Abstract

Predicting post-treatment facial profiles from lateral cephalograms is clinically valuable for orthodontic planning and patient communication, yet remains challenging due to the complex interplay between skeletal, dental, and soft-tissue changes. Conventional approaches either predict discrete clinical indices without visual output or apply direct image-to-image translation that fails to preserve anatomical consistency. We propose OrthoFlow, a two-stage decoupled generative framework that separates geometric prediction from image synthesis. In the first stage, a Flow Matching model with a Graph-Transformer hybrid backbone predicts post-treatment landmark displacements conditioned on textual treatment plans. In the second stage, a conditional Flow Matching model synthesizes high-fidelity lateral cephalograms guided by the predicted anatomical contours. Experiments on a large-scale orthodontic dataset demonstrate that OrthoFlow achieves superior geometric accuracy and visual fidelity compared to existing methods, validated by both quantitative metrics and clinical evaluation by orthodontists. Our code and dataset are available at https://github.com/ShanghaiTech-IMPACT/OrthoFlow.



Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2224_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ShanghaiTech-IMPACT/OrthoFlow

Link to the Dataset(s)

https://github.com/ShanghaiTech-IMPACT/OrthoFlow

BibTex

@InProceedings{WanLey_OrthoFlow_MICCAI2026,
        author = { Wang, Leyuan AND Dou, Yulong AND Yang, Jiancheng AND Song, Guangying AND Cui, Zhiming},
        title = { { OrthoFlow: Decoupled Landmark Prediction and Image Synthesis for Orthodontic Treatment Outcome Visualization } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes OrthoFlow, a novel two-stage generative framework for generating post-lateral cephalograms with post-landmarks from pre-lateral cephalograms. The key contribution lies in explicitly decoupling geometric prediction and image synthesis, which addresses the limitations of prior image-to-image approaches that often fail to preserve anatomical consistency. In the first stage, the authors formulate landmark prediction as a flow-matching-based generative process, where post-treatment landmark displacements are progressively generated from noise, conditioned on pre-treatment landmarks and textual treatment plans. This formulation moves beyond conventional regression-based landmark prediction by modeling the transformation trajectory in a continuous space. In the second stage, a conditional flow matching model is introduced to synthesize high-quality cephalograms guided by predicted anatomical contours, enabling structure-preserving image generation.

  • 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.

    This paper addresses a clinically important problem of predicting post treatment facial profiles, which has strong practical relevance for treatment planning and patient communication in orthodontics.

    A key strength lies in reformulating landmark prediction from a conventional regression task into a flow-matching-based generative process, enabling progressive refinement while preserving structural consistency. The proposed two stage framework that decouples geometry and appearance is well designed, as it improves interpretability and stability by clearly separating structural prediction from image synthesis. Furthermore, the integration of graph-based modeling and Transformer architecture allows the model to effectively capture both local and global relationships among landmarks, which is crucial for maintaining anatomical consistency. The use of text treatment plans as conditional input enhances the method with real clinical workflows. Finally, the paper provides both quantitative evaluations and clinical validation by orthodontists, which strengthens the credibility and practical significance of the proposed approach.

  • 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.

    While the paper is generally well structured, several limitations remain.

    1.the overall framework follows a commonly used two stage pipeline combined with flow matching, resulting in somewhat limited methodological novelty. In particular, the approach of first predicting structural representations and then synthesizing images has been widely explored in prior generative modeling works, making the contribution closer to an integration and application of existing techniques rather than a fundamentally new method.

    2.Although comparisons with diffusion based methods are included, the evaluation against more recent conditional diffusion models appears limited, making it difficult to establish the relative advantages of the proposed approach. Furthermore, while the proposed method shows improvements over existing approaches in Table 1, the performance gains compared to strong baselines such as DiT appear relatively modest.

    3.Since the second stage heavily depends on the predicted landmarks from the first stage, the paper does not sufficiently analyze how errors in landmark prediction affect the final image quality. Considering that orthodontic outcomes can be inherently multi-modal, the lack of evaluation on output diversity is another limitation.

