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

Abstract

Accurate and efficient deformable medical image registra tion remains challenging, particularly for high-resolution 3D volumes where dense multi-scale architectures incur substantial computational and memory overhead. We propose SPEAR, a scalable error-guided sparse refinement framework for 3D deformable registration. It leverages anatomy-aware discrepancy signals to guide targeted sparse refinement, enabling progressive deformation updates without redundant computa tion. Experiments show that SPEAR improves deformation accuracy and regularity over conventional unsupervised models, while achieving com parable performance to recent multi-scale approaches with substantially lower computational and memory cost. Importantly, the proposed design scales effectively to high-resolution 3D volumes under constrained GPU memory. These results demonstrate that structured sparse refinement of fers a scalable and balanced solution for deformable medical image regis tration. Code: https://github.com/ZedKing12138/SPEAR-registration.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4287_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)

N/A

BibTex

@InProceedings{BaiHao_SPEAR_MICCAI2026,
        author = { Bai, Hao AND Hong, Yi},
        title = { { SPEAR: Error-Guided Sparse Refinement for Scalable Deformable Medical Image Registration } },
        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 introduces SPEAR, an unsupervised 3D deformable registration framework that replaces traditional dense multi-scale pyramids with a scalable error-guided sparse refinement mechanism. The primary contributions here lie in a region-wise Top-r routing strategy that uses a segmentation-guided discrepancy signal to identify anatomically misaligned regions. By injecting only these sparse tokens into a dual-stream transformer, the model progressively updates the deformation field through a non-redundant refinement loop. This allows the model to handle high-resolution 3D volumes that typically exceed GPU memory limits for dense multi-scale architectures.

  • 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 structured sparse refinement reduces FLOPs and peak GPU memory usage compared to dense SOTA models such as SACB-Net and CorrMLP, without a significant loss in accuracy.

    The evaluation on IBSR18 is strong, demonstrating the model’s ability to operate at full resolution even when baselines are forced to downsample due to memory constraints.

    The Hard Non-Redundant constraint is an interesting addition that prevents local update oscillations and ensures that successive refinement stages explore complementary anatomical regions.

  • 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 ground truth labels are not used for registration training, the discrepancy signal relies on a pretrained segmentation network trained on the training split. This creates a dependency in which the registration quality is strictly limited by the segmentation model’s quality.

    On the IXI and OASIS datasets, SPEAR does not outperform the heavy multi-scale baselines (CorrMLP and SACB-Net) in terms of Dice score. This needs further performance enhancement. The performance advantage is evident only in the IBSR18 dataset, suggesting that the benefits of sparse refinement may be largely restricted to memory-constrained, high-resolution scenarios.

  • 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 introduces an interesting conceptual framework by replacing dense multi-scale pyramids with error-guided sparse refinement. The main value of this work lies in the high-resolution IBSR18 experiments, where SPEAR achieves a higher DICE score than other baselines, which are forced to downsample due to memory constraints.

    However, the “efficiency” narrative is undermined by results on standard datasets, which fail to provide a significant increase in accuracy. Furthermore, the current framework relies on a frozen auxiliary segmentation network.

    Despite these trade-offs, the architectural novelty of the Top-r non-redundant routing and its demonstrated utility in memory-constrained scenarios are interesting.

  • 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 authors introduce an error-guided, non-redundant sparse refinement mechanism that replaces dense multi-scale pyramids for scalable and efficient 3D deformable medical image registration while achieving competitive registration accuracy.

  • 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 proposes SPEAR, a scalable error-guided sparse refinement framework for 3D deformable image registration that replaces dense multi-scale hierarchies with targeted region-wise updates.

  • 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. I suspect that parts of the model illustration and description are generated by large language models without any subsequent revision. The figure is incomplete, with blurred lines, mismatched content relative to the method description in the main text, and severely inconsistent formatting. 2.The error map is derived from a pretrained segmentation network. Although the segmentation network is frozen and its labels are not directly used to supervise the deformation field, the routing signal still relies on semantic information extracted from ground-truth annotations. This raises a concern: does this design implicitly introduce label-based supervision into the registration task, even if indirectly. 3.In the IBSR18 experiments, baseline methods (CorrMLP, SACB-Net) required downsampling due to memory constraints, while SPEAR operated at full resolution. Comparing methods under different input resolutions is inherently unfair and may bias the results in favor of the proposed method. 4.The authors repeatedly claim high efficiency, but SPEAR does not consistently outperform baselines in GPU memory, FLOPs, or registration accuracy across all settings. 5.In the ablation study, r=9 yields the highest Dice, but the authors choose r=6 based on folding rate. It is questionable to prioritize a secondary metric (folding rate) over the primary objective of registration accuracy.
  • Please rate the clarity and organization of this paper

    Poor

  • 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?

