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

Abstract

Zero-shot self-supervised learning (ZS-SSL) for MRI reconstruction optimizes network parameters directly on the acquired undersampled measurements, eliminating the need for large training datasets. This paradigm mitigates domain shift issues commonly observed in dat-abase trained models when encountering unseen sampling trajectories, acceleration rates, contrasts, or anatomies, while offering improved interpretability relative to other zero-shot approaches. Nonetheless, current ZS-SSL methods face three major limitations: (1) data-splitting strategies introduce training-inference mismatches in effective acceleration rates, leading to suboptimal reconstruction; (2) over-parameterized networks incur substantial computational overhead and increase overfitting risk; and (3) the inherently limited supervision further constrains achievable performance. We propose ECHO (Estimated Composite Hybrid Observation), another ZS-SSL MRI reconstruction framework designed to address these challenges. Instead of partitioning acquired k-space data, ECHO constructs training inputs from intermediate reconstructions that are statistically consistent with inference-time observations, thereby eliminating acceleration mismatch. Moreover, ECHO employs an under-parameterized, physics-informed architecture tailored to the zero-shot regime, significantly reducing computational complexity while maintaining reconstruction accuracy. Extensive experiments demo-nstrate that ECHO consistently improves reconstruction quality and achieves substantial training speedup. Our code is publicly available at {https://github.com/ShanghaiTech-Hu-Lab/ECHO}

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1663_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ShanghaiTech-Hu-Lab/ECHO

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ShaWen_ECHO_MICCAI2026,
        author = { Shang, Wenlei AND Liu, Wenjian AND Hu, Peng AND Zhou, Zijian},
        title = { { ECHO: Estimated Composite Hybrid Observation for Zero-Shot Self-Supervised MRI Reconstruction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16888},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper studies zero-shot self-supervised learning (ZS-SSL) for MRI reconstruction, with the specific goal of reducing the training–inference mismatch caused by k-space splitting in prior self-supervised frameworks such as SSDU-based zero-shot approaches. The main contribution is a new framework, ECHO (Estimated Composite Hybrid Observation), which avoids direct partitioning of the acquired undersampled k-space into separate input and supervision subsets. Instead, the method constructs a hybrid observation from an intermediate reconstruction estimate combined with the acquired measurements, and then re-undersamples this hybrid k-space using a secondary mask drawn from the same distribution as the acquisition mask. This design is intended to make the training input statistically more consistent with the inference-time input while preserving self-supervised training. The paper also argues that, in the zero-shot setting, an under-parameterized physics-informed network is better suited than larger over-parameterized architectures, and provides experiments on fastMRI knee and brain data showing improved SSIM/PSNR and substantially faster training than competing zero-shot baselines. Additional analyses examine training dynamics, the effect of the number of randomized masks, and robustness to backbone hyperparameters.

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

    1.The paper addresses a genuine weakness of existing zero-shot/self-supervised MRI reconstruction methods. The training–inference mismatch induced by data splitting is a real methodological issue in SSDU-style ZS-SSL. The paper clearly identifies this as the central bottleneck and proposes a concrete mechanism to mitigate it. This makes the work well motivated and relevant to the zero-shot reconstruction literature. 2.The proposed idea is conceptually clean and easy to understand. ECHO replaces direct data splitting with hybrid-k-space construction based on an estimated intermediate reconstruction and secondary re-undersampling. This is a relatively intuitive idea, and the paper explains it clearly through both text and the schematic in Fig. 1, which makes the training pipeline easy to follow. 3.The experimental results are consistently strong across the reported settings. On fastMRI knee and brain data, ECHO achieves the best SSIM and PSNR across all reported 4× and 6× settings compared with TV, vanilla DIP, self-guided DIP, SIREN, and Instant-NGP. The gains are not isolated to a single anatomy or acceleration factor, which strengthens the empirical case for the method. 4.The paper includes several useful analyses beyond headline performance. In addition to the main comparison, the authors compare ECHO with SSDU-, SPICER-, and upper-bound-style input constructions, and show that ECHO approaches the upper-bound setting. The paper also includes mechanistic analyses of hybrid k-space evolution, mask संख्या effects, and network-capacity robustness. These analyses make the paper more convincing than a purely benchmark-driven submission. 5.The design choice of using an under-parameterized physics-based network is well aligned with the zero-shot setting. This is not just a computational convenience; it is also a sensible modeling decision because only one undersampled acquisition is available for optimization. The paper’s argument that over-parameterized models are less suitable in this regime is plausible and supported by the reported stability analysis. 6.Clarity is generally good. The writing is understandable, the method section is logically organized, and the paper does a good job of connecting the proposed algorithm to the intuition of alternating optimization.

