List of Papers Browse by Subject Areas Author List
Abstract
Real-time 4D-MRI promises to enhance adaptive radiotherapy through superior soft-tissue contrast and instantaneous motion visualization. Unfortunately, its clinical implementation is confronted by aggressive spatial undersampling to achieve real-time acquisition, which incurs severe image degradation. Super-resolution (SR) offers a viable solution to recover image quality from degraded inputs; however, applying it to real-time 4D-MRI faces two key challenges: how to precisely reconstruct intricate human anatomy (i) from highly corrupted inputs without essential structural details, and (ii) in the absence of ground-truth for supervised training. Regarding these challenges, we propose DuReg-SR, a Dual-Registration reciprocal SR framework. Through dual reciprocal registration, DuReg-SR effectively (i) incorporates routinely acquired diagnostic MRI scans as anatomical priors for recovering authentic anatomy, and (ii) leverages static breath-hold high-resolution (HR) scans as pseudo ground-truth for training. For reciprocal SR and registration, DuReg-SR introduces a Hierarchically Recursive Mutual Network (HRMNet), which performs both tasks in a coarse-to-fine hierarchy to progressively infuse T1w priors for anatomy reconstruction. Further, an Anatomically-Constrained Mutual Training (ACMT) scheme is proposed, which spatially warps a single static HR scan to supervise dynamic 4D-MRI frames while preserving anatomical consistency across temporal sequences. Extensive experiments show that DuReg-SR achieves state-of-the-art performance in both visual quality and quantitative metrics. Code will be released upon publication.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0486_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanYin_DualRegistration_MICCAI2026,
author = { Wang, Yinghui AND Meng, Mingyuan AND Wang, Lu AND Cai, Jing},
title = { { Dual-Registration Reciprocal Super-Resolution for Prior-Guided Real-Time 4D-MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This manuscript introduces DuReg-SR, a dual-registration reciprocal super-resolution framework addressing the severe image degradation caused by aggressive spatial under-sampling in real-time 4D-MRI. Due to the difficulty of collecting perfectly aligned ground-truth data, the authors creatively utilize routinely acquired diagnostic MRI scans as anatomical priors and static breath-hold HR scans as pseudo ground-truth for model training. The proposed framework relies on two main innovations. First, the Hierarchically Recursive Mutual Network (HRMNet) deeply couples SR and image registration across multiple scales, allowing the reference anatomy to be progressively warped and infused into the highly corrupted dynamic frames. Second, the Anatomically-Constrained Mutual Training (ACMT) scheme employs a novel Artifact-Resilient Contrastive (ARC) loss, utilizing a lightweight VoxelMorph registration module to ensure geometric consistency between the generated frames and the pseudo ground-truth without misinterpreting artifacts as valid anatomical structures. Evaluated on an in-house liver 4D-MRI dataset and the public CHAOS dataset, the method significantly outperforms current single-image and multi-contrast SR 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.
- Highly Novel Clinical Integration: The framework identifies a critical bottleneck in real-time 4D-MRI (the lack of perfect ground truth) and elegantly solves it by repurposing underutilized clinical assets—diagnostic T1w scans for anatomical guidance and breath-hold HR scans for training supervision. This marks a significant paradigm shift from conventional single-image super-resolution to prior-guided reconstruction.
- Methodological Soundness: The reciprocal design of HRMNet ensures that registration and super-resolution mutually reinforce one another at each resolution level; improved SR yields better spatial matching, which in turn provides more accurate anatomical priors for further SR refinement. Additionally, the ARC loss intelligently addresses the geometric distortions typical in GAN-based models by assessing similarities on the warped input rather than the direct SR output.
- Strong Quantitative and Visual Results: The method is rigorously benchmarked against state-of-the-art models (AST, ATD, A2-CDic, DANCE), showing consistent and substantial improvements in PSNR, SSIM, and LPIPS, particularly at aggressive acceleration factors up to ×32.Extensive ablation studies (evaluating recursive depths and individual training losses) systematically validate the efficacy of both HRMNet and ACMT.
- 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.
- Limited Clinical Dataset Size: The real-world clinical evaluation relies on an in-house 4D-MRI dataset comprising only 26 liver cancer patients. Validating the model on a larger, multi-centric cohort would be necessary to confidently ensure generalizability across diverse anatomies and cancer morphologies.
- Reliance on Synthetic Deformations for Public Data: For the experiments conducted on the public CHAOS dataset, inter-fraction misalignment resulting from respiration was simulated using synthetic Gaussian deformations. This synthetic approximation may not fully encapsulate the complex, non-linear respiratory and peristaltic motions encountered in real patients.
- Unverified Real-Time Latency: The manuscript explicitly notes that real-time imaging protocols demand a stringent sub-300 ms latency. While the method achieves excellent visual fidelity, the text does not detail the inference speed of the complex, multi-level recursive HRMNet. It remains unclear if the framework can reconstruct images fast enough during deployment to meet the strictly defined real-time radiotherapy gating requirements.
- Errors in Fig. 1.Input of the registration and SR modules should be k+1 instead of k-1.
