Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

We propose a hierarchical multi-node/multi-GPU framework for respiratory motion-resolved reconstruction of 3D non-Cartesian multi-echo gradient-echo MRI. The core idea is a topology-aware, three-level parallelization across the motion, echo, and coil dimensions: 1) we partition motion states and echo times across nodes using a 2D process grid, enabling row-wise (motion-coupled) and column-wise (echo-coupled) communications; and 2) within each node, we further distribute coil computations across GPUs so that communication-intensive coil reductions are carried out intra-node over NVLink via NCCL, while inter-node traffic is limited to nearest-neighbor halo exchanges over CXI via MPI. This design removes the single-node memory ceiling that previously prevented enlarged-FoV reconstructions and reduces end-to-end runtime by scaling to many nodes/GPUs. Experimental results demonstrate that our approach has the potential to make multi-dimensional reconstruction of 3D non-Cartesian abdominal MRI more practical and scalable.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2047_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{ZhaCha_LargeScale_MICCAI2026,
        author = { Zhang, Chao AND Jiao, Xiangmin AND Nash, Michelle AND Schaake, Ryan AND Kee, Youngwook},
        title = { { Large-Scale Distributed GPU-Accelerated Respiratory Motion-Resolved Reconstruction of 3D Non-cartesian mGRE 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 paper presents a hierarchical distributed reconstruction framework for large-scale respiratory motion-resolved 3D non-Cartesian mGRE MRI. The core contribution is a topology-aware parallelization strategy that partitions the echo and motion dimensions across nodes using a 2D MPI process grid, while distributing coils across GPUs within each node. The framework further combines cost-aware communication placement and non-blocking communication/computation overlap to reduce both runtime and per-GPU memory footprint, enabling enlarged-FoV 5D reconstructions that are infeasible on a single node. Experimental results on volunteer and patient data demonstrate strong scaling up to 36 nodes / 144 GPUs while preserving numerical fidelity and downstream quantitative metrics.

  • 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 paper addresses an important practical bottleneck. Free-breathing motion-resolved liver MRI is clinically relevant, but 5D non-Cartesian reconstruction is computationally demanding. Prior work has established the value of motion-resolved PDFF / R2* / QSM, while also making clear that computational burden is a major barrier. 2.The contribution is technically coherent and well scoped. The work does not overclaim a new reconstruction model; instead, it focuses on a well-motivated systems contribution: topology-aware decomposition tailored to the coupling structure of the 5D problem. That makes the paper more credible 3.The empirical gains are substantial and meaningful. The framework bypasses single-node OOM failures, scales to 144 GPUs, reduces runtime dramatically, and lowers per-GPU VRAM significantly. This is not just a modest speedup but a change in feasibility regime. 4.The asynchronous design appears effective at scale. The synchronous vs asynchronous comparison shows limited benefit at small scale but a clear advantage at high GPU counts, which is exactly the right behavior for a communication-hiding design.

  • 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 novelty relative to prior distributed MRI reconstruction frameworks needs sharper positioning. Distributed MRI reconstruction itself is not new: Gadgetron already introduced multi-node distributed reconstruction, and prior work also reported generic multi-node multi-GPU frameworks. The paper should more explicitly articulate what is fundamentally new here beyond “distributed MRI reconstruction,” namely the echo–motion–coil-specific topology-aware decomposition for a tightly coupled 5D non-Cartesian problem. 2.Image-quality evidence is insufficient for a reconstruction paper. The paper mainly reports runtime, VRAM, relative MSE, and ROI-level parameter differences, but provides almost no reconstructed images or quantitative map visualizations in the main paper. Statements such as “images not shown” are not ideal for MICCAI. The reader needs visual confirmation that the distributed implementation preserves image quality locally, not just in aggregate ROI means. 3.The baseline comparisons are somewhat narrow. The experiments compare different grid shapes and asynchronous vs synchronous execution within the proposed framework, but do not directly compare with optimized alternative frameworks or simpler decomposition strategies under matched hardware budgets. This makes it harder to judge how much of the gain comes from topology-aware placement versus distributed execution in general. 4.Scalability analysis is incomplete from a systems perspective. Strong scaling is shown, but weak scaling, communication/computation breakdown, and profiling of halo exchange vs NCCL reduction overhead are not reported. These would strengthen the systems claim significantly. 5.Clinical-practicality claims should be slightly toned down. The study includes only 5 subjects and evaluates downstream fidelity mainly through ROI differences and runtime. This supports feasibility, but not yet broad clinical practicality or deployment readiness. Beside, clinical-practicality may not allow 144 GPUs.

