Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Quantitative cardiac magnetic resonance imaging (MRI) enables non-invasive myocardial tissue characterization but relies on robust motion correction within these variable-length, variable-contrast image sequences. Groupwise registration, which simultaneously aligns all images, has shown greater robustness than pairwise registration for motion correction. However, current deep-learning-based groupwise registration methods cannot generalize across MRI sequences: the architecture typically encodes input data as a fixed-length channel stack, which rigidly couples network design to protocol-specific sequence length, input ordering, and contrast dynamics. At inference time, any change in imaging protocols will render the network unusable. In this work, we introduce Any$^{2}$Reg, a new set-based groupwise registration framework that takes a quantitative MRI sequence as an unordered set. This set formulation fundamentally decouples network design from sequence length and input ordering. By utilizing a shared encoder and correlation-guided feature aggregation, Any$^{2}$Reg constructs a permutation-invariant canonical reference for registration, and learns a permutation-equivariant mapping from images to deformation fields. Additionally, we extract contrast-insensitive image features from an existing foundation model to handle extreme contrast variations. Trained exclusively on a single public $T_1$ mapping dataset (STONE, sequence length $L=11$), Any$^{2}$Reg generalizes to two unseen quantitative MRI datasets (MOLLI, ASL) with variable lengths ($L \in [11, 60]$) and different contrast dynamics. It achieves strong cross-protocol generalization in a zero-shot manner, and consistently improves downstream quantitative mapping quality. Notably, while designed for quantitative MRI sequences, our framework is directly applicable to Cine MRI sequences for inter-cardiac-phase registration.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0220_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/YiZhang025/Any2Reg

Link to the Dataset(s)

ACDC dataset: https://www.creatis.insa-lyon.fr/Challenge/acdc/databases.html STONE dataset: https://pubmed.ncbi.nlm.nih.gov/29314198/

BibTex

@InProceedings{ZhaYi_SetBased_MICCAI2026,
        author = { Zhang, Yi AND Zhao, Yidong AND Toxopeus, Tijmen AND Božić-Iven, Maša AND Weingärtner, Sebastian AND Tao, Qian},
        title = { { Set-Based Groupwise Registration for Variable-Length, Variable-Contrast Cardiac MRI } },
        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 proposes Any2Reg, a set-based groupwise registration framework for cardiac MRI that explicitly models image sequences as unordered sets rather than fixed-length, ordered stacks. This design enforces permutation equivariance and allows the model to generalize across varying sequence lengths and contrast dynamics.

A key technical contribution is the correlation-guided aggregation module, which constructs a canonical reference representation from the input set. Additionally, the method optionally incorporates contrast-insensitive features from a pretrained foundation model to improve robustness under varying contrasts.

The method is trained on a single T1 mapping dataset and demonstrates zero-shot generalization to unseen datasets with different sequence lengths and contrast mechanisms, improving both registration accuracy and downstream quantitative mapping quality.
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.
- Conceptually strong formulation (set-based modeling) The reformulation of groupwise registration as a set-to-set problem is well-motivated and addresses a real limitation of channel-stacked architectures (fixed length, order dependence). This is a meaningful methodological contribution.
- Generalization across protocols The demonstrated zero-shot generalization across MOLLI, ASL, and Cine datasets is compelling and practically relevant for clinical MRI workflows with heterogeneous protocols.
- Solid empirical evaluation The paper includes multiple datasets, baselines (including classical and DL-based methods), and downstream evaluation (e. g. , Dice, TRE, T1 fitting metrics), which provides a comprehensive assessment.
- Consistent performance improvements The method shows consistent improvements over competing approaches across datasets (Table 1, p. 7), especially in Dice and TRE.
- Scalability and efficiency The linear scaling with sequence length and low runtime overhead (p. 8) is a practical advantage over classical optimization-based methods.
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.
- Lack of explicit modeling of tissue properties While the method is applied to quantitative MRI, it does not explicitly incorporate tissue property models or physics-based constraints. Instead, robustness to varying contrast is handled implicitly via entropy-based losses (CTE) and feature aggregation. This limits interpretability and potentially misses opportunities for more principled integration with quantitative MRI signal models.
- Magnitude-only processing The method operates solely on magnitude images. Phase information, which can be relevant for motion estimation and quantitative imaging, is not considered.
- Limited analysis of deformation fields The paper does not provide sufficient qualitative or quantitative analysis of the predicted deformation fields. Visualizations or statistics (e. g. , displacement magnitude, anatomical plausibility) would strengthen confidence in the method.
- Relatively high variance in quantitative metrics Despite improvements in mean performance, the reported standard deviations remain relatively large (Table 1, p. 7), particularly for Dice scores in cross-protocol settings. This raises concerns about robustness and stability across subjects.
- Incremental novelty in components While the overall formulation is novel, individual components (shared encoders, correlation-based aggregation, entropy-based losses) build on existing ideas. The novelty lies more in the integration than in fundamentally new building blocks.
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
- It would strengthen the work to better connect the method to quantitative MRI signal models, potentially integrating physics-informed constraints rather than relying solely on entropy-based similarity.
- Including visualizations of deformation fields and failure cases would improve interpretability and trust in the method.
- A more detailed discussion of failure modes and variance across subjects would be valuable, given the relatively large standard deviations.
- Clarify whether the method could be extended to complex-valued MRI data and what limitations arise from magnitude-only processing.
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 presents a well-motivated and practically relevant contribution by introducing a set-based formulation for groupwise registration, addressing a key limitation of existing deep learning approaches. The demonstrated zero-shot generalization across datasets is particularly compelling.

