List of Papers Browse by Subject Areas Author List
Abstract
Clinical MRI frequently acquires anisotropic volumes with high in-plane resolution and low through-plane resolution to reduce acquisition time. Multiple orientations are therefore acquired to provide complementary anatomical information. Conventional integration of these views relies on registration followed by interpolation, which can degrade fine structural details. Recent deep learning–based super-resolution (SR) approaches have demonstrated strong performance in enhancing single-view images. However, their clinical reliability is often limited by the need for large-scale training datasets, resulting in increased dependence on cohort-level priors. Self-supervised strategies offer an alternative by learning directly from the target scans. Prior work either neglects the existence of multi-view information or assumes that in-plane information can supervise through-plane reconstruction under the assumption of pre-alignment between images. However, this assumption is rarely satisfied in clinical settings.
In this work, we introduce Single-Subject Implicit Multi-View Super-Resolution for MRI (SIMS-MRI), a framework that operates solely on anisotropic multi-view scans from a single patient without requiring pre- or post-processing. Our method combines a multi-resolution hash-encoded implicit representation with learned inter-view alignment to generate a spatially consistent isotropic reconstruction. We validate the SIMS-MRI pipeline on both simulated brain and clinical prostate MRI datasets. The code is publicly available for reproducibility.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6049_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/abhshkt/SIMS-MRI
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KimHee_SingleSubject_MICCAI2026,
author = { Kim, Heejong AND Thanki, Abhishek AND van Herten, Rudolf L. M. AND Margolis, Daniel J. AND Sabuncu, Mert R.},
title = { { Single-Subject Multi-view MRI Super-Resolution via Implicit Neural Representations } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
In this paper, the authors propose SIMS-MRI for multi-image super-resolution (MISR) of multiple anisotropic volumes. Their method leverages implicit neural representations (INR) to approximate the image intensities, conduct multi-view registration, and fuse the results into a high-resolution (HR) volume. They accomplish this in three phases. The first phase fits an INR f to the first anisotropic acquisition. The second phase fits an INR g to deformably register the second anisotropic acquisition to the first, using the frozen network f from the first phase. The third phase freezes the registration network g and fine-tunes f to represent both views. During inference, f is queried with isotropic coordinates to yield the estimated HR volume. The authors make use of the InstantNGP feature encoding technique for f and they make use of SIREN for g. To evaluate the efficacy of their method, the authors compare to interpolation and two contemporary single-image super-resolution techniques SMORE and ECLARE in both simulation and real-world experiments. The authors report superior performance across several quantitative metrics compared to the other methods across all experiments.
- 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.
This paper has many strengths. The incorporation of multi-orientation information for anisotropic super-resolution is well-motivated. The restriction of training data to a single-subject addresses pathology, unique contrasts/modalities, and is clinically useful. The use of an INR to perform implicit deformable registration is very novel. The experimental suite included custom augmentations for comparison methods SMORE and ECLARE (including Table 1, demonstrating the utility of their MISR augmentations) is thoughtful and useful for the field of anisotropic MISR. The number of quantitative metrics is extensive.
