Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

The precise fusion of Cone Beam Computed Tomography (CBCT) and 3D Intraoral Scans (IOS) is essential for constructing digital dental models in stomatology. Despite its importance, automated alignment remains challenging due to the scarcity and limited accessibility of open-source datasets, discrepancies in cross-source resolution, and the complex topology of dental structures. To address these limitations, we introduce a newly curated multimodal dataset comprising 425 patients, along with the Anatomy-Aware Geometric Transformers (AAGT)—a novel registration framework specifically designed for dental imaging fusion. Unlike conventional models, AAGT incorporates anatomy-aware priors to effectively capture high-frequency anatomical features. By integrating surface topologies and spatial attributes through a topology-enhanced attention mechanism, our framework emphasizes discriminative landmarks and robustly mitigates challenges posed by metal artifacts and resolution inconsistencies. Extensive experiments demonstrate that AAGT achieves state-of-the-art performance, and external validation confirms its effectiveness in preserving distinct dental anatomy, enabling reliable and fully automated clinical workflows. The data and code are publicly available at https://github.com/AI4MyBUPT/DIgital-dental.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0251_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/AI4MyBUPT/DIgital-dental

Link to the Dataset(s)

https://pan.quark.cn/s/fd2dc541f9ef

BibTex

@InProceedings{ZhoHan_Towards_MICCAI2026,
        author = { Zhou, Hanqing AND Ning, Xiner AND Ren, Chunxia AND Su, Hong AND Xiao, Li},
        title = { { Towards the Digital Dental Models: High-Fidelity CBCT-to-IOS Fusion via Anatomy-Aware Geometric Transformers } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16889},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper addresses the problem of registration of an intraoral dental scan (IOS) with a CBCT scan. In computer-aided dental treatment planning, this is an elementary step whose accuracy is crucial. A fully automated, reliable registration is important for the ease of use of the planning software, even in complex cases, and for further automating the planning itself. The presented results demonstrate improvements over existing methods, PARE-Net (Yao et al. , 2024) and GeoTransformer (Qin et al. , 2022).

    From a methodological perspective, the authors reformulate the problem of registering a surface mesh and volumetric data as a two-point cloud registration problem. The CBCT scan undergoes preprocessing, during which they extract the crown surfaces and remove soft tissues. The proposed method builds on the previously published concept of so-called vector neurons (PARE-Net) and point transformers. The authors design a novel architecture called the Anatomy-Aware Geometric Transformer (AAGT) by proposing the concepts of the Equivariant Saliency Weighting Module (ESWM), Anatomy-Aware Embedding (AAEmbedding), and Anatomy-Aware Transformer (AATransformer). Authors present the solution as tailored to dental data, but one can imagine the concepts to be useful for other types of point clouds.

    The authors also plan to publish their dataset of IOS-CBCT scan pairs for training and comparing IOS-CBCT registration methods. Such public datasets are rare, and this is a significant contribution to the scientific community. As well as the planned publication of the code.

  • 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.

    Overall, the scientific approach is very good; the authors identified three other existing point-cloud registration methods (two of them represent SotA methods (GeoTransformer and PARE-Net) and compared the proposed approach to them. An ablation study to demonstrate the benefit of the authors’ model design is also included. Experiments present both quantitative and qualitative results demonstrating superior alignment precision compared to the other three methods.

    The paper is well-structured and clearly presents the motivation for the research. The explanation of the model architecture is also quite clear. Figures help explain the method and also visually compare registration results.

  • 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 authors claim, “We release a challenging multi-modal clinical dataset,” but the dataset description provides no details about the types of challenging cases included. How many such cases are present? At least a figure showing a few different cases from the dataset would help to understand the complexity of the cases. It is also unclear how the ground-truth alignment was obtained. Does the data come from different CBCT (IOS) devices?

    While the ablation study and experiments provide an overall insight into the framework and overall performance, a few additional experiments would be needed to support the paper’s claims.

    • There is no experiment that explicitly shows that the equivariant point cloud representation actually brings a benefit. It is not clear what the relative rotation of IOS scans in 3D space is, and whether these were randomised. Can the method properly handle alignment of the lower and upper jaws, or is a proper rough vertical orientation of the IOS scan important for correctly aligning the upper jaw scan with the upper jaw in CBCT?

