Abstract

Multimodal medical image registration remains challenging due to substantial cross-modal intensity discrepancies. While recent deep learning methods attempt to reduce modality gaps, they often underexploit modality-invariant anatomical structures or decouple representation alignment from deformation estimation. In this work, we propose a multimodal registration framework that jointly enhances shared anatomical representations and mitigates cross-modal discrepancies. Specifically, we introduce a Structure-Aware Frequency Enhancement (SAFE) module to decompose features into frequency components and selectively reinforce structure-sensitive high-frequency responses, promoting modality-invariant anatomical encoding. In addition, we develop a Registration-guided Multimodal Contrastive Alignment (RMCA) strategy that embeds contrastive learning into the registration process, enabling joint optimization of feature alignment and deformation estimation without requiring pre-aligned supervision. Extensive comparative and ablation experiments on three public datasets demonstrate that our method consistently outperforms state-of-the-art approaches, confirming its efficacy for multimodal medical image registration. The code is available at https://github.com/47-XIAO/MMreg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2472_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/47-XIAO/MMreg

Link to the Dataset(s)

N/A

BibTex

@InProceedings{XiaSiq_Learning_MICCAI2026,
        author = { Xiao, Siqi AND Feng, Zetian AND Zou, Quanling AND Wang, Yi},
        title = { { Learning Modality-Invariant Structure and Distribution Alignment for Multimodal 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

    The paper addresses the challenging problem of multimodal medical image registration and proposes a unified framework that combines structure-aware feature enhancement and contrastive learning.

  • 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.

    One positive aspect is the attempt to integrate contrastive learning directly into the registration process via the proposed RMCA strategy. By constructing positive pairs based on the predicted deformation field rather than relying on pre-aligned data, the method avoids additional supervision and provides a more task-coupled formulation compared to conventional contrastive approaches. This design is conceptually reasonable and aligns with the goal of learning modality-invariant representations in an unsupervised manner.

    The SAFE module introduces a frequency-based feature decomposition to emphasize structural information, which is intuitively motivated given that anatomical structures are more consistent across modalities than raw intensities. While the idea itself is not entirely new, incorporating it into a registration framework is a sensible design choice that may help improve feature consistency across modalities.

  • 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)The formulation of the contrastive learning component remains somewhat ambiguous. Could the authors clarify how positive and negative pairs are explicitly constructed in practice (e. g. , voxel-level vs. patch-level, intra- vs. inter-sample negatives)? In particular, how do you mitigate the risk of false negatives arising from anatomically similar regions across different subjects? 2)The RMCA strategy relies on the predicted deformation field to establish correspondence. How sensitive is the method to inaccurate deformation in the early stages of training? Although a ramp-up strategy is introduced, it would be helpful to provide a quantitative or ablation analysis demonstrating its effectiveness in stabilizing contrastive learning. 3)For the SAFE module, have the authors analyzed whether the enhanced high-frequency components indeed correspond to anatomical structures rather than noise or modality-specific artifacts? Additional visualization or frequency-domain analysis could strengthen the justification of this design. 4)The contrastive loss is applied only at the finest resolution level. Could the authors elaborate on this design choice? Have you explored multi-scale contrastive alignment, and if so, how does it affect performance? 5)From a methodological perspective, the current framework treats all other samples in the batch as negatives. Would incorporating more structured sampling strategies (e. g. , spatially-aware negatives or hard-negative mining) further improve the effectiveness of the contrastive alignment?

  • 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.

  • 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?

    The paper proposes a unified framework that combines frequency-based feature enhancement (SAFE) and registration-guided contrastive learning (RMCA) for multimodal image registration. However, the overall score is mainly limited by insufficient methodological justification. The SAFE module assumes that high-frequency components correspond to anatomical structures, yet high-frequency features may also include noise and modality-specific artifacts, and the method lacks explicit mechanisms to disentangle these factors. In addition, the contrastive learning formulation relies on the assumption that intra-sample aligned features are more similar than inter-sample features used as negatives, which may not hold in medical imaging due to high anatomical similarity across subjects. This issue is further exacerbated by imperfect deformation fields during training, potentially introducing noisy positive pairs and false negatives. Overall, while the framework is reasonable and shows promising results, the lack of rigorous formulation and analysis weakens the reliability and interpretability of the proposed method.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper introduces a unified multimodal medical image registration framework that jointly enhances modality‑invariant anatomical structures and aligns cross‑modal feature distributions. The method is built upon two key technical components: Structure‑Aware Frequency Enhancement (SAFE) and Registration‑Guided Multimodal Contrastive Alignment (RMCA).