    4.The evaluation of image synthesis quality primarily relies on FID and clinical assessment, while reconstruction based metrics such as PSNR or SSIM are not reported. Including such metrics would provide a more comprehensive and objective assessment of the fidelity of generated images from noisy inputs.

    5.Although Fig. 1 illustrates the use of both KNN and static graphs, the paper lacks sufficient implementation details, such as the number of neighbors k, the coordinate space used for graph construction, how the static graph is defined, and whether the graph is fixed or dynamically updated. These design choices can significantly affect model performance, and clearer descriptions would be necessary to ensure reproducibility.

  • 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.

    (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?

    The paper is rated as “Weak Accept” because it addresses a clinically important problem and presents a reasonably well-designed methodology with convincing experimental results. Generating post-treatment cephalograms has clear practical value in orthodontics, and the proposed two-stage framework that decouples landmark prediction and image synthesis is a sensible design that improves performance. The use of flow matching for landmark prediction, the integration of graph and Transformer architectures, and the incorporation of text treatment plans as conditional inputs are meaningful contributions, particularly in terms of aligning the method with real clinical workflows.

    In addition, the inclusion of both quantitative metrics and clinical evaluations by orthodontists strengthens the results’ fidelity.

    However, in terms of methodological novelty, the approach is closer to a combination and extension of existing techniques rather than a fundamentally new framework.

    The performance improvements over strong baselines are relatively modest, and the paper lacks a deeper analysis of key aspects, including errors in landmark prediction, image synthesis, and the diversity of generated outputs.

    Furthermore, some implementation details are insufficiently described, which raises concerns about reproducibility. Overall, while the paper demonstrates practical significance and solid technical quality, its limited novelty and lack of deeper analysis lead to a Weak Accept recommendation.

  • 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.

    I have carefully considered the rebuttal. The authors provided reasonable clarifications regarding the scope, motivation, and limitations of the work. While some concerns remain, I believe the main contribution remains technically sound and clinically relevant. Therefore, I maintain my Weak Accept recommendation.



Review #2

  • Please describe the contribution of the paper

    The paper proposes OrthoFlow, a two-stage decoupled generative framework for predicting post-treatment facial profiles from lateral cephalograms. The framework synthesizes high-fidelity lateral cephalograms preserving anatomical consistency. Experiments on a large-scale orthodontic dataset demonstrate that the method achieves superior geometric accuracy and visual fidelity compared to existing methods.

  • 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 paper proposes a novel framework for predicting post-treatment facial profiles from lateral cephalograms. The two-stage design is well motivated by the decomposition of the orthodontic prediction task into two sub-problems: estimating the movement of anatomical structures and synthesizing the resulting appearance. This insight leads to a rational and effective framework that addresses key limitations of existing approaches. 2.The incorporation of treatment plans as textual conditions is a meaningful contribution. The paper leverages a large language model (LLM) to generate treatment plans from electronic medical records (EMRs) and uses them as conditioning inputs. This design enhances the clinical plausibility of the predictions, as post-treatment outcomes should naturally depend on the specific treatment plan. 3.The proposed dual-path residual graph neural network (GNN) architecture is also novel. The static graph encodes anatomical priors, while the dynamic graph captures cross-tissue correlations, enabling more comprehensive modeling of structural relationships.

  • 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.In Stage II, the contour map plays a crucial role in image generation. However, the paper does not clearly explain how the predicted landmarks from Stage I are rendered into an anatomical contour map, which is a key component of the overall pipeline. 2.As noted in the paper, device-specific domain shift is a fundamental challenge. Although the contour map is used as structural guidance, this issue does not appear to be adequately addressed, since the training objective in Stage II mainly focuses on pixel-level differences. Furthermore, the paper use the pre-treatment cephalogram as one of the condition in stage 2.I think it’s unreasonable if there exists a domain shift between pre- and post-treatment images

  • 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.

    (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?

    The paper proposes a novel framework for predicting post-treatment facial profiles from lateral cephalograms. To the best of my knowledge, it is the first learning-based approach for this task. The overall framework design, along with several of its components, is well-motivated and inspiring, and the results are promising. However, there are several concerns, as outlined in the weaknesses, that I would like the authors to address in their rebuttal.