    • “General-purpose geometric encoders tend to prioritise low-frequency semantic information, inadvertently smoothing these critical high-frequency anatomical details. ” What exactly do you mean when you talk about high frequencies? What surface details correspond to such frequencies? Visualisation of the weights estimated by the ESWM module may help to support the claims about the importance of high frequencies.

    • Table 2 shows that precision drops when the Anatomy-Aware transformer is used without the ESWM module. The precision is better than the baseline when both AATr and ESWM are used. What about adding just the ESWM module? Isn’t this the crucial part of the design?

    • “Finally, the initial pose is generated via a Hypothesis Proposer and refined by optional ICP refinement. ” What is the benefit of the ICP refinement? Is it included in the results presented or not?

    From a clinical perspective, it would be great to quantify the number of properly aligned and failed cases, e. g. , by defining a maximum allowed alignment error. This would more clearly demonstrate the benefit for clinical practice.

    A detailed analysis of failure cases would allow the authors to identify problematic patients and help steer the scientific community’s focus toward them.

  • 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

    • It would also be interesting to include a comparison with PEAL (Yu et al. , 2023), a model that appears to outperform GeoTransformer.

    • What threshold did you use to generate the dentition surface from the CBCT scan?

    • Why did you choose <0.06 mm for the negligible error and >0.24 mm for the significant misalignment?

    • Spaces are missing between references when citing multiple sources: “[14,19,20]”. Please also fix the missing space in references to sections: “(Sec. 2.2)”.

  • 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?

    Although I have comments on the experiments and the dataset’s unclear content, I consider both the methodological contribution of the article and the publication of the new dataset to be beneficial to the scientific community. Addressing at least some of the comments in the camera-ready version, I believe, would improve the article’s quality.

  • 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.

    Although the authors did not respond to all of my comments, e.g. the missing ablation study to quantify the impact and benefit of individual parts of the model, based on the additional information on the dataset provided, I consider it beneficial to publish the dataset and the method, which can be taken as a baseline for further works.



Review #2

  • Please describe the contribution of the paper

    This paper tackles an important and clinically relevant problem: rigid registration between CBCT-derived dentition and intraoral scans for digital dental model construction. The idea of injecting dental-geometry priors into a point-cloud registration backbone is reasonable, and the reported results on the authors dataset are promising. The paper is also clearly motivated by real clinical challenges such as metal artifacts and cross-modality resolution mismatch.

  • 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 paper addresses automatic CBCT-IOS registration, which is a highly relevant problem for digital dentistry and surgical planning, especially in the presence of modality mismatch and artifacts.

    • The idea of injecting dental/geometric priors into a point-cloud registration framework is sensible, and the experiments suggest it can improve over more generic baselines on the willing-to-release dataset.

    • The dataset, if truly released, would be a major contribution to the community.

  • 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.

    • My main concern is that the paper’s central supervision signal is not described clearly enough. Section 2 introduces ground-truth correspondences and Section 2.4 states that labels are generated by transforming the source using a transformation matrix, but the paper never explains how these ground-truth transforms/correspondences were actually obtained. For CBCT-IOS fusion, producing reliable ground truth is not trivial: manual alignment is labor-intensive and error-prone, and semi-automatic workflows typically require careful initialization, operator correction, and quality control. Because the validity of both training and evaluation depends on these alignments, the paper needs to explicitly describe the annotation pipeline. Was the alignment manual, semi-automatic, or software-assisted? Which tool was used? Was there landmark initialization, RANSAC/ICP refinement, or another registration backend? How many annotators were involved? Was there any inter-operator variability study or quality control protocol? Without this, it is difficult to judge how trustworthy the reported gains are.

    • The paper repeatedly states that code and data “will be made publicly available,” but in the anonymous submission, there is no code, no anonymous repository, no sample data, no split lists, and no model checkpoints. For a paper whose main selling points are a new dataset and a new method, this is a serious weakness. At a minimum, the authors should provide an anonymized repository. In my view, post-acceptance promises are not an adequate substitute for review-time reproducibility.

