Salivary gland tumors, although relatively uncommon, predominantly arise in the parotid glands. Over 80% of these tumors are benign (1). The treatment strategies for benign and malignant parotid tumors (MPTs) differ significantly. Benign parotid tumors (BPTs) are generally managed with local or superficial parotidectomy, whereas MPTs necessitate a more aggressive treatment, including total or subtotal parotidectomy, potentially accompanied by facial nerve resection or postoperative chemoradiation. Thus, precise preoperative diagnosis is pivotal for selecting the appropriate treatment modality and for prognostic assessment. Fine needle aspiration biopsy (FNAB) is widely recognized as the gold standard for diagnosing parotid gland tumors. Despite its high overall accuracy, sensitivity, and specificity nearly 90% (2, 3), FNAB is invasive and associated with complications. Notably, between 5.4% and 19.4% of patients may experience facial palsy as a complication (4). FNAB also faces challenges in sampling tumors located in the deep lobe due to the procedure’s invasiveness. The inherent heterogeneity of tumors may not fully capture. When considering ‘inconclusive’ reports, the sensitivity of FNAB for detecting malignant lesions drops significantly to 48% (5). Additionally, distinguishing between malignant parotid tumors (MPTs) and benign parotid tumors (BPTs) presents significant heterogeneity, as evidenced by a large meta-analysis (6).

The most prevalent BPT is pleomorphic adenoma (PA), followed by Warthin’s tumor (WT) (1, 7). WT is characterized by significant clinical and imaging features. It predominantly occurs in elderly men with a smoking history (7), manifests as multiple nodules in both parotid glands, typically situated in the posterior lower pole of the superficial lobe (8, 9), and exhibits a rapid enhancement and washout pattern on multi-phase enhanced CT scans (10). Consequently, precise preoperative differentiation between PAs and parotid carcinomas (PCAs) becomes critically important in clinical practice. Distinguishing between PCAs and PAs based solely on clinical symptoms, such as pain and palpable masses, proves challenging. Meta-analyses indicate that while conventional imaging techniques offer high specificity, their sensitivity in differentiating between BPTs and malignant parotid tumors (MPTs) is comparatively low (11).

Magnetic resonance imaging (MRI) is favored for its superior tissue contrast and multi-parametric capabilities, establishing it as the preferred method for diagnosing PTs. Recent advancements in MRI techniques, including diffusion-weighted imaging (DWI), arterial spin labeling (ASL), diffusion kurtosis imaging (DKI), and dynamic contrast-enhanced (DCE) MRI, have been explored for PT diagnosis (12, 13). However, MRI presents several key limitations: (1) Longer acquisition times increase patient discomfort and motion artifact risk; (2) Higher costs and lower availability limit accessibility; (3) Contraindications exist for patients with claustrophobia, non-cooperation, or incompatible implants; and (4)Debate persists regarding the repeatability and diagnostic accuracy of some advanced techniques, particularly in PCAs with highly cellularity (13). Computed tomography (CT), in contrast, offers distinct practical advantages for preoperative evaluation: (1) Faster scan times improve patient tolerance, especially for uncooperative or dyspneic individuals; (2) Lower cost and wider availability enhance accessibility; and (3) It avoids MRI-specific contraindications like claustrophobia. Multi-phase enhanced CT effectively captures tumor vascular perfusion information. Critically, CT radiomics presents a unique opportunity: it can be retrospectively applied to existing multi-phase CT datasets. This is particularly valuable for analyzing parotid masses discovered incidentally. Nevertheless, CT has significant drawbacks: (1) Ionizing radiation exposure, especially concerning with multi-phase protocols as used in this study; and (2) Overlapping imaging features between PCAs and PAs (e.g., size, shape, cystic change, calcification, enhancement) frequently complicate diagnosis (10), particularly for atypical PCAs without invasion or metastasis.

Radiomics, distinguished by its ability to extract high-throughput quantitative data from medical images non-invasively, has emerged as a groundbreaking field in precision medicine. This approach, which aims to elucidate tumor heterogeneity, has gained widespread recognition for its applications in tumor classification (14, 15), outcome prediction (16), and therapy response assessment (16). Notably, radiomic models have demonstrated significant diagnostic value in differentiating between benign and malignant lympho-associated PTs (17), as well as in distinguishing between PAs, WTs (18), and basal cell adenomas (19) based on CT images. However, the rarity of PTs and the absence of a large database have limited most radiomic studies on these tumors to single-center efforts with small sample sizes. Variability in the selection of enhancement phases has led to significant discrepancies and controversies in findings. Furthermore, the stability and robustness of existing models remain suboptimal (20), posing challenges for clinical application. To date, studies have generally grouped WTs and PAs into BPTs for differentiation from malignant counterparts, with no multicenter studies specifically addressing PAs and PCAs (especially atypical PCAs) differentiation reported. Atypical parotid gland cancer refers to tumors that do not show significant malignant features, such as invasion of adjacent tissues or lymph node enlargement, on radiological imaging. Such cancers are often misdiagnosed as PAs in clinical practice. Consequently, the effectiveness of phase selection, feature extraction methods, model construction techniques, or their combinations in enhancing differential diagnosis performance remains unexplored.