- 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 method is both interesting and promising. Yet, I still have a few clarifying questions and some comments on the theory and experimental design, ordered by what I consider most important first. 1.How is slice thickness addressed in this work? The authors explicitly mention in Sec 2, sentence 2: “thick slices along the through-plane directions. “ This is correct for 2D-acquired MRI. However, this slice thickness is not addressed in either the methods section nor in the simulation experiment. Rather, as-is, the paper addresses sparse-slice anisotropic MRI. This is fine if the authors intend it to be so, but it should be explicitly mentioned. 2.Related to the previous comment, “resolution” is not strictly defined in this work, and this causes some ambiguity in interpreting the method. Consider two cases. A) the HR image volume has voxels of resolution a x a x a mm. The simulation approach described in Sec 3.1.sounds like the LR image volume still has voxels of resolution a x a x a mm, where slices are sub-sampled with something like cubic B-spline interpolation, linear interpolation, or slice sub-sampling. In this way, the INR network f learns to use complementary information from both views, and the theory of the method makes sense: inference-time isotropic grid sampling should yield an isotropic volume, since for all coordinates the voxel resolution learned is always a x a x a mm. Now consider the second case, B) the HR volume still has voxels of resolution a x a x a mm, but the LR volumes are truly thick-slice, with voxels of resolution a x a x b mm and a x c x a mm, where b > a and c > a. This could be simulated, for example, by convolving in 1D along the slice-selection axis with a rect or Gaussian filter with appropriate full-width-half-max. In this case, the INR network f does not learn a straightforward function. For the first view, it learns to map coordinates \Omega to voxels of resolution a x a x b mm. For the second view, it learns to map coordinates \Omega to voxels of resolution a x b x a mm. On average, one would expect the final inference-time result to produce images of voxels a x (a+c)/2 x (a+b)/2 mm. I believe some theoretical discussion of this phenomenon would greatly strengthen the paper. 3.Based on the description, my assumption is that the experiment in Sec 3.1 uses cubic B-spline interpolation to produce LR volumes 4x lower resolution simply by subsampling, not by simulating slice acquisition. Is this assumption correct? If so, does this cause a mismatch on the reported results for SMORE and ECLARE, both of which assume thick slices? 4.For the experiment in Sec 3.2, how are the pseudo-ground truth (GT) sagittal slices selected? I believe an appropriate comparison between each of the SR methods would be to simulate 3mm-thick-slices by convolving with an appropriate-width rect or Gaussian profile, then selecting the same 3mm-thick sagittal slice for SR method and pseudo-GT. This would produce results with (supposedly) similar SNR and anatomical locations. Otherwise it is difficult to trust the results reported in Fig. 3, since it is unknown were in the 3mm slice the 0.5-thick estimated slice exists. The anatomy may vary greatly in that section. 5.A small comment on the runtime of each method would be useful. In Sec 4, the authors mention “While the per-subject optimization is more computationally demanding than feed-forward approaches,” but all comparison methods here are per-subject. It would be interesting to know approximately how long SIMS-MRI takes compared to SMORE and ECLARE. 6.Some minor weaknesses: 1.Please consider using ~\ref{} (with the tilde symbol) for all in-text references, since this introduces a non-breaking space between the word and citation. 2.Section 1, paragraph 2: Typo in algorithm name (should be ESPRESO with two total S’s).
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
I believe this paper is well-motivated with many fair experiments. I hope my comments listed in the “Major Weaknesses” section lead to a stronger paper after rebuttal. I also personally believe that all super-resolution paper should explicitly write out the forward problem they are studying for clarity. This should describe how the LR image is mathematically derived from the HR image, and should also correspond to how the method works as well as any simulation experiments. This would clarify my comments regarding slice thickness.
- 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?
As is probably obvious by my above comments, my primary concern with this work is the ambiguity around modeling of slice thickness. However, to the best of my knowledge, most (perhaps all) INR-for-super-resolution works that are published do not appropriately model nor discuss slice thickness in MRI. For this reason, I don’t think that this paper should also be punished. However, I think it would do good for the field for the authors to consider commenting on how thick slices should be handled with INR and the implications on the method and results. The rest of the paper’s strengths outweigh this concern, though, so I believe the paper should be accepted.
- 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.
The authors did an excellent job in their response. My previous score was accept and my current score is still accept. I also want to quickly take a moment here to reiterate to the authors that saying “we did 4x downsampling” essentially implies that the original image was subsampled with linear or cubic spline interpolation to have 1/4 the number of voxels along the axes of interest. But in 2D MRI, the appropriate method for “4x downsampling” would first be a blur with rect or gauss (as the authors did in their prostate LR simulation as regarded in the R2R), and then a subsampling. Again, this does not change my score, the paper should be accepted, but I only harp on this for the authors’ edification and to improve future work.
Review #2
- Please describe the contribution of the paper
The authors introduce a single-subject INR-based method for MRI super-resolution from two orthogonal stacks. The proposed framework has three phases during training: regressing intensities onto coordinates based on the first view, registering the second view to the first view, regressing intensities onto coordinates based on both the first and the second view. Then, during inference, there is plenty of flexibility in terms of the lattice on which the final image is sampled. The method is validated on both simulated brain data and in vivo prostate data.