  • Please rate the clarity and organization of this paper

    Good

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    I lean positive on this paper because it solves a real bottleneck in advanced MRI reconstruction and demonstrates substantial gains at a scale that appears genuinely useful. The main contribution is not a new reconstruction objective, but a carefully designed hierarchical distributed implementation that matches the coupling structure of the underlying 5D problem. The experiments convincingly show reduced runtime and VRAM, and the asynchronous communication-hiding strategy becomes more beneficial exactly where one would expect it to. My main reservations concern the limited image-quality presentation, somewhat narrow baselines, and the need for sharper positioning against prior distributed MRI frameworks. Overall, I believe the technical contribution is meaningfull, but the current version falls slightly below the acceptance threshold.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 rebuttal strengthens the paper in several important ways. In particular, it clarifies the numerical-fidelity typo, adds a more relevant topology-oblivious baseline, and provides profiling evidence that better supports the topology-aware and asynchronous execution claims. These responses improve my confidence in the systems contribution.

    However, my final decision remains negative because, based on the submitted manuscript itself, I still find the validation package insufficient for an MRI reconstruction paper at this venue. The paper emphasizes runtime, memory scalability, and numerical agreement, but the manuscript does not provide direct reconstructed-image or quantitative-map visual validation. I am not penalizing the authors for not adding such results in rebuttal, since MICCAI does not allow new experimental results there. Rather, I view this as an unresolved weakness of the original submission that continues to affect my overall assessment.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a hierarchical multi-node, multi-GPU framework for large-scale respiratory motion-resolved reconstruction of 3D non-Cartesian mGRE MRI.

    The key idea is a topology-aware parallelization strategy that distributes computation across motion states, echo times, and coil dimensions using a 2D MPI process grid combined with intra-node GPU parallelization. The framework further introduces asynchronous communication (non-blocking halo exchange and reductions) to overlap communication with computation, thereby improving scalability.

    The method enables reconstruction of high-dimensional (5D) MRI data that would otherwise be infeasible due to memory and runtime constraints, demonstrating strong scaling up to 144 GPUs with substantial runtime reductions.

  • 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.

    • Strong systems contribution with clear practical relevance The paper addresses a real bottleneck in large-scale MRI reconstruction (runtime and memory), and provides a well-engineered distributed solution that is directly applicable in clinical research settings.

    • Well-designed parallelization strategy The hierarchical decomposition across motion, echo, and coil dimensions is well thought out and clearly motivated. The distinction between intra-node (NCCL/NVLink) and inter-node (MPI/CXI) communication is particularly effective.

    • Effective communication–computation overlap The use of asynchronous execution (Fig. 1, p.2) and the demonstrated scaling benefits (Fig. 3, p.8) are convincing and represent a meaningful technical contribution.

    • Impressive scalability results The framework scales to 144 GPUs with substantial reductions in runtime and memory footprint (Tables 1–2, pp.6–7), making previously infeasible reconstructions practical.

    • Comprehensive experimental evaluation (systems perspective) The paper includes detailed runtime, memory, and scaling analyses across multiple configurations and datasets.

  • 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.

    • Incomplete discussion of related NUFFT implementations The paper does not sufficiently acknowledge existing efficient GPU-based NUFFT libraries, such as: o gpuNUFFT (Knoll, F.; Schwarzl, A.; Diwoky, C.; Sodickson DK.: gpuNUFFT – An Open-Source GPU Library for 3D Gridding with Direct Matlab Interface. Proc ISMRM p4297 (2014).) o cuFINUFFT (Shih, Y. H., Wright, G., Andén, J., Blaschke, J., & Barnett, A. H. (2021, June). cuFINUFFT: a load-balanced GPU library for general-purpose nonuniform FFTs. In 2021 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) (pp. 688-697). IEEE.1) o torchkbnufft (M. J. Muckley and R. Stern and T. Murrell and F. Knoll, TorchKbNufft: A High-Level, Hardware-Agnostic Non-Uniform Fast Fourier Transform, ISMRM Workshop on Data Sampling & Image Reconstruction 2020) These are highly relevant baselines for the core computational bottleneck (nuFFT), and their omission weakens the positioning of the work within existing HPC MRI reconstructions.

    • Limited novelty on the reconstruction/modeling side The contribution is primarily engineering and systems-oriented, while the reconstruction formulation (e.g., PDHG-based optimization, TV regularization) follows established prior work. The paper does not introduce new reconstruction models or priors.

