Establishment And Clinical Application Of Classification Model Of Parotid Gland Tumors Based On Clinical And Ultrasound

Objective:Parotid gland is the largest pair of salivary gland tissue in human body.The incidence of parotid tumor,accounting for 70%～80%,is the highest in all salivary gland tumors.For patients with different pathological types of parotid tumors,their clinical manifestations are not the same,and the corresponding treatment and prognosis are not the same.Previous studies have reported that the incidence of postoperative facial paralysis in patients with malignant tumors in the parotid region is higher than that in patients with benign tumors.The occurrence of facial paralysis can lead to the inconvenience of life,such as puffy cheeks,air leakage,crooked mouth,incomplete closure of eyes,and the increase of the incidence of depression caused by facial problems.The five-year recurrence rate of pleomorphic adenoma in benign tumors is higher than that of Warthin’s tumors.Therefore,it is particularly important to improve the accuracy of pathological classification of parotid tumors before an operation.At present,the conventional imaging methods for parotid tumors are ultrasound,CT,and MRI.Ultrasound has the advantages of low price,easy access to images,less time-consuming,and no problems such as radiation,metal incompatibility,and claustrophobia.It is an important way of parotid tumor screening.With the development of high-frequency ultrasound,the demand for ultrasound examination in clinical work is increasing.However,due to the shortcomings of operator dependence and the strong subjectivity of ultrasound,the ultrasound diagnosis levels of parotid tumors by physicians of different ages and different development levels are not the same.This study constructed a risk prediction model for malignant parotid tumors based on clinical and ultrasound features and a ultrasound-based machine learning classification model for pleomorphic adenoma and Warthin’s tumor,aiming to provide an effective and simple method to reduce the subjective dependence on ultrasound diagnosis and provide more accurate and targeted diagnosis and treatment strategies for clinical practice.Materials and methods:Patients with parotid tumors who underwent parotid surgery in Foshan First People’s Hospital from 06 June 2018 to 03 March 2021 were retrospectively analyzed.In the first part,266 cases were included,which were divided into the training set of 214 cases and the verification set of 52 cases.The selected ultrasonic features and clinical data were included in multivariate Logistic regression analysis,and the malignant risk prediction model of parotid tumors was constructed.The column diagram,ROC curve,correction curve,and DCA curve were drawn,and the model was verified externally.The diagnostic performance of the Nomogram prediction model was compared with traditional empirical diagnosis and PIRADS,including ROC curve comparison,AUC,accuracy,sensitivity,specificity,positive predictive value,negative predictive value,positive likelihood ratio,and negative likelihood ratio.The second part is a retrospective analysis of patients with pleomorphic adenoma or Warthin’s tumor who underwent parotid surgery in Foshan First People’s Hospital from June 19,2017,to March 15,2022.A total of 173 patients were included and divided into a training set and validation set with 121 cases and a test set with 52 cases.Through image analysis and segmentation of ROI,feature extraction,and repeated training,the optimal three basic models were selected and fused with the meta-model to construct the multimodal machine learning classification model of pleomorphic adenoma and Warthin’s tumor,and the SHAP value was used to evaluate the importance of features.The classification performance of the constructed multimodal machine learning model,the single machine learning model,and the traditional empirical diagnosis in pleomorphic adenoma and Warthin’s tumor was analyzed and compared by the DeLong test.Results:1、The first part:Multivariate analysis showed that facial nerve function and cervical lymph node abnormalities in ultrasound images,the maximum diameter,boundary,and shape of the tumor were independent predictors of malignant risk of parotid tumors.The nomogram prediction model was established using the above five indicators,and the results showed that the C-index of the nomogram was 0.896(95%CI:0.8340.958).The standard curve showed that the nomogram prediction effect was in good agreement with the actual situation of benign and malignant parotid tumors,and the internal verification C-index was 0.878.The C-index of external validation of the nomogram was 0.850.The DCA curve is far away from the x-axis(threshold probability)and y-axis(net income).Compared with traditional empirical diagnosis,the classification performance of the nomogram prediction model was better(P<0.01).2、The second part:A multi-modal machine learning classification model for pleomorphic adenoma and Warthin tumor was constructed based on ultrasonic gray-scale image radiomics features,clinical features,and ultrasonic features.The AUC value of the machine learning model was 0.912(95%CI:0.799-0.972),compared with the traditional empirical diagnosis(AUC=0.765,95%CI:0.627-0.871).The diagnostic efficiency of the multi-modal machine learning model for the Warthin’s tumor in the parotid region was better than that of the traditional empirical diagnosis(P<0.05),The sensitivity,specificity,negative predictive value,and positive predictive value of Warthin’s tumor in the parotid region can be improved from 82.14%,70.83%,76.7%,and 77.3%of traditional empirical diagnosis to 92.86%,87.50%,89.7%,and 91.3%,respectively.SHAP value results show that the greatest impact on the diagnosis of parotid Warthin’s tumors are:smoking,age,maximum tumor diameter,and radiomics features 1bp-2D_glrlm_RunLengthNonUniformity、original_shape_Maximum2DDiameterColumn.Conclusion:The malignant risk prediction model of parotid tumors based on clinical and ultrasonic characteristics can better evaluate the biological characteristics of parotid tumors and has high accuracy in predicting the malignant risk of parotid tumors.Its clinical practicability is strong,which can help clinicians directly predict the malignant risk probability of parotid tumors before operation and improve the diagnosis and treatment level.The ultrasound-based machine learning classification model of pleomorphic adenoma and Warthin’s tumor constructed in this study has high diagnostic efficiency,which provides more accurate and targeted diagnostic strategies for clinical practice,and better assists clinical classification of pleomorphic adenoma and Warthin’s tumor with large differences in biological behavior and prognosis.It is conducive to the use of artificial intelligence technology to promote the development of ultrasound in the classification and diagnosis of parotid tumors,and further assists the realization of the goal of accurate diagnosis and treatment of parotid tumors.