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

Kun Tang; Xinquan Yang; Mianjie Zheng; Xuefen Liu; Xuguang Li; Xiaoqi Guo; Ruihan Chen; Linlin Shen

cs.CVMar 30, 2026

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

Authors: Kun Tang, Xinquan Yang, Mianjie Zheng, Xuefen Liu, Xuguang Li, Xiaoqi Guo, Ruihan Chen, Linlin Shen, +1 more

Organizations: Shenzhen University. · Shenzhen University General Hospital. · Chongqing University.

Abstract

The scarcity and high cost of expert annotations in dental imaging present a significant challenge for the development of AI in dentistry. DINOv3, a state-of-the-art, self-supervised vision foundation model pre-trained on 1.7 billion images, offers a promising pathway to mitigate this issue. However, its reliability when transferred to the dental domain, with its unique imaging characteristics and clinical subtleties, remains unclear. To address this, we introduce DinoDental, a unified benchmark designed to systematically evaluate whether DINOv3 can serve as a reliable, off-the-shelf encoder for comprehensive dental image analysis without requiring domain-specific pre-training. Constructed from multiple public datasets, DinoDental covers a wide range of tasks, including classification, detection, and instance segmentation on both panoramic radiographs and intraoral photographs. We further analyze the model's transfer performance by scaling its size and input resolution, and by comparing different adaptation strategies, including frozen features, full fine-tuning, and the parameter-efficient Low-Rank Adaptation (LoRA) method. Our experiments show that DINOv3 can serve as a strong unified encoder for dental image analysis across both panoramic radiographs and intraoral photographs, remaining competitive across tasks while showing particularly clear advantages for intraoral image understanding and boundary-sensitive dense prediction. Collectively, DinoDental provides a systematic framework for comprehensively evaluating DINOv3 in dental analysis, establishing a foundational benchmark to guide efficient and effective model selection and adaptation for the dental AI community.

Figures & tables

Dataset	Modality	Data Scale	Description
Object Detection & Instance Segmentation
OralXrays-9 [ 7 ]	Panoramic X-ray	12,688	Multi-object anomaly detection over 9 categories (84,113 instances)
DENTEX [ 23 ]	Panoramic X-ray	1,005	Abnormal-tooth detection with dental enumeration and 4 diagnosis classes
ADCD [ 47 ]	Panoramic X-ray	1,628	Detection of diverse dental conditions
CariesXrays [ 8 ]	Panoramic X-ray	6,000	Dental caries detection (13,783 annotated instances)
AlphaDent [ 57 ]	intraoral Photo	1,320	Pathology instance segmentation over 9 classes

Table 1: Overview of the public datasets included in DinoDental. The benchmark covers panoramic radiographs and intraoral photographs across multiple dental imaging tasks, including classification, detection and instance segmentation, and semantic segmentation, and summarizes the modality, data scale, and task description of each dataset.

Figure 1: Qualitative detection comparisons on the DENTEX dataset.

Figure 2: Qualitative detection comparisons on the AlphaDent dataset.

Figure 3: Qualitative segmentation comparison on the Multi-center Panoramic dataset. Red masks indicate impacted-tooth regions.

Supervised methods
Backbone	DentalAI (Multi-label)				AlphaDent (Multi-label)
	mAP	Prec.	Rec.	F1	mAP	Prec.	Rec.	F1
ResNet-50 [ 27 ]	79.82	69.93	78.89	72.60	64.88	48.32	71.52	56.84
Swin-B [ 42 ]	81.11	67.42	83.38	74.10	64.09	43.73	71.58	52.91
SwinV2-B [ 41 ]	78.78	67.49	83.24	73.83	58.09	38.22	71.62	47.23
ConvNeXt-B [ 43 ]	80.97	64.92	88.03	73.17	58.96	43.63	69.75	52.61

Table 6: Performance comparison of supervised and self-supervised methods on intraoral multi-label classification tasks.

Figure 4: Confusion matrices of the multi-label classification on OralXrays-9 dataset.

Figure 5: Visualization of the different resolution sensitivity of different DINOv3 sizes on panoramic radiographs.

Figure 6: Visualization of the different resolution sensitivity of different DINOv3 sizes on intraoral photographs.

Backbone	Input	DentalAI				AlphaDent
		$AP^{b}$	$AP^{b}_{50}$	$AP^{m}$	$AP^{m}_{50}$	$AP^{b}$	$AP^{b}_{50}$	$AP^{m}$	$AP^{m}_{50}$
DINOv3-S [ 56 ]	$512\times 512$	25.50	42.10	25.40	40.60	17.90	32.80	17.20	30.50
	$768\times 768$	28.60	45.20	28.90	44.40	21.50	38.70	20.90	36.00
	$1024\times 1024$	29.20	49.30	30.30	46.60	23.50	43.80	23.30	42.40
	$1280\times 1280$	28.40	47.20	29.80	46.50	22.10	39.20	22.20	37.60
DINOv3-B [ 56 ]	$512\times 512$	28.20	46.10	27.90	44.30	19.90	35.50	18.70	30.90

Table 8: Performance comparison of different DINOv3 models and input resolutions on intraoral detection and instance segmentation datasets.

Backbone	Input	Periodontitis				Impacted
		mIoU	mDice	Prec.	Rec.	mIoU	mDice	Prec.	Rec.
DINOv3-S [ 56 ]	$512\times 512$	90.84	95.05	94.30	95.83	76.36	84.63	83.78	85.53
	$768\times 768$	91.10	95.20	94.80	95.61	77.26	85.38	84.73	86.06
	$1024\times 1024$	95.81	97.83	98.03	97.63	86.66	92.33	93.01	91.67
	$1280\times 1280$	95.27	97.54	97.72	97.36	85.30	91.42	92.85	90.08
DINOv3-B [ 56 ]	$512\times 512$	90.28	94.72	93.95	95.53	75.93	84.26	84.58	83.95

Table 9: Performance comparison of different DINOv3 models and input resolutions on the semantic segmentation task.

