Salivary gland tumors, which are relatively uncommon, account for 3-12% of head and neck tumors and most often occur in the parotid gland. Among these tumors, parotid gland tumors are the most common type, with the majority being benign. The treatment for these tumors is mainly surgical, with varying strategies and outcomes based on histopathological types. Accurate preoperative diagnosis of benign and malignant parotid gland tumors is crucial for treatment decisions.
Preoperative diagnosis of parotid gland tumors is typically done using ultrasound-guided core needle biopsy or fine-needle aspiration, supplemented with medical imaging. While ultrasound is widely used for diagnosis due to its non-invasiveness and cost-effectiveness, differential diagnosis of benign and malignant parotid masses remains a challenge. Machine learning methods, particularly deep learning, have shown promise in medical imaging diagnosis across various diseases. However, there is limited research on using deep learning to distinguish between benign and malignant parotid tumors based on ultrasound images.
In this study, a model was developed to automatically segment parotid lesions on ultrasound images, utilizing deep learning algorithms. The ResNet18 model was found to outperform other models in classifying benign and malignant parotid tumors, achieving favorable performance in both internal and external test sets. The model provided radiologists with significant assistance in improving diagnostic accuracy, particularly for less experienced radiologists. By combining automatic segmentation and deep learning methods, the study aimed to enhance workflow efficiency and diagnostic accuracy in distinguishing between benign and malignant parotid gland tumors. Further prospective research is warranted to validate these findings and explore the potential clinical applications of the developed model.