acceptodds
Withdrawn from ICLR 2027

Masked Alignment Pretraining from Local to Global for Intraoral Image Representation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Self-supervised pretraining for intraoral photographs remains underexplored because generic objectives overlook the specific anatomy of the dental arch (\eg bilaterally symmetrical and grouped into six sextants). We present a masked alignment pretraining framework for dental-arch-aware representation learning based on intraoral images. Random, single-tooth, and sextant crops are aligned with dense features at corresponding locations in full-mouth images using geometric overlap as supervision. Masked reconstruction of these crops further captures tooth texture and morphology that complement structural cues. FDI notation (the tooth numbering system) and tooth anatomical relations additionally regularize the tooth representation, so that mirror-symmetric or adjacent teeth, or those in the same sextant are closer, while distant teeth are separated. We also introduce YoungDent, a five-view intraoral dataset with over 5K images from more than 1,000 adolescents captured under non-clinical settings, and combine it with public data into a 10K-scale corpus. Compared with generic SSL baselines, our dental-arch-aware pretraining better captures tooth anatomical organization and improves representation quality, showing clear potential as a foundation for downstream tooth segmentation tasks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, dental images, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gzFuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes anatomy-aware pretraining for intraoral photographs by combining crop-to-image overlap alignment, FDI-based tooth-relation regularization, masked reconstruction, and teacher–student consistency. It introduces YoungDent and evaluates localization, feature probing, anatomical relations, and tooth segmentation. The application is relevant, but the methodological advantage over DINOv2 and the clinical significance require stronger evidence.

Critical strengths
  • Useful domain-specific formulation: Tooth-level and sextant-level crops provide a reasonable way to incorporate dental structure into representation learning.
  • Valuable data collection: YoungDent contains 5,377 smartphone photographs from 1,027 adolescents, with five-view coverage and subject-level train/test separation.
  • Promising transfer results: The Swin-based model improves frozen-feature segmentation and transfers to the external Odontify dataset without retraining.
Critical weaknesses
  • The advantage over DINOv2 is not consistently established under the same backbone. In Table A4, the ViT variant improves decoder-based in-domain mIoU only from 0.510 to 0.521 with 10% labels and from 0.574 to 0.578 with full labels. In Table 2, its logistic-probe FDI and tooth-category scores are lower than DINOv2. The larger Swin gains cannot be attributed solely to the proposed objectives.
  • External supervision and evaluation-label quality need clearer treatment. Structural crops and FDI targets rely on SegmentAnyTooth. Appendix C.3 further states that YoungDent segmentation masks are generated using SegmentAnyTooth boxes and SAM-HQ. Without an expert-verified test subset, in-domain results may partly reflect agreement with the annotation pipeline rather than anatomical accuracy.
  • Several evaluations closely follow the training targets. Localization measures the correspondence explicitly optimized during pretraining, while FDI-relation evaluation measures agreement with the imposed anatomical prior. These are useful representation diagnostics but provide limited independent evidence of broader transfer.
  • Clinical relevance remains preliminary. No disease-detection or clinical-workflow task is evaluated. Furthermore, segmentation evaluation merges 32 tooth identities into eight categories, potentially hiding laterality and jaw-identification errors that matter for dental charting.
  • The distinct methodological contribution needs stronger isolation. The framework combines established self-distillation, positional alignment, and masked reconstruction with dental priors. Comparisons against simpler pseudo-label supervision and a dental adaptation of VoCo would help establish the necessity of the full design.
Questions for authors
  1. Can you compare against DINOv2 using identical backbones, initialization, training budgets, resolutions, and structural crops, adding each proposed objective separately?
  2. Were the YoungDent test masks and FDI labels reviewed by dental experts? How accurate are the upstream predictions, and how sensitive is pretraining to their errors?
  3. Does the approach outperform direct tooth-number pseudo-label supervision using the same SegmentAnyTooth outputs?
  4. Can you report 32-FDI-class segmentation and numbering accuracy, rather than only eight-category metrics, with uncertainty across runs?
  5. Does the representation improve an independently annotated disease-related task? Could the symmetry prior suppress clinically meaningful differences between anatomically corresponding teeth?
1 comment

Q1: Detailed comparison with DINOv2.

A1: Thank you for this comment. We acknowledge that the current results do not establish a consistent advantage over DINOv2 under the same backbone.

Q2: Prediction quality of upstream SegmentAnyTooth and verification.

A2: A subset of YoungDent annotations has been reviewed by dental experts. We will report the subset size, review protocol, and the accuracy of the predicted tooth boxes and FDI identities.

Q3: Relation of evluation metrics with training objectives.

