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

Abstract

Real-world MRI reconstruction systems face the open-world challenge: test data from unseen imaging centers, anatomical structures, or acquisition protocols can differ drastically from training data, causing severe performance degradation. Existing methods struggle with this challenge. To address this, we propose BiasRecon, a bias-calibrated adaptation framework grounded in the minimal intervention principle: preserve what transfers, calibrate what does not. Concretely, BiasRecon formulates open-world adaptation as an alternating optimization framework that jointly optimizes three components: (1) frequency-guided prior calibration that introduces layer-wise calibration variables to selectively modulate frequency-specific features of the pre-trained score network via self-supervised k-space signals, (2) score-based denoising that leverages the calibrated generative prior for high-fidelity image reconstruction, and (3) adaptive regularization that employs Stein’s Unbiased Risk Estimator to dynamically balance the prior-measurement trade-off, matching test-time noise characteristics without requiring ground truth. By intervening minimally and precisely through this alternating scheme, BiasRecon achieves robust adaptation with fewer than 100 tunable parameters. Extensive experiments across four datasets demonstrate state-of-the-art performance on open-world reconstruction tasks.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0763_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiuJiy_Open_MICCAI2026,
        author = { Liu, Jiyao AND Gao, Shangqi AND Liu, Lihao AND Ning, Junzhi AND Wei, Jinjie AND He, Junjun AND Zhuang, Xiahai AND Xu, Ningsheng},
        title = { { Open World MRI Reconstruction with Bias-Calibrated Adaptation } },
        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

    The paper proposes a novel calibration-based method to address the distribution shift problem in MRI reconstruction, such that the pre-trained models can be adapted to new datasets without heavy retraining.

  • 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 problem of distribution shift is comprehensively addressed at both the prior side and the hyperparameter side. 2.The proposed solutions are well driven by experimental findings and analysis, including distribution differences in layers and frequencies. 3.

  • 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 compared methods are relatively old and not competitive. 2.The diversity of investigated datasets is not sufficient, given the claims of “open world”. All datasets are retrospectively undersampled. 3.Lack of information for reproducibility, including training details of the compared methods and the autocalibration signal size.

  • 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 provide sufficient information for 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

    1.Diffusion-based models are known to be robust for many distribution shifts in MRI. So from fastMRI knee to Stanford Knee or to fastMRI brain is no surprise. Such transfer has been demonstrated years ago, e. g. in CSGM by Jalal et. al. Using CMRxRecon dataset is a good demonstration, but real challenges can be DWI data, low-field data, and breast/prostate data, which all have open data available. It raises concerns about whether this method works fine on these data. There is even no discussion about the limitations. 2.It is odd to compared with MoDL instead of VarNet, as the latter is shipped with fastMRI official repository. Also, MRI-specific diffusion-based methods, such as CSGM, AdaDiff, and Nila, should be compared.

  • 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?

    Despite the weakness, I am, in general, positive to this study. The idea is evidence-based, and the presentation is clear.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    The authors did not provide a valid justification for comparing with the old methods, especially considering that VarNet, which is the default method recommended by the employed fastMRI dataset, has been intentionally avoided.



Review #2

  • Please describe the contribution of the paper

    This paper proposes BiasRecon, a test-time adaptation framework for open-world MRI reconstruction. The main idea is to adapt a pretrained score-based reconstruction model to unseen domains (cross-anatomy, cross-center, cross-modality) by minimally modifying only the components that are sensitive to distribution shift. The method introduces a parameter-efficient frequency-guided calibration mechanism that modulates low- and high-frequency components of intermediate features using a small number of learnable scalars, combined with a SURE-based adaptive regularization strategy to adjust the data-prior balance at test time without ground truth. The framework is formulated as an alternating optimization procedure over image reconstruction, calibration parameters, and regularization weight. The method is evaluated on multiple datasets and demonstrates consistent improvements over several baselines under distribution shift.

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

    S1: Addresses an important and practical problem. The focus on open-world MRI reconstruction and test-time adaptation under distribution shift is highly relevant for real clinical deployment scenarios. S2: Clear and intuitive design principle. The “minimal intervention” idea (preserve transferable features, adapt domain-specific ones) is conceptually appealing and easy to understand. S3: Parameter-efficient adaptation. The method adapts fewer than 100 parameters, which is a strong practical advantage compared to full model fine-tuning. S4: Self-supervised adaptation. The use of k-space masking and SURE allows adaptation without ground-truth target data, which aligns well with realistic deployment settings. S5: Combination of two complementary adaptation mechanisms. The joint use of prior calibration and adaptive regularization is well-motivated and supported by ablation results. S6: Evaluation across multiple distribution shifts. The experiments include cross-anatomy, cross-center, and combined shifts, which is stronger than many prior works that only consider a single OOD scenario.