- 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 authors demonstrate through ablation studies that using hash encoding and FreeNeRF improves the performance of the proposed method. This provides evidence and guidance for readers should they wish to build their own INR-based super-resolution frameworks. 2.The evaluation metrics includes both conventional metrics like PSNR and SSIM, and also learning-based metrics that capture perceptual similarity. 3.The organization of the paper is great. Notations are consistent and are easy to follow.
- 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 major weakness of the paper is its limited novelty. A similar work, with a different application (fetal imaging), is NeSVoR by Xu J, et al., IEEE TMI, 2023 (https://doi.org/10.1109/tmi.2023.3236216). The task is almost the same: multi-view super-resolution with explicitly modeled subject motion across views. The method is also similar: NeSVoR also uses INR with hash encoding. In addition to NeSVoR, there is another work named IREM that uses INR for MRI reconstruction, which appeared in MICCAI a couple of years ago: Wu Q, et al., MICCAI 2021 (https://doi.org/10.1007/978-3-030-87231-1_7). This makes the novelty of this work limited unless the authors, during rebuttal, make a strong case for why their work demonstrates non-incremental improvement over these existing methods.
Some design choices are poorly justified/not supported by evidence. The authors should consider providing more clarifications during revision. 1.Regarding Phase 2: Implicit Image Registration. 1.a. Why is NCC used as the similarity measure? It is slower compared to MSE and using NCC for within-modality registration is an overkill. 1.b. Why is a non-linear deformation model used for image registration? Typically, bulk motion in brain MRI is modeled as rigid transformation (6 DOF), which is also what the authors claim to use for their simulation experiments. Non-linear deformation might be helpful for prostate MRI due to possible bowel movement which is non-linear, but most of the time affine transformation (12 DOF) is sufficient. Why not just directly learn the 12 affine transformation parameters rather than using an INR to parametrize a dense displacement field which is close to affine anyway due to the strong bending energy constraint? Using INR for an underlying transformation that is near-affine offers marginal benefit and is an overkill. 1.c. Consider reporting what \alpha for the bending energy regularization is used for the in vivo clinical prostate MRI experiment. 2.Additional experiments would be helpful for the readers to better appreciate the work. 2.a. For example, it would be helpful to see a visualization (if space permits) of the results in Table 2.It would be interesting to see what w/o FreeNeRF & w/ HE looks like since it gives slightly worse performance according to conventional metrics but slightly better performance according to learning-based metrics. 2.b. In might be helpful to also report runtime/number of parameters in Table 1 since the authors intend this work to eventually have clinical applications where speed is important. 2.c. Table 1 only reports result on a single simulated data. Consider randomly choosing a couple of (say, 10) different rotation angles and report metrics over these 10 experiments. The rationale for this suggestion is that there is no other result in the paper that test for the robustness of Phase 2: Implicit Image Registration. 3.Regarding the in vivo prostate MRI experiment. 3.a. The authors are aware of the lack of ground truth for in vivo prostate MRI experiment. However, it makes little sense to compare reconstructed slices to the closest slice in the “ground truth” just to avoid blurring introduced by registration. Doing so makes Fig. 3 very hard to interpret. 3.b. In Fig. 3, only perceptual metric from MONAI is reported. It doesn’t hurt to also report other metrics, e.g., MAE, PSNR, SSIM, along with the learning-based metrics, as long as the authors provide explanations of why some of these metrics do not necessarily reflect true visual quality of the reconstruction.
There are some typographical errors, e.g., spaces in front of citations.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
1.It would be helpful to provide the readers with some additional context. Specifically, consider highlighting the fact that the proposed method works on a single-subject and does not require training on a large and carefully curated dataset, making it less likely to hallucinate. 2.In my opinion, the framework is flexible in a sense that additional views (beyond two views) can be incorporated to further improve the performance. It would be helpful to say that as a future direction.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
A score of 2 is given due to the limited novelty of this work. However, a score of 1 is not given because there are no major flaws.