    SAFE decomposes deep features into low‑ and high‑frequency components and selectively amplifies structure‑relevant high‑frequency responses through a channel‑attention mechanism, thereby strengthening modality‑invariant structural representations. RMCA introduces a deformation‑guided contrastive learning strategy that dynamically constructs spatially aligned positive pairs using the predicted deformation fields, enabling feature alignment without requiring pre‑aligned supervision.

  • 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 method demonstrates consistently strong empirical performance across all three benchmark datasets, achieving substantial and statistically significant improvements over both unimodal and multimodal baselines.

    The framework is well‑motivated and conceptually coherent. The paper clearly articulates two central challenges in multimodal registration—capturing modality‑invariant anatomical structures and reducing cross‑modal feature discrepancies—and introduces SAFE and RMCA as complementary components that address these challenges within a unified architecture.

  • 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.

    Multimodal registration frameworks typically adopt a dual‑branch encoder to accommodate modality‑specific feature characteristics. Given that MR and CT exhibit distinct high‑frequency responses, is it appropriate to share parameters when computing SE‑based channel attention within the SAFE module? Would modality‑specific SE parameters provide greater flexibility or better capture modality‑dependent structural cues?

    The contrastive learning component relies on deformation‑guided positive pair construction. However, during early training stages, the predicted displacement field may be inaccurate. How does the contrastive loss ensure that these initially constructed positive pairs are reliable? In principle, even slight spatial misalignment would imply that each voxel could have additional potential positive locations among the 2BHWD-2 candidates. Furthermore, both ARDMR and the proposed method employ contrastive learning, yet the performance gap between them is substantial. What factors contribute to this discrepancy?

    Furthermore, it is unclear how this method prevents positive samples from being overwhelmed when the entire volume data is unfolded into batches. Why not borrow from the approach of “modality-independent structural image representation learning for deformable multimodal medical image registration,” which samples only a small number of negative samples to ensure that the gradients of positive sample pairs are not overwhelmed? Under the current formulation, the InfoNCE loss appears ineffective.

  • 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?

    Overall, this paper presents a solid and effective contribution to multimodal registration. While SAFE and RMCA are not entirely novel as standalone components, their integration is well‑motivated and leads to clear and substantial performance gains across multiple datasets. However, the source of RMCA’s superior performance relative to ARDMR—despite both methods employing contrastive learning—remains insufficiently explained.

  • 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.

    I am inclined to accept the paper, but the rebuttal still leaves several technical questions unresolved. I look forward to seeing these details clarified in the final version or in the released code.



Review #3

  • Please describe the contribution of the paper

    This paper presents a unified framework for multimodal medical image registration, with the goal of jointly addressing two key challenges: enhancing shared anatomical structure representations across modalities and reducing cross-modal feature discrepancies within a single registration pipeline. The overall architecture consists of dual-stream multi-scale encoders and a coarse-to-fine decoder. Fixed and moving images are processed by two encoders with the same architecture but separate parameters, which extract four-scale hierarchical features, and the final deformation field is estimated progressively through multi-scale refinement.

    The method mainly introduces two components. The first is SAFE (Structure-Aware Frequency Enhancement), which operates on the deepest features by decomposing them into low- and high-frequency components and applying channel attention to the high-frequency branch in order to emphasize more structure-relevant responses. SAFE is shared across the fixed and moving branches, with the intention of encouraging more consistent structural representation learning across modalities. The second is RMCA (Registration-guided Multimodal Contrastive Alignment), which embeds contrastive learning directly into the registration process. Specifically, the current predicted deformation field is used to warp the moving feature into the fixed space, thereby dynamically constructing cross-modal positive pairs without requiring pre-aligned supervision, and enabling joint optimization of feature alignment and deformation estimation. The final training objective combines a MIND-based similarity term, a contrastive loss, and a deformation regularization term.

