DINOv2 versus Supervised Visual Representations for Tooth-Level Multi- Label Diagnosis Classification on Panoramic Radiographs: A Frozen-Feature Sample-Efficiency Study
Year: 2026
Type: preprint
Abstract: Objectives. To compare frozen DINOv2-small with supervised ImageNet-pretrained visual representations for tooth-level multi-label diagnosis classification on panoramic radiographs, emphasizing performance under constrained annotation budgets. Methods. We analysed 705 fully annotated DENTEX panoramic radiographs using 224×224 tooth-level regions of interest for four diagnosis labels and a radiograph-grouped 65/15/20 train/validation/test split. Frozen DINOv2-small, ImageNet-pretrained ResNet50, and random ResNet50 representations were evaluated with per-label logistic-regression linear probes. Fifty matched few-shot repetitions used annotation budgets of k = 5–100 positive examples per label with matched negatives. A secondary sensitivity analysis added a supervised ImageNet-pretrained ViT-S/16 comparator. Results. In the prespecified full-data comparison, DINOv2 and ImageNet ResNet50 achieved similar macro-F1 (0.6243 vs 0.6181; 95% CI for the difference, -0.057 to 0.068; permutation p = 0.88), while random ResNet50 achieved 0.4635. Across few-shot budgets, DINOv2 exceeded ImageNet ResNet50 in all 50 repetitions (mean paired difference + 0.0488 to + 0.0732). Supervised ViT achieved full-data macro-F1 0.6088; DINOv2 also exceeded it on average at every few-shot budget (mean paired difference + 0.0139 to + 0.0331). Conclusions. DINOv2 showed no statistically robust full-data advantage over ImageNet ResNet50 but provided the strongest frozen representation under constrained annotation budgets. Its few-shot advantage persisted against a supervised Vision Transformer, indicating that the finding was not solely a CNN-versus-transformer effect. Differences in pretraining data, objective, patch size, and training recipe prevent attribution of the advantage specifically to self-supervision.
Source: Research Square
Language: English
Subfield: Oral Surgery
Field: Dentistry
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7204756228