Artificial intelligence for dental caries detection: A mixup, fine-tuning, and quantization approach on the MobileNetV2 model

Abstract Context: Dental caries is among the most common oral diseases worldwide and often goes undetected in its early stages, particularly in underserved regions with limited access to health care. There is an urgent need for early detection tools that are portable, accessible, and powered by artificial intelligence (AI). Aims: This study aims to develop an efficient and accurate AI-based system for early caries detection, using deep learning alongside optimization strategies suitable for deployment on mobile devices in real-world settings. Settings and Design: A lightweight caries classification system was developed based on the MobileNetV2 architecture, optimized using mixup augmentation, fine-tuning, and quantization-aware training to improve performance and efficiency. Materials and Methods: The primary dataset utilized in this study comprises 500 dental images acquired using a smartphone, while a secondary dataset of 5000 images was sourced from publicly available repositories to enhance model generalization. To further augment the dataset and improve the robustness of the AI-based dental caries detection system, an additional 25,000 images were generated through a series of data augmentation techniques. These included random rotations (±20°), brightness adjustments (ranging from 0.9 to 1.1), zoom transformations (±10%), positional translations, and horizontal flipping. As a result, the final training dataset consisted of 25,500 images. Model performance was assessed using standard evaluation metrics, namely accuracy, precision, recall, and F1-score. Furthermore, model interpretability was examined through Gradient-weighted Class Activation Mapping (Grad-CAM), providing insights into the regions of interest that contributed most significantly to the model’s predictions. Statistical Analysis Used: Comparative statistical analysis was conducted to assess improvements over the baseline model using standard classification metrics. Results: The optimized model achieved 96.0% accuracy and F1-score, with a 72.68% reduction in model size and a 98.28% increase in inference speed. Grad-CAM confirmed anatomically relevant focus. Conclusions: The proposed system provides a practical, accurate, and offline AI solution for caries screening in low-resource settings.