List of Papers Browse by Subject Areas Author List
Abstract
Magnetic Resonance Imaging (MRI) field-strength enhancement holds immense value for both clinical diagnostics and advanced research. However, existing methods typically focus on isolated enhancement tasks, such as specific 64mT-to-3T or 3T-to-7T transitions using limited subject cohorts, thereby failing to exploit the shared degradation patterns inherent across different field strengths and severely restricting model generalization. To address this challenge, we propose \methodname, a unified framework integrating multiple modalities and enhancement tasks to mutually promote representation learning by exploiting these shared degradation characteristics. Specifically, our main contributions are threefold. Firstly, to overcome MRI data scarcity and capture continuous anatomical structures, \methodname departs from conventional methods that treat 3D MRI volumes as independent 2D slices. Instead, we directly exploit comprehensive 3D volumetric information by leveraging pre-trained 3D foundation models, thereby embedding generalized and robust structural representations to significantly boost enhancement performance. In addition, to mitigate the spectral bias of mainstream flow-matching models that often over-smooth high-frequency details, we explicitly incorporate the physical mechanisms of magnetic fields to introduce a Field-Aware Spectral Rectification Mechanism (FASRM), tailoring customized spectral corrections to distinct field strengths. Finally, to resolve the fundamental data bottleneck, we organize and publicly release a comprehensive paired multi-field MRI dataset, which is an order of magnitude larger than existing datasets. Extensive experiments demonstrate our method’s superiority over state-of-the-art approaches, achieving an average improvement of approximately 1.81 dB in PSNR and 9.47\% in SSIM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3466_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/linyiyang98/UniField
Link to the Dataset(s)
https://github.com/linyiyang98/UniField/blob/main/Dataset
BibTex
@InProceedings{LinYiy_UniField_MICCAI2026,
author = { Lin, Yiyang AND Wang, Chenhui AND Peng, Zhihao AND Yuan, Yixuan},
title = { { UniField: A Unified Field-Aware MRI Enhancement Framework } },
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
1.The authors curate and plan to publicly release the largest multi-center paired multi-field MRI dataset in the field to date, with rigorous cross-field registration and standardized preprocessing completed. The dataset is an order of magnitude larger than existing benchmarks, providing a new standardized research benchmark for the field of MRI field-strength enhancement with long-term academic value.
2.For the first time, this paper fully implements the unified multi-task learning paradigm in the field of MRI multi-field-strength enhancement, and verifies the feasibility of improving performance via joint training with shared degradation patterns across modalities and field strengths. It breaks the long-standing isolated training paradigm of “one model for one task” in the field, and provides complete engineering verification for the paradigm transformation of the field.
3.Aiming at the inherent high-frequency over-smoothing problem of flow-matching models in MRI enhancement, the authors design the Field-Aware Spectral Rectification Mechanism (FASRM), which optimizes the detail recovery effect of different field-strength tasks through customized frequency band weights. Preliminary performance verification is completed, providing engineering reference for the application of flow-matching models in medical image enhancement.
4.The authors explore the migration scheme of pre-trained video super-resolution prior to 3D MRI volume enhancement, and achieve lightweight training through “frozen pre-trained encoder-decoder + Low-Rank Adaptation (LoRA) fine-tuning”. It solves the problem that traditional slice-by-slice 2D processing discards inter-slice anatomical structures, and provides a reusable engineering scheme for 3D medical image generation in low-data 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.
1.The multi-center paired multi-field-strength MRI dataset built here is the largest public benchmark in the field to date. It solves long-standing core bottlenecks including scarce paired data and poor generalization of single-center studies. This contribution is independent of the method itself and has irreplaceable long-term value for future algorithm research and benchmarking. It is the most solid and undisputed strength of this paper. 2.Existing MRI field-strength enhancement methods mostly use single-task isolated training. This paper is the first to fully implement unified multi-task learning across modalities and field strengths in this subfield and validates joint training with shared degradation patterns. It improves the usage efficiency of limited paired data and provides a lightweight multi-task deployment scheme for clinical use, offering a new paradigm for the field. 3.The paper addresses three core bottlenecks of existing methods with one-to-one solutions. The FASRM module performs customized frequency-domain optimization for different artifact patterns in 64mT→3T and 3T→7T tasks unlike generic frequency regularization. Visual results clearly show improved detail recovery and artifact suppression with complete and reproducible implementation logic. 4.The frozen pre-trained encoder-decoder and LoRA fine-tuning scheme greatly reduces computing and data requirements for 3D generative model training making engineering adoption easy. 5.The unified framework supports one-stop processing for multi-modal and multi-field-strength conversion tasks matching two key clinical needs including low-field to clinical field and clinical field to ultra-high field. It is practically convenient for clinical deployment.