  • 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 methodological novelty is moderate rather than strong. The core idea of improving self-supervised input construction to reduce the training–inference gap is meaningful, but the overall framework still feels like an incremental extension of prior ZS-SSL / SSDU-style thinking rather than a fundamentally new reconstruction paradigm. The method combines intermediate reconstruction, secondary masking, and Siamese-style parameter copying in a reasonable way, but the conceptual jump beyond existing self-supervised formulations is somewhat limited. 2.The baseline comparison is not fully balanced with respect to the main claimed contribution. The main baselines in Table 1 are TV, DIP variants, SIREN, and Instant-NGP, but the paper’s central claim is specifically about improving zero-shot self-supervised reconstruction relative to SSDU-like ZS-SSL. Although SSDU/SPICER/upper-bound comparisons are discussed later, they are not presented as prominently or as completely as the main benchmark table. A clearer and more direct head-to-head quantitative comparison against strong ZS-SSL baselines would strengthen the paper. 3.The experimental scale is relatively limited. The paper uses twenty subjects per dataset with ten slices per subject, and hyperparameters are selected based on validation in one subject with ten slices. While zero-shot methods are evaluated per subject and do not require large training cohorts, this still makes the empirical evidence narrower than ideal, especially for claims about robustness and general applicability. 4.The architecture itself is fairly simple, and the paper does not fully disentangle how much of the improvement comes from the hybrid-observation strategy versus other implementation choices. The method uses a compact unrolled physics-based network, which is reasonable, but stronger ablations isolating the contribution of each algorithmic ingredient would make the evidence cleaner. In particular, more explicit comparisons between using the same backbone with and without ECHO-style hybrid input construction would be helpful. 5.Some of the mechanistic claims are suggestive rather than rigorous. For example, the alternating-optimization interpretation is presented as an intuition or hypothesis rather than a formal derivation, and the DreamSim/t-SNE visualizations are helpful qualitatively but not a strong substitute for more principled theoretical analysis. This does not invalidate the method, but it limits how strongly one can interpret the mechanistic explanation. 6.Reproducibility details are decent but not exhaustive. The paper reports the backbone structure, training steps, and mask settings, but some practical details of the optimization and implementation remain somewhat concise for exact reproduction.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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?

    My overall recommendation is based on a balance between a well-motivated problem, clear empirical gains, and moderate methodological novelty. The paper identifies a real weakness of existing zero-shot self-supervised MRI reconstruction methods, namely the training–inference mismatch induced by data splitting, and proposes a conceptually clean solution based on hybrid observation construction and secondary re-undersampling. This is a meaningful contribution because it directly targets a limitation of prior ZS-SSL formulations rather than introducing complexity for its own sake.

    The main factors limiting my score are that the contribution feels more like a strong refinement of existing zero-shot self-supervised ideas than a major conceptual leap, and that the evaluation could be more comprehensive and more directly centered on the strongest relevant ZS-SSL baselines. In addition, some of the mechanistic interpretations are plausible but remain somewhat heuristic. Overall, I view the paper as a solid and useful methodological contribution with convincing empirical evidence, but not one with the level of novelty or depth that would place it among the strongest submissions.

  • 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 study presents ECHO (Estimated Composite Hybrid Observations), a zero-shot self-supervised learning framework for the task of MRI reconstruction. The novelty of this work resides in addressing the problem of training/inference mismatch associated with data splitting techniques employed by previous works, such as SSDU, in the context of zero-shot self-supervised learning tasks. In contrast with earlier studies, instead of splitting the data in k-space, this study utilizes hybrid observations derived from intermediate reconstructions that statistically match the inference inputs. Moreover, an under-parametrized physics-informed model architecture has been developed specifically for zero-shot scenarios.

