acceptodds
Under review as a conference paper at ICLR 2027

Fourier-Dual Gaussian Representations for 3D Non-Cartesian Parallel MRI Reconstruction

Abstract

Gaussian primitives are an attractive continuous representation choice for the MRI reconstruction inverse problem because their Fourier transform is available in closed form, enabling fitting directly from the measurement in the spatial Fourier domain (-space). However, existing Gaussian-based MRI reconstruction methods fail to leverage this, as they either rasterize an image during the forward model or use expensive parameterizations and parallel imaging formulations that limit this analytical advantage. In addition, they parameterize each primitive by its amplitude and couples its energy to its spatial extent, which degrades the conditioning of the fit in -space. We propose a Fourier-dual representation for 3D non-Cartesian parallel MRI that removes discretization entirely as follows: (i) Integral-Parameterized Gaussian (IPG) decouples the integral from spatial scale in -space, and we prove better conditioning both theoretically and experimentally; (ii) Coil sensitivities are expressed in a compact Fourier basis for direct measurement at arbitrary -space locations from a single shared Gaussian representation; (iii) A surrogate continuous total variation prior by Monte Carlo sampling of the analytical spatial gradients. We compare against compressed sensing, scan-specific implicit neural representations (INR), and existing Gaussian methods on 19 3D brain volumes, as well as a phantom acquisition. The proposed method obtains the highest PSNR and SSIM across this benchmark.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Reconstruction, Medical Imaging, Inverse Problem, 3D Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_syzfOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

MRI measurements are natively collected in k-space, yet most MRI reconstruction algorithms require some stage of rasterization and nonuniform Fourier transform to translate between image space and non-Cartesian k-space measurements. Given the rise of Gaussian splatting, and some recent works using Gaussian representations for MRI reconstruction, this paper proposes to represent an MRI volume as a mixture of Gaussians and leverage the efficient analytical Fourier transform of each Gaussian to completely avoid rasterization and discrete Fourier transforms. The key technical innovations are (1) reparameterizing the Gaussian primitives so that their covariance and height parameters are decoupled in Fourier domain rather than image domain, and showing that this reparameterization makes gradient-based MRI reconstruction better-conditioned, (2) more efficiently allocating parameters for multi-coil MRI, so that the number of optimizable parameters does not scale with the number of coils, and (3) constructing a complete MRI reconstruction pipeline, including TV regularization, using only analytical evaluation of the Fourier transform and no rasterization, and demonstrating improved performance both in simulation and with a physical phantom.

Critical strengths

The main ideas of the paper are solid and presented clearly. The proposed method is elegant and effective empirically. I appreciate the inclusion of new understanding (mainly around the effect of the Gaussian scale reparameterization) as well as strong validation including comparison with recent and strong baselines, matched parameter count and generous-to-baseline settings, realistic noise levels, and both digital and physical phantoms.

Critical weaknesses

I do not flag any major weaknesses. Minor weaknesses are listed below; addressing these should strengthen the paper:

  • There are scattered minor grammatical mistakes like missing connective words, inconsistent pluralization, missing spaces around citations, and occasional punctuation errors. These generally do not impede comprehension but are worth polishing in the final version.
  • In some places notation/jargon is not explicitly defined, though usually in context I am able to infer what was probably intended. Some examples:
    • Starting from the abstract, the terms “integral” and “scale” are applied to Gaussian primitives without being clear what is meant, since if one were interpreting “Gaussian” as a probability distribution its integral would always be 1, and “scale” could refer to either spatial extent (covariance) or peak height/amplitude. Later on these terms can be inferred especially after seeing Figure 2, but for the first 3 pages they are used without clear definition.
    • Equation 2 uses several notations that are not explicitly defined; I inferred that they are the discretized variants of the corresponding quantities in eq. 1 and that \mathcal{F}_K is the nonuniform Fourier transform operator but it would be better to state these definitions to avoid ambiguity.
    • Line 240 writes out the terms of the loss function without definition, but these terms are not defined until later on the same page. A simple “(defined below)” would help to head off potential reader confusion.
    • On Line 320 it’s not clear what “matched voxel averaging” means (matched between what and what?)
    • On Line 321, the term “complex-LS” is used without definition; is this complex-valued least squares?
    • Line 370 introduces the acronym “CSM” without definition.
  • Some phrases are worded in a way that could invite misinterpretation, though I believe I know what the intended meaning is. Specific examples:
    • Line 116 states “…a single primitive under the Gauss-Newton approximation, … for center-weighted trajectories”. While I am fairly confident that “the Gauss-Newton approximation” refers to the optimization algorithm used for MRI reconstruction, the sentence structure makes it sound like an approximation of the Gaussian primitives themselves. Likewise I believe “center-weighted trajectories” refers to k-space sampling trajectories, but it’s not stated clearly (as stated, this could be referring to optimization trajectories).
    • Line 138 starts “We parameterize …” but is describing the amplitude parameterization approach which is a baseline and not the proposed integral parameterization. Starting with “we” makes it sound like this is the proposed approach.
  • Section 2.2 serves to explain why IPG works better than APG, which is arguably the most intellectually exciting part of the contribution, but most of its content is references to appendix sections. It would be great to include more of that content in the main paper, to make the theoretical component of the contribution more evident and easier to interpret. For example, the notations in Table 1 (e.g. “energy-scale correlation” and “p”) are not defined in-context. In my view it would be a worthwhile trade to move Figure 5 to the appendix to make space for more detail in Section 2.2.
  • I appreciate the inclusion of the ablation study in Section 3.3 and Table 3, but it is not explained clearly enough for me to really tell what was ablated.
Questions for authors
  • The objective function terms in equations 7 and 8 contain several hyper parameters and design choices (like the definition of delta in equation 7 and the inclusion of epsilon_TV in equation 8), that are not really justified or explained. Why are these parameters necessary? How do you choose their values?
  • The optimization is done in stages, with an initial voxel-based optimization followed by the proposed Gaussian primitive optimization. Why is the first stage needed? It would be simpler to use Gaussians from the start; does that not work for some reason?
  • Do the authors have a sense of why B2B’s performance often gets worse when it is allowed more parameters? I realize this is a baseline not the proposed method, but was surprised to see it do worse with more parameters so want to ensure there is no mistake.

