Abstract

Accurate preoperative–intraoperative registration is crucial for transferring patient-specific surgical plans to the intraoperative setting in computer- and robot-assisted orthopedic surgery. Ultrasound (US) has gained increasing attention as an intraoperative modality due to its radiation-free and real-time capabilities. However, the inherent imaging constraints, such as bone shadowing and the limited anatomical coverage during surgery, often result in incomplete US-derived bone surfaces, compromising registration accuracy and robustness. Moreover, the scarcity of paired CT–US datasets restricts the applicability of powerful learning-based approaches to different anatomies. To address these challenges, we propose ComReg, an end-to-end framework for CT–US bone surface registration via instance-specific shape completion, jointly optimizing surface completion and cross-modal registration within a unified network. In addition, we introduce a self-supervised registration strategy based on patient-specific US bone surface simulation to alleviate data scarcity, enabling effective learning of shape priors without requiring cross-instance datasets. Experiments on the publicly-available UltraBones100k dataset demonstrate the effectiveness of our approach, achieving performance comparable to state-of-the-art fully supervised methods trained on cross-instance datasets. Overall, our work provides a promising solution for cross-modality bone surface registration in limited-data settings. The code is available at: https://comreg-page.github.io

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3358_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://comreg-page.github.io/

Link to the Dataset(s)

UltraBones100k: https://github.com/luohwu/UltraBones100k

BibTex

@InProceedings{WuLuo_ShapeAware_MICCAI2026,
        author = { Wu, Luohong AND Cho, Elise AND Ao, Yunke AND Marx, Lennard AND Cavalcanti, Nicola A. AND Yang, Yiru AND Seibold, Matthias AND Fürnstahl, Philipp},
        title = { { Shape-Aware Registration Without Pretraining: CT-Ultrasound Bone Surface Alignment via Instance-Level Shape Completion } },
        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

    This paper presents a method for ultrasound-CT registration specific to bone alignment that jointly estimates the bone surface in the ultrasound and the global rigid transformation that aligns the preoperative CT scan and the intraoperative ultrasound scan. The authors build on existing shape completion methods and registration methods (coarse shape regression MLP followed by a variant of ICP for fine-scale refinement). They demonstrate the promise of the proposed method on a publicly available dataset of paired ultrasound-CT images.

  • 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 is well written and the proposed method is explained clearly. This is a well reasoned method for bone alignment that promises to enhance surgical navigation. The method archives better accuracy than RANSCAC-based alignment procedures.

  • 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 comparison to learning-based methods is unconvincing. It seems when learning-based models are available, they outperform the proposed approach. The authors make a point that such models require paired ultrasound-CT data in a large number of subjects for training, and don’t generalize to new bone structures. No empirical evidence for this lack of generalization is offered. At the same time, the authors rely on the shape completion methods that might also fail for more complex bones, such as vertebrae and ribs. So the advantages of the proposed approach are not clearly demonstrated. Moreover, the methodological contribution is quite limited as most of the components of the proposed method have been demonstrated before.

  • 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?

    Well reasoned approach, limited methodological innovation, not entirely convincing comparison to existing learning based methods.

  • 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

    This paper proposes a self-supervised model to register CT volumes onto ultrasound scans. The proposed method focuses on bone surfaces and relies on an interesting data simulation pipeline. The method is evaluated on the public “UltraBones100k” dataset.

  • 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 tackles a difficult and clinically relevant problem.

    • I appreciate the fact that it attempts to design a method that can deliver acceptable results with a limited amount of data.

  • 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 authors claim that their method provides “consistent improvements in registration robustness and accuracy”, this is not supported by their quantitative results (Table 1). Notably, while average registration errors seem acceptable, standard deviations for the proposed method are an order of magnitude larger than those reported for competing baselines. Looking at the angular registration error (RRE), the “Predator” baseline reached an accuracy of 1.43 +- 1.12 degrees, while the authors report 2.93 +- 19.98 (!) degrees for the proposed method. This suggests that far from being robust, the proposed method often experiences catastrophic failure. It cannot be trusted in a clinical environment.

    • Two of the competing baselines, “RANSAC100M” and “RANSAC250M”, are used without being defined. I could find no discussion of these models in reference [6] provided by the authors. In this context, I cannot assess the discussion about “Runtime and clinical feasibility” at the start of page 8.
    • While I understand that ultrasound scans are expensive to annotate, I am not convinced by the discussion about “Population-scale vs Instance-level” training at the end of page 7.The authors present using n=5 patients as a realistic “limited” scenario, and n=14 patients as a “large, population-scale” setting. This is a bit far fetched.
  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    While the paper is relatively original and interesting, the reported angular errors are too large to support any kind of clinical use.