This study aims to develop and validate a machine learning prediction model that leverages multi-phase CT radiomics features, clinical-radiological characteristics, and their integration to ascertain its efficacy in distinguishing atypical PCAs from PAs.

This study explores the development of more efficient diagnostic methods to differentiate between atypical PCAs and PAs. To achieve this, clinical models, a radiomics signature, and a radiomics nomogram were constructed. These tools were based on the clinical and radiomics features derived from multiphase CT images, utilizing the LightGBM model. The radiomics nomogram demonstrated exceptional diagnostic value, evidenced by its AUC scores of 1.000, 0.854, and 0.783, accuracy rates of 1.000, 0.864, and 0.750, and F1 score of 1.000,0.914 and 0.826 in the training, validation, and independent testing sets, respectively. Consequently, the radiomics nomogram shows great promise as an effective and non-invasive method for distinguishing atypical PCAs from PAs.

Although PAs and PCAs have well-defined classifications of benign and malignant tumors, but the diverse proportions of epithelial and mesenchymal components contribute to the varied morphological characteristics of PAs. In addition, the morphology of MPTs varies greatly among different tissue types. This complexity often makes it challenging to distinguish PCAs from PAs, particularly PAs with unclear margins and lobulated borders, atypical PCAs with clear boundaries, small size, no invasion of adjacent tissues or lymph node metastases, when using conventional CT (22). In this study, patients with invasion of adjacent tissues and lymph node metastasis were excluded, which increased the difficulty of differential diagnosis and had more clinical significance.

The results showed that age, tumor margins, shape, location and size were statistically different between PAs and atypical PCAs in training set. This explains the good diagnostic performance of the clinical model. Our study demonstrates that a clinical model, integrating clinical risk factors and radiological findings, performs satisfactorily, exhibiting AUC values of 0.869, 0.673, and 0.746 in training, validation, and independent testing sets, respectively. In previous studies, the patients with MPTs were elder than those with BPTs, and the sizes of MPTs were larger than BPTs (4). PAs occurred more often in women with the distribution of bimodal age of 38 and 64 (23). The significant radiological findings to indicating MPTs was lobular or irregular contour and ill-defined margin of the mass (24, 25). However, the above image feature extraction is subjective and greatly affected by the knowledge and experience of radiologists. A subspecialist with 30 years of experience in head and neck radiology demonstrated higher sensitivity and comparable specificity to MR radiomics models for PAs diagnosis (26). However, radiomics models potentially reduce the technical barriers in diagnosing parotid tumors.

The radiomics data and features derived from artificial intelligence algorithms quantify tumor heterogeneity, which may not be discernible to the naked eye, thereby enhancing the diagnostic utility of image datasets (27). CT/MR Radiomics has been widely used in the diagnosis of head and neck tumors, gene expression, immune status and efficacy evaluation, etc. In recent years, it has also been partially reported in the diagnosis of PGTs.MR-based radiomics achieves an accuracy of 80.4% in differentiating MPTs from BPTs (28) and 80.43% in distinguishing MPTs from PAs (29). He’s research indicated that the accuracy of the XGBoost model, based on MR images for differentiating MPTs from BPTs, surpasses that of radiologists (89.9% vs 83.1%) (30). Although MRI offers detailed insights, CT scans remain the preferred preoperative examination due to their lower cost, faster scanning, and fewer contraindications. Additionally, MRI- and CT-based radiomics signatures exhibit no significant disparity in differentiating PAs from WTs (31). Qiang discovered that the radiomics model, which combines three-phase enhanced CT imaging, particularly the arterial phase, significantly aids in diagnosing BPTs from MPTs (AUC=0.933, 0.936) (32). The support vector machine (SVM)-based prediction model reported AUCs of 0.854 in the training set and 0.741 in the test set in Xu’s study (33). These studies were constrained by the absence of an independent testing set and the small sample size of MPTs. Our study, incorporating patients from five different institutions, enhances the reliability and reproducibility of our results through verification in both validation and independent testing sets.

