The role of ultrasound texture analysis in the discrimination of pleomorphic adenoma and Warthin tumor in subjects with well-defined tumor borders
Highlight box
Key findings
• Pleomorphic adenoma (PA) and Warthin tumor (WT) are the two most common benign parotid tumors but require different management strategies.
• Conventional ultrasound diagnosis relies on subjective visual assessment, which has limited accuracy due to overlapping sonographic features between PA and WT.
What is known and what is new?
• PA and WT are the two most common benign parotid tumors but require different management strategies. Conventional ultrasound diagnosis relies on subjective visual assessment, which has limited accuracy due to overlapping sonographic features between PA and WT.
• This study identifies and validates a set of eight quantitative ultrasound texture features as well as a machine learning model that can help differentiate PA from WT.
What is the implication, and what should change now?
• The proposed texture analysis provides a non-invasive, interpretable adjunct to conventional ultrasound.
• Future prospective and multi-center studies are needed to validate generalizability before clinical implementation.
Introduction
Parotid gland tumors represent one of the most common types of head and neck neoplasms, among which pleomorphic adenoma (PA) and Warthin tumor (WT) are the two predominant benign pathological subtypes (1). Although both are benign, their clinical management strategies and prognoses differ significantly. PA carries a potential risk of malignant transformation, typically warranting surgical resection, whereas WT may be managed more conservatively with active surveillance or less invasive interventions (2,3). Consequently, achieving accurate, non-invasive preoperative differentiation is of paramount clinical importance for formulating individualized treatment plans, as it directly informs surgical strategy.
Superficial or total parotidectomy, which involves formal dissection of the facial nerve, is typically indicated for PA to achieve complete excision with negative margins, thereby minimizing recurrence risk. In contrast, extracapsular dissection—a nerve-sparing, parenchyma-preserving technique—is often sufficient for WT, given its benign behavior and low recurrence potential, and may even permit observation in select cases. Thus, accurate preoperative characterization can help balance oncologic safety with functional preservation, guiding the extent of surgery and postoperative follow-up.
Ultrasonography has emerged as the primary imaging modality for evaluating parotid tumors due to its advantages: radiation-free, real-time, cost-effective, and widely accessible. Diagnosis typically relies on sonographers’ subjective visual assessment of features such as tumor morphology, margins, internal echogenicity, and vascularity (4). However, the sonographic appearances of PA and WT often overlap; for instance, both can present as well-circumscribed, hypoechoic masses. This inherent similarity leads to diagnostic accuracy that is highly operator-dependent and subjective, posing a significant clinical challenge.
Texture analysis technology enables the high-throughput extraction and quantification of sub-visual features from medical images. These features reveal tumor internal heterogeneity that is imperceptible to the human eye, thereby offering a novel perspective for precision diagnosis (5,6). In recent years, several studies have explored radiomics for discriminating parotid tumors using modalities like magnetic resonance imaging (MRI) and computed tomography (CT) (7-9). He et al. developed a model using dual-center parotid ultrasound data to differentiate PA from WT, which demonstrated promising performance; nevertheless, their work primarily focused on the model itself and lacked an interpretable feature set suitable for clinical evaluation (10).
Therefore, this study seeks to explore an interpretable set of texture features derived from parotid ultrasound images, with the goal of assisting in the differentiation between PA and WT. We attempt not only to develop a machine learning (ML) model for this task but also to emphasize the identification and interpretation of the key imaging features that influence the model’s decisions. By correlating these features with established sonographic characteristics, we hope to offer preliminary insights and a potential adjunctive tool that may contribute to the preoperative assessment of parotid tumors. We present this article in accordance with the STARD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0102/rc).
Methods
Study population
This dual-center, retrospective study included patients from two institutions: The Fifth Affiliated Hospital, Southern Medical University and Beijing Chaoyang Hospital. All patients underwent ultrasound examination, surgical resection, and pathological confirmation at their respective hospitals. The flow diagram is shown in Figure 1, and all included cases were confirmed as either PA or WT. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Boards of The Fifth Affiliated Hospital, Southern Medical University (No. 2025-EBHK-K-001) and Beijing Chaoyang Hospital (No. 2026-science-354). Informed consent was waived for this retrospective study because all data were fully de-identified.