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

    W1: Limited novelty. The method mainly combines existing ideas (score-based reconstruction, feature modulation, self-supervised masking, SURE-based tuning) in a reasonable way, but the individual components are not fundamentally new. The paper does not clearly demonstrate a strong conceptual advance beyond prior test-time adaptation and generative prior works. W2: Missing strong and targeted baselines. The comparison lacks test-time adaptation baselines (e.g., fine-tuning, entropy minimization, parameter-efficient adapters) and stronger modern reconstruction methods. It is unclear whether the proposed approach outperforms simpler or more standard adaptation strategies. W3: No comparison to straightforward fine-tuning. A key missing experiment is adapting the full model (or partial layers) using the same self-supervised loss. Without this, it is unclear whether the proposed parameter-efficient calibration is actually necessary. W4: Weak justification of the core assumption. The claim that deeper/low-frequency components are domain-specific and shallow/high-frequency components are domain-agnostic is mainly supported by qualitative visualization. There is no strong quantitative validation of this assumption. W5: Unclear frequency decomposition. The method relies on splitting features into low/high-frequency components, but the exact implementation (cutoff, definition, consistency across layers) is not clearly described. The connection between feature-space frequency and MRI k-space frequency is also not well justified. W6: Questionable assumptions in SURE. The use of SURE assumes Gaussian noise and certain differentiability properties, which may not hold in accelerated MRI reconstruction with nonlinear sampling and diffusion-based models. This is not discussed or justified. W7: Missing runtime and efficiency analysis. Test-time adaptation introduces additional optimization steps (δ and γ updates plus diffusion sampling), but no runtime, convergence, or computational cost analysis is provided. This is critical for practical use. W8: Potential error in reported results. The LPIPS value for MoDL in Table 2 appears inconsistent and likely incorrect. W9: Insufficient reproducibility details. Important implementation details are missing, including frequency split definition, optimization schedules, hyperparameters, and stopping criteria. W10: No statistical significance analysis. Results are reported as mean ± std, but no significance testing is provided to support claims of improvement. W11: No analysis of limitations or failure cases. The paper does not discuss when the method may fail or how performance degrades under extreme shifts.

  • 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 provide sufficient information for 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

    1- Important and practical problem (open-world MRI reconstruction under distribution shift) 2- Clear and intuitive overall idea (minimal intervention, parameter-efficient adaptation) 3- Needs clearer differentiation from prior work (test-time adaptation, parameter-efficient tuning, generative priors) 4- Missing strong baselines (no comparison to TTA methods, no fine-tuning baseline, no adapter-style methods) 5- Critical missing experiment: full/partial model fine-tuning with same self-supervised loss 6- Frequency decomposition unclear (no definition of cutoff, how applied in feature space, sensitivity not analyzed) 7- Weak justification of core assumption (low-frequency vs high-frequency transferability mostly qualitative) 8- SURE assumptions not discussed (non-Gaussian noise, diffusion model differentiability) 9- No runtime / inference cost / convergence analysis (important for clinical usability) 10- Missing implementation details (optimization schedule, hyperparameters, stopping criteria, CG settings)- Reproducibility limited without code or pseudocode 11- Potential issue in reported metrics (Table 2 value needs verification) 12- No statistical significance testing 13- No discussion of failure cases or limitations Clarify terminology (e.g., “bias” not formally defined)

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

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    1- Important problem and relevant for real-world deployment- Conceptually clean and reasonable framework 2- However, novelty is moderate (combination of existing ideas rather than new formulation) 3- Evaluation is not strong enough (missing key baselines and fair comparisons) 4- Core assumptions not rigorously validated 5- Key components not clearly defined (frequency decomposition, SURE usage) 6- Reproducibility concerns due to missing implementation details 7- Overall promising but not yet at MICCAI acceptance level Could move toward acceptance if rebuttal addresses baselines, clarity, and validation rigor

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    The rebuttal addressed most of my concerns. Although some concerns remain regarding incremental novelty and the heuristic frequency cutoff, I now lean toward acceptance. Thanks.



Review #3

  • Please describe the contribution of the paper

    The paper proposes BiasRecon, a test-time adaptation framework for open-world MRI reconstruction using a pretrained score-based generative prior. The key idea is to perform minimal, frequency-aware calibration of the prior by introducing only a small number of layer-wise low-/high-frequency modulation parameters, together with SURE-based adaptation of the data-consistency regularization parameter. The method targets reconstruction under distribution shifts across anatomy, center, and acquisition/sampling settings without requiring ground-truth images at test time.

