Intraoral dental caries detection based on mask-regularized vision transformer

Accurate detection of dental caries supports early intervention; however, many deep-learning approaches require dense pixel-level annotations for every image, which are costly to obtain and limit scalability. Models trained with limited supervision may also rely on spurious visual cues. We propose a mask-regularized network for dental caries classification based on the DINOv2 Vision Transformer. Sparse lesion masks are used as auxiliary spatial priors during training, and probabilistic mask dropout reduces reliance on mask availability. A dynamic gate then balances mask-derived local features with global visual representations. We evaluated the method on a de-identified clinical dataset of 498 near-infrared intraoral images acquired at West China Hospital, Sichuan University. Original colour NIRI images were represented as single-channel grayscale inputs during preprocessing. Compared with the DINOv2 baseline, the proposed network achieved higher Accuracy (0.9157 vs. 0.8193), Specificity (0.9204 vs. 0.7848), and F1-Score (0.8393 vs. 0.7139), while maintaining Sensitivity (0.9039 vs. 0.9254). Grad-CAM was additionally used as a post-hoc visualization to examine prediction-associated regions. These internal results suggest that mask-regularized training can reduce dependence on dense lesion annotation and may support future development of dental caries screening tools.