Up to 80% of parotid tumors are benign with the two most common types being pleomorphic adenoma (PA) and Warthin tumor (WT) (1, 2). Compared with WT, PA exhibits a higher potential for malignant transformation and recurrence, so the surgical approaches and prognosis are completely different (3, 4). Hence, for the purpose of precisely and individually treating patients with benign parotid tumors, it is crucial to accurately distinguish between PA and WT.

At present, the preoperative diagnosis of PA or WT relies on fine needle aspiration cytology (FNAC) and radiological images. However, FNAC is not always conclusive because of sampling difficulties and the experience of pathologist (5, 6). Furthermore, FNAC is invasive, which may lead to hemorrhage (7), inflammation (8), and dissemination of tumor cells along the needle route (9). Patients with similar clinical factors may have varying outcomes, and it is often difficult to definitively distinguish between WT and PA based solely on clinical factors. In comparison to CT and needle biopsy, MRI offers several advantages such as non-invasiveness, absence of radiation, and excellent soft tissue resolution (10). In the evaluation of parotid tumors, MRI can provide information about the size, location, shape, and characteristics of the tumor, which can help guide treatment decisions (11, 12). Nonetheless, conventional MRI differential diagnosis has not always been adequate because of the substantial overlap in morphological features between PA and WT. In addition, conventional MRI diagnosis may have a subjective component and depend on the expertise and experience of radiologist (13).

Radiomics can extract high throughput of quantitative features by converting images into amenable data, and the analyzing these data for decision support (14). Radiomics can provide much more comprehensive information from medical images than human eyes (15). In recent years, radiomics has been widely used for preoperative diagnosis of parotid tumors (16–18). Some previous studies have tried to discriminate benign and malignant parotid tumors using radiomics (19, 20), but only a few of them have analyzed the differentiation of PA from WT (21, 22). However, there is no research focused on differentiating misdiagnosed or ambiguous PA and WT using radiomics.

Therefore, we use a variety of machine learning methods to establish different MRI-based radiomics models and determine the optimal radiomics models for identifying misdiagnosed or ambiguous PA and WT. By integrating a variety of models combining radiomics and clinical parameters, we evaluate the effect of multimode combined application in differential diagnosis of the disease, so as to improve the accuracy of diagnosis of the disease.

We found that gender, age, and smoking were clinical independent predictors for the differential diagnosis of PA and WT. The LR algorithm model, which was based on nine popular machine learning algorithms, was the best radiomics model with the highest accuracy and AUC. The fusion models—nomogram, stacking, and ensemble—also demonstrated superior diagnostic performance and produced a good net clinical benefit when compared with the clinical model. In comparison with the best radiomics model, the nomogram showed a better AUC. It also outperformed stacking and ensemble models in terms of superior generalization ability and more consistent discrimination efficiency.

Previous studies have reported that gender, age, and smoking history of patients had significance in the identification of PA and WT (30, 31). Our results were similar to those studies. Some studies suggested that duration of smoking was a strong risk factor (32). Because male smokers were more prevalent, WT was more common in men. The pathogenesis may relate to the fact that tobacco contains chemical irritants such as benzopyrene, arsenic, and N-nitrosoguanidine (31).. These irritants leading to secondary tumor change was a lengthy phenomenon, so WT occurred in middle or old age. Some studies suggested that the comparatively significantly greater incidence of WT in men might indicate a hormone dependence, and progesterone receptors have been found in WT (33). The evidence of progesterone receptor in WT implicated a potential role of endocrine factors in the development of this tumor, which might explain the predominance of the male sex regarding this disease (34).

PAs are also known as mixed tumors due to histological heterogeneity, which also suggests that it is represented by various imaging findings (23). When PA has fewer cellular components of mucoid tissue, high signal intensity on T2WI images decreases, which reduces the rate depending on the proportion of cellular components (35).The tumor signal expression of WT depends mainly on the cystic component of the tumor, and as the size of the cystic component increases, the internal structure looks bright at T2WI images that may simulate PA (36). When PA and WT showed similar imaging manifestations, it was difficult to distinguish PA from WT based on conventional MR imaging ( Figure 5 ). The radiologists only paid attention to the imaging manifestations of the tumor while ignoring the clinical characteristics, which was more likely to be misdiagnosed.