- 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 core paradigm of a unified multi-task enhancement framework using shared cross-task degradation patterns is not original. Fei et al. [1] CVPR 2023 proposed a unified image restoration and enhancement framework with generative diffusion priors fully validating shared degradation features. Yang et al. [2] CVPR 2024 proposed CRNet for end-to-end unified restoration confirming the paradigm’s generality. This paper only ports the paradigm to MRI field-strength enhancement without essential structural innovation. This directly challenges the paper’s core novelty claim. The novelty is only a first application in the subfield not methodological invention. The authors must revise all novelty claims otherwise the core contribution cannot stand. 2.FASRM is essentially a manually tuned frequency-band weighted loss. Sub-band frequency-domain regularization for medical image super-resolution and enhancement has been widely used in MRI field-strength enhancement. Cui et al. [3] MICCAI 2024 used frequency sub-band optimization for 3T→7T synthesis. Many prior top-conference works used frequency correction for diffusion and flow-matching spectral bias. This paper only adjusts fixed frequency weights manually for different field strengths without establishing a quantitative mapping between field strength physical artifacts and frequency weights. It does not truly embed magnetic field physical priors. The field-aware claim lacks rigorous physical and mathematical support. The authors only did manual engineering tuning not physical prior integration. Physical modeling and ablation studies must be added otherwise this module’s novelty is invalid. 3.The paper transfers the general video super-resolution model FlashVSR to 3D MRI enhancement. However many native medical 3D pre-trained models such as 3D SwinUNETR and MONAI pre-trained 3D UNet have much stronger anatomical representation for medical volumes. Peng et al. [4] already explored transferring temporal priors to 3D brain imaging. This paper does not compare with medical-native 3D models or verify necessity and superiority via ablations. It is only engineering adaptation not methodological innovation. This issue does not negate existing strengths but means the paper cannot prove video prior is the best choice for 3D structure preservation. Comparisons and ablations must be added otherwise this part’s contribution is invalid. 4.The paper compares UniField fine-tuned from FlashVSR with vanilla FlashVSR without task adaptation. Training setups and adaptability are imbalanced so results are unconvincing. It fails to include field-strength-specific SOTA methods such as Cui et al. [3] 3T→7T and Islam et al. [5] low-field MRI enhancement. Only basic U-Net and GAN baselines are used, which cannot prove state-of-the-art performance. No ablations for the text condition module or FlashVSR pre-trained prior so the necessity of core components is unproven. This violates top medical imaging conference standards. Fair comparisons and full ablations must be added otherwise the main performance claims cannot be accepted. 5.All experiments are random splits under the same distribution. No cross-center and cross-scanner out-of-distribution generalization tests are performed even though solving poor generalization is a core motivation. No clinical validation is conducted including no radiologist double-blind subjective evaluation and no downstream task testing such as brain segmentation and lesion detection. Meanwhile recent works including Baljer et al. [6] and Lucas et al. [7] already include clinical validation. Clinical usability is the core value of medical imaging methods. Without validation potential is just theoretical speculation. Clinical and generalization experiments must be added otherwise practical value is unproven. [1] Generative diffusion prior for unified image restoration and enhancement. [2] Crnet: A detail-preserving network for unified image restoration and enhancement task [3] 7t mri synthesization from 3t acquisitions [4] Geometric-oriented brain transformer for autism diagnosis [5] Improving portable low-field mri image quality through image-to-image translation using paired low-and high-field images. [6] Ultra-low-field paediatric mri in low-and middle-income countries: Super-resolution using a multi- orientation u-net. [7] Multisequence 3-t image synthesis from 64-mt low-field-strength mri using generative adversarial networks in multiple sclerosis
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
please refer to 6 and 7
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the authors’ rebuttal, I recommend acceptance. My initial concerns were mainly about the clarity of the FlashVSR adaptation, the fairness of the FlashVSR baseline, the lack of ablations for several key components, the limited evidence for generalization/downstream utility, and the strength of the claimed methodological novelty. The rebuttal addressed several of these concerns in a reasonably convincing manner. The authors clarified that the vanilla FlashVSR baseline underwent the same fine-tuning process, which alleviates my concern about unfair comparison. They also provided additional ablations on the text condition and FlashVSR pretrained weights, and these results support the necessity of the proposed components. The added comparison with a segmentation-pretrained SwinUNETR baseline also helps justify the use of a video super-resolution prior for this task. In addition, the reported external-center evaluation and downstream segmentation analysis provide preliminary evidence for generalization and practical utility. I still believe that the methodological novelty is moderate rather than strong. In particular, FASRM remains closer to a field-motivated frequency-weighted loss than a rigorously derived MRI physics-based model, since the frequency weights are selected by grid search rather than derived from a principled physical formulation. Therefore, I would encourage the authors to tone down the “physics-aware” claim and more clearly present FASRM as a field-motivated spectral regularization strategy. Nevertheless, the paper addresses an important and clinically relevant problem, proposes a unified framework for multi-modality and multi-field MRI enhancement, and contributes a valuable large-scale paired multi-field MRI dataset. The ablation results in the original paper support the effectiveness of unified modality/task learning and FASRM, and the rebuttal substantially improves confidence in the implementation and evaluation. Therefore, despite the remaining concern about the strength of methodological novelty, I believe the paper is above the acceptance threshold.