  • 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.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper proposes OrthoFlow, a two-stage framework for predicting post-orthodontic lateral cephalograms from pre-treatment images and textual treatment plans. The method first predicts post-treatment landmark displacements using a text-conditioned Flow Matching model with a Graph-Transformer backbone, combining anatomical graph priors, dynamic KNN relations, and global attention. The predicted landmarks are then converted into anatomical contours to guide a second Flow Matching model for post-treatment image synthesis. The framework also uses edge consistency loss, BERT-based treatment-plan conditioning, and masked synthesis losses. Experiments on a large orthodontic dataset and clinical evaluation by orthodontists suggest improved geometric accuracy and visual quality 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.Well-motivated decoupling. The two-stage decomposition of orthodontic outcome prediction into geometric landmark prediction and structure-guided image synthesis is intuitive and helps avoid the pixel-level pitfalls of end-to-end image translation on noisy multi-center data.2.Plan-specific prediction via text conditioning. Using BERT-encoded treatment plans with classifier-free guidance to generate extraction vs. non-extraction predictions for the same patient (Fig. 2) is a clinically meaningful capability that goes beyond deterministic baselines.3.Dual-path graph design. Combining a static anatomical graph with a dynamic KNN graph is a reasonable way to inject structural priors into landmark prediction, and the ablation (w/o Graph) shows a measurable effect on soft-tissue accuracy.4.Scale of the dataset and clinical evaluation. The authors assemble a large orthodontic dataset (44,809 cephalograms / 17,622 patients) and conduct a double-blinded evaluation with three orthodontists (ICC 0.82 / 0.79), which is more rigorous than typical image-synthesis evaluations that rely on FID alone.

  • 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.Page 5 contains an unresolved BibTeX reference “Qwen [?]”, indicating the manuscript was not properly compiled before submission

    2.The paper claims OrthoFlow “achieves the best overall performance” on landmark prediction, but Table 1 shows that the vanilla Flow Matching baseline achieves better S-low (1.69 mm vs. 1.74 mm) than the proposed method.

    3.Table 2 reports quantitative comparisons against Pix2PixHD, UGATIT, and cDDPM, but Figure 3 shows qualitative comparisons against Pix2PixHD, CycleGAN, and cDDPM. UGATIT and CycleGAN are not the same method, and this inconsistency makes the visual evaluation unverifiable.

  • 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.

    (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?

    Please refer to weakness.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    After considering the rebuttal, I view this as a borderline paper. Two concerns raised by other reviewers remain unresolved:1.Limited methodological novelty (R1). R1 noted that the framework follows a commonly used two-stage pipeline combined with Flow Matching, and that the contribution is closer to an integration of existing techniques (Graph-Transformer, text conditioning, contour-guided synthesis) than a fundamentally new method. The rebuttal defends novelty primarily by claiming to be “the first learning-based approach for this task,” which is an application-level rather than methodological contribution and does not address R1’s core concern.2.Unresolved Stage II design concern (R2). R2 questioned the use of pre-treatment cephalograms as conditioning in Stage II given the domain shift between pre- and post-treatment images. The rebuttal confirms that a device-level domain shift exists (“acquired on different devices more than 1 year apart”) and uses it to justify omitting PSNR/SSIM, but offers no mitigation. This is a systemic design issue rather than an evaluation choice, and the rebuttal acknowledges the problem without resolving it.That said, the clinical relevance, large-scale dataset (44,809 cephalograms / 17,622 patients), and rigorous double-blinded orthodontist evaluation (ICC 0.82 / 0.79) are genuine strengths. On balance, I consider this a borderline paper and would not object to acceptance.



Author Feedback

We thank the reviewers for acknowledging the clinical importance of this work, the decoupling rationale, the value of plan conditioned prediction, and the rigor of our clinical study.

Q1.Typos and naming consistency (R4) The Qwen citation and remaining BibTeX errors will be corrected. UGATIT is implemented on the CycleGAN backbone and we cited the two together. The label “CycleGAN” in Fig. 3 in fact refers to the UGATIT model reported in Table 2.We will unify the naming as UGATIT in the text, Table 2 and Fig. 3 so the quantitative and qualitative comparisons align.