    1.I suspect that the model figure and parts of the description are generated by large language models without any subsequent revision. The figure is incomplete, with blurred lines, mismatched with the corresponding method description. 2.The claimed efficiency advantages in GPU memory and computation are not clearly supported by the experimental results. 3.The fairness of the experimental comparisons is questionable.

  • 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.

    The author has resolved all the raised issues; therefore, I recommend acceptance.



Review #3

  • Please describe the contribution of the paper

    This paper presents SPEAR, an error-guided sparse refinement framework for 3D deformable medical image registration, aiming to reduce the computational and memory burden of conventional dense multi-scale methods on high-resolution volumes. The key idea is to estimate anatomy-aware discrepancy signals using a frozen segmentation network and then perform region-wise Top-r sparse routing so that only the most informative local candidate tokens are selected for refinement at each stage, instead of repeatedly processing the entire volume in a dense multi-scale manner. In addition, the paper introduces a hard non-redundant constraint to prevent consecutive stages from repeatedly selecting the same local subregions, thereby improving refinement efficiency and stability. The overall architecture combines global-token-based coarse initialization, sparse local token injection, dual-stream transformer interaction, and residual flow composition into a progressive deformation update process. Experiments are conducted on IXI, OASIS, and IBSR18, and the paper evaluates the method not only in terms of Dice and folding rate but also inference time, memory, and FLOPs.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    A major strength of the paper is that the proposed idea is methodologically distinctive. Compared with the standard dense full-volume modeling paradigm or conventional multi-scale coarse-to-fine registration frameworks, SPEAR explores a different direction by selectively refining error-prone regions. In particular, the combination of region-wise sparse routing and hard non-redundant refinement makes the method more than just a simple top-k sparsification trick, and gives it a recognizable identity within the registration literature.

    A second strength is that the paper provides a relatively complete efficiency-oriented evaluation. In addition to registration accuracy and deformation regularity, the paper systematically reports runtime, peak memory, and FLOPs. This allows the reader to assess the accuracy-efficiency trade-off in a more balanced way. The results on IBSR18 are particularly compelling in this regard: while CorrMLP and SACB-Net require downsampling due to memory constraints, SPEAR can run directly at full resolution and achieves the best overall performance, which provides meaningful support for the paper’s scalability claim.

    Third, the ablation study in Table 2 is informative. Removing error guidance clearly hurts both Dice and folding, removing the hard non-redundant constraint degrades deformation stability, and replacing sparse routing with dense injection yields comparable accuracy but significantly worse efficiency. These results suggest that the core design choices are not merely narrative additions, but do contribute to the claimed sparse-yet-effective refinement behavior.

  • 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.

    Although the proposed sparse refinement idea is interesting, the current version still has several notable weaknesses. First, the paper presents the method as error-guided sparse refinement, but the sparsity is mainly applied at the level of token selection and injection rather than at the level of deformation updates themselves. Each stage still predicts a residual flow for the whole volume, so the paper does not actually establish that the selected high-discrepancy regions are the ones predominantly corrected by the current update. In this sense, the method is better characterized as sparsely guided dense refinement rather than truly sparse, spatially localized deformation correction.

    Second, the method appears to depend strongly on initialization and coarse global alignment. In the experiments, all inputs are affinely pre-aligned before the deformable stage, and the method itself further relies on a global-token-based coarse initialization before sparse local refinement begins. This design may be relatively suitable for brain MRI, where deformations and anatomical variability are more controlled, but the paper does not verify whether the same sparse refinement mechanism remains effective under larger displacements, more complex anatomical variation, or more challenging initialization conditions. Therefore, the current evidence more strongly supports effectiveness under reasonably good coarse alignment than robustness in more difficult registration settings.

    Third, the routing mechanism depends heavily on a frozen, dataset-specific segmentation network to generate the discrepancy signal. While this may be a reasonable engineering prior for brain MRI, the paper does not provide a strategy for applying the method in the complete absence of segmentation labels, nor does it analyze how the routing mechanism would degrade when the auxiliary segmentation model is inaccurate, misses relevant structures, or suffers from domain shift. As a result, the effectiveness of the current method depends to a substantial extent on an additional semantic prior, and this dependency is not sufficiently discussed.