A3: We agree that localization and FDI-relation evaluations are closely aligned with the training objectives. These evaluations were designed to examine the properties motivating our framework: correspondence across spatial scales and anatomical organization across subjects. Their alignment with the objectives makes them useful representation diagnostics, but limits their independence as evidence of broader transfer.

Q4: 32-FDI segmentation results and clinical diagnosis applications.

A4: Thank you for these suggestions. In our study, we followed the eight-category reporting convention used in tooth segmentation but agree that it may conceal laterality and jaw identification errors, a key limitation given that your question concerns full 32-tooth FDI segmentation. We also acknowledge that clinical utility remains preliminary, since the current experiments do not include a disease-related task. The symmetry prior encourages similarity between anatomically corresponding teeth without requiring identical representations; however, it could reduce sensitivity to clinically meaningful asymmetries. A disease-related evaluation, including asymmetric cases, would help assess this limitation.

Q5: Comparison with simpler pesudo-lable-guided supervised learning and VoCo.

A5: Thank you for these suggestions. The random-crop experiments in Table 2 provide a partial comparison, but do not fully isolate the contributions of the proposed components. We agree that comparisons with a 2D VoCo-style random-crop baseline and direct tooth-number pseudo-label supervision would help determine the benefits of our alignment objectives and relation objectives beyond the available external guidance.

Reviewer_DXfdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents an approach and data to pretrain a small scale transformer model for the task of dental image understanding on intraoral images. The proposed method trains from scratch on 5K intraoral images obtained through mobile phones, presumably using a cheek retractor. The data itself is the first contribution of the work. On the data, first pseudolabels are obtained on the images for tooth segmentation and enumeration using [Nguyen et al., 2025]. Two new losses are designed on top of the standard DINOv2 style MIM-based pre-training losses based on the pseudo-labels as additional regularization. The first is the symmetry around vertical axis and the second is the neighboring tooth number prior. The method is trained on the proposed dataset combined with two others after some natural images pretraining on different architectures. Results are provided under three main metrics, including cross subject tooth number relation.

Critical strengths
  • The paper presents a dataset of intraoral images with several improvements over previous ones on raw data attributes. The problem of pre-training (continual pre-training) is a valuable task in specialized domain and once the presentation is improved with the experiments, it may provide valuable insights on the dental optical images.
Critical weaknesses
  1. As claimed in the contribution (2), the proposed approach is not self-supervised. Using pseudolabels, particularly when obtained through methods relying on human annotated training, implies annotated supervision. Although self-supervision may not be the most important aspect of the paper, this is an obvious and fairly consequential error. Even more severe flaw is the positioning of the paper as pretraining. After reading the experiments, one realizes there is a strong initialization from DINOv2 or ImageNet22k (losses not clear here).

  2. The framework of training and comparisons are extremely unfair to say the least. Without going into too much details, here is the summary in 3 sentences:

  • Proposed method is trained on top of DINOv2 init with pseudo labels and external heads.
  • DINOv2/SimCLRv2 - the main compared methods (Tab. 2) are never provided either.
  • Final metric in Tab. 2, is highly related to the pseudo labels.
  1. Conventional wisdom says that one should perform a domain adaptive (e.g., LoRA, low rank mlp, etc.) finetuning or continual finetuning with labels on top of existing pre-trained models. The continual pre-training is chosen here as the main approach but it is in fact using pseudo-labels highly related to the final metrics. What happens if one directly learns linear heads for the relevant tasks on DINOv2 with the 3 different data used?

  2. The presented data does not come with annotations. This is a significant limitation as the value of data is significantly lower without the labels. There are also no experiments which actually show the data has additional value over the existing ones. Tab.1 only shows some qualitative measures.

Questions for authors
  1. The symmetry and FDI number prior implies that the method prefers clean data with no missing teeth or asymmetry. Is that the case? What is the percentage of missing tooth or defect tooth in the data?

Also Q3 in weaknesses

1 comment

Q1: Grounding for self-supervised learning and pretraining.

A1: Thank you for highlighting the need to clarify the rationale for self-supervised learning and pretraining in our study, particularly regarding supervision sources. Our representation-learning stage requires no additional manual annotations for the intraoral training corpus: reconstruction and masked–clean consistency use image-derived targets, while geometric alignment uses crop coordinates. However, SegmentAnyTooth provides predicted tooth bounding boxes and FDI identities for structural sampling, negative weighting, and anatomical regularization. In further revision, we will clarify that the framework combines self-supervised image objectives with externally predicted anatomical guidance. Automatically generated targets are compatible with self-supervised learning, as illustrated by DeepCluster [1] and iBOT [2]. Unlike those methods, our framework additionally uses guidance from an externally supervised tooth-numbering model.