- 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’ responses address some of my concerns. However, I maintain my opinion that the work lacks novelty since the motivation and the methodology resembles that of NeSVoR and the improvement is marginal.
One of the main unaddressed concern is the following. The authors claim that the main contribution/novelty over previous works is (a) incorporation of non-linear motion and (b) integrated motion correction/registration. However, regarding (a), the authors do not have an experiment demonstrating the benefit of using non-linear motion model over using a more simplified linear motion model (12 or 6 DOF). Regarding (b), the authors claim that they will incorporate simulated motion experiment.
The clarification on NCC/MSE and the use of non-linear registration is appreciated. However, the authors claim that “[w]e chose NCC because MRI intensity profiles can shift across acquisitions, making MSE unreliable,” which is in general not true. “For prostate, we used no regularizer (α=0).” Lack of regularization is in general not a good idea since it typically leads to unrealistic deformations.
Review #3
- Please describe the contribution of the paper
The paper introduces SIMS-MRI, a single-subject INR for multi-view super resolution. The paper proposes a 3-stage pipeline to train INRs, register multiview signals, and learn a fused multiview signal representation for isotropic and potentially high-resolution volume reconstruction. Overall, the work provides a stable and good setup for approaching such problems.
- 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.Decoupling the INR training and registration step seems like a good way to handle misaligned data. Furthermore, INRs excel at showing generalizability with homogeneous data – which makes the decoupled registration even more interesting. 2.Qualitative results show that SIMS-MRI has substantial improvements over prior methods. Quantitative results further confirm its performance. A real-world clinical example also shows that SIMS-MRI captures high-frequency details better than prior methods; however, it seems to lack contrast.
- 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 paper sets up the problem that MRIs have poor through-plane resolution, and then goes on to propose a multiview interpolation scheme, mainly using INRs. Results, however, do not show an improvement or explicit evaluation on through-plane MRI slices – demonstrating the effectiveness of the proposed method, particularly with respect to the problem stated in its introduction. 2.It is unclear whether the simulation has realistic in-plane and through-plane sampling rates, and whether the evaluation for interpolation performance is happening on unsampled slices. 3.Can the authors comment on how their model performs when the in-plane: through-plane sampling ratio changes? How much information trade-off / super-resolution capability arise from dense in-plane measurements?
- 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.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents an interesting method for super-resolution of single-subject MRIs. Evaluation seems unclear and lacks key details. Please refer to the weaknesses for further clarification.
- 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 R1, R2, R3, and Rmeta for their constructive feedback and encouraging comments. All suggestions, including minor writing improvements, will be reflected in the camera-ready version as rules and space permit. We address the remaining concerns below.
Novelty (Rmeta and R2): We agree on shared similarities with NeSVOR and IREM. However, NeSVOR addresses only rigid motion, and IREM requires pre-registered volumes. SIMS-MRI incorporates its own registration step, eliminating prior alignment while handling linear/nonlinear misalignment, making it more generalizable for clinical MR settings where anisotropic acquisitions are the norm. We will cite both and clarify.
Slice thickness, spacing, and simulation (R1, R3 and Rmeta): For brain, we 4× downsampled isotropic images (e.g., 1×1×1 mm → 1×1×4 mm); for prostate, native slice thickness is 3 mm (Sec. 3.2). Anisotropic volumes are used as-is (no further subsampling) by converting directly to intensity point sets. The final isotropic inference grid uses the original GT grid for brain and a 1×1×1 mm subsampled grid for prostate. We acknowledge that modeling the physical slice-selection profile and through-plane PSF for simulation would produce a more realistic simulation. We see this as future work. Importantly, evaluation always uses the held-out sagittal view unseen during training, testing recovery of an unseen orientation.
Runtime and parameters (R1, R2 and Rmeta): End-to-end processing could take as little as 20 min/subject on a single A100.Implementation is unoptimized. Multi-GPU support and better implementation could substantially reduce this. Parameter counts (hash encoder (268.4M), MLP (4.7M), and registration model (0.1M)) will be reported in the camera-ready.