Review #2
- Please describe the contribution of the paper
This paper introduced a unified approach for the enhancement of 64mT to 3T and 3T to 7T MRI that supports multiple contrasts. They adopted a 3D volumetric method using pre-trained embedding from a video super-resolution prior with a flow-based approach and field-aware rectification to address distinct spectral corrections in each task. The authors have validated their approach with multiple paired LF and HF MRI datasets and shown better results in comparison to other recent baselines.
- 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 unified nature of the proposed framework towards the enhancement of multiple field strengths is novel and a strong contribution, as it handles many potential challenges and limitations in this domain.
The field awareness of the synthesis for different frequency scales is also an interesting way to handle variant requirements across different tasks.
- 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.
The paper states that FlaskVSR is used as a pre-trained prior without fine-tuning, but key architectural details are missing. In particular, the components of the LQ encoder are unclear, including whether it uses the Proj-In layer as in the original model. It is also not explained how a model pre-trained on large-scale natural videos can adapt to MRI enhancement, given the substantial differences in intensity distributions and normalization ect.
Beyond the proposed field-awareness, the paper does not clearly explain what new components are introduced on top of FlaskVSR.
The term 𝜆_spat is introduced but not used. In Eq. 2, it is also unclear how α dynamically adjusts the error scale, and the equation does not clearly show that this operation is applied to the velocity difference. The paper only reports L_FASFL, but does not clarify the total loss.
There is no ablation or analysis of the selected frequency weights.
The qualitative comparisons show only end slices, which makes it difficult to assess enhancement quality in important brain regions. Also, since this is a 3D method, showing all three anatomical views would provide a more complete evaluation.
In Fig. 4 (second panel-Ours), the enhanced image appears blurry and contains artifacts. Similar issues are visible in Fig. 3, where errors and contrast inconsistencies remain, despite improvement over baselines. These observations are not discussed in the results or as limitations.
The same baselines are used for both tasks, although some are not designed for MRI enhancement. There are many existing recent works that are specifically designed for 7T synthesis and ULF to HF synthesis.
The paper lacks a downstream analysis of the synthetic results. While quantitative and qualitative results are useful, the practical value of synthetic MRI is better demonstrated through downstream tasks.
No discussion on the limitations of the study in terms of methodological aspects and key observations from the evaluation of results.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The main factors that led to this score are the lack of methodological details in certain aspects, the lack of clarity in contribution statements, and limitations in the evaluation of results.
- 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.
I appreciate the author’s effort to address my comments. However, I still feel that the novelty of the proposed framework is relatively incremental, as the method mainly integrates and adapts existing concepts rather than introducing a fundamentally new synthesis strategy. I am also not convinced by the qualitative results presented in the paper. Although the method achieves improved quantitative performance (compared to SOTA), pixel-wise metrics alone are insufficient to validate synthesis fidelity in MRI enhancement tasks, especially when the qualitative examples still appear blurry and contain noticeable artifacts and contrast inconsistencies. While I understand the challenging nature of low-field and ultra-high-field MRI synthesis, the paper still lacks sufficient downstream validation within the manuscript (even some results were reported in the rebuttal) to demonstrate the practical reliability of the generated images.
Furthermore, I still have concerns regarding the overall comparison of results. The study includes limited task-specific state-of-the-art comparisons, with only one comparison specifically targeting the 3T to 7T task and two for the 64mT to 3T task. In addition, some baselines designed primarily for low-field enhancement are also used in the evaluation of ultra-high-field synthesis. Given the growing number of recent works in both low-field MRI enhancement and 7T synthesis, I believe a stronger and more comprehensive comparison is necessary.
