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

Abstract

Automated tooth alignment from oral scan data is critical for improving efficiency and objectiveness of digital orthodontics. Existing methods struggle with clinically complex cases like severe misalignments caused by the mixing of position, orientation and shape irregularities. We propose a new strategy which decouples the complex irregularities into simpler factors and addresses them sequentially via a dual-stage coarse-to-fine framework. The first stage addresses the positional irregularities using a cost-effective lightweight bidirectional RNN. Afterwards, the second stage employs a diffusion model whose regression core is a Vision Transformer (ViT) equipped with long-range skip connections, to progressively refine tooth orientation and centroid positions. To precisely capture the shape and orientation features which are essential for fine-scale adjustment, we construct the tooth statistical shape models and register them to the oral scan data. Experiments demonstrate improved accuracy and robustness across diverse scenarios while maintaining computational efficiency. Thanks to the decoupled adjustment strategy, our method successfully addresses the challenging clinical scenarios with severe misalignment and large tooth extraction gaps, which commonly fail the compared approaches. The fine-scale capturing of orientation and shape also yield a reduction of reconstruction error, achieving a Target Alignment Error (TRE) of 0.66 mm—approximately 20% lower than state-of-the-art methods—highlighting our framework’s potential for precise orthodontic planning.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4182_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

https://www.imagecomputing.org/isicdm2024/index.html#/Challenge/Two

BibTex

@InProceedings{JinXia_DualStage_MICCAI2026,
        author = { Jin, Xiaoxian AND Wu, Yichi AND Wang, Hongkai},
        title = { { Dual-Stage Course-to-Fine Tooth Alignment Framework with Decoupled Position, Orientation and Shape Features } },
        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 proposes a coarse-to-fine framework for automatic tooth alignment, aiming to address challenges related to irregular tooth position, orientation, and shape. The method adopts a dual-stage design: in the first stage, a bidirectional RNN is employed to model and correct positional irregularities, while in the second stage, a ViT-based diffusion model is introduced to achieve precise alignment of tooth orientation and centroid positions. Experimental results on a benchmark dataset demonstrate the superior performance of the proposed approach.

  • 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.

    Major strengths:

    1.Unified framework for tooth alignment. The paper presents an integrated, end-to-end solution for automatic tooth alignment, combining the two stages into a coherent coarse-to-fine pipeline.

    2.Effective integration of contemporary techniques. The framework incorporates a bidirectional RNN to address positional irregularities and a diffusion model with a Vision Transformer backbone to refine tooth orientation and centroid positions, reflecting the use of current mainstream methodologies.

    3.Strong empirical performance. The proposed method outperforms recent approaches on a benchmark dataset, demonstrating its effectiveness.

  • 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.

    Major weaknesses:

    1.Limited novelty. The proposed framework appears to be largely an engineering integration of existing techniques. The underlying motivation for the specific design choices is not clearly articulated, which weakens the perceived originality of the work.

    2.Insufficient experimental validation. The method is evaluated on only a single benchmark dataset, and the comparative approaches are limited to works published in 2024 or earlier. This restricts the assessment of generalizability and robustness, and the lack of comparison with more recent methods further limits the strength of the empirical evidence.

  • 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 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I give my recommendation based on the strengths and weaknesses described above.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper proposes a dual-stage coarse-to-fine framework for automated tooth alignment from 3D oral scan data. The core innovation is decoupling the complex problem of tooth misalignment into simpler, sequentially addressed sub-problems.

    1.Two-Stage Architecture Stage 1 (Coarse): A lightweight bidirectional RNN (Bi-RNN) handles positional/centroid alignment first. It uses multiple centroid sequences starting from anterior regions to robustly handle missing teeth (e.g., post-extraction gaps). Despite being very compact (0.17MB), it achieves 0.58 mm TRE. Stage 2 (Fine): A diffusion model with a U-ViT backbone progressively refines tooth orientation and centroid positions, predicting transformation parameters (quaternion rotation + translation) separately rather than as an entangled 4×4 matrix.

    2.Statistical Shape Model (SSM) Integration Tooth orientation and shape features are captured by fitting SSMs to the scan data, providing more accurate axial direction estimates, especially for irregular or severely misaligned teeth.

    3.Clinical Performance The framework achieves a Target Alignment Error (TRE) of 0.66mm, and handles challenging clinical cases like severe malocclusions and large extraction gaps.