  • 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

    The paper proposes a novel framework for CT-Ultrasound (CT-US) bone surface registration that addresses the challenges of limited paired data and the inherent imaging constraints of B-mode US (e.g., acoustic shadowing, partial coverage). The authors introduce an end-to-end, learning-based approach that jointly optimizes shape completion and registration. The authors rely on a self-supervised, patient-specific synthetic US simulation strategy that mimics clinical US acquisition processes. The model is trained on this simulated data and applied to real intraoperative US data without further fine-tuning at inference time. Experimental results on the UltraBones100k dataset demonstrate performance comparable to state-of-the-art fully supervised, with significantly faster runtime and robustness to lack of data

  • 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.

    S1 Data-Efficient approach : The shift from population-scale learning to a patient-specific, self-supervised strategy is highly relevant for clinical scenarios where paired CT-US data is scarce.

    S2 Joint Optimization: By jointly optimizing shape completion and registration, the framework achieves a more coherent structural reasoning compared to decoupled pipelines.

    S3 Clinical Relevance: The approach directly addresses real-world clinical constraints, such as the need for faster runtime (under 1 second) and the lack of large-scale annotated datasets for specific anatomies.

    S4: Strong Experimental Validation: The ablation studies clearly justify the architectural choices, including the importance of synthetic data realism and the differential refinement module.

  • 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.

    W1 Impact of Segmentation Quality: the method assumes accurate bone surface extraction from CT, and the simulation of segmentation noise in US is simplistic (random false intersections). Additional tests are needed to evaluate the impact of the discrepancy in terms of segmentation errors both in US and CT

    W2 Limited Experiments : The evaluation is restricted to the UltraBones100k dataset (fibula and tibia). It remains to be seen how the method performs on other complex anatomical structures or under more severe intraoperative conditions.

    W3 : missing references. Recent work on differentiable ultrasound simulation should be cited :

    • Fully Differentiable Ultrasound Simulation Utilizing Ray-Tracing by River-spencer et al.
    • DiffUS: Differentiable Ultrasound Rendering from Volumetric Imaging by Bertramo et al.
  • 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
    • Clarification on Inference: Explicitly detail the sensitivity of the model to the initial preoperative bone model segmentation.

    -Refine Discussion on Runtime: While the 1-second runtime is impressive, provide more context on the computational requirements (e.g., GPU memory, inference hardware) to clarify its “real-time” capability in a standard operating room setting.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

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

    The paper provides a joint optimization of shape completion and 3d/2d registration based on synthetic data. The paper is well written with convincing experiments.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    The authors have clarified the high standard deviation observed in their results and explain the annotation bottleneck that limit the size of the test set.



Review #4

  • Please describe the contribution of the paper

    This paper presents an end-to-end framework for CT-US bone surface registration under severe partial observations, where ultrasound-derived bone surfaces are incomplete. The key idea is to combine instance-level shape completion and rigid registration within a unified network, trained using patient-specific synthetic ultrasound generated from preoperative data rather than relying on large scale paired CT-US datasets. The method uses a completion module based on AdaPoinTr and a coarse-to-fine registration module with differentiable ICP, and is evaluated on the UltraBones100k dataset. The main contribution is therefore a practically motivated learning framework for limited-data or single instance level CT-US registration, together with an experimental study comparing the proposed approach against both traditional instance-level registration methods and population-scale learning-based 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.

    The paper addresses a clinically relevant and technically challenging problem: CT-US bone surface registration under incomplete ultrasound observations. This is a meaningful problem setting in computer-assisted orthopedic surgery rather than a benchmark. The overall method is well motivated. Instead of treating shape completion and registration as separate stages, the paper jointly optimizes both within a unified framework. It leads a notable strength is the patient-specific synthetic data simulation strategy. The method does not rely on a conventional large paired CT-US training dataset, which is important in a setting where such data are scarce. The framing against population-scale methods is therefore meaningful. The experimental section is reasonably aligned with the paper’s claim. The authors compare against both traditional instance-level methods and learning-based population-scale methods, and additionally analyze performance under limited-data regimes. This helps support the claimed advantage in data-scarce settings. The runtime is promising from an application perspective. Compared with RANSAC-based baselines, the proposed method is substantially faster while maintaining competitive registration accuracy. The paper’s strength is closer and more CAI than strongly MIC.

  • 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 methodological novelty is moderate rather than strong. The framework mainly integrates existing components, including AdaPoinTr-style completion, MLP-based coarse pose regression, and differentiable ICP refinement. The contribution is more in the problem formulation and system integration than in a fundamentally new learning method. The training part is not communicated clearly enough. Although the text indicates that synthetic observations are used to train the completion and registration modules and that the method relies on patient-specific pre-training, it remains unclear from the main presentation exactly how training is performed for each case, and how this differs operationally from conventional dataset-level learning. This may confuse readers (already made me confusion) because both completion and registration are learned components. The fairness and interpretation of the limited-data comparison in Table 2 could be clarified further. In particular, the paper does not clearly specify whether the population-scale baselines are trained from scratch or fine-tuned under the reduced-data settings, which matters for interpreting the comparison.