I am open to raising my score if these are addressed/answered well (especially the first two bullets).

4 comments
AuthorsOpenReview ↗

Clarifications and Experimental Evidence on Optimization and Hyperparameters

We would like to thank the reviewer again for the thoughtful and encouraging assessment, particularly for recognizing the theoretical contribution of our integral parameterization. Developing a mathematically principled continuous representation that can be fitted efficiently to Fourier-domain measurements has been a central motivation of this work.

To facilitate your assessment of the revised manuscript, we summarize the key clarifications and new experimental evidence below, particularly regarding the two questions you highlighted as most important.

1. Why these hyperparameters, and how are they chosen?

Answer: The Huber threshold δ\delta adapts to the median absolute k-space residual during optimization. With αH=1\alpha_H=1, approximately half of the residual magnitudes fall in the quadratic regime and half in the linear regime. This avoids the behavior of a fixed threshold, which could gradually shift the loss from predominantly L1-like to L2-like as residuals decrease.

The TV parameter ϵTV=10−3\epsilon_{\mathrm{TV}}=10^{-3} is a small numerical smoothing constant that makes the TV penalty differentiable near zero. We also use ϵH=10−12\epsilon_H=10^{-12} to avoid division by zero. These settings were fixed on the development subject.

To further validate the choice of adaptive Huber, we added a controlled comparison against L1 and L2 (Appendix I.1, Table 5), with independently tuned MC-TV weights for each loss and SNR:

SNRHuber PSNRL1 PSNRL2 PSNR
100 dB38.3538.25‾\underline{38.25}37.22
50 dB37.4737.17‾\underline{37.17}36.92
30 dB32.3332.1732.22‾\underline{32.22}

AF-16; mean PSNR (dB) over 19 subjects; Huber vs. L1/L2: Holm-adjusted p<0.01.

These results support the adaptive Huber formulation across different noise levels. Interestingly, L1 is more competitive at higher SNR, whereas L2 becomes more competitive as noise increases, illustrating the trade-off between emphasizing smaller residuals and limiting noise sensitivity. Adaptive Huber provides a favorable balance across these regimes.

2. Why voxel initialization and staged optimization instead of fitting Gaussians directly?

Answer: Direct Gaussian fitting is feasible, but considerably less accurate under the same optimization budget. The difficulty lies in jointly optimizing the coupled coefficients, positions, scales, and orientations of Gaussian primitives against undersampled k-space measurements.

Our voxel-derived initialization provides spatially informed Gaussian geometry, while staged optimization progressively introduces the nonlinear geometric parameters, starting from the linear coefficient subproblem.

We added a dedicated ablation (Appendix I.4, Table 9):

Configuration100-dB PSNR50-dB PSNR
Voxel prior + staged (proposed)38.3537.43
Voxel prior + joint optimization37.96‾\underline{37.96}37.20‾\underline{37.20}
Uniform initialization + staged35.5835.43
Uniform initialization + joint optimization34.3934.39

AF-16; mean PSNR (dB) over 19 subjects; matched 3,000-epoch budgets.

Removing voxel initialization reduces PSNR by 2.0–2.8 dB, while replacing staged optimization with joint optimization reduces it by 0.2–0.4 dB. Removing both leads to a 3.0–4.0 dB decrease (all Holm-adjusted p<0.01).

These results explain why we retain the additional optimization stages: they make the challenging non-convex Fourier-domain fit more tractable, rather than introducing complexity without measurable benefit.

3. Why does B2B sometimes deteriorate with more parameters?

Answer: The additional coil-dependent degrees of freedom appear increasingly difficult to constrain under severe undersampling, rather than larger Gaussian representations being inherently worse.