The inclusion criteria were: (I) preoperative parotid ultrasonography with image quality sufficient for analysis; (II) a definitive pathological diagnosis from surgery or biopsy serving as the gold standard. The exclusion criteria were: (I) poorly defined tumor borders on ultrasound, precluding accurate delineation; (II) a history of prior surgery or treatment in the parotid region; (III) an unclear pathological diagnosis or a diagnosis of another tumor type.
Sample size determination was based on the rule of thumb for clinical prediction model development, which recommends at least 10–20 events per predictor variable to avoid over-fitting. In this study, after feature selection, eight radiomic features were retained. With 80 patients in the positive group (PA), the sample size satisfies the recommended minimum of 10 events per feature (80 vs. 8), ensuring adequate model stability.
Image acquisition and lesion segmentation
Static grayscale ultrasound images of parotid masses, containing at least the largest cross-sectional view of the tumor, were exported in a lossless format. An experienced sonographer with over five years of experience, blinded to the pathological results, manually delineated the entire two-dimensional region of interest (ROI) along the tumor boundary using the open-source software ImageJ. The segmented ROIs were subsequently processed in ImageJ to generate nine distinct color channel representations (as shown in Figure 2). Each ROI lesion was segmented three times, and the mean value of the extracted radiomic features was used for subsequent analysis to reduce intraobserver variability.
Radiomics feature extraction
High-throughput extraction of radiomics features from each ROI was performed using the PyRadiomics library in Python. The extracted features encompassed six categories: (I) first-order statistics; (II) shape-based features; (III) gray-level co-occurrence matrix (GLCM) features; (IV) gray-level run-length matrix (GLRLM) features; (V) gray-level size zone matrix (GLSZM) features; (VI) neighboring gray tone difference matrix (NGTDM) features. A total of 936 texture features were initially extracted per tumor ROI across the color channels. Features containing duplicate or outlier values were identified and removed.
Feature selection strategy
A two-step feature selection process was implemented. First, an independent-samples t-test was applied to identify features with significant inter-group differences. Subsequently, a second selection was performed using least absolute shrinkage and selection operator (LASSO) logistic regression with receiver operating characteristic (ROC) curve optimization. The optimal regularization parameter was determined via grid search, yielding the final set of features for model construction.
Model construction and validation
The selected key features were used as input, with pathological diagnosis (PA vs. WT) as the output label. We manually tuned the hyperparameters of multiple ML algorithms and compared their performance. The algorithm with superior performance was selected to build the final diagnostic model. The model’s comprehensive performance was evaluated on the independent validation set using metrics including accuracy, sensitivity, and specificity. A ROC curve was plotted, and the area under the curve (AUC) was calculated. The model training results were validated on two test sets: an internal test set (split at a 7:3 ratio) and an independent external validation set comprising 41 additional cases.
Model interpretability analysis
To enhance model transparency and clinical trustworthiness, we employed the SHapley Additive exPlanations (SHAP) framework to interpret the best-performing model. By calculating the SHAP value for each feature per prediction, we quantified each feature’s contribution and directional influence (toward PA or WT) on the model’s output, thereby visually illustrating how key features relate to the diagnostic decision.
Cut-off value calculation of each feature
To assess the standalone diagnostic value of each key texture feature, we determined its optimal cut-off value. Using the same preprocessed data (mean imputation and StandardScaler standardization based on the training set) and the same train-test split (random seed 37; 104 training, 45 test samples), we applied a linear discriminant analysis (LDA) model to each of the selected features individually. The optimal cut-off for the predicted probability of each single-feature LDA model was identified on the test set by maximizing the F1-score via the precision-recall curve, aiming to best balance sensitivity and precision for each feature when used alone.
Statistical analysis
All data processing, feature selection, model construction, and analysis were performed using Python (Version 3.9) with libraries including scikit-learn, PyRadiomics, and SHAP. A P value of less than 0.05 was considered statistically significant.