  • 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

    A good CAI paper practically relevant and generally well motivated. In particular, the focus on CT-US bone surface registration under partial observations and limited paired data is meaningful for computer-assisted orthopedic surgery. The integration of completion and registration is coherent, and the runtime is promising from an application perspective. Some main suggestions are about clarity and positioning. Firstly, the manuscript should explain the training protocol more explicitly, especially how the completion and registration modules are trained in the patient-specific setting, and how this differs from conventional dataset-level learning. Besides, the limited-data comparison in Table 2 would be easier to interpret if the training setup of the population-scale baselines were described in more detail.

  • 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 assign a weak accept score because the paper prepose a CAI surgery problem in CT-US bone surface registration, and proposes a coherent framework that combines shape completion, registration, and patient-specific synthetic data generation in a practically meaningful way. The experimental results on UltraBones100k are competitive, and the runtime is encouraging for intraoperative applications. However, the training protocol and the interpretation of the limited-data comparison are not explained as clearly as they could be, and the experimental validation remains limited to a single dataset (I can understand it is a CAI problem). For these reasons, I find the paper above the threshold, but only marginally so.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    The authors could give clear response and admit they will refine the content about how they training and inference. It meet my requirement.

    Very good job could clear answer the question from R2.It make me feel very satisfactory.



Author Feedback

Dear reviewers,

Thank you for very positive (R3), positive (R1, R4), and constructive reviews (R2).

Regarding RRE standard deviation (SD) and resulting clinical reliability (R2): R2 identified this as the primary reason for rejection, and we appreciate to provide a more detailed analysis and concrete manuscript additions that we hope fully addresses this concern. The RRE percentiles are 0.30° (5%), 1.47° (25%), 2.55° (50%), 4.06° (75%), and 10.92° (95%), showing low RRE in the majority of runs. An in-depth analysis reveals that the large SD stems primarily from rare flip cases (~1.46% of runs). Due to long-bone symmetry, shape completion can flip to the opposite cortical side or wrong proximal-distal direction, producing flipped registration and ~180° RRE. Evaluation without flip cases reduces RRE SD from ~19° to ~8°. We believe that this is acceptable for pre-clinical work, given the pressing clinical need and the inherent challenges of ultrasound (US) data collection. Flip cases are a detectable failure mode rather than a general registration instability that we can address through shape priors or flip-consistency checks in future work. We will report these observations and statistics, as well as discuss flip cases explicitly as a limitation that can motivate future MICCAI research.

Regarding population-scale learning and the concern that n=14 is not “large” (R2): We acknowledge that the “large” vs. “limited” framing needs more justification and want to emphasize that our argument is about annotation feasibility, not clinical coverage. Annotation remains the bottleneck for robust US-based registration. First, expert annotation of one US slice takes several minutes, making population-level annotation across anatomies and pathologies not only costly but infeasible. Second, UltraBones100k is, to our knowledge, the only approach providing automatic annotations. However, it does not scale to different anatomies, pathologies, or in-vivo settings, as it requires very costly cadaveric experiments (~$30,000 for 14 specimens). Thus, n=14 is not large for population-level learning, but reflects the current practical ceiling of annotation infrastructure. This motivates our patient-specific formulation which adapts to the target bone model using synthetic US, avoiding dependence on unavailable large annotated datasets. We will rename the table settings from “patients” to “annotations”, using full annotations (n=14), limited annotations (n=5), and instance-based and will explain that the annotation bottleneck is a fundamental infrastructural limitation making population-scale supervision currently out of reach.

Regarding segmentation quality, noise modeling, and generalizability (R3/R4): We agree that advanced noise modeling and US simulation could further improve robustness and generalizability, and will describe them as future work in the revision. We will also weaken broad generalization claims. Although our method is anatomically agnostic in principle, cross-anatomy generalization remains untested, what we will state as a limitation.

Regarding methodological novelty (R1/R4): MICCAI guidelines explicitly state that “a novel algorithm is only one of many ways to contribute” and that “an application of existing methods to a new problem” constitutes a valid contribution. Our main contribution is a novel CAI formulation for a relevant clinical problem. Beyond known components, the paper contributes joint end-to-end optimization of completion and registration under partial US observations, differentiable ICP embedded in the learning pipeline, and patient-specific self-supervised US simulation that adapts to each preoperative bone while avoiding paired CT-US training data entirely. We will clarify this positioning in the Introduction.