  • 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 highlights one of the common issues associated with zero-shot MRI reconstruction, which is the distribution mismatch brought about by the use of data splitting techniques. The hybrid k-space construction approach employed is both logical and appropriate and provides a good alternative to SSDU techniques. The use of the under-parameterized physics-informed architecture is well justified in the zero-shot scenario and enhances efficiency. The experiment conducted is well designed, with experiments being run against a number of baseline models such as DIP, SIREN, and Instant-NGP. The proposed method outperforms these models in terms of SSIM and PSNR metrics. Experiments run to provide an understanding of different aspects of the model, such as performance under the influence of various randomized mask sizes and other relevant aspects, have been done. The finding that ECHO converges to upper bound settings is very impressive.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    This aspect of the paper is somewhat underexplained, making it difficult for readers to reproduce the technique themselves. Although the alternating optimization concept offers a great deal of insight, it still is somewhat theoretical, and an algorithmic approach would help. As for experimental results, their generalization power is only assessed using the fastMRI dataset, which leaves room for doubt concerning their applicability to other problems. In addition, although the approach itself appears to be computationally efficient, no information regarding its comparison to other state-of-the-art methods in terms of time or memory consumption is provided. With regards to the novelty of the presented study compared to other research in the field, such as SSDU and SPICER, it is rather subtle, especially with regards to the idea behind it being data split-avoidant.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    This paper proposes an interesting way towards ZSR reconstruction of MRI images. For future improvements on this topic, it is recommended that the authors clarify their methodology section, especially concerning the hybrid k-space construction process and the training method. The inclusion of pseudocode or a detailed flowchart will make the paper more accessible. In addition, the authors might want to add some computations between the proposed method and other methods as well as evaluate their technique in various sampling or noisy cases. Lastly, the novelty compared to previous ZS-SSL methods could be elaborated further.

  • 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 work considers a significant issue in MRI reconstruction and presents an excellent technical and practical approach to solving the problem. The concept of overcoming the discrepancy between training and inference by means of hybrid observation construction is meaningful and highly substantiated by impressive experiment results. It also shows stable superiority against several competing methods and even provides meaningful ablations. Nevertheless, there are certain drawbacks in the clarity of presentation as well as experimental evaluation and computational complexity analysis. In spite of that, the work makes a valuable contribution. Thus, it can be considered marginally acceptable but not fully acceptable.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper proposes ECHO, a zero-shot self-supervised MRI reconstruction framework that aims to remove the training–inference mismatch caused by data splitting in prior ZS-SSL / SSDU-style methods. The key idea is to construct a composite hybrid observation from an intermediate k-space estimate and the acquired undersampled measurements, and then apply a secondary randomized undersampling mask drawn from the same distribution as the acquisition mask. The paper further advocates an under-parameterized physics-based network tailored to the zero-shot setting, with the goal of reducing overfitting and computational cost. The method is evaluated on fastMRI knee and brain data, and the paper reports better SSIM/PSNR than TV, DIP variants, SIREN, Instant-NGP, and improved performance relative to SSDU/SPICER-style training constructions.

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

    1.Problem formulation is well-motivated and relevant. The paper identifies a concrete weakness of prior zero-shot self-supervised MRI reconstruction: the mismatch between training-time inputs created by k-space splitting and inference-time use of all acquired samples. This is a real methodological issue in ZS-SSL and is clearly articulated.

    2.The adaptation of the Siamese-style training scheme to the zero-shot setting is technically clean. Operating on a single scan without cross-sample statistics or SSDU-style pretraining initialization introduces non-trivial constraints (extreme data scarcity, heightened overfitting risk), and the authors’ design choices—under-parameterized physics-based backbone, hybrid observation with DC projection, fixed training budget—appear well-suited to this regime.

    3.Quantitative gains appear consistent across datasets/settings reported. In Table 1, ECHO is best across fastMRI knee/brain at 4× and 6× acceleration, and the paper also claims a small gap to the “upper-bound” training setting.

    4.The ablation studies (Sec. 5) are reasonably thorough, covering training dynamics of the hybrid k-space, the impact of the number and type of randomized masks, and robustness to network architecture choices.

  • 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.Insufficient discussion of closely related prior work (SiamRecon). The core training strategy of ECHO—using an estimator network to produce an intermediate full k-space estimate, applying random re-undersampling to this estimate, training a target (Siamese) network with stop-gradient, and interpreting the procedure as alternating optimization—closely parallels the SiamRecon framework (Sun et al., “Exploring Siamese Networks in Self-Supervised Fast MRI Reconstruction,” IEEE JBHI, 2025). The manuscript itself acknowledges the Siamese architecture (Sec. 3, “similar to a Siamese network”), yet does not cite or discuss SiamRecon. The specific overlapping components include: (a) the estimator/target Siamese structure with shared weights, (b) k-space resampling of the intermediate reconstruction as the augmented input, (c) stop-gradient on the estimation branch, (d) the alternating-optimization / EM-like theoretical interpretation, and (e) using acquired k-space as the supervision target. We note that SiamRecon appeared on arXiv in January 2025, and the authors may not have been aware of it at the time of writing. We also acknowledge that ECHO operates in a zero-shot (single-scan optimization) setting rather than a dataset-level self-supervised setting, and this distinction is not trivial: zero-shot precludes SSDU-style pretraining initialization (which SiamRecon relies on, cf. Algorithm 1 therein), limits available data to a single scan, and demands careful network capacity control. Nonetheless, the high-level training strategy and its theoretical framing are substantially shared, and the paper does not position itself as an adaptation of such a paradigm but rather as a novel framework. This significantly diminishes the perceived novelty. I encourage the authors to clearly discuss SiamRecon, acknowledge the shared methodological lineage, and articulate what is fundamentally new beyond adapting the paradigm to the zero-shot regime—for instance, the removal of pretraining dependence, the hybrid observation with explicit DC projection, or the under-parameterized design rationale.

