List of Papers Browse by Subject Areas Author List
Abstract
Dynamic 3D Magnetic Resonance Imaging (MRI) is a powerful imaging technique for motion monitoring and tracking, offering both excellent soft-tissue contrast and the ability to capture dynamic changes in tissue. Current reconstruction methods typically assume that multiple spokes share the same motion state. However, this assumption does not align with the complex realities of patient motion and clinical acquisition protocols, often resulting in anatomical discontinuities or blurring artifacts in the reconstructed images. In this work, we propose an unsupervised Single-sPoke motion-compensated Implicit NEural Representation method (SPINER) for dynamic volumetric MRI reconstruction. We address a more challenging yet realistic scenario, single-spoke motion modeling, which assigns a unique motion state for each spoke measurement. To address this highly ill-posed inverse problem, we propose a motion-ignoring static initialization strategy that exploits static anatomical information across all spokes. We find that a good initialization of the canonical volume significantly improves the optimization process and facilitates better dynamic volumetric reconstruction based on implicit neural representation learning. Experiments on abdomen MRI datasets demonstrate that our methods can reconstruct high-quality dynamic volumetric MRI while capturing continuous and accurate motion.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0398_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{CheLix_Singlespoke_MICCAI2025,
author = { Chen, Lixuan and Balter, James M. and Shen, Liyue and Park, Jeong Joon},
title = { { Single-spoke Motion-compensated Dynamic 3D MRI Reconstruction via Neural Representation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {518 -- 528}
}
Reviews
Review #1
- Please describe the contribution of the paper
- The core contribution is the formulation of a single-spoke motion-compensated dynamic 3D MRI framework using neural implicit representations.
- The paper proposes a novel initialization strategy—motion-ignoring static volume estimation—to improve convergence and quality in an otherwise severely underconstrained setting.
- The framework demonstrates that even under extreme undersampling (1 spoke per 3D volume), high-fidelity dynamic reconstructions with temporally coherent motion fields are possible using implicit models and spatiotemporal deformation learning.
- 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.
Addressing single-spoke motion modeling responds to a highly realistic and challenging clinical scenario where traditional motion-grouping assumptions break down. In this regard, the use of INR for jointly representing 3D anatomy and 4D motion fields aligns well with recent trends in neural scene representation, and the proposed motion-ignoring initialization directly tackles the instability of the optimization. The authors conduct quantitative comparisons against multiple baselines (NUFFT, TD-DIP, Naive INR with and without temporal windowing) and include both PSNR/SSIM and qualitative spatiotemporal profile visualizations. The impact of both the initialization strategy and the single-spoke motion formulation is clearly isolated and supported with appropriate visual and quantitative evidence. The method is formulated generally and could, in principle, be adapted to more complex motion beyond respiratory dynamics, such as cardiac applications.
- 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 dataset diversity: The method is only validated on two in-house abdominal DCE-MRI scans. This raises concerns about generalizability across anatomies, motion types, scanner configurations, and vendors. Even limited validation on a publicly available 2D+t or other benchmark datasets would help strengthen the claim of robustness.
- Missing comparison with reference [1]: The paper does not compare with or discuss the recently published “Dynamic-aware Spatio-temporal Representation Learning for Dynamic MRI Reconstruction” (Baik & Yoo, arXiv:2501.09049), which also combines spatio-temporal decomposition, implicit modeling, and motion compensation. While SPINER is more aggressive in modeling per-spoke motion, both works share architectural similarities. A side-by-side conceptual and experimental comparison is necessary to clarify contributions and novelty.
- Unclear scalability: It remains uncertain how the method scales to longer sequences or larger spatial resolutions. The joint optimization of two neural networks (canonical + DVF) could become computationally expensive as temporal resolution increases.
- Caption detail and interpretability: Some visualizations (e.g., Fig. 3) would benefit from clearer captioning and explanation—particularly the meaning and directionality of DVF vectors and how they relate to breathing phases
- 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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.
(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?
The paper addresses a non-trivial and practically relevant inverse problem in dynamic MRI and proposes a method that is both novel in formulation and well-validated. The use of motion-ignoring initialization is insightful and shows real benefits in optimization. However, the lack of evaluation on more diverse datasets limits the strength of its generalization claims. The reviewers would strongly prefer to see at least one additional dataset (public or simulated) to support robustness. Additionally, the comparison with reference [1] (Baik & Yoo) is a missing which is seemingly similar in structure. Clarifying the distinction and demonstrating superiority (or complementarity) would improve the credibility of the novelty claim. Despite these issues, the paper is clearly written, technically correct, and proposes a valuable framework. With a detailed rebuttal addressing the above concerns—especially comparative and generalization limitations—it is likely to be acceptable for publication.
- 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
Review #2
- Please describe the contribution of the paper
The authors present an approach to reconstruct dynamic MRI images by applying motion compensation to individual stacks of spokes in a stack-of-starts MRI acquisition. To that objective, they implicit neural representation networks to encode both the deformation vector fields and the image reconstruction.
- 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.
