EnamelNet-TRiX: A Lesion-Aware Dual-Transformer With Cross-Attention for Early and Advanced Enamel Caries Diagnosis

Purpose: Dental caries is one of the most prevalent oral diseases worldwide, and predicting Early Enamel Caries (EEC) and Advanced Enamel Caries (AEC) in intraoral imaging is a significant clinical research challenge. To overcome the challenges, this research presents an EnamelNet-TRiX, an automated diagnosis system for enamel caries based on a lesion-aware dual-transformer framework with across-attention guidance. Methods: The proposed framework is developed using a shallow convolutional lesion-aware module (LAM) that incorporates on the lesions of the enamel combined with a Swin Transformer that zooms in on textures of the lesions, and a Vision Transformer (ViT) that captures the overall global structure to provide contextual global structural guidance followed by a multi-scale feature fusion stage that augments cross-stream attention with concatenation and unification for classification. The model is trained on the Caries-Spectra dataset, a set of internal images comprising 2000 intraoral images with three diagnostic classes (EEC, AEC, and No Enamel Caries [NEC]), which are cross-validated using an 80-20 split on a patient basis for training and testing. Results: The model is externally evaluated on the DentRT-2 dataset that consists of 300 real-world intraoral images with diverse diagnostic conditions. The model achieved accuracy: 99.25%, precision: 98.94%, recall: 99.12%, and F1-score: 99.03%, respectively, for the Caries-Spectra dataset, while confirming 96.33% accuracy on DentRT-2. The simulation results show the solid domain generalisation of caries diagnosis. Conclusion: The proposed lesion-aware dual-transformers with cross-attention, a multiscale fusion mechanism incorporating local, global, and lesion-prior features, and exhaustive internal and external clinical testing on a real-time dataset shed light on EnamelNet-TRiX as a trustworthy framework for the diagnosis of enamel caries.