Previous researches utilized multiphasic CT scanning to analyze the enhancement patterns of PGTs, which are indicative of their pathological characteristics. There exists a debate over the optimal phase of enhancement to effectively differentiate between BPTs and MPTs. A notable observation is the limited efficacy of delayed enhancement in distinguishing PAs from PCAs (10, 34). However, findings from early contrast phase CT images (35) and radiomics models (19) have demonstrated superior performance in differentiating PCAs from PAs. These may be due to pathological features of PCAs: increased number of blood vessels, increased vascular permeability and abnormal arteriovenous fistula (36). Consequently, our study developed a radiomics model (model _A ) using the arterial phase features as the basic model. Additionally, we evaluated radiomics models that incorporated combinations of different phases. In the training set, radiomics signatures model _A+P, model _A+V and model _A+P+v exhibited enhanced predictive capabilities compared to model _A . The model _A+P demonstrated superior diagnostic performance in the external test group, as indicated by its higher recall and F1 score, suggesting that the venous phase contributes marginally to the diagnosis of PGTs in radiomics model. This finding aligns with previous studies, where radiomics models based on the arterial phase outperformed those using the venous and plain phases in a single-phase analysis. Moreover, the best predictive performance was achieved by models that integrated all phases(model _A+P+v) (32 ). Our radiomics analysis indicates the potential to optimize multiphase CT protocols for parotid tumor assessment. The high discriminative power of the model _A+P for differentiation of PA and atypical PCA suggests that reduced protocols may maintain diagnostic accuracy while addressing critical limitations. This approach promises significant radiation reduction, which is particularly vital for head and neck imaging under the As Low As Reasonably Achievable (ALARA) principles, as well as enhanced clinical feasibility through shorter scan times. Importantly, this approach enables the retrospective application of radiomics to the abundant clinical CT archives containing incidentally detected parotid masses from routine examinations, substantially expanding its real-world applicability. However, future validation is necessary to confirm performance parity across multicenter cohorts with varying acquisition parameters. Both the arterial phase and the multiphase combined radiomics models demonstrated excellent and comparable effectiveness in differentiating PAs from other BPTs (19). Similarly, in a multicenter radiomics study, the arterial phase was selected for distinguishing PAs from WTs (18). In contrast to previous studies, our research innovatively combined radiomics signatures from plain or venous phases with those from the arterial phase, comparing these combinations to identify the most effective diagnostic approach for PGTs. The model_A+P demonstrated superior diagnostic performance, achieving a significantly higher AUC value in the training set compared to model _A . While the AUC values for model_A+P remained higher than those for model _A in both the validation and independent testing sets, the differences were not statistically significant. This suggests that while model_A+P offers an improvement in diagnostic performance, further investigation is needed to fully understand the clinical impact of these findings.

To enhance the efficiency of the diagnostic model, we estimated radiomics nomograms based on clinical risk factors combined with radiomics signatures model _A+P . This integrated approach yielded AUC values ranging from 0.750 to 1.000, significantly improving the differentiation between PAs and atypical PCAs compared to the standalone radiomics model in the independent set (p < 0.05). Previous studies corroborate our findings, showing that radiomics nomograms that include both clinical factors and a radiomic signature outperform models based solely on radiomic scores (37) or clinical factors (38). Our study aligns with these results, demonstrating the enhanced diagnostic performance of this comprehensive approach.

To underscore the significance of various features in the model’s predictions and to augment model interpretability, key elements of this model were identified using the SHAP interpretation methodology (39). This study highlighted the pivotal role of the cluster shade, run length non-uniformity, and first-order mean features through wavelet or exponential transformation in diagnosing PAs and atypical PCAs. These features are used to quantify image inhomogeneity and texture directionality, which may reflect tumor heterogeneity to some extent. Echoing findings from prior research (32), it was found that sphericity and wavelet features derived from multiphasic CT images are closely associated with the diagnosis of benign and malignant PGTs. Sphericity features quantitatively assess the tumor region’s roundness in comparison to a sphere (40), while wavelet features are believed to capture the tumor’s heterogeneity (41) and are valuable in indicating tumor proliferation (42). In other texture analysis studies, the Gray Level Co-occurrence Matrix (GLCM) features, which represent the spatial relationship between neighboring pixels or voxels, have also been shown to depict tumor heterogeneity. These features have been utilized in constructing radiomics models for differentiating MPTs from PAs (33) and other subtypes of BPTs (19, 43). However, due to differences in incidence rates, even though this study included data from five centers, there was still an imbalance in the proportion of PCA and PA. Consequently, the calculation of SHAP values may have underestimated the significance of PCA characteristics, resulting in a bias in the local interpretation of PCA. Expanding the sample size of parotid gland cancer cases in future studies may resolve this issue.

Consequently, based on existing literature, there are no universally acknowledged image features for model construction (20), and the repeatability and reliability of these models warrant further investigation in future research.

Despite the promising findings, our study has limitation which require attention. Ever though we collected the patients from 5 different institutions, the sample size for parotid carcinoma is still not enough because of the low incidence rate. Therefore, the subtypes of parotid carcinomas were unable to be taken into consideration. The small sample size of PCAs also has an impact on the generalizability of the constructed model. Therefore, the sample size, especially for PCAs, will be further increased in the next study. Second, due to the large number of sites included, differences in machine and scanning protocols may cause image instability. Finally, deep learning which may provide more efficient compared with the traditional machine learning should be applied in the following study.

In conclusion, the radiomics nomogram developed in this study, leveraging data from both arterial and plain CT phases, exhibits superior performance in accurately distinguishing between atypical PCAs and PAs. Consequently, it holds significant promise as an indispensable tool for enhancing clinical decision-making, offering a pathway toward more tailored and effective management strategies for patients with parotid gland disorders. Looking ahead, the integration of this nomogram into clinical practice could revolutionize the diagnostic approach, encouraging further research and development in this promising field.