A Comparative Study of YOLOv5 and YOLOv8 for Automatic Detection of Dental Lesions in Panoramic X-rays

Abstract

This study addresses automatic detection and localization of dental lesions in radiographic images. We systematically compare YOLO-family detectors (YOLOv5/YOLOv8) using the public Kaggle dataset “Teeth Segmentation on Dental X-ray Images” (panoramic & periapical X-rays; 598 images with pixel-level masks converted to axis-aligned bounding boxes) under a unified pipeline. Models are trained and evaluated with identical protocols; we analyze mean average precision (mAP@0.5/0.5:0.95), precision, recall, and inference efficiency (FPS/latency), revealing architecture-specific trade-offs between accuracy and throughput. The results provide practical guidance for model selection in AI-assisted dental diagnosis and establish a reproducible baseline for future multimodal detection integrating 3D CBCT.