However, the work has some limitations, including the lack of explicit integration of tissue property models, limited analysis of deformation fields, and relatively high variance in results. The methodological novelty is primarily in the formulation rather than individual components.

Overall, the paper is above the acceptance threshold, but not without weaknesses that could be addressed in future work.
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 proposes Any2Reg, a registration framework that models cardiac MRI sequences as unordered sets, enabling permutation-invariant, variable-length, and cross-protocol generalizable registration.
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’s main strengths include a well-motivated set-based formulation that alleviates dependence on sequence length and ordering, a reasonable correlation-guided aggregation design for handling contrast variation, and solid empirical results showing generalization across different MRI datasets.
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.Limited novelty over prior groupwise/PCA methods. The proposed set-based formulation is well-motivated, but it largely aligns with existing groupwise registration paradigms. In particular, the correlation-guided aggregation is quite close in spirit to PCA-based approaches (e.g., Huizinga et al., MedIA 2016; PCA-Relax, MICCAI 2024), and the paper does not clearly highlight a substantial methodological advance beyond these prior works.

2.Weak support for zero-shot generalization. The claim of strong generalization is somewhat under-supported. The evaluation on external datasets relies on nnU-Net-generated segmentations rather than ground truth, which makes it hard to disentangle registration performance from segmentation quality. In addition, the experiments do not cover more challenging domain shifts, so the generalization claim feels a bit overstated.

3.Incomplete comparison with recent strong baselines. The comparisons are mostly limited to older or closely related methods. It would strengthen the paper to include more recent and competitive registration approaches, as it is currently unclear how the method compares to the latest state of the art.
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.

(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 presents a well-motivated and technically sound approach with solid empirical results, particularly in handling variable-length sequences. However, the overall contribution appears somewhat incremental, and the evidence for strong generalization is not fully convincing, leading to a weak accept recommendation.
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 presents a group-wise registration method for quantitative cardiac MRI. The proposed method seeks to improve generalization across MRI sequences by: i) decoupling the methodology from sequence length and order, and ii) extracting contrast-insensitive features using a foundation model. Experiments highlight the model’s generalization to other datasets with acquisition protocols distinct from the training dataset. The proposed method outperforms 5 previous methods in terms of DICE and registration errors.
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 methodology is clearly explained, and the experimental evaluation is relevant and convincing: several recent methods are compared, and experiments are conducted on various datasets, including an out-of-distribution dataset.

The results show that the method’s major strength is its generalization across different acquisition protocols. The results indicate that this is mainly achieved thanks to the auxiliary anatomical features extracted from the foundation model in [30]. The fast execution times are also a practical advantage of the method.
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.
I) The model with FM outperforms the ablation w/o FM, showing that integrating the auxiliary features is at the core of the success of the proposed model, however: 1/The impact of the associated loss (choice of lambda_f) is not discussed. 2/Wouldn’t the best performing model be the one trained solely with contrast-insensitive auxiliary features (second term in the loss)? 3/ This major contribution does not appear in Figure 2 summarizing the proposed framework. For example, it is unclear how h’_theta and g are integrated into the framework described in Figure 2. 4/ The method is not compared to other methods using contrast-insensitive features, e.g., [17], to validate the other key elements of the methodology.

II) It is unclear whether the model performance is independent of the training sequence length L: 1/ T is constructed in the training phase through the L times L correlation matrix. 2/ Experimental results suggest that generalization is less effective for longer sequences (ASL, L=60). 3/ Performance is not reported for the much longer L=512 sequence presented in the scalability analysis.

III) Although the experimental results are convincing, it is unclear whether the proposed method is sufficiently novel: both the correlation-based canonical sequence and integrating contrast-insensitive features have been studied in the literature ([3], [9],[12],[23] ).

Other remarks:
- S_L is undefined.
- What is d_theta (undefined) in Figure 2?
- Why are the execution times for the best-performing model (IO variant) not reported for L=512?
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.

(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 experimental validation shows that the proposed method is promising in terms of generalization, a crucial factor for clinical applicability. However, the true impact of sequence length is not sufficiently discussed, and the contrast-insensitive approach is not compared with other methods using a similar approach. Finally, the novelty of the proposed method appears limited, as explained above.
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 #4

Please describe the contribution of the paper

The paper proposes a novel Groupwise Cardiac MRI registration method. The method is trained on one dataset and validated on two datasets.
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 idea of reformulating the group registration from multi-channel regression to set prediction is inspired.
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.Better validation of the effectiveness of correlation-guided aggregation: Simply comparing the mean is insufficient. The correlation matrix should be compared with attention mechanisms; otherwise, its breadth is limited.