By allowing motion compensation on a stack-of-spokes basis, the authors potentially allow reconstruction of more rapid and complex dynamics. Model used is interesting and shows promising 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.
- use of synthetic data as well as fairly simple dynamics (almost no respiratory motion in the sequences shown).
- missing comparison against practical dynamic MRI methods such as GRASP. NUFFT is clearly ill-suited for this task.
- no details on training parameters nor reconstruction time
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- 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.
(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?
Single spoke reconstruction could provide good image quality even in cases of rapid dynamic changes. The model presented is interesting and might allow for reconstruction of non-repeating motions such as deep breaths, However, it is unclear that the model can actually capture rapid dynamics that make single-spoke reconstruction useful. The evaluation is limited to relatively smooth dynamics which can be reconstructed with more standard reconstruction with methods such as GRASP.
- 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
Review #3
- Please describe the contribution of the paper
This paper addresses the challenging single-spoke motion modelling for dynamic 3D MRI reconstruction. They propose a motion-ignoring static initialization strategy that exploits static anatomical information across all spokes, which shows great improvement. Extensive experiments demonstrate their effectiveness over previous approaches.
- 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 paper proposes a new motion-ignoring static initialization strategy for dynamic 3D MRI reconstruction, and demonstrate that a good initialization can significantly improve the performance of this task.
++ The paper’s visualization of canoniocal volumes and DVFs clearly show the learned results of their method.
++ The paper is well written and organised.
- 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.
– In the Motion-ignoring Static Initialization section, the paper disregards motion and exploit all aggregated spoke measurements to initialize the canonical network. It is suggested to evaluate the performance under various motion patterns. The method might degrade a bit when handling large or complex motions.
– It is suggested to compare the training and inference time of the proposed method against baselines as well.
– Minor weakness: Since the modelling approaches are MLPs instead of explicit volumes, it would be better to avoid using explicit volumes terms to avoid confusion in the paper writing.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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.
(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?
The paper’s main contributions are interesting and reasonable, and they have been demonstrated to be effective in dynamic 3D MRI reconstruction. Therefore, the reviewer is leaning towards acceptance of this paper.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Limited dataset diversity (R1Q1&R2Q1&R3Q1): We thank the reviewers for the valuable suggestions. We agree that evaluating the method under diverse motion types and larger or composite motions is important, and we will extend our experiments accordingly to further assess generalizability. Regarding R1’s suggestion of using 2D+t data, we believe that validation on 3D+t data is more appropriate for our formulation, as decomposing 2D+t sequences into a canonical volume and DVFs may not be meaningful due to out-of-plane motion causing structure appearance/disappearance.
Compared methods: We thank the reviewer for pointing out the missing comparison
- GRASP (R2Q2): We agree that GRASP provides a stronger baseline than simple NUFFT and will include it in our comparison. Compared to GRASP, which groups several consecutive spokes to form undersampled dynamic frames and may introduce motion blur, our single-spoke motion modeling offers a more accurate representation of real-world motion.
- DA-INR (R1Q2): We thank the reviewer for pointing out the recent work [1]. While both our method and DA-INR share the idea of combining spatio-temporal decomposition with implicit modeling, our approach differs in several key aspects: 1. We focus on 3D+t MRI data, where decomposing into a canonical volume and DVFs is more appropriate for capturing volumetric motion, compared to the 2D+t setting used in DA-INR. 2. DA-INR relies on semantic features extracted from reconstructed undersampled images to guide network optimization, which can be unreliable under extremely sparse sampling like single-spoke motion modeling. In contrast, our motion-ignoring static initialization strategy leverages the temporal redundancy of static tissue to robustly guide optimization, even in highly ill-posed cases. 3. Our single-spoke motion modeling captures finer motion variations and avoids motion artifacts caused by grouping spokes under shared motion assumptions.
Unclear scalability (R1Q3): We thank the reviewer for the valuable comment. We agree that increasing spatial resolution and sequence length can lead to higher computational cost. We will include a discussion on how the performance and reconstruction time of our method vary with longer sequences, to better clarify its scalability.
Training parameters & reconstruction time (R2Q3&R3Q2): We use the Adam algorithm with default hyperparameters to optimize the model. For the motion-ignoring static initialization, the learning rate for the canonical network is set to 0.001 and decays by half every 300 epochs, for a total of 900 epochs. In the subsequent joint training, the learning rates for both the canonical and DVF networks are initialized to 0.001 and decay by half every 15 epochs, with a total of 50 epochs. The motion-ignoring static initialization takes about 0.5 hours, and the joint training takes about 3 hours on a single NVIDIA A100 GPU. We will add this description to the paper.
Presentation comments (R1Q4&R3Q3): We thank the reviewers for the helpful suggestions. We will revise the captions, particularly for Fig. 3, to better clarify the meaning and directionality of the DVF vectors and their relationship to breathing phases. We will also avoid using terms related to explicit volumes where they are not appropriate, in order to prevent confusion.
[1] “Dynamic-aware Spatio-temporal Representation Learning for Dynamic MRI Reconstruction” (Baik & Yoo, arXiv:2501.09049)
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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”.
N/A