    • The paper state that RANSAC computational cost is prohibitive for high-dimensional dental point clouds. A 2023 dental registration paper explicitly aligned segmented CBCT and IOS crowns using robust RANSAC followed by ICP [1]. More generally, official Open3D documentation still presents global registration with RANSAC followed by ICP as a standard coarse-to-fine workflow, and describes ICP as a long-standing mainstay in research and industry. Moreover, runtime may depend on the correspondence set and implementation, but the paper provides no timing evidence for this claim. The method itself still uses optional ICP refinement, so the discussion of classical registration methods is somewhat one-sided. I strongly suggest adding a well-implemented coarse global registration + ICP baseline and reporting both accuracy and runtime.

    • Table 1 mentions PIR in the caption, but PIR is not actually reported. Several metrics are labeled as percentages, but the table entries are shown as decimals.

    • Section 2.4 refers to a “bidirectional” weighted BCE loss, but the equation as written does not clearly show the bidirectional formulation and i could not find in the literature any reference regarding a bidirectional BCE.

    • Registration Recall (RR) is the fraction of point-cloud pairs whose rotation and translation errors fall below specified thresholds, but the threshold used is not specified in the paper.

    [1] Kim, Y.-J., Ahn, J.-H., Lim, H.-K., Nguyen, T. P., Jha, N., Kim, A., & Yoon, J. (2023). Novel Procedure for Automatic Registration between Cone-Beam Computed Tomography and Intraoral Scan Data Supported with 3D Segmentation. Bioengineering, 10(11), 1326.https://doi.org/10.3390/bioengineering10111326

  • 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 provide sufficient information for 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.

    (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?

    The paper proposes both a dataset and a method; however, the dataset lacks essential details regarding its collection and creation, which undermines its validity and limits the strength of the method as a well-supported contribution. See weaknesses for details.

  • 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.

    Thanks for the response to my concerns. I lean towards acceptance while hoping that the code and the dataset will be publicly released for real, and that they will be easy to access.



Review #3

  • Please describe the contribution of the paper

    The paper’s main contribution is the proposed Anatomy-Aware Geometric Transformer, a new framework for registering CBCT and IOS data. This is a meaningful contribution because paired multimodal CBCT-IOS datasets in dentistry are scarce, which has limited progress in this area. As a result, the field has largely relied on conventional non-deep-learning methods such as iterative closest point (ICP) for registration. In this context, the paper contributes both a learning-based registration framework and a paired multimodal dataset of 425 patients, which together provide an important step toward advancing data-driven dental registration.

  • 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.

    A major strength of the paper is that it tackles CBCT-IOS registration, which has not been extensively studied, partly due to the lack of suitable paired datasets. The introduction of a registered dataset is, therefore, a meaningful contribution that may facilitate further research in digital dentistry. In addition, the paper proposes an anatomy-aware registration framework, which is well motivated because accurate dental registration should preserve anatomically meaningful correspondence. However, as stated in the weaknesses section, this does not seem to be truly anatomy-aware. Finally, the method demonstrates strong performance relative to existing registration approaches, suggesting that the proposed framework is well-suited to the dental setting.

  • 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) A major weakness is the paper’s framing of the proposed Anatomy-Aware Embedding (AAEmbedding) as encoding “intrinsic dental topology” or “anatomy-aware priors,” when the actual representation appears to rely on generic geometric descriptors: curvature, surface normals, Euclidean distances, and angular relationships. These are standard differential geometric features applicable to any 3D surface, not specifically to dental anatomy. In other words, the method does not encode semantic anatomical structure, such as cusps, gingival margins, or occlusal surfaces. Rather, it captures local shape variations that are useful for registration because dental surfaces exhibit distinctive geometric patterns. The practical contribution, therefore, seems to be better preservation of high-frequency geometric detail, not genuine modeling of dental anatomy. As a result, the term “anatomy-aware” feels overstated, since the mechanism is fundamentally geometry-driven rather than anatomically informed in a semantic sense.

    (2) The paper does not provide enough detail about the dental status of the patients in the dataset. For example, it is unclear how many teeth each patient has, whether the cases mainly involve full dentition, or whether patients with missing teeth are included. This is important because performance may vary substantially across these groups. A model evaluated mostly on complete dentition may not generalize equally well to patients with missing teeth or more complex dental conditions.