Our extended comparison (Appendix G) reveals a clear dependence on acceleration:

AccelerationB2B, 2,400 primitivesB2B, 8,000 primitives
AF-6429.5124.84
AF-1634.5935.78
AF-435.0937.95

Mean PSNR (dB) at 100-dB SNR over 19 subjects.

The larger model performs worse at AF-64 but better at AF-16 and AF-4, consistent with an interaction between representation capacity and the available measurement information. We have clarified this observation in the revised manuscript.


We have also substantially revised Sections 2.1–2.2 to bring the key conditioning analysis into the main text, clarified the notation and ablations, and corrected the presentation issues you identified.

We hope this concise summary facilitates your assessment of the revised manuscript. We would be grateful if you could reconsider your evaluation in light of the additional evidence, and we remain happy to clarify any remaining questions. Thank you again for your time and constructive feedback.

AuthorsOpenReview ↗

Point-by-Point Clarifications and Evidence Addressing The Questions

We sincerely thank the reviewer syzf again for the encouraging and thoughtful assessment. We particularly appreciate your recognition that the main ideas are solid, the method is mathematically elegant and empirically effective, and that the theoretical understanding of Gaussian reparameterization is a valuable contribution. This recognition means a great deal to us!

Since you kindly indicated openness to reconsidering your score, we provide a concise point-by-point guide to the revised manuscript to facilitate your reassessment.

1. Hyperparameter choices and data-consistency loss

Section 2.4, Eqs. (9)–(11) (p. 6); Appendix I.1, Table 5 (p. 25); Appendix F.5, Table 4 (p. 21).

We clarified the adaptive Huber threshold, defined by median residual magnitude, and explained the numerical safeguards used in the data-consistency and continuous TV terms.

More importantly, we added a controlled Huber/L1/L2 ablation across three noise levels, independently tuning regularization weights on the development subject. Huber achieves the highest PSNR in all three conditions (Holm-adjusted paired tests, p < 0.01), providing direct empirical justification for this design choice.

2. Why voxel initialization and staged optimization?

Section 2.5 (p. 6); Appendix C (pp. 13–14); Appendix I.4, Table 9 (pp. 27–28).

We added controlled ablations comparing voxel-derived initialization against initializing Gaussians directly with uniformly sampled centers, and staged optimization against joint optimization from the beginning.

Starting directly from uniformly initialized Gaussians is feasible but less effective, as the strong coupling between Gaussian coefficients and geometric parameters makes joint optimization particularly challenging.

Voxel-derived initialization provides a spatially informed starting point, while staged optimization progressively releases the coupled parameter groups, first optimizing coefficients, then centers, and finally scales and orientations. This reduces the complexity of the early optimization stages and facilitates a more stable transition to joint refinement.

3. Why can B2B perform worse with more parameters?

Section 3.1 (baseline definitions); Appendix F.4 (parameter budgets); Appendix G, Figures 8–9; Appendix I.5, Table 10 (pp. 28–29).

We clarified the distinction between parameter-matched and primitive-count-matched comparisons. B2B assigns coil-specific coefficients to each Gaussian, so increasing the primitive count from 2,400 to 8,000 increases its trainable parameters from 96k to 320k.

The additional results reveal a clear capacity–sampling interaction: the larger B2B model performs worse under severe undersampling (AF-64), but substantially better when more measurements are available (AF-4). This is consistent with the additional coil-dependent degrees of freedom being harder to constrain under limited sampling, rather than a general failure of increased representation capacity.

4. Theoretical presentation, notation, and ablation clarity

Sections 1.3 and 2.1–2.2, Figure 2 and Table 1 (pp. 2–5); Section 3.3, Table 3 (pp. 9–10).

We carefully addressed your detailed presentation suggestions. In particular, we moved the key conditioning analysis into Section 2.2, explicitly defined the Gaussian integral, scale, sampling-density parameter, and relevant notation, and clarified the assumptions underlying the theoretical results.

We also revised the component analysis to distinguish the effects of Gaussian parameterization, Fourier coil-sensitivity encoding, and continuous MC-TV regularization, and polished terminology and formatting throughout the manuscript.

We believe these revisions make the theoretical contribution more self-contained and accessible, particularly the conditioning advantage that you identified as one of the most intellectually interesting aspects of the paper.


We are especially grateful for your recognition of the mathematical elegance, technical soundness, and empirical effectiveness of our approach. We believe these qualities reflect the central scientific value of the work: a principled continuous formulation that improves optimization conditioning while enabling efficient, grid-free multi-coil Fourier reconstruction.

We also appreciate that you expressed openness to reconsidering your evaluation if the questions were addressed. We believe the additional theoretical explanations and controlled experiments now provide comprehensive answers to those questions, while preserving the core methodology. As the discussion deadline approaches, we would be sincerely grateful if you could revisit the revised manuscript and consider raising your assessment in light of the new evidence, as well as the broader methodological novelty, mathematical foundations, and practical significance of the proposed framework.