Our method constitutes a domain-specific pretraining stage initialized from generic visual checkpoints. Continuing training from an existing checkpoint is consistent with domain-adaptive pretraining, and REMEDIS [3] provides a related medical-imaging example. We will clarify the initialization sources and distinguish this stage from downstream supervised adaptation in the revision.

Reference:

[1] Caron, Mathilde, et al. "Deep clustering for unsupervised learning of visual features." European conference on computer vision. Cham: Springer International Publishing, 2018.

[2] Zhou, Jinghao, et al. "IBoT: Image bert pre-training with online tokenizer." arXiv preprint arXiv:2111.07832 (2021).

[3] Azizi, Shekoofeh, et al. "Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging." Nature Biomedical Engineering 7.6 (2023): 756-779.

Q2: Initialization from general-porpose pretraining model and comparison fairness on evaluation metrics.

A2: Initializing from generic visual checkpoints supports our goal of adapting representations to intraoral photographs. As described in the appendix, DINOv2 and SimCLRv2 are also adapted to the same intraoral training corpus using their respective objectives. We compare their dental representations both qualitatively (Fig. 5, Fig. 7) and quantitatively (Tab. 2, Tab. 3), including their ability to distinguish tooth categories.

Our metrics assess spatial correspondence and anatomical organization, properties that our method also targeted. We therefore interpret them as diagnostic measures of the learned representation rather than independent evidence of general transferability.

Q3: Why not use DINOv2 for direct linear probing and task adapatation?

A3: We used adapted DINOv2 baseline to test whether continuing its original learning objective on intraoral images produces dental representations we need. As shown in Fig. 5, its representations remain less distinct across tooth categories than those of our method. Based on these findings, direct linear probing with frozen DINOv2 features may not provide better results.

Q4: Value of YoungDent Dataset.

A4: YoungDent is intended to complement existing clinical datasets by providing smartphone-captured intraoral photographs collected outside standardized clinical conditions. Its contribution is the acquisition setting and the resulting variation in devices, illumination, and viewpoints. We agree, however, that the absence of comprehensive manual annotations limits its immediate use as a supervised benchmark. However, in future studies, we will develop and release a complete dataset with annotation labels for dental plaque, gingivitis, malocclusion, and caries, which will support the broader objectives of AI-assisted diagnosis for patient-uploaded intraoral photos.

Q5: Missing tooth or defect tooth problem.

A5: YoungDent was collected from an adolescent population (12 to 18 years old) and contains relatively few missing-tooth or defect-tooth cases. We acknowledge that this limits the evidence for generalization to incomplete dentitions and substantial dental abnormalities. However, our formulation does not require a complete or bilaterally symmetric dentition: FDI-based relations are constructed among the teeth detected in each image, so no counterpart needs to be present for every tooth. Missing teeth therefore do not inherently prevent the objective from being computed. Instead, accurate localization and numbering remain important. We will explicitly discuss the limited coverage of these cases as a limitation.

Reviewer_5M56OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a dental-arch-aware self-supervised pretraining framework for intraoral photographs. The method combines bidirectional local-to-global feature alignment using geometric crop–patch overlap, FDI-based anatomical relation regularization, masked reconstruction, and masked–clean feature consistency in a momentum teacher–student architecture. The paper also introduces YoungDent, a five-view dataset containing 5377 intraoral images from 1027 adolescent subjects, and combines it with the public AIDC dataset for pretraining. The method is evaluated through crop localization, tooth and sextant feature probing, cross-subject FDI relation analysis, and downstream tooth segmentation on YoungDent and Odontify. The reported results suggest improvements over several generic self-supervised learning baselines, particularly in localization and frozen-encoder segmentation.

Critical strengths
  1. A clear motivation for the proposed method. The paper considers the anatomical structure of teeth and uses it to guide self-supervised pretraining. In particular, combining local-to-global alignment with FDI-based tooth relations is a reasonable way to make the learned features more suitable for intraoral images.
  2. Experiments on several aspects of representation learning. The paper introduces the YoungDent dataset and evaluates the method using localization, feature probing, tooth relation analysis, and downstream segmentation. The experiments also include an out-of-domain dataset and frozen-encoder settings, which help show how well the learned features transfer to other tasks and data.
Critical weaknesses
  1. The novelty of the method is not very clear. The method combines several existing ideas, including local-to-global alignment, teacher–student learning, masked reconstruction, and anatomical relation constraints. Using FDI tooth relations is reasonable for dental images, but the paper does not compare the method with a closely related position-aware SSL method, such as VoCo. Therefore, it is difficult to tell how much improvement comes from the FDI constraints and how much comes from the other components. It would also be helpful to test whether the results change when different FDI relation scores are used.
  2. The segmentation evaluation needs further clarification. Appendix C.3 states that the YoungDent tooth masks are generated using SAM-HQ with predicted tooth boxes. This raises a concern that the segmentation results may mainly reflect agreement with automatically generated masks, rather than performance on manually annotated ground truth. The authors should clarify how these masks were evaluated and whether the training data of the tooth-numbering or segmentation models overlap with YoungDent. In addition, results from multiple runs would help show whether the reported improvements are consistent.
Questions for authors
  1. Appendix C.3 states that the YoungDent tooth masks are generated using SAM-HQ with predicted tooth boxes. Are the segmentation results evaluated using these automatically generated masks or manually annotated ground truth? If possible, could the authors provide results on a manually checked test set? It would also be helpful to clarify whether the models used to generate the tooth boxes and masks were trained on any data from YoungDent.