    Experimentally, the paper evaluates the proposed method on three public datasets, namely Learn2Reg, AMOS, and SR-Reg, and compares against SyN, VoxelMorph, TransMorph, Fourier-Net, and several recent multimodal registration approaches. According to the reported results, the proposed method achieves the best DSC on all three datasets while maintaining low folding ratios, and the ablation studies further suggest that both SAFE and RMCA contribute positively to the final 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.

    A clear strength of the paper is that the problem setting and methodological motivation are both well aligned with the challenges of multimodal registration. The difficulty of multimodal registration indeed comes from both anatomical correspondence and modality discrepancy, and the paper explicitly attempts to improve the registration process from these two perspectives rather than relying only on a new backbone or a different similarity metric. As a result, the overall narrative of the paper is coherent and the method is conceptually easy to follow.

    The roles of SAFE and RMCA are also conceptually distinct. SAFE is designed to emphasize structure-relevant information at the deepest level, while RMCA introduces contrastive learning into the deformation estimation loop through registration-guided positive-pair construction. In particular, the idea of dynamically constructing cross-modal positives based on the current registration result is more task-specific than decoupling multimodal representation learning from registration, and RMCA is arguably the more distinctive technical component of the paper.

    The experimental section is reasonably convincing. The paper does not rely on a single benchmark, but instead evaluates the method on three public datasets covering both abdominal and brain MR-CT registration scenarios. The set of baselines is also relatively comprehensive, including a classical method, generic deep registration models, and several recent multimodal approaches. Most importantly, the reported gains are consistent across datasets, and the proposed method achieves both improved DSC and low folding ratios. This makes the empirical evidence itself fairly persuasive.

    In addition, the ablation results provide at least some support for the proposed components. SAFE alone leads to consistent improvements, RMCA alone yields larger gains in DSC, and the full model performs best across the three datasets. While this does not fully validate all of the paper’s mechanistic interpretations, it does indicate that both modules contribute positively to the final performance rather than the gains being caused purely by training tricks or chance.

  • 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.

    Despite the promising overall results, some methodological aspects remain insufficiently clarified. First, the proposed “frequency enhancement” in SAFE is, in practice, closer to a high-frequency enhancement module based on low/high residual decomposition than to a fully established mechanism for learning modality-invariant structural representations. The paper does not directly demonstrate that SAFE truly learns the kind of modality-invariant structure emphasized in the claims. Moreover, SAFE is only applied at the deepest scale, while finer-scale registration stages still rely on the original encoder features. If the main value of SAFE is to strengthen structural consistency, the paper should better justify why this mechanism is only used at the bottom level rather than applied more systematically across scales. This design choice is not sufficiently validated.

    RMCA is the more interesting part of the method, but its robustness also deserves further discussion. According to the paper, RMCA constructs positive pairs dynamically using the current deformation field, while negatives are taken from other features within the same mini-batch; at the same time, the reported batch size is only 2.Under such a setting, the number of available negatives is very limited, which means that the contrastive signal may be weak and potentially sensitive to batch composition. Given that the overall dataset sizes are not very large, this raises some concern about how robust the claimed cross-modal alignment effect of RMCA really is. The paper does not provide a batch-size sensitivity analysis, nor does it discuss alternative negative sampling strategies, so the mechanistic conclusion about RMCA should be interpreted with caution.

    In addition, the implementation details of RMCA remain somewhat ambiguous. The text refers to “spatially aligned feature maps as positive pairs,” whereas the loss is written in a standard sample-level InfoNCE form. It is therefore not entirely clear whether the actual implementation performs contrastive learning on globally aggregated sample-level embeddings or on finer-grained local features. Since this detail directly affects the strength of the contrastive supervision, the positive/negative definitions, and reproducibility, a clearer explanation would strengthen the paper.

