List of Papers Browse by Subject Areas Author List
Abstract
Intraoral 3D scans are fundamental in digital dentistry but
often contain geometric artifacts such as holes and non-manifold structures due to acquisition constraints and complex dental anatomy. We introduce SS-IoSR, a self-supervised hybrid framework for intraoral scan repair that combines masked autoencoder-based geometric representation learning with a hybrid explicit–implicit reconstruction strategy. The model operates on localized point cloud patches guided by clinicianselected seed points, enabling anatomically consistent reconstruction of missing regions. A differentiable Poisson-based refinement module further improves surface continuity and geometric fidelity. Experimental evaluations on public and in-house intraoral scan datasets show that SS-IoSR consistently outperforms state-of-the-art completion methods across multiple metrics, demonstrating improved reconstruction accuracy
and higher structural similarity for clinically relevant dental regions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5293_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/crns-smartvision/iosr
Link to the Dataset(s)
BibTex
@InProceedings{FarMan_SSIoSR_MICCAI2026,
author = { Farhat, Manel AND Ben-Hamadou, Achraf AND Rekik, Ahmed AND Abida, Ons AND Smaoui, Oussama},
title = { { SS-IoSR: Self-supervised Intraoral Scans Repair } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a self-supervised point cloud completion method for intraoral 3D scans. The approach consists of two stages: in the first stage, a Masked Autoencoder (MAE) is pre-trained by masking parts of the point cloud and reconstructing them; in the second stage, the pre-trained network is further utilized to refine the reconstruction of input point clouds. Experimental results demonstrate the effectiveness of the proposed method.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The main contribution of this work lies in demonstrating the effectiveness of the proposed framework for intraoral 3D scans.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper has the following limitations:
The overall level of novelty is limited. The adopted framework has already been applied in self-supervised point cloud completion for general scenarios, and the authors do not provide sufficient task-specific improvements, making the novelty unclear. There are several unclear aspects in the method description. For example, in the first stage, it is not specified whether the masked point cloud is fed into the MAE as a whole or processed patch-by-patch. The concept of the masked token in Figure 1 is not clearly explained. Additionally, when the input point cloud contains missing regions, it is unclear how the seed points are selected. The writing quality needs improvement, as the paper contains several issues such as improper paragraph organization and grammatical errors.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper has issues in method design, writing quality, and novelty. Therefore, I recommend rejection.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors’ clarification regarding the novelty of the work has convinced me. Therefore, I have decided to raise my score to Accept.
Review #2
- Please describe the contribution of the paper
The authors propose a two-stage model training method to learn detailed completions of partial intraoral scans. The first stage is fully self-supervised, as they train transformer-based encoders and decoders on uniformly distributed patches of the scan, of which ca. 65% are masked out. The second stage then uses clinically chosen landmarks around which a fixed-radius region is masked out, completed with the model pre-trained in the first stage. The completed point cloud is then refined with an implicit representation to remove noise and to extract a watertight mesh.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The authors show convincing results, and the application is quite relevant in the dental field. I have just very recently taken an intraoral scan myself and can confirm that it can be quite difficult to always get an accurate surface reconstruction of all teeth. Hence, a prior-based completion model could prove beneficial.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
There is a big overlap in the proposed method with the prior work “IoSR: End-to-End Intraoral Scans Repairing”. The authors don’t try to hide this, since they cite this previous work and since the naming is also a clear extension. However, it is not quite obvious to me what the exact innovation is in this work compared to this prior work. If I’m not mistaken, the authors propose to use a PointNet instead of a DGCNN in the pre-training, but I fail to see how this is any more self supervised (as suggested by the added “SS” in front of IoSR) than the previous work. The authors should describe in more detail how exactly they improved upon the prior work instead of describing their method as if it was completely new.
The authors don’t mention any plans to make their code or trained model public. Since the methodology does not seem to be very original, but their results indeed show some improvement and potential for actual clinical implementation, I strongly suggest the authors consider publishing of their code and their trained model.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The authors include a short subsection about implicit models in their related work section. They include DeepSDF, which is quite a famous and relevant work, but if I’m not mistaken, the implicit representation the authors employ is optimized on each case, not a pre-trained shape model like DeepSDF. The other work on implicit representations that the authors mention seems a bit random to me. Additionally, the authors state that “most existing implicit reconstruction approaches are formulated for supervised global shape completion, aiming to recover overall object geometry from partial observations. In contrast, intraoral scan repair targets localized geometric defects on otherwise complete meshes, requiring spatially localized reasoning and tighter integration with explicit surface representations rather than purely global volumetric modeling.” I do think there’s some truth in that, but there are definitely some more recent implicit models, like NPHM, which use localized models in addition to the global model and together with a geometric initialization of the model weights, this allows such models to capture a lot of geometric details. Another highly relevant work would be “An Implicit Parametric Morphable Dental Model” which is not cited by the authors. The authors could argue here that these two works would require manual preprocessing for identifying the individual teeth, but otherwise, I fail to see why their method should perform any better, at least there’s no comparison in the results.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although the results look promising, I cannot confirm their validity without the authors sharing their trained model, and the authors fail to describe their exact contribution compared to the prior work “IoSR: End-to-End Intraoral Scans Repairing”.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors further specified some key differences between their approach and prior work. I encourage them to specify this appropriately in the paper as well. Furthermore, the authors promise to release source code, models, and even the dataset upon acceptance, which makes their contribution considerably stronger.
