Abstract
Early detection of dental caries in panoramic radiographs remains challenging due to subtle radiographic features and complex anatomical structures. This study develops a Transformer-based intelligent detection model specifically optimized for identifying early-stage carious lesions in panoramic dental images. The proposed architecture integrates enhanced multi-scale feature fusion mechanisms, spatially-aware attention optimization, and improved two-dimensional positional encoding to capture global contextual relationships while maintaining fine-grained feature discrimination. A comprehensive dataset comprising 3,856 panoramic radiographs with 12,847 annotated carious lesions across severity grades (D1-D4) was constructed for model development and validation. The model achieved 87.3% mean average precision (mAP) across all caries stages, with notable sensitivity of 81.3% for D1 lesions and 84.7% for D2 lesions, surpassing conventional CNN-based approaches and average dentist performance. The system processes images in real-time (70 milliseconds per radiograph). This research demonstrates the efficacy of domain-adapted Transformer architectures for early dental caries detection and establishes its potential utility as a decision support tool for enhancing diagnostic accuracy and screening efficiency in dental practice.