Abstract

Accurate orthopedic fracture reduction planning is essential for ensuring successful postoperative recovery and improving patient outcomes. However, current methods are challenged by the complex and irregular fracture geometries and the scarcity of annotated training data. To address these challenges, we propose a novel approach that integrates learning-based shape restoration and fracture simulation. A transformer-based model is developed, which utilizes patch-to-patch restoration and recursive fragment registration to iteratively refine fracture reduction poses. To generate diverse and anatomically realistic fractured datasets for model training, we develop a fracture data simulation approach that combines statistical shape modeling with clinically representative fracture patterns, reducing reliance on annotated samples. Tested on extensive clinical data with hipbone and sacrum fractures, the proposed method achieved mean translational and rotational errors of 2.34 mm and 4.54°, respectively, outperforming both template-based and existing learning-based methods. Our approach enhances learning and generalization for automated fracture reduction by connecting synthetic and real-world fracture data.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2742_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2025/supp/2742_supp.zip

Link to the Code Repository

https://github.com/Sutuk/FracFormer

Link to the Dataset(s)

CTPelvic1K dataset: https://github.com/MIRACLE-Center/CTPelvic1K PENGWIN dataset: https://pengwin.grand-challenge.org/

BibTex

@InProceedings{YibSut_SimtoReal_MICCAI2025,
        author = { Yibulayimu, Sutuke and Liu, Yanzhen and Sang, Yudi and Zhu, Gang and Li, Hui and Lu, Hao and Zhao, Chunpeng and Wu, Xinbao and Wang, Yu},
        title = { { Sim-to-Real Transformer-Based Shape Reconstruction for Automated Orthopedic Fracture Reduction Planning } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15962},
        month = {September},
        page = {595 -- 605}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a novel transformed-based network for the estimation of shape reconstruction on fractures related to the hipbone and sacrum. The method outperformed currently available approaches and in the ablation study justified the selection of every proposed component, including novel proposed FAPE (Fragment-aware patch encoding) for extracting and encoding feature patches.

  • 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 major strength of the paper is the method which is robust and works on both the Unilateral and Bilateral hipbone as well as on the Sacrum. The ablation study pointing out the benefits of each component is useful for further studies. The limitations of the proposed method are stated (for instance irregular fractures which are out of the scope of the conducted research).

  • 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 are that the manuscript is hard to follow and the presentation of the method is complex. I am aware that the method has many components, but for instance, the input data is not clear: How are the initial Fractures (marked as fracture input in Fig 1.) represented? If the input is 3D-CT scan, how is the initial estimation of the fractures made?

    What is also missing is the ability to expand the prposed method to the other body parts. Can the authors provide a brief oppinion on the possibilities to expand the proposed methods to shoulder or knee?

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    A link to the source code and samples, as well as a video of the data generation process, has been provided; however, the link to the source code and samples is not working.

  • 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 — could be accepted, dependent on rebuttal

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

    The method is complex, but its complexity and large number of components gives impression that it can be easily tuned to many other problems. Other than the issues mentioned in the weaknesses, I have no concerns.

  • Reviewer confidence

    Not confident (1)

  • [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

    This paper presents an integrated approach for fracture reduction planning and fracture simulation to address the challenges of learning fracture-specific features and the scarcity of training data. Specifically, a fragment-aware patch encoding is designed to make the transformer-based model realize accurate fracture reduction, a recursive refinement strategy is introduced for improving the accuracy further, and a simulation pipeline is developed to generate diverse and realistic synthetic fractures. Extensive experiments demonstrate the superiority of the proposed approach over traditional and learning-based 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. Meaningful themes: Accurate orthopedic fracture reduction planning is important, but there is relatively little research on it.
    2. Beautiful figures: The figures are pleasing to the eye, clearly showing the details of the model and the results of fracture reduction.
    3. Elegant writing: The overall structure of this paper is well-organized, and most of the descriptions are clear.
  • 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. More details or clarification are needed: (1) In Equation (5), The ground truth values of the CD loss function are the same, but the predictions are different, with one prediction representing the patch center C and the other representing the restored shape R. The ground truth corresponding to different predictions should be different. (2) Clarifying the specific dimensions and physical meanings of the patch centers C and restored shape R would be better. Does the center represent coordinates, and does the shape represent rotation and offset? (3) What are encoder ME and decoder MD like?
    2. Concern about sampling: According to the description, “we form a local patch Pi by selecting its k nearest neighbors (kNN) within the its entire fragment”, this process may lose global information and have a negative impact on shape prediction.
    3. Further highlight the contribution of this paper: The overall structure of this paper adopts many modules/strategies from previous studies (e.g., Pointr, DGCNN, SSM), which may cause readers to confuse which modules come from existing ones and which modules are newly proposed.
  • 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    Suggestions for rebuttal:

    1. Can the authors add more details about realizing the fracture reduction planning?
    2. Can the authors analyze the impact of generating each fragment patch by sampling?

    Suggestions for improving this paper in the future (not for rebuttal):

    1. Adding more details about the proposed method and focusing on its own design is beneficial.
    2. Minors: (1) Variable “I” has not been defined. (2) Splitting Table 2 into two tables for presentation would be more concise. (3) “within the its entire fragment” should be “within its entire fragment”.
  • 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 — could be accepted, dependent on rebuttal

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

    The arrangement of this paper is good, with smooth logic and clear writing. The proposed method is relatively simple but effective.

