Abstract

Ultra-low-field (ULF) Magnetic Resonance Imaging (MRI) improves accessibility and affordability but suffers from lower image quality compared to high-field MRI. This study proposes a novel enhancement framework that integrates Implicit Neural Representations (INRs) with Neural Style Transfer (NST) to improve ULF MRI quality by transferring high-resolution structural details from 7T MRI. Unlike conventional methods, our approach does not require paired datasets or extensive pre-training, leveraging INR’s continuous representation and NST’s perceptual refinement to enhance contrast, sharpness, and noise suppression. Quantitative evaluations on T1-weighted ULF MRI demonstrate significant improvements in perceptual quality (PIQE), contrast-to-noise ratio (CNR), and structural consistency (MLC/MSLC), outperforming state-of-the-art methods. These findings underscore the potential of INR-driven learning for advancing MRI reconstruction, enabling higher-quality imaging in resource-limited settings. Our method is fully unsupervised and operates in an unpaired setting, requiring no voxel-wise correspondence or labeled training data. The implementation of our proposed method and model hyperparameters is publicly available at https://github.com/khtohidulislam/ULF-MRI-Enhance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1054_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/khtohidulislam/ULF-MRI-Enhance

Link to the Dataset(s)

N/A

BibTex

@InProceedings{IslKh_UltraLowField_MICCAI2025,
        author = { Islam, Kh Tohidul and Ekanayake, Mevan and Chen, Zhaolin},
        title = { { Ultra-Low-Field MRI Enhancement via INR-Based Style Transfer } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15975},
        month = {September},
        page = {605 -- 615}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes an enhancement framework that incorporates the INR and NST to improve the quality of ULF MRI, facilitating higher-quality imaging in resource-limited settings. Quantitative experiments on T1-weighted ULF MRI datasets demonstrate the effectiveness of the proposed framework.

  • 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 major strengths are:

    1. A framework that incorporates the INR and NST for quality enhancement of ULF MRI.

    2. The proposed framework has a practical application scenario for enhancing quality in resource-limited settings.

    3. Quantitative experiments and qualitative visualizations for evaluating the effectiveness of the proposed framework.

  • 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 main weaknesses are:

    1. There are no experiments or theoretical justifications for ensuring the accuracy of enhanced results. The styles typically refer to high-frequency components in NST, and the framework is based on unpair style transfer. Does the reference style from the 7T MRI introduce extra artifacts to the results?

    2. Missing citations in the method description. In Sec2. Method “Theoretical foundation of INR-Based Style Transfer,” there are no proper citations for “2. Feature Matching via Gram Matrices” and “3. Residual Connections for Stability”.

    3. There are limited baselines for evaluating the effectiveness of the proposed framework. The main results are reported in Table 2, where Neighbor2Neighbor is designed for the image denoising task, and SIREN is for image representation (also for general image restoration). And these two are not targeted at this enhancement task. How about the performance of other ULF MRI enhancement methods? Does the proposed framework still outperform them?

    4. There are limited variations of the main components of the proposed framework in the ablation study. In Table 2., the majority of variation is about the NST, and another key component INR is rather neglected, which is crucial in the proposed frame as demonstrated in the paper. The INR in the proposed framework is based on SIREN. Does SIREN is the only choice? How about other INR methods, such as WIRE, does other variants lead to better result?

    5. There are some typos and missing details in the method description that hinder understanding of the proposed method.

    Eq.(2) lacks of detailed formla of loss components: $L_{\content}$, $L_{\style}$ and $L_{\recon}$. If they are well known in NST, a clear citation should be provided.

    Formula typos in Eq.(1) where “f_\theta = \arg\min_\theta \mathcal{L}{\text{total}} (I{ULF}, I_{HF})”, the “argmin” operates on parameter theta (and the theta should be placed on its bottom-center as well), then the whole equation should be “\hat{I}{ULF}(x,y) = f{\theta^\star}(x,y), \quad \text{where} \quad \theta^\star = \mathop{\mathrm{argmin}} \theta \mathcal{L}{\text{total}} (I_{ULF}, I_{HF})”.

    Besides, in Fig 2. and Fig. 3. A, the abbreviation “N2N” is not mentioned before, it might refer to Neighbor2Neighbor based on the context.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/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 — could be rejected, 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 major strengths and weakness.