Results
Baseline demographics and feature selection outcome
Ultimately, 149 patients were enrolled between November 2015 and November 2025, including 80 with PA, 69 with WT, 106 males, and 43 females. The age ranged from 19 to 83 years, with a mean age of 57.46 years and a median age of 60 years. All patients were randomly split into a training set and an internal validation set at a 7:3 ratio to ensure independence between model development and performance evaluation. In addition, we added 41 cases (18 WT and 23 PA, from November 2025 to April 2026) as an independent test set.
An independent samples t-test was first performed, identifying 143 features with statistically significant differences (P<0.05) between the two tumor groups. Subsequently, a second round of feature selection was conducted using a ROC-oriented LASSO logistic regression. This process culminated in the selection of nine key features: “a_star_energy”, “Hue_gldm_LDHGLE”, “Hue_MeanAbsoluteDeviation”, “a_star_RootMeanSquared”, “Hue_glrlm_RunEntropy”, “Hue_glszm_ZoneEntropy”, “Hue_10Percentile”, “Hue_glcm_Idmn”, and “Sat_glcm_MCC”. Since the values of the two features Hue_glrlm_RunEntropy and Hue_glszm_ZoneEntropy were completely identical, we removed the latter, ultimately retaining only eight effective features. The comparison of inter-group differences of the selected parameters is shown in Table S1.
Comparison of ML algorithms
Eleven distinct ML algorithms were employed and compared for model construction. Their performance metrics on the validation set are summarized in Table 1. The LDA algorithm achieved the best performance, with an AUC of 95.8% [95% confidence interval (CI): 89.9–99.6%], a sensitivity of 90.5% (95% CI: 76.9–100.0%), and a specificity of 83.3% (95% CI: 66.7–96.3%). Results of the DeLong test are provided in Figure S1. On the independent external validation cohort, the model yielded an accuracy of 85.4% (95% CI: 73.2–95.1%), a sensitivity of 72.2% (95% CI: 50.0–92.9%), a specificity of 95.7% (95% CI: 86.4–100.0%), and an AUC of 82.4% (95% CI: 67.7–95.7%). The corresponding confusion matrices and ROC curves for the LDA model are presented in Figure 3: Figure 3A,3B show the external validation ROC curve and confusion matrix, respectively; Figure 3C,3D show the internal validation ROC curve and confusion matrix, respectively.
Table 1
| Algorithm | Accuracy (95% CI), % | Sensitivity (95% CI), % | Specificity (95% CI), % | AUC (95% CI), % |
|---|---|---|---|---|
| SVM | 88.9 (80.0–97.8) | 95.2 (84.6–100.0) | 83.3 (66.7–96.3) | 94.2 (86.2–99.2) |
| Random forest | 75.6 (62.2–86.7) | 71.4 (52.2–90.0) | 79.2 (61.9–95.5) | 90.7 (80.6–97.6) |
| LR | 82.2 (71.1–93.3) | 81.0 (64.0–95.8) | 83.3 (68.2–96.4) | 95.2 (88.7–99.6) |
| KNN | 84.4 (73.3–93.3) | 85.7 (69.6–100.0) | 83.3 (66.7–96.0) | 86.9 (74.0–98.0) |
| LDA | 86.7 (75.6–95.6) | 90.5 (76.9–100.0) | 83.3 (66.7–96.3) | 95.8 (89.9–99.6) |
| Decision tree | 66.7 (53.3–80.0) | 57.1 (35.3–77.3) | 75.0 (56.0–91.7) | 77.4 (62.4–89.5) |
| AdaBoost | 80.0 (66.7–91.1) | 81.0 (61.5–95.5) | 79.2 (61.9–95.5) | 80.1 (66.7–.91.1) |
| Gradient boosting | 80.0 (68.9–91.1) | 81.0 (63.6–95.8) | 79.2 (61.9–95.5) | 89.7 (79.0–97.0) |
| XGBoost | 82.2 (71.1–93.3) | 81.0 (63.6–95.8) | 83.3 (68.2–96.0) | 91.1 (81.8–97.8) |
| LGBM | 80.0 (68.8–91.1) | 85.7 (68.7–100.0) | 75.0 (58.3–91.7) | 87.5 (75.2–97.2) |
| MLP | 68.9 (55.6–82.2) | 66.7 (47.0–87.0) | 70.1 (50.0–88.2) | 82.1 (67.2–93.3) |
AdaBoost, adaptive boosting; AUC, area under the curve; CI, confidence interval; KNN, K-nearest neighbors; LDA, linear discriminant analysis; LGBM, light gradient boosting machine; LR, logistic regression; MLP, multilayer perceptron; SVM, support vector machine; XGBoost, extreme gradient boosting.
