Diagnosis of Salivary Gland Tumors Using Ultrasound Radiomics

Abstract

Objective

The current diagnosis of salivary gland tumors (SGTs) is dependent on subjective ultrasound features. Here we aimed to develop an objective method using ultrasound radiomics.

Methods

We collected 248 benign and 46 malignant images and divided them into training (80%) and testing (20%) groups, with 105 radiomic features extracted from each image. Data re-sampling, feature selection and classification were conducted. The diagnostic accuracy of different combinations was evaluated.

Results

After data re-sampling using the Synthetic Minority Over Sampling Technique (SMOTE) and feature selection with LASSO+ANOVA, 10 radiomic features were selected. Using the Random Forest classifier, the testing set achieved an area under the receiver operating characteristic curve of 0.85, accuracy of 90%, sensitivity of 78% and specificity of 92% for diagnosing SGTs. It maintained an accuracy of 85% in a separate internal validation set.

Conclusion

This study offers significant insights into the use of radiomics for the diagnosis of SGTs. When selected properly and paired with a suitable classification model, radiomics can be used to differentiate between benign and malignant SGTs.