Review #3
- Please describe the contribution of the paper
This paper presents SS-IoSR, a self-supervised framework for repairing defective 3D intraoral scans. The method first pre-trains a masked autoencoder on point cloud patches, then uses it with an implicit refinement module and differentiable Poisson surface reconstruction to complete missing regions. A sees point was provided by a clinician to indicate the defect location. The method is evaluated on public Teeth3DS and an in-house dataset, showing improvements over existing baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The problem is clinically relevant. 2.The framework design is reasonable and well-motivated. 3.The paper is well-written, and the improvement is considerable.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The semi-automatic setting requires a clinician to select a seed point and a fixed radius. The sensitivity to seed point placement and radius choice is not analyzed. 2.The in-house dataset is used for evaluation but will only be released upon acceptance. Reproducibility is currently not possible. 3.The paper lacks discussion on failure cases or limitations of the method.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant problem with a reasonable hybrid design, and the experimental results show clear improvements over baselines.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers for their constructive feedback and are encouraged that R3 recognizes the clinical relevance and effectiveness of our results. This rebuttal addresses concerns regarding novelty, methodology, and reproducibility. #Novelty & IoSR Comparison (R1, R2): SS-IoSR introduces a novel Self-Supervised paradigm for intraoral scan repair, addressing fundamental limitations in existing literature:
- Data & Learning Novelty: Unlike supervised frameworks relying on paired “clean/defective” scans, our Phase 1 Masked Autoencoder (MAE) learns the dental geometry manifold directly from raw, unlabeled clinical data. Unlike MAE frameworks for general 3D shapes, our method is specifically adapted to intraoral scans, where highly non-uniform densities and complex structures make standard KNN-based patching unreliable. We introduce a density-aware fixed-radius patch strategy to preserve anatomical coherence across irregular surfaces. This eliminates synthetic defect bias and demonstrates that self-supervised priors, when tailored to dental topology, outperform supervised synthetic training.
- Conceptual & Architectural Shift: While IoSR uses a global strategy, SS-IoSR focuses on localized repair to preserve high-frequency details (e.g., incisal edges) that global methods oversmooth. We leverage a Transformer-based MAE for long-range geometric reasoning across patches. Crucially, our ablation studies show that self-supervised methods without implicit geometric modeling are insufficient; thus, we introduce a DPSR-based implicit refinement block (IRB) to recover fine-grained inter-dental details and maintain surface continuity, ensuring clinical fidelity where explicit reconstructions fail.
#Methodological Details (R1, R3):
- Patch-wise Processing: In Stage 1, the point cloud is processed patch-by-patch. We utilize a fixed-radius strategy (3mm) rather than KNN to ensure a consistent geometric scale, critical for dental scans where point density varies significantly based on scanner hardware.
- Masked Tokens: Masked tokens are learnable geometric placeholders. Combined with positional embeddings, they serve as spatial queries that allow the decoder to identify the locations of missing regions and reconstruct their coordinates based on the surrounding anatomical context.
- Seed Point Selection: Seed points are clinician-selected to indicate the centroid of a defect (e.g., the centroid of a hole’s boundary polyline). This human-in-the-loop (HITL) paradigm is the most effective strategy for clinical safety; it ensures repairs are targeted and avoids “hallucinations” common in fully autonomous models. This is vital as real-world defects are typically limited in scale; repairing massive gaps is often unnecessary and lacks patient-specific anatomical reference. Our sensitivity analysis confirms robustness to selection variability up to 1mm.
#Literature & Implicit Models (R2): We thank R2 for the citations (e.g., NPHM, Zhang et al.), which we will include. Unlike DeepSDF, our implicit module performs per-case optimization to preserve clinical fidelity rather than global class-based modeling. Unlike Zhang et al., which requires explicit tooth-level identification and occupancy vectors, SS-IoSR operates on raw, arbitrary surfaces, bypassing the need for prior segmentation or tooth-level labeling.
#Sensitivity, Limitations, & Reproducibility (R1, R2, R3): The 3mm radius was defined with dental experts to balance context and precision. We will add a dedicated section on limitations: SS-IoSR may struggle with extensive structural loss (e.g., missing an entire crown) where local context is insufficient. Future work could integrate global shape priors as a guide. Finally, we officially commit to releasing the source code, models, and our paired dataset upon acceptance. This will establish the first standardized benchmark for dental scan restoration. A professional linguistic revision (R1) will be performed to clarify Stage 1/2 transitions.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
While the paper demonstrates strong applicability and compelling results in the dental field, its technical novelty is currently unclear, as the framework appears to be a direct application of general self-supervised point cloud completion methods. The authors must clarify the specific task-based contributions that distinguish this work. Additionally, several methodological details require clarification. The authors are encouraged to address these technical and presentation concerns in their response.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper introduces a self-supervised framework named SS-IoSR for repairing defective three-dimensional intraoral scans. The method is highly innovative and the experimental results on the public Teeth3DS and internal datasets demonstrate significant improvements over the baseline.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
There are 3 Accepts after rebuttal. Reviewers agreed that the application is clinically relevant and the results are promising. The main initial concerns were novelty, relation to prior IoSR work, and reproducibility. The rebuttal clarified the differences from prior work and the self-supervised/local repair setting, and the authors also committed to releasing code, models and data. Some details should still be made clearer in the final version, but the main concerns seem addressed. I recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Great work