YOLO-DentSeg: A Lightweight Real-Time Model for Accurate Detection and Segmentation of Oral Diseases in Panoramic Radiographs

Abstract

Panoramic radiography is vital in dentistry, where accurate detection and segmentation of diseased regions aid clinicians in fast, precise diagnosis. However, the current methods struggle with accuracy, speed, feature extraction, and suitability for low-resource devices. To overcome these challenges, this research introduces a unique YOLO-DentSeg model, a lightweight architecture designed for real-time detection and segmentation of oral dental diseases, which is based on an enhanced version of the YOLOv8n-seg framework. First, the C2f(Channel to Feature Map)-Faster structure is introduced in the backbone network, achieving a lightweight design while improving the model accuracy. Next, the BiFPN(Bidirectional Feature Pyramid Network) structure is employed to enhance its multi-scale feature extraction capabilities. Then, the EMCA(Enhanced Efficient Multi-Channel Attention) attention mechanism is introduced to improve the model’s focus on key disease features. Finally, the Powerful-IOU(Intersection over Union) loss function is used to optimize the detection box localization accuracy. Experiments show that YOLO-DentSeg achieves a detection precision (mAP50(Box)) of 87%, segmentation precision (mAP50(Seg)) of 85.5%, and a speed of 90.3 FPS. Compared to YOLOv8n-seg, it achieves superior precise and faster inference times while decreasing the model size, computational load, and parameter count by 44.9%, 17.5%, and 44.5%, respectively. YOLO-DentSeg enables fast, accurate disease detection and segmentation, making it practical for devices with limited computing power and ideal for real-world dental applications.

Topics & Concepts