis; TN, thyroid nodule. Table 4The predictive performance of the DLRN model in assessing the nature of TNs in the context of LT in different datasets at recommended optimal cutoffsModel/cohortAUC (95% CI)Accuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)Training set Clinical model0.84 (0.80–0.87)0.85 (0.82–0.89)0.98 (0.96–0.99)0.33 (0.23–0.44)0.86 (0.82–0.89)0.79 (0.61–0.91) DLR model0.90 (0.87–0.93)0.89 (0.85–0.92)0.96 (0.93–0.98)0.58 (0.47–0.69)0.91 (0.87–0.94)0.78 (0.65–0.88) DLRN model0.93 (0.91–0.96)0.93 (0.90–0.95)0.98 (0.95–0.99)0.73 (0.62–0.83)0.94 (0.91–0.96)0.88 (0.78–0.95)External test set Clinical model0.81 (0.74–0.87)0.84 (0.77–0.89)0.98 (0.94–1.00)0.38 (0.23–0.55)0.83 (0.76–0.89)0.88 (0.64–0.99) DLR model0.90 (0.84–0.94)0.84 (0.77–0.89)0.94 (0.88–0.97)0.54 (0.37–0.70)0.87 (0.80–0.92)0.72 (0.53–0.87) DLRN model0.94 (0.90–0.97)0.88 (0.82–0.92)0.94 (0.89–0.98)0.67 (0.50–0.81)0.90 (0.83–0.95)0.79 (0.61–0.91)Prospective validation set Clinical model0.82 (0.71–0.90)0.82 (0.71–0.90)0.91 (0.79–0.97)0.50 (0.23–0.77)0.87 (0.76–0.95)0.58 (0.28–0.85) DLR model0.82 (0.71–0.91)0.87 (0.76–0.94)0.96 (0.87–1.00)0.50 (0.23–0.77)0.88 (0.77–0.95)0.78 (0.40–0.97) DLRN model0.92 (0.83–0.97)0.90 (0.80–0.96)0.96 (0.87–1.00)0.64 (0.35–0.87)0.91 (0.80–0.97)0.82 (0.48–0.98)AUC, area under the curve; CI, confidence interval; DLR, deep learning and radiomics; DLRN, deep learning and radiomics nomogram; LT, lymphocytic thyroiditis; NPV, negative predictive value; PPV, positive predictive value; TN, thyroid nodule. The predictive performance of the DLRN model in assessing the nature of TNs in the context of LT in different datasets at recommended optimal cutoffs AUC, area under the curve; CI, confidence interval; DLR, deep learning and radiomics; DLRN, deep learning and radiomics nomogram; LT, lymphocytic thyroiditis; NPV, negative predictive value; PPV, positive predictive value; TN, thyroid nodule. The ROC, DCA, and stratification plot across diverse datasets. (A-C) ROC of different predictive models in training set, external test set, and prospective validation set. (D-F) DCA of different predictive models in training set, external test set, and prospective validation set. The violin plots display the data distribution in the training set (G), external test set (H) and prospective validation set (I), incorporating a box plot within it. DCA, decision curve analysis; DLR, deep learning radiomics; ROC, receiver operating characteristic. Discussion In this multicenter study, we developed and validated a DLRN model integrating clinical characteristics with DLR features, including peri-nodular information. The DLRN model achieves remarkable predictive performance, facilitating personalized prediction of benign versus malignant TNs associated with LT. In our study, univariate analysis revealed that nodule size was a significant factor and showed a negative correlation with malignant pathology. However, this variable did not demonstrate statistical significance in the multivariate regression analysis. This finding aligns with the results reported by Zhang et al. (25), whose study similarly demonstrated that size failed to reach statistical significance in their multivariate logistic regression analysis. We speculate that this discrepancy may be related to the characteristics of the enrolled population: benign nodules only require regular follow-up, while surgical intervention is usually considered only for cases with suspected malignant transformation, cosmetic concerns, or compressive symptoms. This practice inevitably introduces a degree of selection bias into our study. In addition, our model was developed and validated exclusively in a surgical cohort, with malignancy prevalence around 76–81%, substantially higher than in general clinical practice or screening populations. Large epidemiological studies show that the overall malignancy rate among TNs detected in health checkup populations ranges from 7% to 15% (26), highlighting a substantial disparity and the challenge of model generalization. To quantify the impact of prevalence shifts on performance, we applied Bayes’ theorem using the model’s actual sensitivity (94%) and specificity (67%). In the surgical dataset (prevalence ≈80%), the PPV reached 91.9%. If the same model were applied to a general screening population with a prevalence of 10%, the PPV would drop to 24.0%, meaning over 75% of positive predictions would be false positives. A nationwide analysis reports malignancy rates of 36.2% for Bethesda III and 36.7% for Bethesda IV (27). Accordingly, in our Bayesian analysis, we set the malignancy prevalence for Bethesda III–IV indeterminate nodules to roughly 35–40%; at this prevalence, the PPV rises to 60.5–65.5%, with the model showing greater applicability in clinically challenging subgroups. The essence of radiomics lies in extracting many quantitative features from medical images (28). Traditional radiomics typically requires precise delineation of tumor boundaries and other tissues of interest for feature extraction (29). While DL automates feature extraction, eliminating the need for clinical indicator collection, semantic feature interpretation, or manual