Overall, I appreciate the contribution of this work, as unified MRI enhancement across multiple field strengths is a promising research direction. However, the current methodology, experimental validation, and presentation of results are still not sufficiently convincing for me to change my score.
Review #3
- Please describe the contribution of the paper
The main contribution is UniField, a unified framework that jointly handles multiple MRI modalities (T1, T2, FLAIR) and multiple field‑strength enhancement tasks (64mT→3T and 3T→7T) within a single model. Unlike prior work that trains isolated networks per task, UniField exploits shared degradation patterns across different field strengths to improve generalization and data efficiency. Key innovations include: (1) leveraging a pre‑trained 3D video super‑resolution prior (FlashVSR) to capture continuous volumetric anatomy; (2) a Field‑Aware Spectral Rectification Mechanism (FASRM) that adaptively modulates frequency‑domain optimization based on the physical characteristics of the source and target fields; and (3) the release of the largest registered multi‑field MRI dataset to date, an order of magnitude larger than existing benchmarks.
- 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.UniField learns a single conditional model that benefits from implicit data augmentation. The ablation in Table 4 clearly shows that unified modalities and unified tasks consistently outperform task‑specific or modality‑specific models. This is both practically useful (one model for multiple clinical scenarios) and scientifically interesting (shared degradation patterns do exist across field strengths). 2.The use of a pre‑trained 3D video prior is well‑justified. Training 3D generative models from scratch on limited MRI data is difficult. By freezing a FlashVSR model (originally for video super‑resolution) and adapting it with LoRA, the authors bypass data scarcity while preserving volumetric continuity. This is a practical transfer learning success that other medical imaging works could follow. 3.FASRM addresses a real problem (spectral bias in flow matching) with field‑aware frequency weighting. The idea of dynamically adjusting high‑frequency vs. low‑frequency penalties based on whether the task is extreme low‑to‑high (where hallucination risk is high) vs. high‑to‑ultra‑high (where artifact propagation is a concern) is novel and physically grounded. The qualitative results in Fig. 4 show that FASRM clearly restores sharp boundaries and reduces blurring compared to the same model without it.
- 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.Table 1 notes that several datasets (e.g., Leiden, KCL, BNU) were originally unregistered and required registration, but no details are given about the registration algorithm, quality control, or potential residual misalignment. Given that paired training relies on precise voxel‑wise correspondence, this should be clarified. 2.The paper reports training details but does not compare inference speed or model size against baselines. Since clinical deployment cares about runtime, adding a small table or sentence on inference time per volume would strengthen the paper.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper solves a practically important problem (MRI field‑strength enhancement) with a unified framework that is both novel and well‑executed. The unified modeling paradigm is a real improvement over fragmented prior work; the FASRM module is physically motivated and shown to improve high‑frequency detail; and the released dataset is a significant community resource. The experiments are compelete(multiple modalities, two tasks, several baselines, ablation studies, and visual comparisons with error maps). Overall, this is an impactful paper.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I recommend accepting this paper.
Author Feedback
We sincerely thank the reviewers for recognizing UniField as a novel framework (R2, R3) and highlighting our large-scale dataset (R1, R3). Note that all results are averaged across tasks and modalities, with our method achieving [9.29, 20.77, 72.35, 0.1817] for [NRMSE↓, PSNR↑, SSIM↑, LPIPS↓]. (R1Q1) Though unified restoration exists for natural images, unifying multi-field MRI enhancement remains an under-explored yet promising topic. Since 64mT→3T and 3T→7T share many MR-physics-driven factors, unifying them enables mutual benefit by mining shared patterns, and FASRM is proposed to further inject physical priors. (R1Q2/R2Q4) In FASRM, the weights are grid-searched in [0.0, 1.0] (step 0.1) for optimal performance, yielding the weights listed in the paper. In contrast, random weights yielded worse performance: [9.34, 20.70, 71.45, 0.1902] (mean of five runs), confirming the effectiveness of our strategy. Future work will explore dynamic, physics-informed adaptive weighting. (R1Q3/R2Q1) We added a segmentation-pretrained SwinUNETR baseline, but it yielded poor results: [15.10, 18.97, 39.52, 0.3790], likely due to the large gap between segmentation and enhancement. Since both SR and our task need to recover high-frequency details from degraded input and remove artifacts, the SOTA pretrained SR weights (FlashVSR [CVPR’26]) are leveraged in our work. Additionally, training without these weights drastically degraded performance to [26.18, 11.84, 47.84, 0.5858], confirming the effectiveness of natural-SR priors. (R1Q4/R2Q7) In our comparison, vanilla FlashVSR also underwent identical fine-tuning process as ours for fairness. Furthermore, we conducted