Multi-Modal Artificial Intelligence for Early Caries Detection: Integrating Radiographs and Intraoral Images to Enhance Diagnostic Accuracy and Public Health Impact

Early detection of dental caries remains a persistent challenge in preventive dentistry, particularly in low-resource settings where delayed diagnoses often lead to advanced interventions and higher treatment costs. In this study, we developed and evaluated a multi-modal artificial intelligence (AI) framework that integrates radiographs and intraoral photographs to improve diagnostic accuracy for early caries detection. Our dataset comprised 6,700 annotated images drawn from institutional dental archives and publicly available sources. Separate deep convolutional neural network (CNN) encoders (ResNet-50 for radiographs and EfficientNet-B4 for intraoral photos) were trained, and features were fused using both late-fusion and attention-based strategies. Experimental results demonstrated that single-modality models achieved moderate accuracy (radiograph-only: 86.7%, intraoral-only: 84.9%), whereas the multi-modal attention fusion model significantly outperformed them, achieving 94.6% accuracy, 95.9% sensitivity, 93.1% specificity, and an AUC-ROC of 0.97 (p < 0.01). These improvements not only enhanced early caries detection but also carried substantial clinical and public health implications, enabling cost-effective preventive interventions. The integration of AI-assisted diagnostics into primary care, rural mobile units, and tele-dentistry platforms offers an accessible and economically sustainable solution to global oral health disparities.