DOVE
Dental Optical Visual Encoder
A Self-Supervised Vision Foundation Model for Oral Photography Understanding
Model will be released upon acceptance of the paper.
Learning from Unlabeled Oral Photographs
Oral photographs contain rich clinical information and hold significant potential for artificial in-telligence applications; however, most models are trained for individual tasks and rely on costly expert annotations. Whether training on unlabeled oral photographs can produce transferable visual representations for heterogeneous vision tasks remains an open question. To address this question, we developed DOVE (Dental Optical Visual Encoder) by domain-adaptive self-supervised pretraining on approximately 530,000 unlabeled oral photographs collected from publicly accessible sources. DOVE was benchmarked against three established baselines across 5 classification datasets and 4 segmentation datasets, covering both frozen backbone evaluation and end-to-end finetuning. DOVE achieved the strongest overall performance across these bench-marks. Its representations generalized well across heterogeneous tasks and diverse oral diseases. DOVE also showed strong few-shot transfer and high label efficiency. With far fewer labeled im-ages, it still matched or surpassed baselines trained with much more supervision. Qualitative re-trieval and local-feature analyses further indicated better organization of image-level and spatial representations. These findings support DOVE as a generalizable and label-efficient backbone for oral vision research.
Large-Scale Oral Photography Corpus
DOVE is pretrained on various unlabeled clinical photos with different semantics.