    Moreover, although the experiments cover three public datasets, they are all fundamentally inter-subject T1-weighted brain MRI registration settings. This is sufficient to demonstrate effectiveness and some scalability within the brain MRI domain, but it is still not strong evidence that the proposed sparse refinement mechanism will naturally generalize to more complex anatomies, larger deformations, or settings with weaker structural priors. At present, the paper more convincingly supports the claim that the method works for brain MRI registration than a broader claim about a generally applicable registration paradigm. Finally, the choice of baselines remains somewhat limited. The paper mainly compares SPEAR against relatively heavy multi-scale pyramid-style methods, such as CorrMLP and SACB-Net, and then emphasizes its advantages in runtime, memory, and FLOPs; on IBSR18, these baselines cannot even run at full resolution and therefore require 2× downsampling followed by flow upsampling for evaluation. Such comparisons do support the claim that SPEAR scales better than heavy multi-scale architectures, but they are still insufficient to establish superiority over lightweight pyramid-style alternatives. In particular, the paper does not include a direct comparison against a lightweight pyramid-based registration method such as IIRP-Net [1], and therefore the current evidence more strongly supports a better trade-off relative to heavy multi-scale models rather than a broader conclusion that SPEAR outperforms all efficient pyramid-style baselines. [1] Ma, T., Zhang, S., Li, J., Wen, Y. IIRP-Net: Iterative Inference Residual Pyramid Network for Enhanced Image Registration. CVPR 2024.

  • 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?

    I lean toward a weak accept. The main reason is that the paper proposes a genuinely interesting direction for deformable registration, namely segmentation-guided sparse refinement with non-redundant routing, which goes beyond the common pattern of incrementally modifying existing dense multi-scale architectures. The empirical results, especially Table 1 and Table 2, provide meaningful evidence that this sparse-guided refinement strategy can deliver a strong accuracy-efficiency trade-off and scale better to high-resolution 3D volumes under memory constraints.

    At the same time, I do believe the paper has several nontrivial shortcomings, particularly the lack of transparency in how the token representations are converted into residual flow updates, the fact that sparsity is not enforced directly at the deformation level, and the dependence on a frozen segmentation prior. These issues make the mechanistic interpretation weaker than the empirical story and limit the strength of the paper’s broader generalization claims.

    Nevertheless, on balance, I think the work reaches the bar for a weak accept. The idea is novel enough, the empirical validation is reasonably complete, and the efficiency-oriented contribution is practically meaningful. The paper would be significantly strengthened if the final version clarified the implementation details of global initialization and token-to-flow mapping, and discussed more explicitly the limitations imposed by the segmentation-guided discrepancy prior.

  • 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 appreciate the authors’ clarifications. The rebuttal confirms several limitations that I had already considered in my original score: the method uses a frozen semantic prior for routing, the sparse operation is mainly at the token routing/injection level rather than at the deformation update level, and the strongest evidence is concentrated on the high-resolution IBSR18 setting. I still view these as important limitations, and I agree with the meta-reviewer that the fairness and generality claims should be substantially tempered. However, the core idea remains methodologically interesting, and Table 2 provides useful evidence for the proposed routing and non-redundant refinement. Since these concerns were already reflected in my initial weak-accept score, I would keep my score at 4 after rebuttal.



Author Feedback

We thank reviewers/meta-reviewer for constructive feedback and address prioritized concerns below.

Q1: Segmentation prior / fairness [All, AC] The frozen segmenter is an auxiliary semantic prior, not direct supervision. During training it is frozen; the registration network receives no GT labels, seg/overlap loss, deformation, landmark, or correspondence target. The loss is image-based (NCC + smoothness). The segmenter outputs soft predictions on fixed/warped images; their discrepancy ranks tokens. Thus, the prior affects where refinement is allocated, not how deformation is supervised. Dense Injection uses the same discrepancy without sparse selection and gives similar Dice but much higher time/memory/FLOPs, showing SPEAR converts this prior into efficient sparse refinement rather than merely adding it. We will avoid “fully unsupervised” and describe SPEAR as image-based unsupervised registration with frozen semantic-prior routing. Inaccurate segmentation/domain shift may degrade token prioritization. Since predictions are not the matching criterion or deformation target, they do not hard-bound accuracy, but may reduce routing effectiveness. We will state this limitation.