Thank you again for your thoughtful, encouraging, and constructive review.

Thanks for the thorough response. I am raising my score.

For what it's worth, I have also experienced cases where voxels are more stable to early optimization compared to Gaussians, though I haven't tried it in k space before.

I also notice that other reviewers requested more real-data or real-patient acquisition. Personally, I am fully satisfied by the inclusion of both digital phantoms and a physical phantom to demonstrate the value of the proposed parameterization and optimization strategy.

AuthorsOpenReview ↗

Sincere Thanks for Your Encouraging Feedback and Support

We sincerely thank you for revisiting our responses and raising your score. Your recognition and support truly mean a great deal to us.

Indeed, we believe that carefully designed initialization and staged optimization are key practical components that translate our theoretical formulation into an effective and computationally efficient reconstruction framework.

We are also sincerely grateful for your recognition of the value of combining digital and physical phantom validation. This means a great deal to us, particularly because we are already preparing further in-vivo acquisitions, which involve substantial practical and experimental challenges. We believe that the current combination of controlled simulations and real scanner measurements already provides strong experimental support for the core methodological contributions of this work.

More broadly, we are genuinely proud of this work, particularly the combination of mathematical elegance, theoretical understanding, computational efficiency, and empirical effectiveness. Our goal has been not only to introduce a continuous representation, but to develop a principled and practical reconstruction framework that makes analytical Fourier-domain optimization genuinely useful.

Your thoughtful recognition of these contributions is deeply encouraging and motivates us to continue developing this research direction.

Thank you once again for your time, careful consideration, constructive engagement, and generous support. We truly appreciate it.

Reviewer_ygdzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Fourier-dual Gaussian representation for 3D non-Cartesian parallel MRI reconstruction. The method introduces integral-parameterized Gaussians (IPG) to improve optimization conditioning and enables direct k-space evaluation without intermediate image rasterization. It further combines Fourier-based coil sensitivity encoding with Monte Carlo estimation of continuous total variation (TV) regularization. Experiments on simulated 3D brain data and a real 3T phantom demonstrate improved reconstruction accuracy over conventional iterative reconstruction, INR-based methods, and Gaussian-based baselines.

Critical strengths
  1. The integral-parameterized Gaussian formulation is mathematically elegant and provides a principled alternative to conventional amplitude-parameterized Gaussians. It enables direct Fourier-domain optimization without repeated rasterization, with theoretical analysis supporting its improved conditioning.

  2. The integration of Fourier-based coil sensitivity encoding and continuous Monte Carlo TV regularization is technically appealing. These designs enable shared multi-coil Fourier evaluation and spatial regularization while preserving a continuous image representation.

Critical weaknesses
  1. Limited experimental validation and generalizability. The evaluation relies primarily on analytical simulations and a single phantom acquisition, without in-vivo validation. Although different noise levels are considered, all experiments focus on 3D radial trajectories at relatively modest spatial resolutions. The proposed formulation is largely dimension-agnostic, while the theoretical conditioning advantage of IPG depends on the sampling density and dimensionality. It therefore remains unclear whether the reported improvements generalize to 2D non-Cartesian acquisitions or other sampling trajectories (e.g., spiral or cones). Further validation on higher-resolution in-vivo data and diverse acquisition protocols would strengthen the generalizability claims.

  2. Insufficient sensitivity and scalability analyses. The method consistently uses 8,000 Gaussian primitives without investigating how the number of primitives affects reconstruction accuracy, achievable resolution, computational cost, or memory consumption. The choice of the adaptive Huber data-consistency loss also lacks sufficient empirical justification relative to simpler alternatives.

Questions for authors

1. Generalization across non-Cartesian acquisitions. How well does the proposed method generalize beyond 3D radial sampling to other commonly used 2D and 3D non-Cartesian trajectories, such as 2D radial, spiral, 3D cones, stack-of-stars, and stack-of-spirals? Given that the conditioning advantage of IPG depends on the sampling density, evaluating different sampling patterns, ideally with higher-resolution and in-vivo data, would help establish the broader applicability of the proposed method.

2. Data consistency and density compensation. What motivates the adaptive Huber loss over standard L1 or normalized L2 losses? Although density compensation is not strictly necessary for iterative non-Cartesian reconstruction, how do different density weightings or preconditioning strategies affect convergence, high-frequency detail preservation, and reconstruction accuracy, particularly under variable-density sampling?

3. Gaussian representation capacity. How was the number of Gaussian primitives selected? How does it affect the trade-off among reconstruction accuracy, spatial resolution, and computational efficiency? Could adaptive splitting, merging, or pruning provide a more efficient representation?

4. Extensibility to other image priors. How readily can the framework incorporate more general image regularizers beyond TV, such as sparsity-based or learned priors, while preserving grid-free optimization? A discussion of the flexibility and limitations of the continuous Gaussian representation in supporting these priors would be valuable.