    (3) The evaluation does not include comparison with common baselines such as ICP (iterative closest point)-based registration methods. Since ICP and related geometric matching approaches are standard baselines for 3D alignment and shape correspondence, including them would help clarify how much benefit the proposed learning-based method provides over conventional approaches.

    (4) The paper mentions patients with orthodontic brackets or metal artifacts as an external setting, but the description is not sufficiently clear. It is unclear whether these cases were collected from a different source or whether they are part of the same 425 patient cohort. This distinction matters for interpreting the strength of the external evaluation and the claimed robustness of the method.

    (5) The manuscript states that nnU-Net was used to segment the CBCT data, but it does not explain whether segmentation labels were available, how those labels were obtained, or what anatomical structures were annotated. This is important because the quality and source of the segmentation labels directly affect the validity and reproducibility of the pipeline.

    (6) The preprocessing of CBCT and IOS appears inconsistent. The CBCT is segmented, whereas the IOS is not. The paper should justify why the IOS segmentation was not also performed. In particular, the gingiva may act as noise during registration, so it is unclear whether keeping it is beneficial.

  • 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 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.

    (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?

    Weak reject. Although the paper introduces a promising framework and a valuable paired CBCT-IOS dataset, I am not convinced by the current presentation and validation. The methodological claim of being “anatomy-aware” appears overstated, as the formulation seems more geometry-aware than explicitly anatomy-driven. In addition, the paper provides limited evaluation and insufficient detail on the dataset and experimental setup. Taken together, these issues weaken the empirical support for the paper’s claims and lead me to a weak reject recommendation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    The authors have addressed most of the concerns, but the key concern regarding the “anatomy-aware” claim remains unresolved. The empirical results show improved geometric modeling, but they do not directly demonstrate that the model captures anatomy-specific features. While dental anatomy is partly encoded in geometric structures such as cusps, grooves, ridges, and curvature patterns, this only supports a weak form of anatomy awareness. In my view, a strongly anatomy-aware algorithm should be guided by explicit anatomical structures, such as anatomical labels, landmarks, regions, topology, spatial relationships, or biologically meaningful priors. The proposed Anatomy-Aware Transformer appears to be more geometry-aware than anatomy-aware. Therefore, my decision remains borderline. I would not object to acceptance if the authors temper the anatomy-aware claim accordingly.



Author Feedback

Dear Meta-Reviewer and Reviewers, We sincerely thank you for your constructive comments and for recognizing our work’s clinical relevance. To address your concerns, we provide a systematic summary and response to key issues raised by multiple reviewers, based strictly on our existing experimental results and preliminary evaluation data.

1.Dataset Construction and Ground Truth. We appreciate the emphasis on data transparency. We used two independent datasets for training and external validation. The training set (425 orthodontic patients, 850 pairs, 41.9% male, mean age 18.5, range 10-40) was acquired using the DCT PRO Dentofacial CBCT system and 3Shape TRIOS 5 intraoral scanner. Clinical status: 72.1% normal, 26% severe crowding, 2.4% missing teeth, and 0.5% with brackets. The external validation set (90 patients, 180 pairs, 40% male, mean age 25.2, range 15-35) is evenly divided into well-aligned, metal brackets, and severe crowding groups. Our model’s strong performance on metal brackets demonstrates robustness. For ground truth, a dentist and technician performed initial fine manual alignment using Geomagic Studio and Mimics Research. A senior dentist then conducted rigorous quality control; unqualified cases were strictly re-annotated. We will adhere to MICCAI guidelines and publicly release the code and training dataset upon acceptance, along with a detailed dataset description in the repository.

2.Baselines and ICP Analysis. Adopting your valuable suggestions, we supplemented a “global coarse + ICP fine-tuning” baseline analysis using existing evaluation data (previously omitted due to space constraints). This substantiates that traditional ICP struggles with precision and speed on complex dental point clouds, and adding ICP yields marginal improvements. Specifically, evaluated under unified conditions, traditional ICP required 54.725 seconds with an average RMSE of 5.624 mm. Conversely, our method achieved 0.104 mm RMSE in 0.154 seconds. Utilizing our method for coarse registration followed by ICP fine-tuning took 0.610 seconds with a 0.112 mm RMSE. Regarding PEAL, our baseline experimentally outperforms it; additionally, its closed-source repository makes fair reproduction unfeasible.