- Typo. The first sentence of subsection 2.1 should be “As shown in Fig. 1,”
- 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.
(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?
I scored this paper as a weak accept because the it focuses on an important topic and results are promising, with a reasonably clear presentation and potential interest to the MICCAI community. The main limitation is that all results on based on synthetic data instead of realistic data, but I felt the combination of practical relevance and generally solid evaluation was sufficient to place it slightly on the accept side.
- 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 rebuttal, I am satisfied that the authors addressed the main concerns raised in the initial reviews, especially for concern about the limited clinical data size and reliance on synthetic deformation.
Review #2
- Please describe the contribution of the paper
Please see the major strengths.
- 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) This work addresses the SR problem for 4D MRI, which represents a clinically significant task with substantial practical implications. To my knowledge, prior research specifically targeting 4D MRI SR (and image restoration) remains limited.
(2) The proposed HRMNet architecture and ACMT training strategy exhibit good methodological innovation. The experimental results on two datasets show performance improvements over existing approaches. Moreover, the ACMT strategy a general-purpose framework that could potentially enhance other multi-contrast SR networks beyond the specific application presented here.
- 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) Multiple variables and loss functions lack explicit mathematical definitions, e.g., \phi_i^{S\toH}, L_adv, L_smt, although they can by inferred from the context. The authors should be aware that excessive use of inline expressions can affect the readability of paper.
(2) Also, the figures contain many arrows and symbols. Figure 1 includes the “K-2” case, which appears unnecessary and could be simplified.
(3) Questions for experimental design: 3A) The paper appears to process each frame (i) independently. Have the authors considered incorporating temporal correlations between frames, or estimating inter-frame deformation fields? Typically, single-modality registration within the same MRI sequence is easier than cross-modality (e.g., between R and S) and cross-scale (between R and L) registration. 3B) Is the number of patients (26) of the in-house dataset sufficient for model optimization? 3C) The backbone network utilized for LPIPS metrics is not specified. 3D) The undersampling method of MRI data in CHAOS is not mentioned.
(4) Based on point (3), this work appears to address a broader image restoration (IR) problem, rather than super-resolution (SR), i.e., the input I^L has a spatial resolution of (HWD) that is equal to I^H, although anatomical details are corrupted during the undersampling.
- 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 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?
Despite the clinical relevance and methological innovations, the paper suffers from organizational and presentation issues: the over-use of inline formulas and unclear variable definitions may cause significant barriers to reader comprehension. If the authors could give a clear revision plan (yet, NOT violating the page limit) addressing these readability concerns in the rebuttal, I may reconsider the overall recommendation score.
- 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 still recommend that the authors should pay attention to distinguishing between Image Restoration and Super-Resolution. It has been noted that the author has mentioned several references for multi-contrast SR [3, 5, 11-14], but their downsampling methods involves retaining low-frequency signals in the central region of K-space and truncating high-frequency outlier regions, which is equivalent to reducing spatial resolution. Other undersampling methods (e.g., radial undersampling in CHAOS of this paper) do not strictly equate to spatial resolution reduction.
Nevertheless, I believe this is a good paper with satisfactory scientific merit. The authors’ rebuttal has addressed most of my concerns on writting format and experimental design.
Review #3
- Please describe the contribution of the paper
This paper proposed a super-resolution approach for 4D MRI. The method itself is novel and offers a potential solution for training when ground truth is unavailable. Experiments show the benefit of the proposed approach.
- 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.For the challenges of no ground truth for supervision, the paper proposed to use the breath-hold scans to serve as the pseudo-ground-truth. 2.The paper proposes to use anatomical scans (T1 scans) as a reference to enhance the SR details. 3.The paper proposes to jointly train SR and registration modules in the model.
- 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 training and evaluation are both on the pseudo-ground-truth by the HR breath-hold scans, which may raise concerns about the validity of the results. Although it is understandable that true ground truth is generally unavailable, the authors should at least discuss the reliability of using such pseudo-labels or conduct a reader study to assess the clinical quality of the SR results. 2.The qualitative results should show different time frames across the scan to show the effectiveness of the SR results, considering this is a 4D scan. 3.Some more recent generative-based SISR methods like diffusion models should be included in comparison.
- 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.
(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?
Although the framework has some weaknesses as mentioned in the previous section, it shows novelty, and the experiments demonstrate superior image quality.
- 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 discuss the limitations of true ground truth and explain this using the evaluation protocols in the paper. I suggest that the authors add results from other time points to demonstrate the effectiveness of the 4D SR approach and explore reader-study evaluation in future work. I suggest accepting this paper.
Author Feedback
We thank all reviewers (R1, R3, R4) for their constructive comments. We provide our responses to major concerns below and will correct all minor issues (typos, Fig.1) in the camera-ready version.
1: Limited 4D-MRI Data (R1, R3) As 4D-MRI is not yet clinical routine, it’ s extremely scarce and current studies typically enrolled ≤30 cases (e.g., Refs [1, 25, 27] in the paper). Nevertheless, each 4D-MRI contains 72 temporal frames, providing sufficient data for model optimization, as confirmed by converged validation loss. For rigorous evaluation, all experiments were repeated five times with Wilcoxon signed-rank tests in Table 1, and a public 3D MRI dataset (CHAOS) was used for additional evaluation.
