(A531) Deep Learning-Based Web Application for the Detection and Classification of Salivary Gland Tumors

Introduction: Ultrasound is the primary tool for evaluating salivary gland tumors (SGTs), but its dependence on subjective interpretation can result in diagnostic inconsistencies. This study aims to present a deep learning-based web application designed to automate the diagnosis of SGTs.

Methods: This retrospective study was conducted at a tertiary medical center. A comprehensive dataset of ultrasound images from patients with pathology-confirmed SGTs was compiled and divided into two cohorts: a training set consisting of 687 images collected from 2007 to 2020, and a testing set of 181 images from 2021 to 2022. The YOLOv11 model was employed for SGT detection, while a modified ResNet50V2 model was utilized for classifying tumors as benign or malignant. Following training and integration, a user-friendly web application was developed using Streamlit for online deployment. The models underwent validation using two distinct datasets: an internal set of 56 images from 2023 and an external set of 57 images sourced from online repositories.

Results: For tumor detection, our model achieved bounding box precision and recall rates of 0.94 each on the testing set. For tumor classification, it resulted in an accuracy of 80%, a sensitivity of 71%, and a specificity of 83% on the testing set. The web application demonstrated consistent diagnostic performance, with accuracies of 75% and 79%, sensitivities of 75% and 67%, and specificities of 75% and 86% among separate internal and external validations.

Conclusions: This study introduces a deep learning-based web application for automated SGT diagnosis. The results indicate effective performance in both tumor detection and classification across various cohorts, suggesting that our model could assist the diagnostic processes in clinical settings.