2.Insufficient comparative experiments: Since foundation model-assisted segmentation is used, it should be compared with SAM-based registration methods.

3.The auxiliary anatomy-driven feature stream is not shown in Figure 2, thus lacking data flow and making it difficult to understand.

4.Lack of more details regarding ablation methods: What is Any2Reg w/o FM? Is that “without foundation model”? What is Any2Reg IO? Is that “with instance-optimization”? What are these specific details?
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?

Refer to the weaknesses above.
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

Author Feedback

We sincerely thank the reviewers and the meta-reviewer for their insightful comments and encouraging feedback. We appreciate the positive remarks on the novelty of Any²Reg’s set-based formulation, practical relevance for cross-protocol cardiac MRI registration, and computational scalability. We also appreciate the meta-reviewer’s summary that the paper addresses a key limitation of current learning-based groupwise registration: the hard coupling between network design, sequence length, and acquisition order. We respond to the main comments below.

Core contribution. We would like to echo that the major contribution of Any²Reg is indeed the reformulation of learning-based groupwise cardiac MRI registration as a set-to-set prediction problem with built-in permutation equivariance. Existing methods mainly use fixed-length channel-stacked inputs, coupling the architecture to protocol-specific length and order. This is intrinsically mismatched to quantitative MRI, where sampled observations can vary arbitrarily in length and order. Changing the modelling unit from an ordered stack to a set addresses this mismatch, both mathematically and practically. As such, our work is not a combination of existing components, but a reformulation of groupwise registration with broad applications. We note that standard pairwise registration can also be seen as a special case where L=2. We would like to clarify that numerical gains from the auxiliary foundation-model stream are not claimed as the core contribution. In Any²Reg, FM provides anatomy-related signals that are more stable under severe contrast variation, thereby improving robustness and accuracy. As stated in the paper, this feature stream is auxiliary. The core contribution remains the set-based formulation, independently validated by the w/o-FM model, which already generalises across unseen protocols and variable sequence lengths (Tab 1) while FM further improves performance by adding anatomical priors.

Evaluation and baselines. Our evidence is not based on automatic segmentations in most experiments: we used clinical expert annotations to evaluate registration on STONE, landmarks on STONE/MOLLI, and further evaluated T1 fitting accuracy, a clinically relevant proxy for registration quality. nnU-Net segmentations were used only for evaluation of external datasets to keep evaluation metrics consistent.

Our baselines cover a broad spectrum of groupwise registration strategies: classical optimization, fixed-length learning, PCA/correlation-based designs, and anatomy-guided variants. Correlation-based methods motivate correlation-aware representations, but traditional uses are not learnable and existing DL variants remain protocol-tied. Any²Reg instead incorporates correlation-aware input and loss features into a learnable, permutation-equivariant, variable-length framework.

Sequence length in training. The correlation matrix is computed dynamically per input set, and all learnable modules are frame-shared and independent of L. Thus, changing L changes only the input set and correlation matrix. The smaller ASL Dice gain reflects better initial breath-hold alignment, whereas cine with larger L shows consistent improvement when initial alignment is poor.

Other comments. R1: Physics-informed models or phase inputs are valuable extensions, but this study targets generic cardiac MR registration without protocol-specific assumptions, which can be unknown or inexact. We report std(log det J) with controlled folding (<1e-5) for deformation plausibility. R3: The L=512 test used pseudo data as a scalability pressure test. R4: SAM-style prompting is not directly comparable, since our FM stream provides ablated features rather than prompts.

Overall, Any²Reg is a simple, generic set-based reformulation that decouples sequence length, frame order, and protocol-specific contrasts from the learning model, demonstrating cross-protocol generalisation.

Meta-Review

Meta-review #1

Your recommendation

Provisional Accept
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 tackles a practical limitation of existing deep learning groupwise registration methods for quantitative cardiac MRI: their architectures hardcode sequence length and acquisition order, making them protocol-specific and non-transferable at inference time. Any²Reg reframes the problem as set-to-set prediction, enforcing permutation equivariance through shared encoders and a correlation-guided canonical reference, and optionally incorporating contrast-insensitive features from a pretrained foundation model. Trained on a single T1 mapping dataset and tests on T1 mapping under a different acquisition scheme, arterial spin labelling, and Cine MRI which are three datasets with meaningfully different sequence lengths and contrast behaviour.

Three reviewers recommended weak accept and one weak reject. The weak reject was driven largely by presentation gaps and requests for additional baselines rather than fundamental disagreement with the core contribution, and I weight it accordingly. The method is technically sound and the cross-protocol generalisation results are the strongest aspect of the paper, therefore I recommend acceptance

back to top

Set-Based Groupwise Registration for Variable-Length, Variable-Contrast Cardiac MRI

Author(s):