2 comments
AuthorsOpenReview ↗

Additional Evidence on Generalization, Robustness, and Scalability

We sincerely thank the reviewer again for the thoughtful feedback and for recognizing the mathematical elegance of our integral parameterization and the technical appeal of our continuous reconstruction framework. We believe that direct Fourier-domain optimization of continuous representations is a promising direction for inverse problems, and your comments have helped us strengthen the experimental validation.

To facilitate your assessment of the revised manuscript, we summarize below the key changes and experimental evidence addressing your concerns.

1. Does IPG generalize beyond 3D radial sampling?

Yes. We evaluated 3D-cones spiral trajectories on the same 19 held-out subjects (Appendix I.2, Table 6, Fig. 12).

MethodAF-64 PSNRAF-16 PSNR
CS-TV30.8033.20
HashINR30.8933.61
Raster Gaussian35.1837.29
B2B31.4634.78
IPG (ours)36.5838.55

Mean PSNR (dB), 100-dB SNR, 19 subjects; all differences from IPG: Holm-adjusted p<0.01p<0.01.

IPG improves over Raster Gaussian by 1.40 dB (AF-64) and 1.26 dB (AF-16), confirming that its advantage extends beyond radial sampling. We focus on fully 3D acquisitions, allowing undersampling across all three spatial dimensions. Validation on 2D and in-vivo acquisitions remains an important extension.

2. Why adaptive Huber, and what about density compensation?

Adaptive Huber balances L1 and L2 residual weighting across noise levels. We added a controlled comparison (Appendix I.1, Table 5), independently tuning MC-TV weights for each loss and SNR.

SNRHuber PSNRL1 PSNRL2 PSNR
100 dB38.3538.25‾\underline{38.25}37.22
50 dB37.4737.17‾\underline{37.17}36.92
30 dB32.3332.1732.22‾\underline{32.22}

AF-16; mean PSNR (dB), 19 subjects; Huber vs. L1/L2: Holm-adjusted p<0.01p<0.01.

L1 is more competitive at high SNR, whereas L2 becomes more competitive as noise increases. Adaptive Huber balances these regimes through a threshold following the median residual magnitude.

We further evaluated Pipe–Menon DCF with both L2 and Huber, using trajectory-derived weights:

SNRL2L2 + DCFHuberHuber + DCF
100 dB37.2238.3738.3638.63
50 dB36.9237.3937.4337.62
30 dB32.2124.7232.2527.57

AF-16; mean PSNR (dB), 19 subjects; paired DCF effects: Holm-adjusted p<0.01p<0.01.

DCF improves reconstruction and high-frequency accuracy at 100/50 dB, but substantially degrades both at 30 dB, where convergence curves indicate increasing noise sensitivity beyond approximately 800 epochs. This noise-dependent trade-off motivates unweighted Huber as a robust default, while showing that DCF can help at higher SNR.

3. How does Gaussian count affect accuracy and efficiency?

Increasing capacity improves accuracy, with approximately linear runtime scaling and nearly constant GPU memory (Appendix I.3, Table 7).

GaussiansPSNRTime (s)Memory (GB)
2,00034.583201.21
4,00036.645001.21
8,000 (ours)38.358401.25
16,00039.491,5351.26
32,00039.492,9951.26

AF-16, 100-dB SNR; 19 subjects, NVIDIA H100.

The 8,000 primitives used in our main experiments offer a practical accuracy–efficiency balance and enable parameter-matched comparisons. IPG is also 4.9× faster than the parameter-matched Raster Gaussian baseline (840 vs. 4,145 s), with higher PSNR (38.35 vs. 37.19 dB) and lower peak memory (1.26 vs. 1.65 GB; Appendix I.3, Table 8).

We also explored adaptive density control but observed no consistent improvement. Since Gaussian primitives contribute globally to Fourier measurements, developing splitting and pruning strategies tailored to k-space fitting remains a promising direction.

4. Can the framework incorporate other image priors?

Yes, with qualifications. Analytical spatial derivatives allow local continuous regularizers to be evaluated through Monte Carlo integration without rasterization. Higher-order smoothness penalties are natural extensions, while TGV requires additional continuous auxiliary variables.

Voxel-based learned priors generally require rasterization. Learning cohort-based priors directly over Gaussian representations could instead preserve continuous reconstruction and is a promising direction.


We have also revised the theoretical presentation and clarified the notation and methodology throughout the manuscript.

We hope this summary makes the additional evidence easier to assess and addresses your main concerns. We would be grateful if you could reconsider your evaluation in light of these revisions, and we remain happy to discuss any remaining questions.

Thank you again for your time and constructive feedback.

AuthorsOpenReview ↗

Point-by-Point Clarifications and References to the Revised Manuscript

We sincerely thank the reviewer again for the thoughtful assessment and for recognizing the mathematical elegance of our representation and the technical appeal of our continuous reconstruction framework. We greatly value this recognition, as mathematical elegance and computational efficiency are central to the scientific contribution and broader significance of our continuous framework for Fourier-domain inverse problems.