Feature importance analysis
The SHAP analysis revealed the relative importance of the selected features in the model’s decision-making process (see Figure 4). The feature “Hue_10Percentile” was identified as the most influential contributor to the model’s predictions, followed in descending order by “Hue_gldm_LDHGLE”, “a_star_RootMeanSquared”, “Hue_glcm_Idmn”, “Sat_glcm_MCC”, “a_star_energy”, “Hue_glrlm_RunEntropy”, and “Hue_MeanAbsoluteDeviation”. A detailed interpretation of these features and their clinical relevance is provided in the discussion section.
Cut off value of each feature
The performance of the eight key texture features in differentiating PA from WT using individually optimized cut-off values is summarized in Table 2. Among all features, Hue_10Percentile demonstrated the highest standalone diagnostic efficacy, achieving an F1-score of 0.837, accuracy of 84.4%, sensitivity of 85.7%, and specificity of 83.3% at its optimal cut-off of 0.498. Hue_MeanAbsoluteDeviation also showed strong and balanced performance (F1 =0.800, specificity =0.875). In contrast, a_star_energy and a_star_RootMeanSquared exhibited high sensitivity (≥0.905) but low specificity (≤0.458), correctly identifying most PA cases while misclassifying many WT. Sat_glcm_MCC achieved perfect sensitivity (1.0) but zero specificity, rendering it ineffective as a standalone diagnostic tool. For reference, the integrated model using all eight features achieved a superior F1-score of 0.878 and specificity of 0.917 at its optimal threshold.
Table 2
| Feature name | Optimal cut-off | Max F1 score | Precision | Recall (sensitivity) | Specificity (1−FPR) | Accuracy |
|---|---|---|---|---|---|---|
| a_star_energy | 0.355 | 0.717 | 0.594 | 0.905 | 0.458 | 0.667 |
| a_star_RootMeanSquared | 0.281 | 0.737 | 0.583 | 1 | 0.375 | 0.667 |
| Hue_10Percentile | 0.498 | 0.837 | 0.818 | 0.857 | 0.833 | 0.844 |
| Hue_MeanAbsoluteDeviation | 0.453 | 0.8 | 0.842 | 0.762 | 0.875 | 0.822 |
| Hue_glcm_Idmn | 0.401 | 0.698 | 0.682 | 0.714 | 0.708 | 0.711 |
| Hue_glrlm_RunEntropy | 0.367 | 0.766 | 0.692 | 0.857 | 0.667 | 0.756 |
| Hue_gldm_LDHGLE | 0.357 | 0.745 | 0.633 | 0.905 | 0.542 | 0.711 |
| Sat_glcm_MCC | 0.398 | 0.636 | 0.467 | 1 | 0 | 0.467 |
FPR, false positive rate; glcm, gray-level co-occurrence matrix; gldm, gray-level dependence matrix; glrlm, gray-level run-length matrix; LDHGLE, low dependence high gray-level emphasis; MCC, maximum correlation coefficient.
Discussion
This study sought to investigate ultrasound-based texture features for differentiating between PA and WT of the parotid gland. By extracting radiomic features from multiple color channels, we developed a preliminary diagnostic model and focused on identifying key texture characteristics associated with this discrimination. Our findings suggest that texture analysis may help capture subtle image patterns difficult to perceive visually, and these features could provide an objective reference for non-invasive differentiation between the two tumor types.