  • 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 an important and clinically relevant problem. 2.It addresses two types of mismatch: target distribution mismatch and hyperparameter mismatch. 3.The method is well motivated, and the use of frequency guidance is intuitive and reasonable. 4.The unified alternating optimization framework, which jointly performs prior calibration and SURE-based regularization adaptation in a fully self-supervised manner without requiring ground truth, making it well aligned with realistic clinical scenarios.

  • 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 proposed frequency-guided calibration is rather coarse-grained, as it only introduces two scalar variables per layer to modulate low- and high-frequency components, which may be insufficient to capture complex domain shifts. 2.Although the number of tunable parameters is very small, the paper lacks a careful analysis of runtime and convergence curve. 3.It is somewhat counter-intuitive that, in Table 1, the Cross-Anatomy setting consistently yields higher PSNR/SSIM than the in-distribution FastMRI-Knee setting across almost all methods. Could the authors clarify whether this is due to intrinsic differences in dataset difficulty or reflecting the actual severity of distribution shift? Minor: 1.In Table 2, the LPIPS value reported for MoDL under 8× Uniform 1D sampling is unusually low compared with the other results, and is even lower than the corresponding 4× setting. The authors are encouraged to verify whether this value is correct and, if so, provide an explanation for this behavior. 2.In the “Robustness against undersampling” section, the mention of “Gaussian2D” appears to be a typo and should likely be “Gaussian1D.”

  • 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.

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

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

    The paper tackles an important real-world problem, and the proposed method is technically well motivated, practically relevant, and supported by promising results across several cross-domain MRI reconstruction settings.

  • 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

Reviewer # (R1) Thank you for reviewing. 1) Baselines: MoDL and VarNet are unrolled supervised reconstructions; we use MoDL as representative. For diffusion methods, we evaluated AdaDiff. On fastMRI-Brain cross-center (8×, PSNR/SSIM/LPIPS): DDIP 31.02/0.825/0.161, AdaDiff 31.92/0.775/0.226, ours 33.69/0.885/0.139.CSGM and Nila apply a diffusion prior without test-time adaptation, conceptually equivalent to DDS. 2) Open-world scope: Experiments cover representative cross-anatomy, cross-center, and combined shifts, though not all MRI domains. 3) Reproducibility: Score model trained 1M iterations, lr=1e-4.Inference uses DDIM 100, CG 5, delta/gamma Adam lr 1e-4/0.3, with 8%/4% ACS lines for 4x/8x.

Reviewer #2 (R2) Thank you for reviewing our paper; we address each concern below. W1) Novelty: We study how distribution shifts affect reconstruction and propose a minimal-intervention test-time framework that jointly calibrates the source-domain score prior and adapts data-prior balance with <100 tunable parameters. W2/W3) Baselines & fine-tuning: See R1-1.Besides, on FastMRI-Brain cross-center 8× (PSNR/SSIM/LPIPS): full-model fine-tuning 31.12±3.85 / 0.843±0.071 / 0.201±0.063, LoRA 32.34±2.47 / 0.861±0.048 / 0.178±0.039, ours 33.65±1.42 / 0.885±0.019 / 0.136±0.025, showing lower variance and higher mean performance. W4) Frequency decomposition and t-SNE. Although t-SNE is qualitative, it reveals a consistent pattern: deeper layers and low-frequency components exhibit domain-specific characteristics, while shallow layers and high-frequency components remain domain-agnostic (as in Sec. 2.2). We further conducted a calibration parameter evolution analysis: in bottom layers, α (low-frequency modulation) decreases while β (high-frequency modulation) increases as denoising progresses. This quantitatively confirms the t-SNE observation. W5)We decompose features via 2D FFT with a fixed 50% cutoff uniformly across all layers. This is consistent with the standard k-space interpretation: low frequencies encode global structure and contrast, while high frequencies encode fine details and edges. Domain shifts primarily affect the former. W6) SURE assumptions. Monte-Carlo SURE [3] supports parameter selection for black-box/iterative denoisers without analytic Jacobians, and deep MRI [2] and diffusion MRI reconstruction [1] have also used SURE for risk estimation and hyperparameter tuning. The Gaussian assumption follows the standard complex Gaussian k-space noise model. The divergence/Jacobian term is estimated by Monte-Carlo finite perturbations of the fixed mapping h(gamma; .), so a closed-form Jacobian through the diffusion sampler is not required. References: [1] 10.1007/978-3-031-43898-1_20; [2] 10.1109/TMI.2020.3025065; [3] 10.1109/TIP.2008.2001404.W7/W9) Runtime & implementation: See R1-3.On A100, batch 1: DDS ~10s/slice, BiasRecon ~20s/slice. Delta updates 70 steps, gamma 100 (stabilizes ~60). ESPIRiT via sigpy Python package. W8) Table 2 typo: MoDL LPIPS 8x Uniform1D corrected to 0.200.W10) Significance: Paired t-tests show statistical significance and will be reported. W11) Limitations. BiasRecon may be less effective under extreme OOD shifts or very high acceleration, e.g., 16x, where the self-supervised signal is weak. Terminology: “Bias” = test-time mismatch between pretrained source-domain score prior and target distribution.