    2.The “alternating optimization” interpretation is heuristic rather than rigorous. The paper explicitly says “we hypothesize that ECHO behaves similarly to alternating optimization” and then gives an informal derivation. This is useful intuition, but not a theoretical guarantee. A similar EM-like interpretation was provided in SiamRecon; however, we note that the zero-shot setting does introduce a different optimization landscape (single-sample, no pretraining warm-start), and the paper could have explored whether the AO/EM convergence properties differ in this regime. As it stands, the theoretical contribution remains underdeveloped for a methodology paper.

    3.Empirical validation is still somewhat limited in scope. The experiments use fastMRI knee and brain, with 20 subjects and 10 slices per subject, and hyperparameters are selected on one subject with 10 slices. This is enough for an initial study, but still feels narrow for a top-tier MICCAI methods paper, especially given the claim of broad robustness across sampling settings and architectures. External datasets, more anatomies, or more realistic acquisition diversity would strengthen the case.

    4.The computational efficiency claim is not documented strongly enough. The abstract claims “substantial training speedup,” and the method advocates an under-parameterized model, but the paper does not provide a clear runtime/memory table, wall-clock comparison, or parameter-count benchmark against the most relevant competing methods. Without explicit quantitative evidence, the efficiency claim is weaker than it could be.

    5.Reproducibility details are incomplete. The manuscript gives high-level architecture and training steps, but important details appear missing or under-specified: optimizer, learning rate schedule, sensitivity map estimation procedure, exact random-mask generation protocol, random seeds, stopping/selection details beyond fixed 5,000 steps, and whether code/models will be released.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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

    I identified SiamRecon (Sun et al., IEEE JBHI, 2025) as a highly relevant prior work after completing my initial review. The degree of methodological overlap between ECHO and SiamRecon is substantial, and I believe the authors should thoroughly discuss this work, clearly delineate the differences, and reposition the contribution accordingly. If the authors can convincingly demonstrate that the adaptation to the zero-shot setting introduces non-trivial technical challenges or insights beyond what SiamRecon addresses—such as the elimination of pretraining dependence, the hybrid observation construction with explicit DC projection, or the principled under-parameterized network design for single-scan optimization—this would help strengthen the novelty claim. I remain open to revising my assessment based on the authors’ response.

  • 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?

    My original assessment leaned toward weak accept, as I considered the training strategy a meaningful—if incremental—contribution, supported by competent experiments and consistent quantitative gains. However, upon identifying SiamRecon (Sun et al., IEEE JBHI, 2025) as a closely related and already published work that shares the core methodological components (Siamese estimator/target architecture, k-space resampling of intermediate reconstructions, stop-gradient, and alternating-optimization interpretation), the novelty of the proposed framework is substantially diminished. The contribution is now better characterized as adapting an existing paradigm to the zero-shot setting rather than proposing a new reconstruction strategy. I acknowledge that the zero-shot adaptation is not entirely trivial—it removes the pretraining dependence, operates under extreme data scarcity, and requires deliberate network capacity control—but these aspects are not sufficiently highlighted or analyzed in the paper to constitute a strong standalone contribution. Combined with the previously noted limitations—heuristic theoretical justification, narrow experimental scope, insufficient efficiency documentation, and incomplete reproducibility details—the overall contribution no longer meets the acceptance threshold in my assessment. I am adjusting my score to weak reject, but I remain open to reconsidering if the authors can clearly differentiate their work from SiamRecon and demonstrate that the zero-shot adaptation introduces substantial new methodological insights.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Author Feedback