Backbone	Tuning	DentalAI				AlphaDent
		$AP^{b}$	$AP^{b}_{50}$	$AP^{m}$	$AP^{m}_{50}$	$AP^{b}$	$AP^{b}_{50}$	$AP^{m}$	$AP^{m}_{50}$
DINOv3-S [ 56 ]	Frozen	28.40	46.80	29.60	46.20	21.50	40.50	21.60	38.00
	Finetune	33.90	53.70	33.60	51.50	25.00	45.70	24.40	43.00
	LoRA	33.00	52.30	33.10	50.70	25.40	44.60	24.20	41.70
DINOv3-B [ 56 ]	Frozen	28.90	48.50	30.00	46.90	21.90	40.60	22.20	37.90
	Finetune	35.70	55.30	34.50	51.90	25.50	43.80	24.60	42.60

Table 12: Detection and instance segmentation results with Frozen / Fine-tuning / LoRA on two intraoral datasets.

Backbone	Tuning	DentalAI (Multi-label)				AlphaDent (Multi-label)
		mAP	Prec.	Rec.	F1	mAP	Prec.	Rec.	F1
DINOv3-S [ 56 ]	Frozen	76.02	61.53	84.49	70.14	58.65	49.46	71.46	52.61
	Finetune	82.52	72.16	83.10	76.99	63.18	45.48	70.54	54.21
	LoRA	84.34	76.39	82.30	79.16	68.19	51.42	70.82	59.15
DINOv3-B [ 56 ]	Frozen	74.16	60.99	83.32	69.43	58.98	43.05	70.44	51.98
	Finetune	84.49	71.15	86.78	77.68	69.65	54.51	65.59	59.18

Table 15: Multi-label classification results with Frozen / Fine-tuning / LoRA on two intraoral datasets.

Explore similar work

Innovative Tooth Segmentation Using Hierarchical Features and Bidirectional Sequence Modeling
Feb 25, 2026Xinxin Zhao, Jian Jiang, Yan Tian +5Medical Image Segmentation
School of Computer Science and Technology, Zhejiang Gongshang University, 310018, Hangzhou, China · School of Computer and Cyber Sciences, Communication University of China, 100024, Beijing, China · Center of Big Data, China Digital Culture Group Co., Ltd, 100176, Beijing, China +6
Large AI Models in Dental Healthcare: From General-Purpose Systems to Domain-Specific Foundation Models
Jun 1, 2026Sema Helali, Lina Abu Nada, Sausan Al Kawas +3Medical Imaging Foundation Models Medical VLMs
Department of Computer Science and Software Engineering, College of Information Technology, United Arab Emirates University, Al Ain, 15551, UAE · Department of Oral and Craniofacial Health Sciences, College of Dental Medicine, University of Sharjah, Sharjah, UAE · Weill Cornell Medicine-Qatar, Doha, Education City, Qatar +1
MetaDent: Labeling Clinical Images for Vision-Language Models in Dentistry
Apr 16, 2026Meng-Xun Li, Wen-Hui Deng, Zhi-Xing Wu +6VLM Evaluation Medical Image Analysis
Department of Prosthodontics, State Key Laboratory of Oral & Maxillofacial Reconstruction and Regeneration, Key Laboratory of Oral Biomedicine, Ministry of Education, Hubei Key Laboratory of Stomatology, School and Hospital of Stomatology, Wuhan University, Wuhan, Hubei, China · School of Artificial Intelligence, Wuhan University, Wuhan, Hubei, China · School of Computer Science, Wuhan University, Wuhan, Hubei, China +2

Feb 25, 2026J/K move · Enter open · S save

Innovative Tooth Segmentation Using Hierarchical Features and Bidirectional Sequence Modeling

Xinxin Zhao, Jian Jiang, Yan Tian, Liqin Wu+4

School of Computer Science and Technology, Zhejiang Gongshang University, 310018, Hangzhou, China · School of Computer and Cyber Sciences, Communication University of China, 100024, Beijing, China · Center of Big Data, China Digital Culture Group Co., Ltd, 100176, Beijing, China+6

Medical Image Segmentation

Open paper AskPDF arXiv

Tooth image segmentation is a cornerstone of dental digitization. However, traditional image encoders relying on fixed-resolution feature maps often lead to discontinuous segmentation and poor discrimination between target regions and background, due to insufficient modeling of environmental and global context. Moreover, transformer-based self-attention introduces substantial computational overhead because of its quadratic complexity (O(n^2)), making it inefficient for high-resolution dental images. To address these challenges, we introduce a three-stage encoder with hierarchical feature representation to capture scale-adaptive information in dental images. By jointly leveraging low-level details and high-level semantics through cross-scale feature fusion, the model effectively preserves fine structural information while maintaining strong contextual awareness. Furthermore, a bidirectional sequence modeling strategy is incorporated to enhance global spatial context understanding without incurring high computational cost. We validate our method on two dental datasets, with experimental results demonstrating its superiority over existing approaches. On the OralVision dataset, our model achieves a 1.1% improvement in mean intersection over union (mIoU).

Abstract

Explore similar work

Innovative Tooth Segmentation Using Hierarchical Features and Bidirectional Sequence Modeling

Large AI Models in Dental Healthcare: From General-Purpose Systems to Domain-Specific Foundation Models

MetaDent: Labeling Clinical Images for Vision-Language Models in Dentistry