    Although the paper evaluates on three datasets, the test sizes of some datasets are still relatively limited, and some reported standard deviations are fairly large. Therefore, the current experiments more convincingly support the conclusion that the method is empirically effective under the present setup than the stronger mechanistic interpretation that SAFE and RMCA fully realize modality-invariant structural learning and robust multimodal alignment. In other words, the empirical results are convincing, but the evidence for the intermediate mechanism is still weaker than the evidence for the final performance.

  • 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 marginal accept. Although the paper still leaves some methodological questions unresolved, particularly regarding the interpretation of SAFE and the robustness of RMCA under a small-batch, limited-data setting, the overall work is reasonably complete and the empirical evidence is fairly convincing. Compared with methods evaluated on only a single benchmark, this paper provides results on three public datasets, which gives a better sense of generalization potential. In addition, the comparison set includes classical, general-purpose, and recent multimodal registration methods, making the reported improvements more meaningful.

    More importantly, the reported gains are not limited to a single dataset or a single metric. The proposed method achieves consistent DSC improvements across all three datasets while maintaining low folding ratios, and the ablation studies further indicate that both SAFE and RMCA contribute positively to the final performance. Therefore, even though I remain somewhat concerned that the sample-level contrastive design with batch size 2 may make RMCA sensitive to data scale and batch composition, I view this as weakening the strength of the mechanistic claims rather than invalidating the empirical findings themselves.

    Overall, I believe this paper meets the bar for a borderline acceptance. The method contains a reasonably interesting idea, and the experimental validation is sufficiently broad and persuasive to support its practical value. The paper would be further strengthened if the final version clarified the implementation details of RMCA and discussed its sensitivity to batch size and dataset scale in more depth.

  • 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.

    After reading the rebuttal, I still have concerns about the actual role of RMCA. The clarified implementation suggests that RMCA is closer to a MIND-guided feature-map consistency regularizer than dense local contrastive correspondence learning. The positive pairs depend on the current deformation field, and the negative pool is limited under batch size 2.Thus, the mechanistic justification remains weaker than the empirical results. However, the rebuttal does clarify the implementation, and the reported gains across three datasets remain reasonably strong. Since my original score already accounted for the methodological concerns, I would keep my score at 4.



Author Feedback

We appreciate the reviewers’ insightful suggestions and recognition of our work.

1) Formulation of the contrastive learning component (R1, R2, R3) RMCA is applied at the finest-resolution feature level because these features preserve richer spatial and anatomical details, while the learned alignment naturally propagates through the encoder-decoder architecture. Empirically, multi-scale contrastive alignment provided only marginal improvements. The motivation of RMCA is that moving features warped by the predicted deformation field exhibit substantially higher structural consistency with fixed features. In contrast, unaligned features within a batch contain noticeable structural discrepancies caused by motion and inter-subject anatomical variations, as also reflected by the low initial DSC in Table 1.Accordingly, RMCA constructs positive pairs based on spatial correspondence established by the predicted deformation. Specifically, warped moving features and their corresponding fixed features are treated as positive pairs, while unaligned features are regarded as negatives. All positives and negatives are constructed at the inter-sample level. RMCA computes voxel-wise cosine similarity followed by global averaging to measure pairwise feature similarity in the contrastive loss. Unlike ARDMR, which constructs positive pairs from the same patient, RMCA defines positive pairs through deformation-based spatial correspondence. Therefore, our method remains fully unsupervised and does not require pre-aligned or paired data, improving its generalizability. For clarity and reproducibility, we will release our code and the splits of three public datasets.

2) Robustness of RMCA (R1, R2, R3) We agree that deformation fields may be inaccurate during early training, which is consistent with our preliminary observations. To improve the robustness of RMCA, we employ a Gaussian ramp-up strategy for the contrastive loss. Specifically, the contrastive weight is kept low during the early stages when the predicted deformation remains unstable, thereby reducing the influence of unreliable positive pairs. As registration accuracy progressively improves, the contrastive constraint is gradually strengthened, enabling more stable and reliable feature alignment.