  • 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 #2

  • Please describe the contribution of the paper

    The authors develop a novel and interesting method for fusing deep neural network features and implicit neural representations, leading to sharper reconstruction and super-resolution for low-resolution and ultra-low-field scans.

  • 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 authors develop a novel and interesting method for fusing deep neural network features and implicit neural representations, leading to sharper reconstruction and super-resolution for low-resolution and ultra-low-field scans.

  • 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.

    Main weaknesses of the paper:

    In its current form, the paper lacks many details, omits (discussing!) relevant concurrent work, and lacks clarity - issues that must be addressed before assessing the soundness and validity of the method and its claims. I have listed my concerns below and kindly ask the authors to make these points crystal clear in their manuscript:

    Pre-Training and Input to DNN:

    • The authors claim that their method can operate in a non-paired dataset setting, i.e., “our approach does not require paired datasets or extensive pre-training,” but it is unclear from the methods and experimental section how the VGG-Net is pre-trained (as suggested in Fig. 1) and what its input is. What constitutes “extensive pre-training”?
    • This could be problematic if (i) the VGG-Net is actually pre-trained with 7T scans (resulting in train/test set leakage) and (ii) a high-resolution image is used as input to the VGG-Net during test time, allwoing for histogram matching with the DNN.
    • It is unclear how the VGG-Net is pre-trained. Is it trained using the 7T scans? Is a VGG-Net pre-trained on natural images (e.g., ImageNet) used, or one trained on medical images? Why was VGG selected over other CNNs? Is it due to the default patch size of 224×224?
    • What exactly is the input to the VGG-Net at test time/inference?

    Processing of the Dataset

    • Why do the authors register the 7T scans to the low-field scans? Why do they resample the low-field scans? Is this to achieve the same resolution for both images? The authors state that their approach does not require voxel-wise correspondenc - so why are the scans registered? This ultimately boils down to the central question: What is the input to the VGG-Net? Please also clarify this in Fig. 1, which is currently unclear.

    Experimental Design

    • If the low-field MR scans have a typical voxel size of, e.g., 2.5 × 2.5 × 5 mm³, wouldn’t established super-resolution methods like SMORE [1] be appropriate baseline comparisons?
    • The authors frame their work as a low-field MRI enhancement problem, but do not contextualize it within the broader MRI super-resolution research field. Much of the existing research may be transferable to ultra-low-field settings. It is unclear why the authors do not compare against established MRI super-resolution methods, even if they need to be trained from scratch. Paired datasets of low-/high-resolution scans are available and of sufficient size.

    For example, the authors could compare against:

    (1) https://github.com/ExtremeViscent/SR-UNet (2) SMORE [1] (3) SynthSR by Iglesias et al. [2] (4) mri_synthsr_hyperfine [3] (if feasible—though it may require both contrasts, and the authors may only have one)

    To know if the algorithm should be seen as a “single-subject” or “cohort-based” SR approach, we would need to know the pre-training method of the VGG-Net.

    Interestingly, the authors cite several relevant baselines (e.g., [3]) to motivate their work, but do not explain what these methods do. This is unfortunate, as it would place their work in the broader research context and establish why these are relevant baselines. Also, I does not detail their limitations, letting the reader know why their method would be relevant.

    Current baselines are not competitive for this problem setting:

    • The vanilla SIREN network has no incentive to interpolate or transfer details from another scan, as it is only trained on the ULR scan using MSE loss. Moreover, it appears quite blurry, raising concerns about whether the hyperparameters were properly optimized. A properly tuned vanilla INR should nearly replicate the low-res scan, as INRs can faithfully reconstruct images when omega is correctly set. Since omega = 30 is the default for SIREN, did the authors tune this baseline at all?
    • N2N is a denoising network designed for natural images. While effective for denoising, it is not specifically designed for super-resolution.

    2D vs 3D Approach Why do the authors choose to approach this problem from a 2D perspective? INRs have been shown to perform well in 3D super-resolution.

    Absence of Medical INR Papers The authors do not place their work in the context of existing INR literature, even though this is the core architecture of their method. There are INR super-resolution approaches that are subject-specific but not trained across subjects. It would be more appropriate to relate this paper to existing MRI super-resolution INR work than, for instance, to DINER.

