Deep Learning-Based Diagnosis of Parotid Gland Tumors on CT Images: A Multi-view Approach for Preoperative Differentiation of Benign and Malignant Lesions.
Authors
Affiliations (3)
Affiliations (3)
- Department of Oral and Maxillofacial Surgery, Peking University School and Hospital of Stomatology, National Center for Stomatology, National Clinical Research Center for Oral Diseases, Beijing Key Laboratory of Digital Stomatology, NHC Key Laboratory of Digital Stomatology, China.
- School of Biomedical Engineering, ShanghaiTech University, Shanghai, China.
- Department of Oral and Maxillofacial Surgery, Peking University School and Hospital of Stomatology, National Center for Stomatology, National Clinical Research Center for Oral Diseases, Beijing Key Laboratory of Digital Stomatology, NHC Key Laboratory of Digital Stomatology, China. Electronic address: [email protected].
Abstract
Accurate preoperative differentiation between benign and malignant parotid gland tumors is essential for guiding surgical planning and treatment decisions. However, early-stage malignant tumors often lack distinctive imaging features on contrast-enhanced CT, making diagnosis challenging. This study aimed to develop an artificial intelligence (AI)-assisted diagnostic system based on a multi-view learning framework and evaluate its diagnostic performance on contrast-enhanced CT images. In this retrospective study, contrast-enhanced CT images were collected from 578 patients (472 benign, 106 malignant) with pathologically confirmed parotid gland tumors at Peking University School and Hospital of Stomatology (2015-2023). An AI system was developed comprising a U-Net-based segmentation network and a multi-view classification model that uses multi-angle reconstructed images and a majority-voting strategy for case-level classification. The diagnostic performance was evaluated using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC). The nnU-Net model achieved a mean Dice score of 0.90 and IoU of 0.83 across 115 testing cases, providing precise segmentation with minimal outliers. For the diagnostic classification task on the test set, the ResNet model outperformed the other evaluated neural networks, achieving the highest accuracy of 89.7% and an AUROC of 85.7%. It correctly identified 57.1% (8/14) of malignant and 96.9% (62/64) of benign tumors. The proposed AI-assisted diagnostic system demonstrates promising performance in preoperative differentiation of benign and malignant parotid gland tumors. This system has the potential to serve as a valuable decision-support tool for clinical decision-making and surgical planning, ultimately improving patient outcomes.