    • Performance gains tied to iterative reconstruction paradigm While the distributed framework significantly accelerates iterative reconstruction, it should be noted that larger performance gains could potentially be achieved by replacing iterative reconstruction with learned approaches. This is outside the scope of the paper, but worth acknowledging in the discussion by referencing some review works, e.g.:

    Hammernik K, Küstner T, Yaman B, Huang Z, Rueckert D, Knoll F, Akçakaya M. Physics-Driven Deep Learning for Computational Magnetic Resonance Imaging: Combining physics and machine learning for improved medical imaging. IEEE Signal Process Mag. 2023 Jan;40(1):98-114.doi: 10.1109/msp.2022.3215288

    • Lack of comparison to alternative acceleration strategies The work focuses exclusively on distributed scaling, but does not compare against: o optimized single-node GPU implementations o alternative algorithmic accelerations (e.g., low-rank, learned recon)

    • Limited clinical/image-quality evaluation While numerical fidelity (e.g., RMSE, <1% parameter deviation) is reported (p.7), the evaluation lacks qualitative image comparisons or clinical validation demonstrating impact on diagnostic quality.

  • 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
    • Please include and discuss existing NUFFT libraries (gpuNUFFT, cuFINUFFT, torchkbnufft), as they are highly relevant to your computational bottleneck and would strengthen the positioning of your contribution.
    • It would be helpful to include qualitative reconstruction results to complement the system-level evaluation.
    • A brief discussion on how this framework could integrate with or compare to learning-based reconstruction methods would provide useful context for future directions.
    • Clarify whether your framework could directly leverage optimized external NUFFT libraries or if it relies on custom implementations.
  • 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?

    This paper presents a strong systems and engineering contribution addressing a critical bottleneck in large-scale MRI reconstruction. The proposed topology-aware distributed framework and asynchronous execution strategy are well-designed and convincingly demonstrated.

    However, the work is primarily focused on scalability rather than methodological novelty in reconstruction, and the omission of relevant NUFFT libraries weakens its positioning. Additionally, the evaluation focuses more on computational performance than clinical or image-quality impact.

    Overall, the paper is above the acceptance threshold due to its practical importance and solid technical execution, but would benefit from improved contextualization and broader evaluation.

  • 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.

    The authors have satisfactorily addressed my concerns.



Review #3

  • Please describe the contribution of the paper

    This work proposed and validated a distributed GPU framework for solving optimization problem for MRI reconstruction that involves 3D non-uniform Fourier Transform (nuFFT). The implementation can be flexible and scalable such that certain data dimensions (e.g., echo and motion state dimensions in this work) can be partitioned across nodes and be scaled up to more than 100 GPUs. This work has also validated this approach in reconstructions with different optimization problem formulation and compared time and VRAM usage.

  • 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.

    Strengths:

    • This work has potential to solve an important bottleneck for large multi-dimensional datasets that involves repeated calculation of forward and adjoint nuFFT.
    • Distributed GPU strategy was clearly explained, implemented, and executed. The reduction in reconstruction time is significant.
  • 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.

    Weakness:

    • Even though the evaluation was done in different optimization problems, they were from the same k-space trajectory and same data dimension. Evaluation in multiple different non-Cartesian trajectories will be more important.
    • The error and difference in images and qMRI quantification accuracy with different distributed GPU settings were only briefly mentioned and results not systematically compared.
  • 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

    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?

    It proposed a promising approach to solve an important bottleneck in reconstruction of certain MRI applications.

  • 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

We thank the reviewers for their constructive feedback.

R1/R2/R3 - Validation/image quality We apologize for an exponent typo: relative MSE 1e-4 should be 3.4e-15, corresponding to a relative RMSE of 5.8e-8.This was potentially misleading. The relative RMSE is consistent with FP32 round-off from reordered reductions/halo exchanges/async node/GPU execution, not algorithmic approximation/degradation. Mean absolute ROI-based element-wise differences in nonlinear least-squares-fitted R2*/PDFF/QSM are ~1e-5 for matched reconstructions and robust to ROI placement inside the object. Maps are visually identical except in low-SNR lungs/outside-object regions, where fitting noise gives expected differences. We omitted them due to space/indistinguishability, but will consider including representative images upon reviewers’ final suggestions.