We believe we have comprehensively addressed your main questions through theoretical clarifications and additional controlled experiments, as detailed in our previous response. To facilitate your reassessment as the discussion deadline approaches, we provide a concise point-by-point guide to the relevant sections, figures, and tables of the revised manuscript.

1. Generalization across non-Cartesian acquisitions

Appendix I.2, Table 6 and Figure 12 (pp. 25–26); Section 2.2 and Table 1 (pp. 4–5).

We extended the evaluation beyond 3D radial sampling to 3D-cones spiral trajectories, where IPG consistently outperforms the evaluated baselines at both acceleration factors. The conditioning analysis further characterizes how the theoretical advantage depends on sampling density.

2. Adaptive Huber loss and density compensation

Section 2.4, Eq. (10); Appendix I.1, Table 5 (p. 25); Appendix I.6, Table 11 and Figure 13 (pp. 28–29).

We added controlled comparisons of Huber, L1, and L2, with independently selected regularization weights, demonstrating the consistent PSNR advantage of adaptive Huber across noise levels.

We further evaluated Pipe–Menon density compensation under both Huber and L2 objectives. The results reveal a noise-dependent trade-off: DCF improves reconstruction accuracy and high-frequency recovery at higher SNR but substantially degrades both at 30 dB. Figure 13 illustrates the corresponding convergence behavior. Appendix E.2 additionally discusses the relationship between density weighting and optimization conditioning.

3. Gaussian representation capacity and computational efficiency

Appendix I.3, Tables 7–8 (pp. 26–27).

We systematically evaluated 2,000–32,000 Gaussian primitives, quantifying the trade-off among reconstruction accuracy, runtime, and GPU memory. The results explain our choice of 8,000 primitives as a practical, parameter-matched configuration.

We also provide direct computational comparisons, showing that IPG is approximately 4.9× faster than the parameter-matched Raster Gaussian baseline, with higher reconstruction accuracy and lower peak memory.

We additionally explored adaptive splitting and pruning but observed no consistent improvement. Developing refinement strategies tailored to Fourier-domain fitting remains a promising direction.

4. Extensibility to other image priors

Section 2.4 and Appendix D (analytical spatial derivatives).

The analytical derivatives of Gaussian primitives allow local continuous regularizers, including higher-order smoothness penalties, to be evaluated through Monte Carlo integration without rasterization. More general priors such as TGV would require additional continuous auxiliary variables, while conventional voxel-based learned priors generally reintroduce rasterization. Learning priors directly over Gaussian representations is a promising direction for preserving the continuous formulation.


We sincerely appreciate your recognition of the mathematical elegance of our formulation and the technical appeal of the proposed framework. We believe these are central contributions of the paper: a principled, computationally efficient continuous reconstruction approach, supported by theoretical conditioning analysis and empirical improvements.

We also appreciate that your questions primarily concern the scope of experimental validation, the justification of specific design choices, and the framework's extensibility. We have taken these concerns seriously and addressed them through additional controlled experiments and theoretical clarifications, while preserving the core methodology.

We believe the resulting evidence substantially strengthens the paper and supports the broader significance of its original contribution. As the discussion deadline approaches, we would be sincerely grateful if you could reconsider your overall assessment in light of both the additional results and the methodological novelty, mathematical foundations, and practical efficiency of the proposed framework.

We remain fully open to further questions and would be happy to provide any additional clarification during the discussion period.

Thank you again for your thoughtful feedback, recognition of our work, and consideration.

Reviewer_xQG6OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a Gaussian Splatting inspired method to reconstruct 3D MRI images from non-Cartesian parallel raw k-space (spatial-frequency) measurements, rather than from formed images as is more conventionally done. They argue that since Gaussian primitives have a closed form Fourier transform, it makes them well suited to representing an image, as well as comparing against the measured frequency signal. Since the Fourier transform of a Gaussian is exact, it can be evaluated directly at any k-space point, without requiring the NUFFT on a voxel grid as previous methods do. They test on 20 simulated MRI brain scans, and achieve slight improvements in PSNR (~1 dB) and SSIM (~0.03) between the final rendered image and the ground truth image.

Critical strengths

From the ablation Table 3, it is clear that the authors' proposition to use Intergal-Parameterized Gaussians (IPG) as opposed to the conventional Amplitude Parameterized Guassians (APG) helps improve image quality, and Figure 2 effectively conveys the difference. Similarly, the motivation behind using Monte Carlo Total Variation for regularisation is sound, and Table 3 again demonstrates that it has a noticeable improvement to the reconstruction quality.

Figure 1 is helpful to understand how the proposed method is different to previous ones.

Critical weaknesses

It would be helpful to include more background and related work on how MRI is formed in the k-space/from the frequency domain, and what the significance of the coils and the different sampling density and shapes mean, as not all ICLR readers will be very familiar with this, and it is central to understanding the methodology and motivation of this paper. The related works section is also a bit thin.