3.Segmentation and Preprocessing. We detailed the nnU-Net pipeline for CBCT segmentation, pre-trained on 1583 training and 396 validation samples. Annotations used a semi-automatic strategy: 3D Slicer initial inference + precise refinement by two doctors/technicians + senior doctor review. This model achieved a 95.3% Dice score before inference on our dataset. We omitted IOS segmentation because manual annotation is resource-intensive and cannot eliminate bracket artifacts. Crucially, visualizing super-point matching from our existing AAGT framework confirms the network inherently performs anatomy-aware soft segmentation (previously omitted due to space constraints), automatically and precisely concentrating on tooth surfaces, cusps, and gingival margins.

4.Anatomy-Aware Definition, Loss Function, and Others 1) Anatomy-Aware Definition: Anatomical structures like cusps and gingival margins exhibit distinct geometric features (curvature, normal vectors). Existing experiments empirically prove the network actively focuses on these landmarks, demonstrating that geometric features clearly represent anatomical characteristics. 2) Loss Function: The “bidirectional” BCE is essentially a standard weighted BCE loss generating labels via distance thresholds. We will revise this ambiguous term to strictly align with our formula. 3) Others: We corrected formatting inconsistencies. The Registration Recall threshold is 1 mm. Equivariant representation advantages were fundamentally validated in baseline studies. Limited by practical collaboration constraints, our data source is the orthodontics department; thus, we are collecting data for complex implant prosthodontic scenarios (e.g., massive tooth loss, splinted crowns) for future optimization.



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.

    The paper addresses an important and clinically relevant problem, and all reviewers acknowledge the promise of the proposed method as well as the potential value of the paired CBCT-IOS dataset. At the same time, there is substantial reviewer concern about whether the current version provides sufficient detail and validation to fully support its claims, particularly regarding dataset/ground-truth construction, reproducibility, the “anatomy-aware” framing, and the strength/completeness of the experimental comparisons.

    While the paper is promising, rebuttal should specifically clarify the dataset construction and supervision pipeline, since the current manuscript does not sufficiently describe how ground-truth alignment/correspondences were established or what kinds of clinical cases are represented in the dataset. The authors should make clear the dental status distribution of patients, the prevalence of challenging cases, whether multiple devices are involved, and the annotation/quality-control procedure, as these points are central to assessing the reliability and generalizability of the method. Ideally, all the MICCAI’s Reproducible Research guidelines for dataset contribution should be followed (https://conferences.miccai.org/2026/en/PAPER-SUBMISSION-GUIDELINES.html#reproducibleresearch).

  • 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.

    Although some methodological concerns remain unresolved after the rebuttal, the dataset has been further detailed and may be valuable to the community. For this reason, I recommend acceptance.

    Please ensure that the dataset is released on a public website with long-term access and is accessible to the entire community when submitting the camera-ready version.



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.

    This paper proposes an Anatomy-Aware Geometric Transformer for rigid CBCT-to-IOS registration, together with a paired clinical dataset of 425 patients. The reviewers agree that the problem is clinically important, the released data fills a long-standing gap for this modality pair, and the reported performance is competitive against recent baselines. The rebuttal adequately clarified the dataset construction, ground-truth alignment workflow, and comparison against classical registration pipelines. The authors are encouraged to temper the “anatomy-aware” framing in the camera-ready version, as the mechanism relies primarily on generic geometric descriptors, and to follow through on the promised public release of the dataset and code.



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.

    This submission prorposes a anatomy-aware geometric transformer and a paired dataset for CBCT-to-IOS dental registration. R1 liked the scientific approach. R2 liked the clinical relevance. R3 disliked the overstatement of the “anatomy-aware” claims. The rebuttal clarified the main methodological concerns with pipeline and data details, while adding an ICP baseline to further support the superiority of speed and accuracy. New results were not permitted for rebuttal (all should be available at submission deadline, unfair to other submissions). The consensus indicates that this is a valuable work for the digital dentistry community. For all these reasons, and with respect to the other submissions, recommendation is towards Weak Acceptance.



back to top