We thank the reviewers and AC for their thorough review. Several major concerns are addressed below. 1.Comparison with SiamRecon (Sun et al., arXiv 2025) Although the two frameworks appear similar, the underlying motivation, implementation design, and theoretical basis are all different. Our re-undersampling mask shares the same distribution as the acquired mask; SiamRecon employs a 2D uniform mask and treats the acceleration factor as a hyperparameter. However, Millard & Chiew (weighted SSDU, TCI 2023) show that 2D masks can cause unstable training, likely explaining why SiamRecon requires pre-training. The mismatched acceleration (unified SSDU, Xu et al, arXiv 2026) and subsampling type (ours Fig. 3) can also degrade reconstructions quality in SiamRecon. Moreover, Siamese frameworks are not the theoretical basis for self-supervised reconstruction, Noise2Noise is. The “homology” prior that SiamRecon invokes is also not unique, in SSDU, the network input is already homologous to the supervision target (both are partial k-space of the same image), meaning fully-sampled estimates are not essential. In ECHO, by contrast, the estimate serves a specific, well-defined purpose: aligning the training-inference distributions. Finally, SiamRecon draws an analogy to EM, but its “E-step” is a deterministic estimate rather than a posterior expectation. Their analogy to SimSiam (Chen & He, CVPR 2021) is also inexact: SimSiam’s latent variable is the network output (admitting a closed-form solution under mean-squared-error), whereas SiamRecon’s latent variable is the network input (requiring the Jacobian). SiamRecon bypasses the Jacobian and adopts the closed-form update, yet its derivation does not justify for the introduction of fully sampled k-space. Nor does the “homology prior” necessitate it. On the other hand, ECHO is based on a solid derivation from previous work (weighted SSDU), which shows that: (1) ideal inputs need fully-sampled k-space, and (2) the self-supervised loss minimizer approximates the supervised one. Together, these support that a self-supervised network can learn a prior that is sufficient to map undersampled k-space to an approximation of the fully-sampled image, grounding our use of the network output as a surrogate for the fully sampled image. We further show its self-consistency: at convergence, if the network parameters are ideal, then its output using the undersampled k-space is the fully sampled image, and the approximation becomes exact (R1.5, R3.2) Therefore, even without considering the zero-shot situation, ECHO is fundamentally different from SiamRecon, with different motivation, mid-step implementation, and theoretical deviation. 2.Conceptual novelty We acknowledge that compared with DIP and INR networks, ECHO represents a relatively incremental extension of mask manipulation. Nevertheless, it is the first that uses a latent variable as a fully sampled image estimate for zero-shot self-supervised reconstruction, addressing the long-standing distribution gap in SSDU. It was also directly compared with DIP and INR in both reconstruction quality and computation efficiency, providing a comprehensive comparison across zero-shot approaches. 3.Experimental Scope ECHO optimizes per sample without any priors, making it inherently target-agnostic. Our evaluation on 200 independent images exceeds published zero-shot studies (e.g., 25 in Self-Guided DIP, Liang et al., TCI 2025; 50 in AeSPa, Joo et al., CVPR 2025) (R1.3, R2, R3.3, AC). We do not directly compare with existing ZS-SSL methods as they are prohibitively slow (~45 min/sample shown in AeSPa). The training time of ECHO is under 2 minutes, comparable to DIP-based methods, though a gap remains with INR-based methods (~20s) due to tiny-cuda-nn (R1.2, R2, R3.4). Using the same backbone, our SSDU/SPICER comparison directly isolates the benefit of ECHO’s training strategy (R1.4). We will promote this to the main results table (R1.2). Code will be open-sourced



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 paper presents a well-motivated approach to zero-shot self-supervised MRI reconstruction, with promising empirical results and generally positive feedback from reviewers. However, concerns remain regarding novelty, experimental scope, and positioning with respect to closely related prior work. I invite the authors to address these points in the rebuttal to clarify the contribution and strengthen the submission.

  • 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.

    I think the authors have adequately addressed the reviewers’ comments.



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.

    The main concerns prior to rebuttal were related to published work (SiamRecon). The author’s response is convincing and clarified their methodology’s difference. I recommend acceptance.



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.

    I recommend acceptance. The paper addresses a clear limitation in zero-shot self-supervised MRI reconstruction: the training-inference mismatch introduced by k-space splitting in SSDU-style methods. ECHO proposes a clean hybrid-observation construction strategy to reduce this mismatch, and the reported results show consistent improvements over several zero-shot baselines on fastMRI knee and brain.

    The main concern after review was the relationship to closely related prior work, especially SiamRecon. The rebuttal provides a reasonable clarification: ECHO targets the zero-shot setting and the distribution mismatch caused by self-supervised splitting, whereas SiamRecon is positioned around dataset-level self-supervised training and a different masking/resampling setup. The contribution remains incremental rather than a major new reconstruction paradigm, but the problem is important, the method is practical, and the empirical evidence is convincing enough for acceptance.



back to top