Image quality metric comparison against alternate methods on real scans would add more weight to the submission, rather than testing purely on (only 19) synthetic scans. Sometimes when the physics is not obeyed as perfectly and there is more noise (which is often the case with real scans versus synthetic scans), methods which are more general actually perform better than those that are designed specifically to a certain problem setting. Whilst this may well not be the case here, and IPG would continue to reconstruct better, an image-quality metric validation on real scans would help determine this.

Questions for authors

Is there any time or computational penalty to the proposed method versus Raster Gaussian (your sole Gaussian-primitive based baseline)? Whilst you have fixed the number of parameters to 96k in Table 2 between IPG and Raster Gauss., for a truly fair comparison, time and/or computational resources needed should also be provided.

NB. You have a few missing spaces between the full stop and the start of the following sentence (e.g. Line 224) or a bracket etc.

2 comments
AuthorsOpenReview ↗

Clarifications and Additional Evidence on Presentation, Real-Data Validation, and Computational Efficiency

We sincerely thank the reviewer for the constructive feedback and for recognizing the benefits of our integral parameterization, continuous TV regularization, and overall framework illustrated in Fig. 1. We agree that the original manuscript could be more accessible to readers outside the MRI community.

We have substantially revised the manuscript and added new experiments. To facilitate your reassessment, we summarize the main changes and evidence below.

1. How have we improved the presentation?

We expanded the introduction to explain the MRI Fourier forward model, receiver-coil encoding, and non-Cartesian sampling, and strengthened the related work with additional MRI references.

We also substantially revised Sections 2.1–2.2, bringing key conditioning results from the appendix into the main text and clarifying the distinction between amplitude- and integral-parameterized Gaussians. We corrected the notation, terminology, and formatting issues noted in your review.

Beyond the analytical Fourier transform itself, our contribution is a continuous reconstruction framework combining better-conditioned Gaussian parameterization, shared analytical multi-coil encoding, and grid-free spatial regularization.

2. Has the method been validated on physical MRI data?

Yes, through a physical 3T phantom acquisition (Fig. 4). Our evaluation included the NIST/ISMRM phantom, where IPG achieved the lowest held-out k-space error among the evaluated methods:

MethodHeld-out k-space error
B2B (next best)0.0834
IPG (ours)0.0599

This provides quantitative evidence on real scanner measurements, beyond the 19 simulated brain volumes.

Unlike simulations, measured MRI data do not provide exact continuous ground truth. Reference reconstructions remain affected by noise, finite resolution, and acquisition differences; comparisons between separate scans may also be confounded by sequence-dependent contrast and, in vivo, motion. We therefore use independent held-out measurement consistency rather than treating a reference reconstruction as exact ground truth for PSNR/SSIM.

3. Is there a computational penalty compared with Raster Gaussian?

No. IPG is substantially faster and uses less GPU memory than the parameter-matched Raster Gaussian baseline.

We added a computational comparison across methods (Appendix I.3, Table 8), measured on the same NVIDIA H100 GPU:

MethodTime (s)Memory (GB)PSNR
Raster Gaussian4,1451.6537.19
APG8791.2637.07
IPG (ours)8401.2638.35

AF-16, 100-dB SNR; 96k trainable parameters per method; means over 19 subjects.

IPG is 4.9× faster than Raster Gaussian, despite the latter using a CUDA-accelerated NUFFT implementation, while achieving higher PSNR and lower peak memory usage.

We also investigated scalability with the number of Gaussian primitives (Appendix I.3, Table 7):

GaussiansPSNRTime (s)Memory (GB)
2,00034.583201.21
4,00036.645001.21
8,00038.358401.25
16,00039.491,5351.26
32,00039.492,9951.26

AF-16, 100-dB SNR; means over 19 subjects.

Runtime scales approximately linearly with primitive count, while peak GPU memory grows only marginally. The 8,000-primitive configuration offers a practical accuracy–efficiency balance and enables parameter-matched baseline comparisons.


More broadly, we believe that directly fitting continuous Gaussian representations to Fourier measurements offers a promising direction for inverse problems beyond MRI.

We hope these revisions and additional experiments address your concerns and make the manuscript easier to assess. We would be grateful if you could reconsider your evaluation in light of the new evidence, and we remain happy to clarify any remaining questions.

Thank you again for your time and constructive feedback.

AuthorsOpenReview ↗

Point-by-Point Clarifications and References to the Revised Manuscript

We thank the reviewer xQG6 again for the assessment and for recognizing the contributions of our framework. We particularly appreciate your observation the value of Figure 1, and your positive assessment of the paper's technical soundness.

We believe we have comprehensively addressed your concerns through targeted clarifications and additional experimental evidence. To facilitate your reassessment as the discussion deadline approaches, we provide a concise point-by-point summary with precise references to the revised manuscript.

1. MRI background, presentation, and related work

Addressed in Sections 1.1–1.3 (pp. 1–3), Figure 1, and Sections 2.1–2.2 (pp. 3–5).

