Comparison of Deep Learning-based Models for Oral Disease Detection

Abstract

Background: In contemporary dentistry, oral object detection is essential for a variety of uses, including automated dental caries diagnosis and orthodontic treatment planning. To find the best strategy for dental image analysis, this study provides a comprehensive comparison of several oral object detection techniques. Method: Three cutting-edge deep learning models, including You Only Look Once or YOLO (especially YOLO V8 and YOLO-NAS), Detection Transformer or DETR, and Detectron2 were implemented, and their performances were compared and contrasted to select the most effective model for dental radiograph image datasets. An opened dataset of 936 oral X-ray images along with the expert annotations were applied in the experiment and the model performances were evaluated in terms of precision, recall, and F1-score. Results: Experimental findings demonstrated that Detectron2 outperforms both YOLO and DETR in terms of detection accuracy, achieving an accuracy of 0.97 and a total loss of 0.3716 while maintaining real-time inference capabilities. Furthermore, a mobile application was also developed to port the model into it and deploy it with Android Studio.