annotation, its performance remains highly sensitive to data quality and source, where even minor perturbations can cause erroneous outputs (30). Given the respective strengths and limitations of both approaches, integrating handcrafted radiomics features with DL features can enhance model diagnostic performance (30-32). Furthermore, a growing body of research confirms that the peritumoral region can provide valuable additional information. Zhang et al. demonstrated that radiomics models for lymph node metastasis in thyroid cancer, based on two US elastography modalities, achieved superior diagnostic performance compared to models using intratumoral radiomics features alone (33). Consequently, our study incorporated both intratumoral and peritumoral regions. We fused radiomics features with DL features, employing the selected fused features to construct the imaging model. Recognizing that different algorithms exhibit varying sensitivity to data characteristics, we incorporated the fused features into 10 distinct ML algorithms for optimal model selection. Although some models, such as Bagging and random forest, outperform logistic regression on the training dataset, their performance declines markedly on test and validation datasets. This difference reflects an inherent trade-off between model flexibility and generalization ability. More flexible algorithms are more prone to overfitting the training data, capturing idiosyncratic patterns and noise that do not generalize to new data. This phenomenon is particularly pronounced in validation dataset with limited sample sizes, where small samples amplify the variance of high-variance models. By contrast, logistic regression with L1 regularization reduces variance by imposing sparse constraints, thereby improving external validity. Given this, logistic regression was ultimately chosen for modeling. Analysis of the SHAP summary plot clearly illustrates the relative contribution of individual radiomics and DL features within the DLR model. Handcrafted radiomics features contributed substantially more to the DLR model. This discrepancy may stem from significant variations in imaging protocols across different centers. Data heterogeneity poses a major challenge in applying DL to oncology, potentially leading to reduced feature contribution in multicenter studies (17). DL requires substantial datasets to fully realize its potential; consequently, set size may constrain the effective application of DL architecture (34). As sample sizes expand in future studies, DL features may demonstrate superior predictive power. The DLRN model demonstrated consistent AUC improvements over C-TIRADS alone across three independent datasets, confirming its clinically meaningful added value. We further observed that the DLR score alone already significantly outperformed C-TIRADS in the training and test datasets (P<0.05), and the full DLRN model achieved further improvement by incorporating molecular mutation status. In the prospective validation cohort, although a similar trend was observed, the difference did not reach statistical significance, which may be attributable to the limited sample size. It must be acknowledged that this study has several limitations. First, potential selection bias exists due to the reliance on postoperative histopathology as the diagnostic standard for TNs classification, resulting in class imbalance between benign and malignant cases. Second, although prospective validation was performed, the sample size remains limited and the study predominantly utilized retrospective data. Future validation through multicenter studies with larger sets is warranted. Finally, constrained by data availability across some participating centers, we were unable to comprehensively collect and incorporate laboratory parameters into either the clinical model or the DLRN model. Conclusions In summary, we developed an integrated predictive model by combining DLR and radiomics features with clinical characteristics. The DLRN model facilitates personalized therapeutic decisions for clinicians managing this patient population. Supplementary The article’s supplementary files as Acknowledgments None. 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 Committee of The First Affiliated Hospital of Anhui Medical University (No. PJ2024-12-72). All participating hospitals were informed of and agreed to this study. Written informed consent was obtained from all participants in the prospective dataset, while a waiver of informed consent was granted for the retrospective parts. Footnotes Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0539/rc Funding: This work was supported by the Anhui Provincial Natural Science Foundation (No.2308085MH278 to C.Z.) and the Health Research Program of Anhui (No. AHWJ2023A10017 to C.Z.). Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0539/coif). The authors have no conflicts of interest to declare. Data Sharing Statement Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0539/dss References Associated Data Supplementary Materials The article’s supplementary files as Data Availability Statement Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0539/dss