IntraDG-MoEFormer: Multi-Level Domain Adaptation with Prompts, Adapters, and Sparse Experts for Intraoral Imaging
Year: 2026
Type: conference-paper
Abstract: Intraoral disease recognition models often degrade when deployed outside the training domain due to acquisition variability across clinics, devices, lighting conditions, and backgrounds. To address this domain shift, we present IntraDG-MoEFormer, a domain-adaptive vision transformer that integrates Domain-Prompt Tokens (DPTs) to condition representation learning on dataset-specific context, Style-Invariant Adapters to suppress appearance-driven bias, and a Sparse Mixture-of-Experts (SMoE) classification head that promotes expert specialization via top-k routing. We evaluate the method on three publicly available intraoral image datasets (OI, OC, and MOD) under both in-domain and cross-domain protocols. IntraDG-MoEFormer achieves strong indomain performance, reaching 97.4%, 96.7%, and 94.5% accuracy on OI, OC, and MOD, respectively. More importantly, it maintains high robustness under dataset transfer, achieving 94.2% F1 on the challenging OI→MOD setting, indicating substantially improved generalization compared to conventional CNN and transformer baselines. When calibration is reported, IntraDG-MoEFormer also produces reliable confidence estimates with low ECE (~0.021), supporting safer decision-making in real clinical workflows. These results suggest that prompt-conditioned adaptation and sparse expert routing are effective mechanisms for robust intraoral disease classification across domains.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204459318