R1 - Baselines Key optimizations are asynchronous communication and topology-aware partitioning; the comparison focused on the former. For the latter, a topology-oblivious baseline (coils uniformly across MPI ranks, no intra/inter-node traffic separation) saturated at 171.19 s on 4 nodes/16 GPUs due to MPI bottlenecks, whereas our topology-aware framework achieved 57.97 s (~3x). We will include this simple-strategy comparison.

R1 - Scaling/profiling Per-iteration profiling separated compute, exposed inter-node halo wait, and intra-node NCCL coil reduction. At 36 nodes/144 GPUs: mean 89 ms; NUFFT/adjoint-NUFFT 76.0%, NCCL reduction 2.74 ms/iter (3.1%), exposed halo wait 9.48 ms/iter (10.6%). Thus most communication is overlapped. Our design keeps coil reductions intra-node via NCCL/NVLink, while only light motion/echo boundary exchanges use MPI. We will add them inline/table. We focus on strong scaling since clinical MRI sizes are fixed by protocols/hardware; weak scaling would decouple from realistic workloads. We prioritize fixed-task time-to-solution.

R2 - Single-node GPU The reviewer made a valid point. Yet, the challenge is to overcome the OOM barrier of enlarged-FoV 5D respiratory motion-resolved reconstruction. At the target scale, memory exceeds our single-node GPU config, preventing end-to-end execution. The single-node implementation is already optimized; our contribution is enabling a scale single-node execution cannot handle, so we do not emphasize it.

R2 - Low-rank/learned reconstruction Low-rank models work well when adjacent temporal states are highly correlated, but respiratory motion-resolved states may have larger anatomical displacement, making such assumptions less applicable. Learned reconstruction is promising but outside scope. Our parallelization could strongly scale time-resolved reconstruction; we will mention this as future work, adding Hammernik et al.

R3 - Trajectory/scanner The shared core of full 3D non-Cartesian trajectories is the nonseparable 3D NUFFT operator, not specific geometry/scanner. Replacing 3D Cones with 3DPR, FLORET, or Yarnball changes only k-space coordinates/density weights; algorithm and communication remain unchanged. We will clarify trajectory/scanner agnosticism for full 3D non-Cartesian encoding.

R1 - Existing frameworks The reviewer made an excellent point; we will articulate contributions. Gadgetron mainly pipeline-dispatches independent jobs (e.g., slice/repetition) to workers, lacking a native abstraction for splitting one tightly coupled 5D optimization over a 2D process grid. Our framework instead performs algorithm-internal model/data parallelism: node-local echo-motion subdomains, nearest-neighbor halo exchanges for regularization, and coil reductions confined to intra-node NVLink/NCCL.

R2 - NUFFT libraries We apologize for the limited discussion. We used SigPy/CuPy NUFFT, but the proposed parallelization is independent of the GPU NUFFT library and can leverage gpuNUFFT, cuFINUFFT, or torchkbnufft. We will add these references.

R1: Clinical practicality As suggested, we will tone down claims on broad clinical deployment readiness.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    the reviewers all agree that this paper presents an interesting and novel method for distributed MR reconstruction. They, however, also point out similar concerns:

    • Limited meaningful validation (eg only one type of trajectory, scanner, quantitative and qualitative evaluation of derived images…) and -better explanation of the difference to existing frameworks (esp other distributed frameworks and NUFFT advanes)
  • 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.

    Despite one reviewer reinforcing the “reject” recommendation, some of their concerns are being answered (evidence for topology-aware and asynchronous execution claims, relevant baseline) and the other reviewers recommend accept. The major strengths of the paper highlighted, an important problem being addressed and a clearly explained and executed distributed GPU strategy with both a significant reduction in time and meaningful methodological contributions, outweigh in my opinion the remaining concerns and could be relevant and interesting for a good part of the MICCAI community.



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.

    The reviewers largely maintained their initial recommendations after the revision, and the majority recommend acceptance. Overall, the reviewers find the paper practically important and technically solid, while noting remaining concerns about image-quality evaluation, baseline comparison, and positioning against prior distributed MRI frameworks. Taking these reviews together, 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 presents a technically solid and practically motivated systems contribution: a topology-aware hierarchical distributed framework for large-scale respiratory motion-resolved 5D non-Cartesian MRI reconstruction.

    After rebuttal, R2 and R3 maintained or moved to accept; R1 maintained rejection because the submitted manuscript lacks direct image-quality visualization. This concern is legitimate, and the absence of reconstructed images or quantitative map comparisons in the main text remains a weakness of the original submission. However, overall, there is clear system novelty and the practical significance of enabling previously infeasible 5D reconstruction. The balance of evidence supports acceptance.



back to top