    Metrics Please explain why the selected metrics are more appropriate than SSIM, PSNR, and LPIPS. Many of the cited papers (e.g., [4]) report these metrics. While these may not be ideal, they are standard, and should be included for completeness—especially since PIQUE and CNR, which are cited, are not reported in the paper you reference for metrics [5].

    Lastly one common theme of this paper is that it cites many relevant papers adjacent to this paper, but does not explain what these works are actually used for or why they are relevant. For instance, the authros cite 6 Ultra Low Field MRI papers in the Intro, but just to make the point that it Ultra Low Field MRI is important in ressource constrained enviornments. I believe this is something the community already knows and greatly appreciates, but it would be much more interesting to discuss the methods these papers actually present.

    Limitations:

    [1] Zhao C, Dewey BE, Pham DL, Calabresi PA, Reich DS, Prince JL. SMORE: a self-supervised anti-aliasing and super-resolution algorithm for MRI using deep learning. IEEE Trans Med Imaging, 2020;40(3):805–17. [2] Iglesias JE et al. SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images. Science Advances, 2023;9(5):eadd3607. [3] Iglesias JE et al. Accurate super-resolution low-field brain MRI. arXiv preprint arXiv:2202.03564, 2022. [4] Lin H et al. Low-field magnetic resonance image enhancement via stochastic image quality transfer. Med Image Anal, 2023;87:102807. [5] Dohmen M et al. Similarity and quality metrics for MR image-to-image translation. arXiv preprint arXiv:2405.08431, 2024. [6] Wu Q, Li Y, Xu L, Feng R, Wei H, Yang Q, Yu B, Liu X, Yu J, Zhang Y. IREM: High-resolution magnetic resonance image reconstruction via implicit neural representation. InMedical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VI 24 2021 (pp. 65-74). Springer International Publishing. [7] Wu Q, Li Y, Sun Y, Zhou Y, Wei H, Yu J, Zhang Y. An arbitrary scale super-resolution approach for 3d mr images via implicit neural representation. IEEE Journal of Biomedical and Health Informatics. 2022 Nov 18;27(2):1004-15. [8] McGinnis J, Shit S, Li HB, Sideri-Lampretsa V, Graf R, Dannecker M, Pan J, Stolt-Ansó N, Mühlau M, Kirschke JS, Rueckert D. Single-subject multi-contrast MRI super-resolution via implicit neural representations. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention 2023 Oct 1 (pp. 173-183). Cham: Springer Nature Switzerland.

  • Please rate the clarity and organization of this paper

    Poor

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

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/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 — could be rejected, dependent on rebuttal

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

    The paper in its current form lacks details that make it hard to assess if the evaluation is fair and meaningful, especially with respect to the absence of important ULF-MRI baseline methods and traditional SR work.

  • 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.

    Most of my concerns during the review stemmed from the use of a somewhat outdated DeepSDF-style architecture, based on concatenating latent codes with coordinates, as well as from an insufficient description of the authors’ own/novel contributions, and the lack of contextualization within relevant prior work, and a few unaddressed limitations.

    The authors have addressed many of these issues to a reasonable extent in their rebuttal, and I trust they will incorporate these clarifications into the final version of the paper.



Review #3

  • Please describe the contribution of the paper

    This paper describes a method for enhancing the quality of ultra-low-field (ULF) MRI acquired at field strengths in the milli-Tesla range. The proposed approach is to combine an Implicit Neural Representation (INR) network with a Neural Style Transfer (NST) technique. In this way the authors propose to make use of the INR’s ability to form a continuous representation of a discrete image to enhance resolution, while the NST contributes the enhanced contrast and noise characteristics of the unpaired Highfield (HF) style conditional image.

  • 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 specific combination of an INR with NST is something I have not seen before. Whilst INR use for super-resolution of low-field (LF) scans has been presented in previous works, the ability to use a single unpaired high-field (HF) scan to impart the contrast properties of more resource intensive imaging to ULF MRI is certainly powerful. In addition, the paper is generally well written and easy to follow, although at times relevant details require some additional thought to piece together and some details are not explicitly explained. The authors also perform evaluation on real ULF scans where many papers limit themselves to simulations, increasing the credibility of the results.