3) Analysis of SAFE (R1, R2, R3) High-frequency (HF) components are widely associated with structural and edge information in both natural and medical images [1-2]. In our preliminary experiments, multimodal images reconstructed using only HF components still preserved highly consistent anatomical structures across modalities, supporting the assumption that HF responses contain modality-invariant structural cues. Similar observations have also been reported in [3]. We acknowledge that HF features may also include modality-specific noise. Thus, SAFE operates on features already refined by RMCA and employs a shared SE channel attention to selectively enhance consistent structural responses while suppressing modality-specific activations. Sharing SE parameters encourages the extraction of modality-invariant features activated consistently across modalities, whereas modality-specific SE parameters may reduce cross-modal consistency. Empirically, our preliminary experiments showed that shared SE parameters achieved better registration performance. SAFE is applied only at the deepest scale because high-level features contain stronger modality-invariant semantic and structural information, making them more suitable for cross-modal enhancement. In contrast, finer-scale stages preserve more modality-specific appearance details, where RMCA is used to reduce cross-modal representation gaps while maintaining local details for accurate deformation estimation. The efficacy of SAFE is further validated by consistent improvements across all three datasets in the ablation studies.

[1] ICCV2023, 10.1109/ICCV51070.2023.01928 [2] CVPR2025, 10.1109/cvpr52734.2025.02222 [3] MICCAI2025, 10.1007/978-3-032-04927-8_15




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 received mixed recommendations from the reviewers, including two Weak Rejects and one Weak Accept.

    The paper addresses the challenging problem of multimodal medical image registration and demonstrates consistent performance gains across three public datasets. However, the reviewers identified several substantive concerns, including the ambiguous formulation of the contrastive learning component (R1, R2, R3), questions regarding the robustness of the RMCA strategy (R1, R2, R3), and the lack of analysis in the frequency domain (R1, R2).

    Overall, the paper shows some merits when comparing to existing methods, and the meta-reviewer therefore recommends that it proceed to rebuttal. The authors are encouraged to focus their response on the major concerns outlined above.

  • 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, this paper received mixed recommendations. Reviewer 2 upgraded the rating from Weak Reject to Accept after the authors clarified the complementarity between frequency-domain structural information and dynamic positive-pair construction, and Reviewer 3 maintained an Accept recommendation. However, despite these positive ratings, the reviews collectively continued to raise overlapping concerns regarding the methodology and the extent of the technical contribution. After revisiting the paper and rebuttal, I find that the experimental evaluation does not include sufficiently strong competitive baselines, and that the core components, SAFE and RMCA, do not appear sufficiently novel when considered as standalone contributions. The paper would be strengthened by including comparisons to ConvexAdam and by providing a more detailed analysis of the learned features. In light of the post-rebuttal discussion and the overall reviewer consensus, I recommend Reject.



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.

    This work proposed an image registration approach for multimodal registration. The approach makes use of structure-aware frequency enhancement (SAFE; essentially focusing on high frequency components) and multimodal contrastive alignment (RMCA). The method was tested on three different datasets (two abdominal datasets and one for the brain) and achieves uniformly good results.

    Reviewers appreciated the overall approach, the extensive evaluation, and its good performance. There were some concerns regarding the SAFE approach (specifically, if focusing on high frequencies indeed is expected to produce modality-invariant representations) and RMCA (specifically, if this contrastive approach suffers from instabilities during training when registration are not accurate at the beginning). The rebuttal addressed these questions and reviews were largely positive after the rebuttal.



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.

    After the rebuttal phase, all reviewer’s opinions appear to lean toward acceptance. I believe this work addresses an important problem in multimodal registration, where contrastive learning shows strong potential but requires careful and dedicated design to be effective. I believe the manuscript would benefit from addressing the technical concerns raised by the reviewers. In particular, further clarification is needed regarding the behavior of the model during the initial stages of training, as well as a more detailed explanation of the SAFE module, especially in terms of how it distinguishes between structural information and noise.



back to top