# Diagnosis of parotid gland tumors using a ternary classification model based on ultrasound radiomics

## TL;DR

This study shows that AI models using ultrasound images can accurately distinguish between benign and malignant parotid tumors, outperforming expert sonographers.

## Contribution

The study introduces two-step radiomics models that improve non-invasive diagnosis of parotid gland tumors.

## Key findings

- The LASSO-BNB model achieved 91.0% AUC in differentiating benign from malignant parotid tumors.
- The RFE-Voting model achieved 96.2% AUC in distinguishing pleomorphic adenomas from Warthin’s tumors.
- Both models outperformed experienced sonographers in diagnostic accuracy.

## Abstract

This study aimed to evaluate the diagnostic value of two-step ultrasound radiomics models in distinguishing parotid malignancies from pleomorphic adenomas (PAs) and Warthin’s tumors (WTs).

A retrospective analysis was conducted on patients who underwent parotidectomy at our institution between January 2015 and December 2022. Radiomics features were extracted from two-dimensional (2D) ultrasound images using 3D Slicer. Feature selection was performed using the Mann–Whitney U test and seven additional selection methods. Two-step LASSO-BNB and voting ensemble learning modeling algorithm with recursive feature elimination feature selection method (RFE-Voting) models were then applied for classification. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), and internal validation was conducted through fivefold cross-validation.

A total of 336 patients were included in the study, comprising 73 with malignant tumors and 263 with benign lesions (118 WT and 145 PA). The LASSO-NB model demonstrated excellent performance in distinguishing between benign and malignant parotid lesions, achieving an AUC of 0.910 (95% CI, 0.907–0.914), with an accuracy of 86.8%, sensitivity of 92.5%, and specificity of 66.7%, significantly outperforming experienced sonographers (accuracy of 61.90%). The RFE-Voting model also showed outstanding performance in differentiating PA from WT, with an AUC of 0.962 (95% CI, 0.959–0.963), accuracy of 83.0%, sensitivity of 84.0%, and specificity of 92.1%, exceeding the diagnostic capability of experienced sonographers (accuracy of 65.39%).

The two-step LASSO-BNB and RFE-Voting models based on ultrasound imaging performed well in distinguishing glandular malignant tumors from PA and WT and have good predictive capabilities, which can provide more useful information for non-invasive differentiation of parotid gland tumors before surgery.