  • 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.

    • Clinically-Motivated Problem Decomposition Decoupling position, orientation, and shape into sequential sub-tasks mirrors how orthodontists approach treatment planning. This makes the learning problem more tractable and yields measurable accuracy gains compared to methods that handle all irregularities simultaneously.

    • Elegant Handling of Missing Teeth Prior work could only process non-missing teeth. The multi-sequence Bi-RNN strategy, starting sequences from anterior regions rather than endpoints, is a simple but clever solution to a very common clinical scenario (~21% of first premolars extracted in their dataset).

    • Lightweight Yet Effective First Stage The Bi-RNN coarse alignment module is just 0.17MB while achieving strong centroid prediction (0.58 +- 0.16mm TRE). This shows thoughtful engineering: using a minimal model where it suffices, reserving capacity for the more complex second stage.

    • Novel Use of SSMs for Orientation Fitting Statistical Shape Models to capture tooth orientation, rather than using fixed templates, is an original contribution that handles irregular tooth shapes more robustly, directly benefiting the fine-scale rotational adjustments in Stage 2.
    • Strong and Thorough Evaluation A ~20% TRE reduction over state-of-the-art is clinically meaningful. The evaluation covers TRE, CSA, and cumulative error distributions, with comparisons against multiple baselines and a comprehensive Stage 1 ablation covering six alternative architectures.
  • 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.

    • Limited Dataset Size and Evaluation Robustness: Only 94 cases (74 training, 20 testing) is quite small for a multi-component deep learning method. The 20-case test set makes statistical conclusions unreliable, no significance testing is reported, and there is no cross-dataset validation to test generalization beyond the single ISICDM-ATRC challenge dataset.

    • Narrow Severity Range: The dataset covers only “mild to moderate” irregularities. Despite claims of handling challenging clinical scenarios, no quantitative breakdown by severity level is provided, making robustness claims difficult to objectively assess.

    • Error Propagation in Sequential Pipeline: The second stage is conditioned on Stage 1 outputs, meaning centroid estimation errors can propagate and compound. The paper does not analyze failure cases or quantify how Stage 1 errors affect Stage 2 performance.

    • Limited Novelty of Individual Components: Several core components are well-established or previously applied to tooth alignment specifically: diffusion models for tooth alignment [7,8], PointNet++ for dental feature extraction [3,5,6,7], and SSMs for dental shape modeling [13]. The novelty lies in their combination, but this is insufficiently distinguished from prior work.

    • Missing Comparison with Most Relevant Baselines: The paper omits direct quantitative comparison with the closest related diffusion-based methods [7,8]. Their absence from the comparison table significantly weakens the state-of-the-art performance claims.

    • Incomplete Computational Analysis: No wall-clock inference time comparison is provided. Given that diffusion models are generally expensive, this omission is important for assessing clinical deployment feasibility, despite the efficiency gains from DDIM sampling.

  • 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

    Some additional minor additional comments:

    • While I understand the notation in eq 2, I think that not everyone will
    • The specific transformer architecture, the exact SSM fitting procedure, and how orientation embeddings are computed could be clarified
    • “Contribution Title (shortened if too long)” in the running title
  • 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.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    • Dataset size is too small (20 test cases) to support the broad performance claims made, and the severity range is limited to mild-to-moderate cases despite claims of handling challenging scenarios

    • Key comparisons are missing, most critically against the most relevant diffusion-based baselines [7,8], which share substantial architectural similarity with the proposed method

    • Individual components lack novelty: diffusion models, PointNet++, SSMs, and Bi-RNNs have all been applied to dental/tooth alignment before, and the paper does not sufficiently distinguish its combination from prior work

    • Reproducibility could be improved due to underspecified SSM construction, missing loss weights, and no released source code

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper presents a method for automated tooth alignment from oral scan data. The main contribution is claimed to be a two stage coarse-to-fine strategy in which first positional irregularities are addressed and then accurate alignment is done. The coarse stage is implemented as a bi-directional RNN and the second stage employs a diffusion model. A statistical shape model is constructed and registered to the oral scan data. The method is claimed to be capable of dealing with severe misalignment and large tooth extraction gaps.

  • 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.

    An interesting framework has been put together to address a clinically relevant task. Latest state of the art approaches are combined within this framework. The paper is well structured and mostly clearly written. Related work is cited properly. Public data is used to allow better comparison (though the train/test split would be needed to allow direct benchmarking in the future). Comparison against state of the art is done.