Pseudo-GT and metrics for prostate (R1, R2): We are aware of the pseudo-GT limitation and will discuss it explicitly in camera-ready. Following R1’s suggestion, we simulated 3 mm thick slices via rect and Gaussian convolution. Results were consistent with reported findings. On MAE/PSNR/SSIM, rankings are mixed across methods. In addition to misregistration, these metrics are sensitive to contrast changes, which might explain the mixed results. DISTS and MONAI metrics show SIMS-MRI significantly outperforming baselines.
NCC vs MSE; non-linear deformation; bending energy α (R2): We chose NCC because MRI intensity profiles can shift across acquisitions, making MSE unreliable. Results with MSE had misalignments in neck/facial regions. For deformation, we began with 12-parameter affine registration for brain (yielding similar results), but adopted non-linear deformation to maintain a general-purpose architecture suitable for all experiments and to handle misalignments in non-brain area. For prostate, we used no regularizer (α=0). These settings are application-specific and will be noted as a tunable choice in the discussion.
Additional experiments (FreeNeRF/HE visualization; additional simulation scenario) (R2): Both sets of experiments are complete. We tested randomly chosen rotation+translation combinations for brain, with results close to those reported. We will include these and the visualizations as rules permit.
Additional view for better SR (R2): Great point. As follow-up work, we are working on a method to utilize all available series (views, contrasts, etc) in a study. We will mention this as future work.
Through-plane clarification; anisotropy ratio study (R3): The paper reports results on through-plane slices: only LR axial and coronal views were used for training, and performance is evaluated on the held-out sagittal GT. We will revise the text to clarify. Regarding varying in-plane: through-plane ratios, conceptually, SIMS-MRI benefits most when orthogonal views provide complementary high-resolution in-plane information, and should degrade as anisotropy increases or view complementarity decreases. A systematic anisotropy ratio study is an important direction we will include in future work.
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.
This paper received mixed reviews that should be addressed in a rebuttal. The split among the reviewers is mainly about novelty and problem formulation rather than whether the method works. The paper addresses a meaningful MRI SR setting from multiple anisotropic acquisitions using only the anisotropic image. The reviewers found the decoupled INR registration and reconstruction setup promising and the results convincing with extensive analysis. However, there are substantial concerns about the novelty of the approach, especially with regards to NeSVoR. The rebuttal should focus on this aspect. There are other, more minor, concerns that should be addressed such as the forward model definition, clarification between sparse slices and thick slices, and runtime.
- 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 maintains mixed reviews after rebuttal, but the authors have provided clear and detailed explanations that address the majority of the reviewers’ concerns. Novelty remains a concern, but the motivation for the methodology is quite clear and supported by clinical and physical reality. I would argue that non-rigid motion can be accepted between modalities, especially outside of the brain, and should be considered a known quantity. However, this should be directly evaluated at a later time. The author’s rationale for using NCC is frequently cited in MRI registration literature. Given the authors’ rebuttal and their addressing of the concerns, I will recommend acceptance for this paper.
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.
After rebuttal, I recommend Accept. The authors addressed many key technical and presentation concerns, including slice thickness/anisotropic spacing, held-out sagittal evaluation, prostate pseudo-GT limitations, runtime/parameter counts, and the rationale for NCC and nonlinear registration. They also added requested experiments on realistic slice-profile simulation and additional motion settings. Some weaknesses remain, particularly regarding comparison with NeSVoR/IREM. However, the paper presents a useful single-subject multi-view MRI super-resolution framework that integrates registration and fusion without large supervised datasets. Given the positive critiques from R1 and R3, improved clarity after rebuttal, and clinical relevance, I recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes a clear single-subject multi-view INR framework with integrated registration and reconstruction. The rebuttal reasonably clarifies the relation to NeSVoR/IREM, the slice-thickness setting, runtime, and additional experiments.
I also do not find the use of NCC problematic, as NCC is commonly used in MRI registration when intensity profiles may vary across acquisitions. Although the novelty is somewhat incremental and the physical slice-profile modeling could be stronger, the method is technically sound, experimentally convincing, and useful for the community, so I recommend acceptance.