The effectiveness of the texture features is rooted in the distinct histopathological characteristics of PA and WT as reflected in their sonographic appearances. Key features identified in this study, such as Hue_gldm_LDHGLE, a_star_energy, and Hue_glcm_Idmn, are highly associated with image homogeneity. The higher values of these features in PA align with its typical ultrasound presentation. Pathologically, PA is composed of a mixture of epithelial and mesenchymal components forming a relatively compact and homogeneous structure, which translates into more uniform internal echogenicity and regular textural patterns on ultrasound. Conversely, WT, characterized by abundant lymphoid stroma and epithelial-lined cystic spaces, often presents with heterogeneous echotexture and frequent cystic changes (11,12).
This structural heterogeneity and complexity directly result in significantly higher values for features representing randomness and heterogeneity, such as Hue_glrlm_RunEntropy, in WT. Furthermore, the lower value of Hue_10Percentile in WT may correspond to the presence of multiple small, hypoechoic cystic areas within the tumor, which is consistent with the “reticular” or “honeycomb” echo pattern frequently observed in clinical practice. Therefore, the discriminative logic of our model is highly consistent with conventional ultrasound diagnostic experience, effectively quantifying the fundamental distinction between “homogeneity” and “heterogeneity” to achieve classification.
The analysis of single-feature diagnostic cut-offs offers preliminary insights into the discriminative capacity and potential clinical relevance of individual texture features. For instance, the relatively balanced performance of Hue_10Percentile (F1=0.837)—a feature associated with hypo-echoic regions—appears consistent with the pathological profile of WT, which often presents with cystic, low-echo areas. This may suggest its usefulness as an auxiliary visual indicator during sonographic evaluation. On the other hand, features such as a_star_RootMeanSquared and a_star_energy, while showing high sensitivity, exhibited limited specificity, implying that they might be more suitable for excluding PA rather than confirming it, thus constraining their standalone diagnostic applicability.
The aforementioned features are designed to be applied within the ML model established in this study. Although we attempted to define cutoff values for individual features—such as using a specific threshold of Hue_glslm_RunEntropy to indicate PA or WT—this approach proved oversimplified and led to reduced sensitivity in practical diagnostic scenarios. In contrast, integrating these features into a ML model mitigates this limitation. The SHAP algorithm clearly visualizes the contribution of each feature to the model’s predictions, enabling a more nuanced and comprehensive decision-making process. Therefore, a model-based application strategy offers a more reliable means of leveraging these texture features to support clinical diagnosis.
From a clinical perspective, distinguishing PA from WT preoperatively directly influences surgical planning and follow‑up strategies. PA carries a risk of recurrence if incompletely excised, typically warranting superficial or total parotidectomy to achieve negative margins. In contrast, WT is often managed with more conservative approaches such as extracapsular dissection or even active surveillance in selected cases, given its benign nature and low recurrence risk. The texture‑based model proposed in this study provides a noninvasive preoperative prediction that can help guide these decisions—suggesting parotidectomy for lesions predicted as PA and considering limited resection or observation for those predicted as WT. Prospective studies are warranted to further validate this integration into routine clinical decision-making.
While fine-needle aspiration biopsy (FNAB) remains the standard for preoperative diagnosis, it is invasive and carries a risk of needle tract seeding, especially for PA (13). Ultrasound texture analysis offers a non‑invasive alternative that can complement existing workflows. It may serve as a prebiopsy screening tool to reduce unnecessary invasive procedures or provide adjunctive information when FNAB yields indeterminate results (13). Integrating this repeatable, risk‑free method into the diagnostic pathway can enhance preoperative risk stratification while mitigating the risks associated with tissue sampling.
Several limitations of this study should be acknowledged. First, while the retrospective dual-center design was beneficial for initial validation, prospective, multi-center studies with larger sample sizes are necessary to further confirm its generalizability. Second, although the manual delineation of ROIs was performed by an experienced sonographer, it may introduce operator dependency. The implementation of automated or semi-automated segmentation techniques in future work would help improve the reproducibility and objective standardization of the results. Third, although our model effectively differentiates between PA and WT, its performance in discriminating other parotid tumors remains to be explored.
Additionally, our study did not directly compare texture analysis with human assessment or evaluate its clinical decision impact. Future research should aim to expand the classification spectrum and examine the temporal robustness of these texture features across different ultrasound machines and settings. Notwithstanding these limitations, the interpretable feature set identified in this study represents a significant step forward in the development of a reliable computer-aided diagnostic tool for the preoperative assessment of parotid tumors.