Q2: IBSR18 scalability and accuracy-efficiency trade-off [All, AC] IBSR18 is a fixed-memory scalability evaluation, not a same-resolution accuracy comparison. It is processed at 225^3.CorrMLP/SACB-Net exceed 48GB memory, so inputs are downsampled 2x and flows upsampled for evaluation, with overhead included. SPEAR runs full resolution, supporting better scalability than dense multi-scale baselines under constrained memory. IBSR18 has few subjects, but evaluation is pair-level: within each subject-wise split, ordered non-self moving-fixed pairs are generated, so n subjects yield n(n-1) pairs (e.g., 4 test subjects -> 12 pairs), not only 3–4 registrations. This follows common IBSR18 practice. Still, pairs are not fully subject-independent, so we will report subject/pair counts and temper claims. SPEAR is not universally most efficient/accurate. On IXI/OASIS it is not highest in Dice, but competitive with CorrMLP/SACB-Net with lower time/memory/FLOPs. Savings may be modest for smaller GPUs; our claim is improved trade-off under 48GB, not deployment on all modest GPUs.

Q3: Dice evaluation protocol [R2, AC] Dice is computed by warping moving segmentation to fixed space and comparing with fixed segmentation at evaluation resolution. For IBSR18 baselines with downsampled inputs, the deformation field is upsampled before Dice computation. Thus, downsampling affects model input, not final Dice evaluation. We will make this explicit.

Q4: Routing budget r [R2] Although r=9 gives slightly higher Dice, the gain is marginal while time/memory/FLOPs increase. SPEAR targets balanced accuracy-efficiency-regularity, so r=6 was selected for the lowest folding rate with lower cost. We will clarify this.

Q5: LLM/Figure 1 [R2, AC] We apologize that Fig. 1 is dense and hard to read after size reduction. It illustrates Sec. 2: hierarchical tokenization, discrepancy estimation, Top-r routing, non-redundant masking, sparse injection, cross-attention, and residual flow composition. We will redraw it as a clearer manually prepared diagram. No LLM was used for the method, experiments, results, analysis, or scientific claims; language editing, if any, was only grammar/polishing and disclosed per policy.

Q6: Scope, baselines, and remaining limitations [R3, AC] Current evidence mainly supports affine-prealigned inter-subject T1 brain MRI registration; we will avoid overgeneralizing to other anatomies, large-deformation, CT, or microscopy settings. SPEAR is sparse at token routing/injection, while dense residual flow preserves spatial continuity; thus “sparse-guided dense residual refinement” is more precise. For literature-reported ANTs/FireANTs, SynthMorph, or efficient pyramid-style methods, we will avoid direct numerical claims unless evaluated under the same protocol.



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 presented work develops an image registration approach that uses sparse refinements with the intent of making the registration approach more scalable (while achieving high accuracy and regularity) so that it can be used also for higher resolution images (which may be precluded for competing approaches due to memory limitations).

    Main strenghts:

    • The reviewers considered this an interesting approach that differs conceptually from existing registration approaches.
    • Two reviewers appreciate the obtained reductions in memory and computational cost.
    • Two reviewers also highlighted the good performance on IBSR (though this meta reviewers has some questions regarding these results – see below).

    Main weaknesses:

    • The proposed approach uses segmentations in training (to compute the error signal) whereas the competing approaches are not afforded this advantage. Hence it is unclear if the comparisons are fair.
    • While the main argument is high accuracy and regularity with reduced memory consumptions and less computational requirements these savings are relatively modest and would still not allow computations on a more modest GPU.
    • The approach underperforms on IXI and OASIS.
    • While strong accuracy improvements are reported for ISBR18 over the baseline approaches that are being tested it is unclear what drives this difference: is it indeed the fact that the method can work at higher spatial resolution? The test size based on the reported 7:1:2 split for ISBR18 is likely only 3 or 4 images and hence it is not clear how reliable these results are.
    • There is no comparison to more classical registration approaches which also would not require downsampling. For example, the “FireANTs: Adaptive Riemannian Optimization for Multi-Scale Diffeomorphic Matching” paper by Jena et al. reports much higher Dice scores for IBSR18 for ANTs and FireANTs than what is reported in the MICCAI submission. See Figure 15 ( https://arxiv.org/pdf/2404.01249v5 ). And SynthMorph results also appear to outperform the proposed method on IBSR18.- There are concerns that at least Figure 1 (and possibly some of the text) have been generated by an LLM.
    • The evaluation is limited to brain MRIs. To make the case that this works well at high resolution it would have been useful to show results on high resolution CT or microscopy for example. Or a synthetic experiment exploring the targeted saving in memory and computational cost more explicitly.