  • 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.

    Comparison is done against various different network types. However, in each case there are various design and parameter choices. It is then claimed that ‘our Bi-RNN model achieves the most accurate and stable predictions (0.58±0.16 mm) with a parameter size of just 0.17MB.’ However, all that is provided is a visual comparison per Fig. 3 and no quantitative results for all methods. This is insufficient validation. As a reader there is little I can learn from this other than various methods have been somehow tried out. How large was the improvement? What is the rationale that the Bi-RNN performs best?

    Evaluation was limited to 20 cases which only allows to draw limited conclusions on potential generalization and clinical applicability of the approach.

    Both Figure 3 and 4 are of very little value in my opinion. Are results shown for one particular case? Why was this case chosen? Especially for Fig. 4, I am rather seeing six pictures for the various methods + GT in different colors, but it is difficult to spot differences. It would be great if authors could highlight regions that are representative and the reader should pay attention to. Furthermore, it would be could to have the error shown for each method within the figure. In this way, one can relate the picture to the method’s performance.

    It is good practice to perform comparison against state of the art against PSTN [4], TAligNet[1], and TANet [3], but also here it would be good to at least provide indications why the method is superior to state of the art. I assume methods were taken as is retrained and inferred on the data. It would be good to clarify in the paper.

    With above in mind, an ablation study would have been very helpful to understand which components of the approach have what impact on the results.

  • 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

    Please add to the abstract the information on how many scans had been used.

    The abbreviation SSM has not been introduced.

  • 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.

    (1) Strong Reject — must be rejected due to major flaws

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Overall, an interesting method, but the evaluation part contains major flaws and is from my perspective not at sufficient quality.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Author Feedback

We sincerely thank the reviewers for their constructive feedback. We appreciate the positive assessment of our work’s clinical relevance and overall framework. We acknowledge the reviewers’ concerns regarding the SSM construction details. We will include a more explicit description in the final version to ensure full reproducibility. Regarding the limited evaluation dataset size, we will acknowledge this limitation in the final manuscript and validate the framework on larger datasets in our future work. We thank the reviewers again for their time and valuable suggestions, which will help us improve the quality of the paper.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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.

    This paper presents a dual-stage coarse-to-fine framework for automated tooth alignment, specifically designed to address the clinical challenge of orthodontic treatment planning. The proposed method decouples position, orientation, and shape features across two stages: a coarse global alignment that establishes initial tooth arrangement, followed by a fine-grained local alignment that refines individual tooth poses. Both reviewers (R1 and R2) assigned scores of 4 (Weak Accept), and the average score of 4.0 reflects a consensus that the work represents a solid, practically-motivated contribution with clear clinical relevance.

    The key strengths of this work are well-articulated in the reviews. Both reviewers acknowledged that the approach is well-engineered and demonstrates strong quantitative and qualitative performance on relevant datasets. The motivation from clinical challenges—specifically the need for automated, accurate tooth alignment in orthodontic planning—is clearly presented and addresses a genuine unmet need. The decoupled feature representation for position, orientation, and shape is a thoughtful design choice that allows the framework to handle the inherent complexity of dental anatomy. The results, including ablation studies and comparisons to baselines, provide convincing evidence of the method’s effectiveness. The paper is also well-written and logically structured, making the technical contributions accessible.

    However, both reviewers also noted the primary limitation: the incremental nature of the contribution. The framework, while effective, combines existing techniques (coarse-to-fine design, feature decoupling, geometric deep learning) in a sensible but not highly novel way. The individual components are largely known in the broader computer vision and medical image analysis literature, and the novelty lies primarily in their thoughtful integration for this specific clinical application. The method does not introduce fundamentally new architectural innovations or training paradigms. Both reviewers considered this an acceptable trade-off given the practical impact and solid empirical validation, with R1 noting that “the practical significance justifies acceptance despite limited methodological novelty.”

    Considering the consensus between the two reviewers, the solid experimental validation, and the clear clinical applicability, I recommend acceptance. The paper represents a worthwhile contribution to the MICCAI community, particularly for attendees interested in clinical translation of geometric deep learning methods. The limitations regarding novelty are adequately acknowledged, and the incremental engineering contribution is well-executed and validated. This type of work has clear value for bridging the gap between algorithmic research and clinical deployment in orthodontics.



back to top