Q2.Methodological novelty (R1) As R2 notes, OrthoFlow is, to our knowledge, the first learning-based framework for orthodontic outcome visualization on lateral cephalograms. Beyond the two stage decomposition, our contributions include (a) formulating post-treatment landmark prediction as a Flow Matching generative process conditioned on textual treatment plans(Fig. 2), enabling plan-specific outcomes.(b) a new Graph-Transformer hybrid backbone, and (c) contour guided Flow Matching for structure preserving synthesis.We are the first to jointly explore these components for this task.

Q3.Performance against strong baselines (Meta-reviewer, R1, R4) Our gains over DiT and vanilla Flow Matching are not modest. On hard tissue (incisor angle), which is clinically decisive for orthodontic planning, OrthoFlow improves accuracy by 9% to 13% (UT 3.90 vs 4.39 and 4.51, LT 3.74 vs 4.12 and 4.16). On the upper soft tissue, the gain is 3.6% to 8.3%. The only metric on which vanilla Flow Matching edges us is S_low (1.74 vs 1.69 mm, under 3%)(R4), well within the inter-rater variability of cephalometric annotation. Aggregated over all metrics OrthoFlow attains the best overall score and balance.

Q4.Output diversity (R1) Diversity in our setting is driven by the clinical plan rather than random sampling, which mirrors clinical practice. Fig. 2 already demonstrates plan-controlled diversity for the same patient, producing clinically meaningful changes in the lateral profile.

Q5.Domain shift and PSNR or SSIM (R1, R2) Pre- and post-treatment images are acquired on different devices more than 1 year apart, so a device-level domain shift indeed exists(R2). This is exactly why Stage II conditions on the pre-treatment image style and uses the predicted contour as structural guidance, so the synthesized image inherits the input device appearance while expressing the predicted anatomy. For the same reason, PSNR or SSIM against the post-treatment ground truth would conflate prediction error with device-induced appearance shift and would not reflect anatomical fidelity. FID with paired clinical scoring (ICC 0.82 and 0.79, R4) is the more faithful evaluation.

Q6.Implementation details (Meta-reviewer, R1, R2) For the Graph Network, k equals 16, the KNN graph is built in the learned high dimensional feature space and updated per layer, while the static graph encodes fixed anatomical adjacency (Sec. 2.1). For the contour map (Sec. 2.2), predicted landmarks are connected via centripetal Catmull Rom splines for skeletal and soft tissue boundaries and rendered with fixed per tooth templates for the dentition. We will expand both descriptions and release the code.

Q7.Effect of landmark error on image quality (Meta-reviewer, R1) Cephalometric landmark detection is a well-studied problem and reaches clinically reliable accuracy with existing methods. In our pipeline the mean detection error is 0.91 mm, while orthodontic interventions typically induce displacements above 5 mm. The detection error is therefore an order of magnitude smaller than the treatment signal and well below the line width used in contour rendering, so its influence on synthesis is marginal, which is consistent with the orthodontist scores.




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 was evaluated by reviewers with expertise in the field who have mixed enthusiasm for this work. The reviewers raised several questions regarding methodological details and analysis of the method performance that should be addressed in rebuttal.

  • 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.

    The reviewers unanimously agree this work passes the bar for acceptance and presentation at the conference.



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.

    The paper addresses a clinically important problem with clear relevance to the MICCAI community. The work introduces the first large-scale learning-based framework for post-treatment facial profile prediction and demonstrates its feasibility on a substantial real-world dataset with expert evaluation. While some design concerns remain, the combination of strong clinical motivation, validation, and practical potential outweighs these limitations. Overall, I believe the paper provides sufficient clinical and scientific value and merits acceptance at MICCAI.



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.

    The rebuttal effectively addressed the reviewers’ concerns, providing sufficient clarification and additional evidence where requested. Following the rebuttal, all reviewers expressed support for acceptance, indicating a clear consensus that the paper meets the standard for acceptance at MICCAI.



back to top