    For the rebuttal I recommend focusing on the following questions:

    • Discuss if the evaluation is fair as only your method is using segmentations.
    • Discuss the low sample size for testing for IBSR18 and comparisons to other approaches in literature.
    • Discuss how the evaluations are being done. Are Dice measures all computed at the original image resolution?
    • Discuss if LLMs were used for this paper and if so how.
    • Discuss any remaining review concerns as space permits.
  • 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 presented work develops an image registration approach that uses sparse refinements with the intent of making the registration approach more scalable (while achieving high accuracy and regularity) so that it can be used also for higher resolution images (which may be precluded for competing approaches due to memory limitations).

    Main strenghts:

    • The reviewers considered this an interesting approach that differs conceptually from existing registration approaches.
    • Two reviewers appreciated the obtained reductions in memory and computational cost.
    • Two reviewers also highlighted the good performance on IBSR.

    Main weaknesses:

    • The proposed approach uses segmentations in training (to compute the error signal) whereas the competing approaches are not afforded this advantage. Hence it is unclear if the comparisons are fully fair.
    • While the main argument is high accuracy and regularity with reduced memory consumptions and less computational requirements these savings are relatively modest and would still not allow computations on a more modest GPU.
    • The approach underperforms on IXI and OASIS.
    • While strong accuracy improvements are reported for ISBR18 over the baseline approaches that are being tested it is unclear what drives this difference: is it indeed the fact that the method can work at higher spatial resolution?
    • There is no comparison to more classical registration approaches which also would not require downsampling. For example, the “FireANTs: Adaptive Riemannian Optimization for Multi-Scale Diffeomorphic Matching” paper by Jena et al. reports much higher Dice scores for IBSR18 for ANTs and FireANTs than what is reported in the MICCAI submission. See Figure 15 ( https://arxiv.org/pdf/2404.01249v5 ). And SynthMorph results also appear to outperform the proposed method on IBSR18.- The evaluation is limited to brain MRIs. To make the case that this works well at high resolution it would have been useful to show results on high resolution CT or microscopy for example. Or a synthetic experiment exploring the targeted saving in memory and computational cost more explicitly.

    For the rebuttal the recommendation was to focus on the following questions: 1.Discuss if the evaluation is fair as only the proposed method is using segmentations. 2.Discuss the low sample size for testing for IBSR18 and comparisons to other approaches in literature. 3.Discuss how the evaluations are being done. Are Dice measures all computed at the original image resolution? 4.Discuss if LLMs were used for this paper and if so how.

    The rebuttal addressed 1 and 2 partially and provided satisfying answers for 3 and 4.Comparisons to more standard segmentation approaches are not included and not properly addressed in the rebuttlal; if results for competing methods would change if segmentations were available to them also remains somewhat unclear.

    After the rebuttal all reviewers recommended acceptance.



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.

    After the rebuttal process, two reviewers agreed in the post-rebuttal decision to accept the manuscript. The third one remained irresponsive, although their review appeared to lean toward acceptance with low confidence. After a thorough review of the manuscript and all associated evaluations, I find the reviewers recommendations somewhat surprising. While the complexity of the proposed methodology is noteworthy, the reported results do not convincingly demonstrate its superiority with respect to the claimed attributes. In particular, the selection of comparison methods is limited to VoxelMorph and TransMorph (which are not pyramid-based), as well as CorrMLP and SACB-Net. Based on my experience, CorrMLP does not represent a particularly strong benchmark for non-rigid registration. Furthermore, the use of the IBSR18 dataset raises concerns. Although it has historically been employed for evaluating non-rigid image registration methods, its quality is relatively poor by current standards, and more suitable and higher-quality datasets are now available for this task. In addition, qualitative results were shown in an easy example from IXI dataset, that involves registration pairs from controls. Transformation grids were not amazing.

    Overall, I believe the principal meta-reviewer provided a well-reasoned assessment, effectively highlighting the key shortcomings and limitations of both the proposed method and the experimental evaluation. Their summary accurately reflects the reviewers initial impressions, which tended toward rejection, and in my view, this assessment was justified.



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 paper received consistently positive recommendations following the rebuttal. Reviewer 2 raised their score from Weak Reject to Accept, indicating that their concerns had been satisfactorily addressed. Reviewer 3 also upgraded their assessment to Accept, noting that routing-level sparsity remains a recognized trade-off rather than a fundamental weakness. Reviewer 1 maintained a Weak Accept recommendation. In view of the post-rebuttal discussion and the overall reviewer consensus, I recommend acceptance.

    For the camera-ready version, the authors are encouraged to further improve the clarity of the presentation and to incorporate the reviewers’ constructive suggestions where appropriate.



back to top