  2. The proposed local-to-global alignment is related to position-based self-supervised learning methods such as VoCo. Could the authors include a comparison with VoCo or a similar method under the same training settings? This would help clarify the contribution of the proposed alignment and FDI relation objectives.

  3. Were the main results obtained from multiple runs with different random seeds? Also, would the results change if different FDI relation scores were used? Finally, will the YoungDent dataset be made publicly available, or will the authors provide access to the dataset for future research?

1 comment

Q1: Novelty of proposed method and comparison with VoCo.

A1: Thank you for this comment. VoCo was developed for volumetric medical images, whereas we adapt its crop-overlap-based spatial correspondence learning to 2D intraoral photographs. Our main novelties and contributions are the alignment of anatomically structured single-tooth and sextant crops with full-image patch representations, and FDI-based guidance through relation-aware negative weighting and patch regularization. Teacher–student learning and masked reconstruction are supporting components. The random-crop experiments in Table 2 provide a partial comparison with VoCo, although they do not fully isolate our contributions. We acknowledge that a controlled 2D VoCo-style baseline and an additional analysis of alternative FDI relation scores would help clarify their individual effects. We will amend this comparison study further.

Q2: Reliablity and stability of SegmentAnyTooth.

A2: Thank you for raising this point. The reference masks in the YoungDent segmentation task were generated automatically by SAM-HQ, using tooth bounding boxes predicted by SegmentAnyTooth; they are not manually annotated ground truth. The original training data of SegmentAnyTooth do not overlap with YoungDent. However, dental experts helped check and validate the accuracy using a small YoungDent subset to analyze potential domain mismatch. We will clarify the size and partition of the adaptation subset in the revision.

Q3: Open access for YoungDent Dataset.

A3: We intend to pursue controlled, request-based access to YoungDent Dataset following acceptance, subject to participant consent and applicable institutional approvals. Because the dataset contains sensitive intraoral photographs, access will depend on the scope of the consent obtained and the permitted research uses.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear respectful PCs, SACs, ACs and reviewers,

We thank all reviewers for their constructive feedback. We appreciate reviewers 5M56 and gzFu for recognizing the motivation, dental-arch-based formulation, YoungDent dataset, and breadth of our representation evaluations. We also thank reviewer DXfd for recognizing the value of domain-specific pretraining for dental images. Below, we clarify the main concerns.

Self-supervised learning and domain-specific pretraining (DXfd). Our image alignment, reconstruction, and consistency objectives use self-supervised targets. Structural sampling and FDI regularization additionally use tooth bounding boxes and FDI codes predicted by SegmentAnyTooth, an externally supervised model. Thus, the representation-learning stage therefore requires no additional manual annotations, although the full pipeline includes indirect supervision. Our domain-specific pretraining starts from general-domain pretrained weights, and we compare our method with DINOv2 adapted to intraoral images.

Reliability of SegmentAnyTooth (5M56, gzFu). We agree that the current description provides insufficient evidence of upstream prediction quality. In particular, the reference masks used for YoungDent segmentation evaluation are generated automatically rather than manually annotated ground truth. Dental experts have checked and validated predictions on a subset of YoungDent. We will clarify the review protocol and results, and distinguish clearly between automatically generated references and expert-verified annotations.

Baseline comparisons and evaluation (5M56, DXfd, gzFu). We agree that comparisons with direct pseudo-label supervision and a dental adaptation of VoCo would help isolate the contributions of our objectives. Our localization and FDI-relation metrics are closely related to the training objectives and primarily diagnose the intended representation properties. Downstream tooth segmentation provides complementary evidence of transfer, while the absence of a disease-related task limits the evidence for clinical utility.

We thank the reviewers again for helping us clarify the scope and limitations of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.