Finally, we will add RANSAC details (R2), clarify training from scratch and patient-specific training/inference (R4), report runtime/hardware, and add US simulation references (R3).




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.

    This paper proposes a unified framework for CT–ultrasound bone surface registration that jointly performs shape completion and rigid alignment, leveraging patient-specific synthetic ultrasound simulation to address the scarcity of paired data. The work is positioned in a clinically relevant CAI setting, with an emphasis on robustness under partial observations and efficiency at inference time. 
The paper is well motivated and addresses an important problem in computer-assisted surgery. Its associated strengths include: (i) the joint optimisation of shape completion and registration, which provides a coherent alternative to multi-stage pipelines; (ii) the use of patient-specific synthetic data to alleviate limited data constraints; (iii) promising runtime characteristics compatible with intraoperative use; and (iv) generally clear writing and structured presentation.

    However, consistent limitations have been raised by the reviewers. The primary concern is about the reported results. Indeed, the comparison to learning-based baseline is not convincing, and the justification might overlook some limitations of the proposed method in terms of generalisation. Moreover, large standard deviations raise serious concerns about potential instability (including catastrophic failures) that limit clinical reliability. I would also like to add that no statistical testing is performed whatsoever. Another weakness is the limited methodological novelty, with the framework largely integrating existing components rather than introducing fundamentally new techniques. Finally, other concerns include missing details about baselines, overstatements, studying the impact of erroneous input segmentations, evaluation limited to one dataset, missing citations, and commenting on generalisation issues.

    Overall, this CAI paper presents a well-motivated method with practical advantages. However, the authors need to address important concerns raised about comparison to learning-based methods, robustness, and clarifications of many issues.

  • 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 authors have convincingly answered concerns about the size of the test dataset (very expensive) and why they haven’t tested on other organs (organ-specific problem). In my opinion, they also addressed concerns about novelty, by introducing a framework that takes on a problem with a new, promising technique. Finally, they have clarified a few methodological points, and agreed to soften the unsupported generalisation claims.



    However, the main concern about empirical results remains unconvincingly addressed. The authors claim that the much larger standard deviations obtained by their method wrt previous works primarily stem from an easily detectable failure mode. However, when they remain the corresponding cases, the std devs remain at a much higher level wrt other baselines (8 vs. 1.12 degrees). This adds up to mean scores that are on average worse than the baselines (2.93 vs. 1.43 degrees). Thus, evaluation results do not support gains in terms of accuracy or robustness.

    Overall, this CAI paper is borderline accept/reject. Despite the limited empirical gains, I still recommend acceptance (in agreement with 3 out of 4 reviewers) due to the interesting proposed framework and the discussions it will foster at MICCAI during a CAI session. I also strongly encourage the authors to add the much needed statistical tests.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    This is a very painful decision to me as the idea and the clinical problem are both very good. After a very careful consideration, I believe the present shape of the manuscript can’t support its headline claims.

    This is intraoperative ultrasound to preoperative CT bone registration, and it trains without a costly paired CT-US dataset, which is a real strength. But the paper fails on one hard point. Its selling claim is that it is “more robust,” yet its own Table 1 reports a rotation error of 2.93 +/-19.98 degrees against the baseline’s 1.43 +/- 1.12.The mean is worse and the standard deviation is about ten times larger, which means the method breaks down badly on a real share of cases.

    On top of that, there is no significance test anywhere in the paper. The rebuttal says most of that huge spread comes from about 1.46 percent of “symmetry flip” failures, and that removing them brings the std down to about 8 degrees. But that analysis is not in the submitted paper, so under MICCAI no-new-experiment rule I do not credit it. Even if I did, 8 degrees still loses badly to the baseline’s 1.12, so the “more robust” claim is still false.

    The present results cannot prove a method is reliable by deleting its own failure cases. The two Faculty reviewers who read the original table were not convinced, while the two who scored it high never addressed the variance problem at all. So this is the submission failing on its own evidence, not the rule blocking it. The headline claim is contradicted by its own data, and nothing is backed by statistics, which puts it below the MICCAI bar.

    I do want to encourage the direction, and the quick fix is cheap: drop the overclaim, disclose the symmetry-flip failure mode, and run a significance test on the results they already have.

    As a result the decision is reject, but with strong encouragement to resubmit a clean version to top tier confs and journals.



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 proposes a joint shape completion and registration framework for CT-ultrasound bone alignment. R3 and R4 liked the clinically relevance. R1 and R2 noted a limited methodological novelty and high error variance. The rebuttal has successfully clarified that the standard deviations mostly comes from rare symmetrical flip bones, justifying their limited dataset size. Consensus is on a possible clinical impact, despite limited novelty. For all these reasons, and with respect ot the other submissions, the recommendation is towards Acceptance.



back to top