Reviewer #3 (R3) Thank you for reviewing our paper. 1) Coarse calibration: Two scalars is minimal-intervention to reduce overfitting. 2) Runtime/convergence. Please see R2-W7/W9 and W4. 3) Cross-Anatomy higher than ID. This is due to dataset difficulty rather than weaker shift. Even zero-filling on FastMRI-Brain has about 3.85 dB higher PSNR than FastMRI-Knee, likely because brain slices are smoother and contain less fine texture. Minor) MoDL Result. Please see R2-W8 , and “Gaussian2D” should be “Gaussian1D”.

We hope the above responses have adequately addressed the raised concerns.



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 reviewers are mixed but generally recognize the relevance of the problem, with scores of 4, 3, and 5.The paper addresses an important practical issue in MRI reconstruction: adapting pretrained reconstruction models to distribution shifts without target-domain ground truth. The proposed parameter-efficient calibration and SURE-based regularization adaptation are well motivated and potentially useful for clinical deployment.

    However, several points require clarification before a final decision. The rebuttal should address the lack of strong adaptation baselines, including standard test-time adaptation, partial or full fine-tuning with the same self-supervised loss, and adapter-style approaches. This is important to establish whether the proposed small-parameter calibration is necessary and advantageous over simpler alternatives.

    The authors should also clarify the frequency decomposition procedure, justify the assumptions behind low-/high-frequency calibration, and discuss the assumptions required for SURE in this nonlinear reconstruction setting. Additional details on optimization schedules, stopping criteria, runtime, convergence, and autocalibration settings are needed for reproducibility and practical assessment. Finally, the authors should verify the potentially inconsistent table entries and better discuss the scope and limitations of the “open-world” claim.

  • 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 an important and practical problem in MRI reconstruction: adapting pretrained reconstruction models to distribution shifts without target-domain ground truth. The proposed BiasRecon framework combines parameter-efficient prior calibration with SURE-based adaptation of the data-prior balance, and the reviewers generally recognized the relevance and promise of this direction.

    The rebuttal clarified several important concerns, including the frequency decomposition procedure, SURE assumptions, optimization/runtime details, ACS settings, and the unusual table value. Reviewer #2 changed to accept after rebuttal, and Reviewer #3 was already positive. Some limitations remain, especially the moderate novelty, heuristic frequency cutoff, limited “open-world” scope, and incomplete comparison to all possible modern reconstruction baselines.



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 rebuttal clarified several important points, including implementation details and limitations under extreme OOD shifts or very high acceleration. These clarifications are useful and should be incorporated into the final manuscript. Overall, this is a promising and timely paper with a clean adaptation idea and supportive empirical trends.



Meta-review #3

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

    Accept

  • Please justify your recommendation.

    This paper addresses an important and clinically relevant challenge in MRI reconstruction: adapting pretrained reconstruction models to unseen domains without access to target-domain ground truth. The proposed BiasRecon framework introduces a parameter-efficient adaptation strategy that combines frequency-guided prior calibration and adaptive regularization within a self-supervised test-time optimization framework. The reviewers acknowledged the practical significance of the problem, the sound motivation behind the proposed design, and the strong empirical performance across multiple distribution-shift scenarios. Although concerns were raised regarding the novelty of individual components, the strength of the baseline comparisons, and reproducibility details, the authors provided a thorough rebuttal addressing most major issues. In particular, they clarified implementation details, corrected reported inconsistencies, provided additional evidence comparing against full-model fine-tuning and LoRA-style adaptation, and discussed runtime, assumptions, and limitations. Reviewer #2 was satisfied by these clarifications and changed their recommendation to acceptance. While some concerns remain regarding the breadth of comparisons to recent MRI reconstruction methods and the heuristic nature of certain design choices, I do not view these limitations as sufficiently strong to outweigh the paper’s practical relevance, technical quality, and demonstrated performance gains. Overall, I believe the paper makes a valuable contribution to open-world MRI reconstruction and merits acceptance.



back to top