  • 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 paper does present some weaknesses, although most are relatively minor. For instance, as with many papers in this field, the authors focus on in plane resolution enhancement rather than 3D reconstructions. In contrast the hyperfine swoop typically has an in-plane resolution between 1-2 mm and an out of plane resolution of 5mm or larger. While the authors do state that their source ULF images had been resampled to 1mm^3 isotropic resolution, they do not state which plane was used for acquisition (although from their images I suspect it was the axial plane). Hence the authors have not shown that they can mitigate the more pressing resolution limitation.

    There are also some other implementation and experimental explanations which could do with being more detailed. The mathematical formulation of the three loss functions would be appreciated as currently only a qualitative overview of the intent behind the losses are presented. For example, without knowing what form the reconstruction loss takes, we cannot know if it is sufficiently robust to the contrast change induced by the style transfer. I also assume that the pretrained VGG-19 network, used for feature extraction in the content and style losses, is using the publicly available weights trained on ImageNet, but this should be stated and cited accordingly.

    Finally, I have some concerns about the structural validation of the author’s methods. While the quantitative metrics used are appropriate, they only assess noise features and relative resolution features such as sharpness. They do not inherently tell us anything about the data-consistency of the method, leaving the possibility of hallucinated features such as shifted boundaries or transfer of possible pathology features from the unpaired style condition. Additionally, the author’s statement that they compare to state-of-the-art methods is only true in that they compare to the natural-image methods related to their proposed approach rather than ULF specific methods. For example, SynthSR which I believe is cited as reference [10] in the current draft. Alternatively, a more comparable approach may be the combination of INR networks and score matching diffusion networks presented X. Lin et. al. at MICCAI 2024 in “Zero-Shot Low-Field MRI Enhancement via Denoising Diffusion Driven Neural Representation”.

  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    I enjoyed reading this paper and hope that the authors will continue to develop such approaches for the enhancement of ULF MRI. While I believe that the weaknesses identified in the section above must be addressed before publication there are also some small corrections that would enhance the paper and some points to investigate if moving towards publication of a larger work.

    Minor corrections:

    1. A Spectral bias mitigation strategy is mentioned in the methods section but not cited or explained.

    2. It is implied that the style condition can be provided from a scan of a different subject and unregistered (use of word unpaired). However, the text would benefit from this being stated more explicitly.

    3. The authors do talk about training the INR model over 1000 iteration. However, it is currently unclear that the network requires re-training for every new subject at inference time. This should just be more clearly stated in the conclusions section rather than just highlighting increased time requirements.

    Additional investigations for future work (not necessary for MICCAI25):

    1. I would be interested to see if there is any sensitivity of the method to the content of the style companion. For example, using a slice that is mismatched or one containing undiagnosed pathology. Some pathology like signs of degenerative brain disease may not be easily visible but can cause changes in contrast or minor shifts in small regional volumes that are picked up by downstream analysis.

    2. The authors should absolutely compare to state-of-the-art MRI specific methods. I believe they would perform well in quality metrics compared to techniques like SynthSR but possibly not the more recent “foundational model” approach proposed by Dr. Iglesias’ group at CVPR last year. Similarly, they might find that the INR and score matching approach has comparable quality metrics but is likely to be slower due to the nature of diffusion based models.

    3. When evaluating the method for further publication there needs to be some metric used that makes sure the same structure and pathological features are present in both the ULF input as well as the inference result. In the absence of a paired HF image this could be done through some kind of degradation model applied to the output, similar to X. Lin et. al. MICCAI 2024, through a degradation model that more directly simulates contrast changes such as the one in H. Lin et. al. Medical Image Analysis 2023, through something like a cycle consistency, or through downstream analysis.

  • 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 — could be accepted, dependent on rebuttal

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

    This paper should be accepted to MICCAI this year. The approach taken is novel, mostly well explained and the evaluations are appropriate, if not complete. Having said this, I cannot accept the paper without addressing some of the concerns listed in the weaknesses section above. This should amount to some minor additions to the text and acknowledgement of limitations in the conclusion (which should be relabelled “Discussion and conclusion”. No further experiments or corrections to results need to be implemented.

