CGD-Det: A Lightweight C2f-ghost-dynamic Detector for Pediatric Dental Caries Detection and ICDAS Grading from Smartphone Images

Pediatric dental caries is a widespread global health issue where early detection is crucial for preventing severe complications and enabling minimally invasive care. Traditional diagnostics rely on specialized equipment and expertise, limiting accessibility, especially in underserved areas. This study introduces a lightweight AI system using deep learning and computer vision to facilitate preliminary caries screening via smartphone images, empowering parents and caregivers to perform timely assessments. We developed CGD-Det, an optimized YOLOv8 model integrating Ghost and Dynamic modules into the C2f structure to improve feature extraction while reducing computational costs. The model was trained on 6260 high-resolution RGB images (5634 for training, 626 for validation), derived from 1241 clinical and community photographs augmented via flipping, rotation, brightness/contrast adjustments, and noise addition. Annotations followed ICDAS criteria across seven severity grades (0–6), validated by expert consensus. CGD-Det achieved a mean Average Precision (mAP@0.5) of 97.60%, recall of 93.30%, and precision of 97.00%, using only 2.1 million parameters. It outperformed baseline YOLOv8 (mAP: 97.50%) with fewer parameters (3.0M to 2.1M) and lower computational load (8.1G to 5.8G GFLOPs), demonstrating superior efficiency over models like Faster R-CNN and enabling real-time use on mobile devices. This study presents an efficient AI tool that democratizes early caries detection, improving accessibility for parental screening and potentially alleviating the global burden of pediatric oral disease. By combining high accuracy with minimal resource requirements, the system supports broader AI applications in public health, with future potential for mobile integration and multi-center validation. • AI-powered detection of pediatric caries with ICDAS classification support. • Employs smartphone imagery for user-friendly and accessible screening. • Enhances YOLOv8 with Ghost and Dynamic modules for optimized performance. • Delivers 97.60% mAP@0.5 with reduced computational resource demands.