ablation experiments by removing text condition and FlashVSR weights separately, which yielded [9.34, 20.44, 70.62, 0.1963] and [26.18, 11.84, 47.84, 0.5858], respectively, confirming the necessity of the proposed components. Note that our original manuscript already includes SOTA methods for 64mT→3T (MO-U-NET and LowGAN) and 3T→7T (MSFA). (R1Q5/R2Q8) We tested our model’s generalization on 102 patients at a local clinical center, achieving similar performance to internal sets: [14.62, 20.99, 74.03, 0.2923]. Furthermore, we used MedSAM3 [CVPR’26] to segment the thalamus, corpus callosum, and midbrain from all input low-field images and their corresponding synthesized high-field images, yielding [Dice↑, IoU↑] scores of [0.8464, 0.7566] and [0.8540, 0.7820], respectively, suggesting better preservation of clinically relevant boundaries. Double-blind radiologist evaluation is future work. (R2Q1/Q2) The novelties of our method compared to FlashVSR are three-fold: i) replacing the causal Proj-in with a non-causal bidirectional LQ encoder (containing customized 2D pixel-shuffle, 3D convolutions, and MLPs) to better perceive the 3D spatial relationships in MRI data; ii) adding a T5 text encoder to specify modality and field transitions; and iii) introducing the FASRM module to effectively integrate field-transition characteristics. (R2Q3) The total loss is exactly L_FASFL, consisting of a standard velocity difference loss and our FFT-based term. We will remove the typo λ_spat, as the velocity difference weight is fixed to 1.The FFT-based term computes the error between the FFT-transformed real and predicted velocities under binary masks. Within this term, the component containing α adjusts the error scale to focus training on hard frequencies, while α (default=1) solely controls the sharpness of this focal region. (R2Q5/Q6/Q9) Axial images usually have the highest resolution, while full views and regions aid complete evaluation, and thus we will add them in the final version. While we acknowledge minor residual artifacts in Figs. 3–4 stemming from the inherently ill-posed nature of the task (high-field textures often lack structural evidence in low-field inputs), our method still clearly outperforms baselines. (R3) We will add registration details & inference speed/model size of all methods.
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 proposes an MRI field strength enhancement framework for 64mT to 3T and 3T to 7T synthesis using a video super-resolution prior, LoRA fine-tuning and a field-aware spectral loss. The reviewers found the method practically relevant and unifying multiple enhancement tasks interesting, but they raised several important limitations. One concern is the limited methodological novelty, noting that unified restoration frameworks and frequency-domain correction strategies have been explored in prior work. Reviewers also questioned the physical justification and formulation of the proposed FASRM module, arguing that the field-aware spectral weighting appears manually designed rather than rigorously derived from MRI physics. Additional concerns included insufficient architectural clarity in adaptation of FlashVSR to MRI, missing ablations for key components such as text conditioning and frequency weighting, incomplete and potentially unfair baseline comparisons against field-specific state-of-the-art methods, and the absence of stronger discussion on registration quality, generalization, and downstream clinical utility. Based on these critiques, the recommendation is rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After rebuttal, the paper is borderline but above the acceptance threshold. The authors provided useful clarifications and new evidence addressing several concerns, including fair FlashVSR fine-tuning, ablations for text conditioning and pretrained weights, a SwinUNETR comparison, external-center evaluation. These additions improve confidence in the method and practical relevance. However, the methodological novelty remains moderate, as the approach largely integrates unified restoration, pretrained video SR priors, LoRA adaptation, and empirically tuned frequency-domain regularization. The FASRM component should be presented more cautiously as field-motivated spectral regularization rather than a physics-derived module. Reviewer 2’s concerns about qualitative artifacts and limited task-specific SOTA comparisons remain valid, but they do not outweigh the dataset contribution, unified formulation, and improved post-rebuttal evidence. Therefore, I would recommend accept.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The rebuttal contains new/additional experimental results and added ablations that were not part of the submitted manuscript, which violates the rebuttal policy and warrants desk rejection. Nevertheless, even aside from this issue, Reviewer #2 remained unconvinced after rebuttal, citing limited methodological novelty, blurry qualitative results with visible artifacts, insufficient downstream validation, and an incomplete comparison against stronger task-specific baselines. These concerns remain unresolved in the post-rebuttal phase.
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.
While the reviewers are still mixed, there is a clear trend towards acceptance based on the initial reviews and post-rebuttal comments. The authors have provided substantial clarification and sufficient results, in my opinion. Even with some remaining concerns, I believe these to be outside the scope of a MICCAI submission.