DinoDental : Benchmarking DINOv3 as a Unified Vision Encoder for Dental Image Analysis

Abstract

ABSTRACT The scarcity and high cost of expert annotation remain major barriers to developing reliable AI systems for dental imaging. DINOv3, a state‐of‐the‐art self‐supervised vision foundation model pretrained on 1.7 billion images, offers a promising way to reduce this dependence on labelled data. However, its transferability to dental imaging has not been systematically evaluated, because dental radiographs and intraoral photographs exhibit distinctive acquisition characteristics, domain shifts and clinically subtle findings. To address this gap, we introduce DinoDental, a unified benchmark for systematically evaluating whether DINOv3 can serve as an off‐the‐shelf encoder for comprehensive dental image analysis without domain‐specific pretraining. Built from multiple public datasets, DinoDental covers classification, object detection, instance segmentation, and semantic segmentation across panoramic radiographs and intraoral photographs. We further analyse transfer performance by varying model size and input resolution, comparing frozen features, full fine‐tuning, and parameter‐efficient Low‐Rank Adaptation (LoRA), and examining LoRA rank sensitivity, computational efficiency and run‐to‐run stability. Our experiments show that DINOv3 is competitive across panoramic and intraoral dental imaging tasks, with particularly strong results in intraoral image understanding and boundary‐sensitive dense prediction, while also exhibiting clear task‐ and modality‐dependent behaviour. Collectively, DinoDental provides a systematic framework for evaluating DINOv3 in dental image analysis and establishes a practical benchmark to guide model selection and adaptation for dental AI research. The code and benchmark configurations are publicly available at https://github.com/tangkunszu/DinoDental .