Radiomics is a non-invasive technique that builds models from digitized medical images and uses clever computation to convert them into high-dimensional, quantitative data that can be used to improve medical decision-making and provide useful diagnoses (14). Liu et al. (37) reported that there were no appreciable variations between MRI and CT in radiomics characteristics for diagnosing parotid malignancies. In this study, the diagnostic efficacy of the radiomics model was not as good as that of the clinical model. Potential explanations for these results included the subjective impact of individual clinical experience as well as a single imaging index. T2WI provided the vital features for the optimal radiomics model. PA contains mucoid tissue and usually shows a high signal on T2WI (38). In comparison, WT are epithelial tissues with lymphoid hyperplasia that contain cystic components of approximately 30% protein liquids or viscous colloids, and they usually show a hypointense/with hyperintensity signal on T2WI (36). Additionally, this study found that GLCM features could help discriminate between PA and WT, similar to the results of Gabelloni (39). The coarseness of the texture was represented by the zone percentage of GLCM features, which may more accurately capture the heterogeneity of various tumor types.

Recognizing the best machine learning techniques for radiomics models is essential (40). Thus, we employed nine common classification algorithms in model construction. LR outperformed other classifiers, which were consistent with the results of Lu et al. (41). More training samples may have been needed for sophisticated models, which could be the cause (42). The optimal radiomics model based on LR did not have a higher AUC than the clinical model. This result also illustrates that when there is a problem with the observation of traditional imaging, radiologists should combine clinical data. Evidence-based clinical decision support systems can be produced with accuracy and dependability by combining radiomics features with clinical parameters and other pertinent data (43). In this study, the clinical or radiomics model did not perform as well in terms of diagnostic performance and clinical net benefits as the nomogram, stacking model, and ensemble model, which are instances of fusion models constructed utilizing clinical parameters and radiomics features. Additionally, the nomogram exhibited a highest AUC when compared with the other models. Zheng et al. (44) constructed a radiomics nomogram based on MRI that had good prediction efficiency in distinguishing PA from WT, obtaining a similar conclusion as this research. The ensemble strategy has the advantage of being able to reduce the variance and bias of the model while also enhancing its robustness and generalization in classification and prediction, by using a strong majority voting or group average method (45). A recent report had proposed that the stacking ensemble model obtained excellent diagnostic performance and showed good stability of the calibration plot (46). While AUCs for the ensemble and stacking models were less than those of the nomogram in the current study, their diagnostic performance in the average validation groups was comparable with and superior to that of the radiomics models. As a result, the nomogram demonstrated better and more consistent differential diagnosis efficiency with superior reproducibility and reliability when compared with stacking and ensemble models.

This study’s limitations were the fact that it only included participants from two centers, and the sample of external test data was relatively small. Additional patients from more centers must be included to expand the universality in clinical applications, in the future. Second, this was a retrospective study, which might cause potential selection bias. In the future, prospective validation will be performed. Third, there were variations in the MRI scanner and parameters, which could have an impact on the models’ output. We performed the N4 bias field correction. Fourth, we only studied conventional MRI sequences, with limited interpretability. Other quantitative functional MRI sequences, such as DWI and DCE-MRI, still need to be further explored.

The MRI-based radiomics models can be accomplished to preoperatively differentiate misdiagnosed or ambiguous PA and WT, and the LR algorithm-established model is the optimal radiomics model. The nomogram is an effective tool for preoperative and non-invasive distinguishing PA and WT, which can be challenging for radiologists and surgeons to ascertain prior to surgery. In daily work, it is necessary to combine with clinical parameters such as gender, age, and smoking when radiologists are difficult to distinguish PA from WT.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by the First People’s Hospital of Yunnan Province. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because this is a retrospective study, so informed consent was waived by the ethics committee.

JY: Methodology, Writing – original draft. QB: Conceptualization, Methodology, Writing – review & editing. YJ: Validation, Writing – review & editing. YY: Data curation, Writing – review & editing. JD: Data curation, Writing – review & editing. HZ: Writing – review & editing. KW: Writing – review & editing.