Conclusions
In conclusion, this study suggests that ultrasound texture analysis may offer a quantifiable approach for differentiating PA from WT. By identifying and interpreting a set of key textural features, we developed a ML model with promising preliminary diagnostic performance and enhanced interpretability. Given the current limitations, these findings should be considered exploratory. Further validation through prospective and multi-center studies is needed before clinical translation. Future work will focus on improving feature stability across imaging platforms and expanding the model to include a broader spectrum of parotid lesions.
Acknowledgments
We would like to thank all the individuals who participated in this study.
Footnote
Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0102/rc
Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0102/dss
Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0102/prf
Funding: None.
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2026-1-0102/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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Boards of The Fifth Affiliated Hospital, Southern Medical University (No. 2025-EBHK-K-001) and Beijing Chaoyang Hospital (No. 2026-science-354). Informed consent was waived for this retrospective study because all data were fully de-identified.
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
- Zeng J, Li J, Saad M, et al. A Review of Surgical Incisions Used for the Excision of Benign Parotid Tumors. Ann Plast Surg 2024;93:S69-74. [Crossref] [PubMed]
- Lu H, Liu S, Zhu W, et al. Surgical treatment of benign parotid gland tumors: From functional surgery to personalized surgery. Oral Oncol 2025;161:107172. [Crossref] [PubMed]
- Fisher R, Ronen O. Cytologic diagnosis of parotid gland Warthin tumor: Systematic review and meta-analysis. Head Neck 2022;44:2277-87. [Crossref] [PubMed]
- Wakonig KM, Lerchbaumer MH, Arens P, et al. Multiparametric Ultrasound for Preoperative Assessment of Parotid Tumors: A Novel Diagnostic Pathway. Head Neck 2026;48:392-406. [Crossref] [PubMed]
- Giganti F, Antunes S, Salerno A, et al. Gastric cancer: texture analysis from multidetector computed tomography as a potential preoperative prognostic biomarker. Eur Radiol 2017;27:1831-9. [Crossref] [PubMed]
- Palani D, Ganesh KM, Karunagaran L, et al. Statistical Analysis on Impact of Image Preprocessing of CT Texture Patterns and Its CT Radiomic Feature Stability: A Phantom Study. Asian Pac J Cancer Prev 2023;24:2061-72. [Crossref] [PubMed]
- Chang YJ, Huang TY, Liu YJ, et al. Classification of parotid gland tumors by using multimodal MRI and deep learning. NMR Biomed 2021;34:e4408. [Crossref] [PubMed]
- Gunduz E, Alçin OF, Kizilay A, Yildirim IO. Deep learning model developed by multiparametric MRI in differential diagnosis of parotid gland tumors. Eur Arch Otorhinolaryngol 2022;279:5389-99. [Crossref] [PubMed]
- Li J, Weng J, Du W, et al. Machine learning-assisted diagnosis of parotid tumor by using contrast-enhanced CT imaging features. J Stomatol Oral Maxillofac Surg 2025;126:102030. [Crossref] [PubMed]
- He Y, Zheng B, Peng W, et al. An ultrasound-based ensemble machine learning model for the preoperative classification of pleomorphic adenoma and Warthin tumor in the parotid gland. Eur Radiol 2024;34:6862-76. [Crossref] [PubMed]
- Prasad RS. Parotid Gland Imaging. Otolaryngol Clin North Am 2016;49:285-312. [Crossref] [PubMed]
- Mantsopoulos K, Tschaikowsky N, Goncalves M, et al. Evaluation of preoperative Ultrasonography in the Differentiation between Superficial and Deep Parotid Gland Tumors. Ultrasound Med Biol 2020;46:2099-103. [Crossref] [PubMed]
- Stoia S, Ciurea A, Băciuț M, et al. Comparative Diagnostic Accuracy of Ultrasound, MRI, and Fine-Needle Aspiration Biopsy in the Preoperative Evaluation of Parotid Gland Tumors. J Clin Med 2025;14:1342. [Crossref] [PubMed]