2: Synthetic Deformations in CHAOS (R1) We wish to clarify that our main results (in-house 4D-MRI dataset) were obtained with real inter-scan deformations, validating performance on actual respiratory motion. Further incorporating public CHAOS data with synthetic deformations provides pixel-wise ground truth for quantitative evaluation, which is a widely used approach in related literature (Refs [3, 11] in the paper).
3: Real-Time Latency (R1) We apologize for missing this information. Our method meets real-time requirements (total ≤500 ms, MRI imaging ≤300 ms to allow reconstruction) [25]. Ours finished reconstruction in 252 ms (k=4) for ×4 4D-MRI (~173 ms), totaling ~425ms latency—well within the limit.
4: Unclear Presentation (R3) Given the extra half-page space in the camera-ready version, we plan to explicitly define ambiguous variables (e.g., Φ_R→L, Φ_S→H) and avoid inline-formatted losses (e.g., L_adv, L_smt, L_recon) to improve readability. Also, we will provide additional implementation details with our open-source code for better reproducibility.
5: Incorporating Temporal Correlations (R3) We’d like to note that our work focuses on prior-guided SR, which is crucial for real-time 4D-MRI under heavy undersampling because it is extremely difficult to recover lost structures based solely on 4D-MRI information. We agree that incorporating temporal correlations is also promising for 4D-MRI, but there also exists a concern that reliance on inter-frame correlations may lead to error accumulation, where reconstruction errors from previous frames can accumulate to the current one. We will further explore this in future studies as this is a non-trivial topic beyond the scope of this work.
6: LPIPS backbone & MRI Undersampling Method (R3); More Visualization (R4) Backbone network for LPIPS is AlexNet, and undersampling method in CHAOS is radial undersampling, which will be added to the camera-ready version. Visualization of different time frames will be provided in online resource with our code.
7: Image Restoration (IR) or Super-Resolution (SR)? (R3) We respectfully suggest that undersampled MRI reconstruction is typically defined as a SR problem, consistent with existing RefSR and multi-contrast SR works (e.g., Refs [3, 5, 11-14] in our paper).
8: Pseudo Ground-Truth (GT) of 4D-MRI (R4) Hardware limitations prevent acquiring fully aligned GT for 4D-MRI. To address inevitable misalignment in pseudo GT, we proposed ACMT for training (mitigating SR-HR spatial mismatches via a registration module) and used Elastix for rigorous pre-registration during evaluation. The latter evaluation protocol is widely used in related literature where GT is inherently missing (Refs [23, 27] in the paper). In addition, as mentioned in Q1/Q2, we also adopted the CHAOS dataset with simulated deformations for rigorous quantitative assessment against real pixel-wise GT. We agree that reader study would strengthen clinical validation and will include this as future work.
9: Compare with Diffusion-based SOTA (R4) As suggested, we add a comparison to Diffusion-based InvSR (CVPR’25) and observe that our method still outperforms it with statistical significance (e.g., 20.11 vs 22.04 in PSNR, 85.01 vs 92.34 in SSIM for ×4 4D-MRI, p < 0.001).
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 recommendations from the reviewers, including one Weak Reject, one Weak Accept, and one Accept.
The paper proposes a 4D-MRI super-resolution approach using bidirectional registration constraints. At the same time, the reviewers identified several substantive concerns, including insufficient evaluation (R1, R2), reliance on synthetic deformations (R1), unverified real-time latency claims (R1), a mathematical formulation that was viewed as insufficiently rigorous (R2), and evaluation based on pseudo-ground truth (R3).
Overall, the paper shows methodological merit, and the meta-reviewer therefore recommends that it proceed to rebuttal. The authors are encouraged to focus their response on the major concerns outlined above.
- 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 received consistently positive evaluations following the rebuttal. Reviewer 2 revised the rating from Weak Reject to Accept, noting that the rebuttal satisfactorily addressed most of the major concerns and recognizing the scientific merit of the work. Reviewers 1 and 3 also provided final Accept recommendations, indicating that the authors’ responses adequately addressed the key issues, particularly those concerning the limited size of the clinical dataset and the inherent difficulty of establishing true ground truth.
Based on the post-rebuttal discussion and the overall reviewer consensus, I recommend Accept. For the camera-ready version, the authors are encouraged to further improve the presentation by incorporating the reviewers’ constructive suggestions, such as clarifying the IR-versus-SR terminology and improving Figure 1.
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.
All reviewers recommend accepting this work post-rebuttal.
This is a solid paper with a convincing rebuttal. I share the recommendation to accept this paper:
- The authors successfully addressed the concern regarding dataset size, synthetic deformations and latency. The dataset sizes and mix of internal clinical data and synthetic public data is sensible.
- the problem and clinical application is comparably novel and of interest to the community
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 authors have addressed the reviewers’ comments.