    The points I would like addressed are:

    1. Acknowledge that the 3D problem, particularly anisotropic resolution, needs to be addressed.
    2. State acquisition parameters of the imaging used.
    3. Give details of the losses used to the extent that a reader may implement the same losses themselves without having to examine the code.
    4. Explain or cite the training of the VGG network.
    5. Acknowledge in discussion that there is a need to evaluate data-consistency between the input ULF and inference result through further work.
    6. Acknowledge that performance needs to be compared to LF MRI specific state-of-the-art methods through further work.
  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A




Author Feedback

We thank the reviewers and meta-reviewer for their constructive and detailed feedback.

Unpaired Framework and VGG Pretraining (R1, R2, R3): Our method is fully unsupervised and operates in an unpaired setting, requiring no voxel-wise correspondence or labeled training data. The 7T dataset is publicly available and used solely as a style reference. Rigid registration was applied to match resolution for perceptual feature extraction, not for supervision. The VGG-19 network used for perceptual loss is pretrained on ImageNet and not fine-tuned. “No extensive pretraining” refers to the absence of large-scale end-to-end training on paired MRI data. We will clarify this and update Fig. 1 accordingly.

2D Design and 3D Structural Consistency (R1, R3): We acknowledge that a 3D INR formulation could further improve volumetric continuity and will highlight this as a valuable future extension. However, we opted for a 2D formulation to accommodate anisotropic ULF resolution (1.6×1.6×5mm³) and reduce training cost. Volumes were resampled to 1mm³ and processed slice-wise. Evaluation was performed on full volumes. Improvements in 3D segmentation-derived CNR and visual consistency across slices (Fig. 3) support our method’s structural coherence.

Loss Components and Hallucination Mitigation (R1, R2): Our total loss combines (i) an L2-based content loss between VGG features of output and ULF input, (ii) a Gram matrix-based style loss with the unpaired 7T reference, and (iii) an L1 reconstruction loss with the ULF input. These were chosen to balance perceptual adaptation with anatomical fidelity, following Gatys et al. and Johnson et al. While we did not include explicit hallucination detection, the reconstruction loss and residual INR design anchor structural consistency. We agree that the formulation in Eq. (1) should follow the standard argmin expression as suggested. This will be clarified.

Baseline Selection and Evaluation Constraints (R1, R2, R3): Our framework is inherently unpaired and unsupervised, trained without voxel-level alignment or annotated supervision. In this setting, SIREN provides a strong baseline for implicit representation, and Neighbor2Neighbor (N2N) offers a representative unsupervised denoising comparator. SIREN was carefully tuned (ω = 30) and evaluated under consistent conditions. While we acknowledge the strength of methods such as SMORE, SynthSR, and SR-UNet, they either require paired or synthetically downsampled high-resolution data during training, which is not feasible in our unpaired and real-world 64mT T1-only setting. We recognize this limitation and plan to explore adaptations in future work.

Absence of Medical INR Papers (R2, R3): We agree that situating our method more explicitly within the context of medical INR literature will strengthen the manuscript. Our work specifically addresses the underexplored problem of ULF MRI enhancement in a fully unsupervised, unpaired setting—without voxel-level supervision or aligned training data. While DINER is cited for architectural context, our focus is more aligned with recent subject-specific and self-supervised INR approaches in MRI, which similarly avoid cohort-based training. Moreover, as reported in the original paper, DINER incurs approximately 2× longer training time than SIREN, making SIREN a more computationally efficient choice for single-subject modeling in resource-constrained ULF applications. We will revise these.

Metric Selection and Contextualization of References (R3): We emphasize that SSIM, PSNR, and LPIPS are not applicable in our unpaired setting, as they require spatially aligned ground truth images. Our framework operates without voxel-wise correspondence, so we adopt PIQE (no-reference), CNR (from 3D segmentations), and MLC/MSLC (structural consistency) to evaluate perceptual and anatomical fidelity. Additionally, we appreciate the suggestion to better contextualize the cited ULF MRI works. In the revision, we will concisely describe.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    The reviewers have expressed divergent views on this paper, resulting in a broad range of scores. Accordingly, the authors are invited to provide a rebuttal to help clarify and support their contributions. It is especially important to respond to concerns regarding the adequacy of comparisons with relevant state-of-the-art and baseline methods specific to the MRI domain, rather than relying heavily on techniques from natural image processing. Additionally, the rationale behind the pre-training strategy should be more clearly justified. Please focus on the most significant issues, as the rebuttal length is limited and should be used strategically to address the primary points raised.

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top