  • 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 work presents a learning-based hip fracture reduction planning framework based on a transformer architecture. One of the key design is the fracture-aware patch encoding. This method is trained on simulated anatomies and shows its generalization ability on real dataset. This work is compared with classic methods and other learning-based methods in the literature, and shows superior performance.

  • 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 key novelty of this work is the proposed fragment-aware patch encoding, which learns the relative information across fragments using the transformer’s key-query-value architecture. It is certainly an interesting idea. This work also proposes the combined label embedding and center embedding, and the morphing and distortion in data synthesis to improve its generalization ability. The overall workflow is designed nicely. Plus, the runtime of 1-2 seconds is certainly appealing compared to the conventional methods which take much longer.

    This paper is well written and of great clarity. The experiment design is comprehensive, which covers both the baseline methods and other learning-based methods in the litearture, together with ablation studies. The presented results are solid.

  • 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.

    This manuscript does not discuss much about the failure cases. Although the performance in general is superior over comparison methods and it is great to have the sim-to-real generalization ability, it is equally important to know the unexpected failure/outlier cases and whether there is a method to mitigate thea failures, for example, due to network overfitting.

    Can the authors discuss the feasibility of applying this method to patient-specific fracture simulation and model fine-tuning? How much time does the training process take?

    This work does not disucss the clinical acceptable metrics/threshold. This work can be stronger if the authors can discuss whether the proposed method has achieved clinically feasible performance, or if not, how much more accurate does it require to be clinically acceptable?

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/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 fragment-aware patch encoding is a novel contribution and is worth sharing with the MICCAI community. This quality of this manuscript is excellent, the experiments are well-designed, the results are solid.

  • 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




Author Feedback

We thank all reviewers for their constructive feedback. Below, we address the main concerns raised:

Response to Reviewer #1

Comment1: The input representation in Fig. 1 is unclear. How are fractures initially estimated from CT scans?

We clarify that our model takes point cloud data derived from fracture segmentation. The preprocessing pipeline is: CT scan (DICOM) → segmentation mask → fragment mesh (STL) → vertex sampling → labeled point cloud. Initial fracture segmentation is performed using previous work [10], which segments individual fragments from pelvic CT scans.

Comment2: The paper lacks discussion on applicability to other anatomies.

Our method can generalize to other rigid bones because: (1) SSMs can model diverse anatomies and be built for different bones; (2) the entire pipeline is point-cloud based and does not rely on modality-specific features like intensity; and (3) the architecture is not dependent on bone-specific traits such as symmetry. We have shown robustness on both symmetric (hipbone) and asymmetric (sacrum) cases.

Response to Reviewer #2

Comment1: Failure cases are not discussed.

We appreciate this point. A key failure mode involves small fragments that lack geometric features. Our fragment-aware encoder improves robustness by reducing their interference with larger fragments, but accurate alignment remains challenging. A potential strategy to address this issue is hierarchical registration—first aligning larger fragments, then using the refined shape as a template to guide the matching of smaller ones.

Comment2: Can the method support patient-specific simulation and fine-tuning? What is the training cost? Technically, Our framework supports patient-specific simulation by conditioning the SSM-based generation on individual anatomy. However, in practice, the lack of pre-fracture anatomy limits its feasibility. Full training takes ~30 hours on an RTX 4070Ti.

Comment3: Is the method clinically acceptable?

Yes. Clinical studies suggest that 5–10 mm translational error is acceptable for fracture reduction (e.g., Smith et al., JBJS 2005). Our method achieves 2.34 mm, well within this range. The ~4.54° rotational error also aligns with typical values in orthopedic literature. Smith et al., “Clinical outcomes of unstable pelvic fractures in skeletally immature patients,” JBJS, 87(11), 2423–2431, 2005.

Response to Reviewer #4

Comment1: Are the same ground truths for C and R appropriate?

Yes. Since Chamfer Distance does not require point-wise correspondence, using the same dense ground-truth point cloud G for both the sparse centers (C) and dense reconstruction (R) provides consistent and stable supervision. This avoids additional variance that could be introduced by separately subsampling targets.

Comment2: Clarify the meaning of C and R.

C are 3D coordinates representing sparse patch centers, and R is the reconstructed dense point cloud around them. Both are in Euclidean 3D space (XYZ). They do not explicitly encode rotation or translation. Instead, we recover fragment poses by aligning the input fragments to R using rigid ICP.

Comment3: Clarify which parts are novel.

(1)A novel patch-to-patch translation approach on point clouds for fracture reduction, where ME and MD follow the transformer design from PoinTr [22]; (2) FAPE is based on DGCNN [16] but includes fragment-aware aggregation and label fusion (ours); (3) the recursive registration loop is our design; (4) the data simulation combines SSMs with real-mode labeling, morphing, and distortion (ours).

Comment4: Concern about the patching strategy and its effect on model performance.

We use an intra-fragment patching strategy to avoid pose-induced noise during feature extraction, helping separate fragment shape from pose. Global context is then captured through inter-patch attention in the transformer. As shown in Table 2, removing FAPE significantly degrades performance, confirming the effectiveness of our patching and encoding design.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A



back to top