We expanded the MRI background to explain the Fourier-domain acquisition model, the role of receiver coils, and the motivation for non-Cartesian sampling (Eqs. 1–2). We also strengthened the related work, including relevant MRI reconstruction and continuous-representation methods.

We further clarified the fundamental distinction illustrated in Figure 1: unlike conventional voxel-based reconstruction or continuous representations rasterized before data consistency, our framework directly optimizes Gaussian primitives against Fourier measurements without intermediate rasterization. The revised Sections 2.1–2.2 and Figure 2 explain the IPG parameterization, while Table 1 brings the key conditioning analysis into the main text.

These revisions aim to make both the methodology and its contributions more accessible to the broader ICLR community.

2. Why 19 evaluation subjects rather than 20?

Clarified in Section 3.1 (p. 6) and Appendix F.1 (p. 19).

We use all 20 BrainWeb anatomies: one subject (s54) is reserved exclusively for development and hyperparameter selection, while the remaining 19 held-out subjects are used for evaluation. All selected hyperparameters remain fixed across these 19 subjects, avoiding evaluation-set tuning.

3. Validation on real MRI measurements

Addressed in Section 3.1 (p. 7) and Figure 4 (p. 9).

Beyond simulation, we evaluated a physical NIST/ISMRM phantom acquired on a 3T MRI scanner. IPG achieves the lowest relative held-out k-space error among the evaluated methods (0.0599, versus 0.0834 for the next-best B2B baseline).

We agree that image-quality evaluation on real acquisitions is valuable. However, measured MRI data lack exact continuous ground truth, and independently acquired reference scans can differ due to noise, resolution, and acquisition conditions. We therefore report held-out measurement consistency rather than treating a reference reconstruction as exact ground truth for PSNR/SSIM.

4. Computational cost versus Raster Gaussian

Directly answered in Appendix I.3, Tables 7–8 (pp. 26–27).

We added a controlled runtime and GPU-memory comparison on identical NVIDIA H100 hardware:

MethodTime (s)Memory (GB)PSNR (dB)
Raster Gaussian4,1451.6537.19
APG8791.2637.07
IPG (ours)8401.2638.35

AF-16, 100-dB SNR; 96k trainable parameters per method; means over 19 subjects.

IPG is approximately 4.9× faster than the parameter-matched Raster Gaussian baseline, despite the latter using CUDA-accelerated NUFFT, while achieving higher reconstruction accuracy and lower peak memory.

Table 7 additionally shows approximately linear runtime scaling with Gaussian count and only marginal memory growth from 2,000 to 32,000 primitives.

5. Typographical and formatting issues

We have revised the manuscript to improve notation, terminology, and readability, and corrected the formatting issues identified in the review.


We are grateful that you recognized the value of the proposed parameterization and regularization, as well as the conceptual distinction illustrated in Figure 1. We believe the revised presentation and additional evidence provide comprehensive answers to your questions and further clarify the contribution of this continuous Fourier-domain reconstruction framework.

As the discussion deadline approaches, we would be sincerely grateful if you could revisit the revised manuscript and consider reassessing your evaluation in light of these clarifications and additional results, while also taking into account the broader theoretical contributions, methodological novelty, and practical significance of the proposed continuous reconstruction framework.

We remain fully open to further questions or discussion and would be happy to address anything that remains unclear.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Consolidated Reviewer Responses

We sincerely thank all three reviewers for their thoughtful and constructive feedback, and for recognizing the mathematical elegance, technical contributions, and empirical effectiveness of our continuous reconstruction framework.

We have substantially revised the manuscript and conducted additional experiments to address the concerns raised. We have also refined and consolidated our individual responses, presenting the key results and supporting evidence directly to make the revisions easier to assess. A diff PDF highlighting the manuscript changes is provided in the Supplementary Material.

The main improvements include:

  • Presentation and theoretical clarity: Expanded the MRI background and related work, substantially revised Sections 2.1–2.2, and brought key conditioning results into the main text.
  • Generalization: Extended evaluation to 3D-cones spiral trajectories, confirming the advantages of IPG beyond radial sampling (Appendix I.2).
  • Data fidelity and noise robustness: Added Huber/L1/L2 comparisons and investigated the effects of density compensation across different noise levels, revealing a trade-off between high-frequency recovery and noise sensitivity (Appendix I.1 and additional rebuttal experiments).
  • Scalability and computational efficiency: Evaluated representation capacity, runtime, and GPU memory. IPG is approximately 4.9× faster than the parameter-matched Raster Gaussian baseline, with higher reconstruction accuracy (Appendix I.3).
  • Optimization strategy: Added controlled ablations demonstrating the contributions of voxel-domain initialization and staged optimization (Appendix I.4).

We believe these revisions and additional experiments comprehensively address the main questions raised by the reviewers and strengthen both the theoretical presentation and empirical validation of the work.

We are grateful for the reviewers' recognition of the potential of this approach. We believe that continuous Gaussian representations optimized directly against Fourier-domain measurements offer a promising direction for MRI reconstruction and potentially broader inverse problems.

We hope the revised manuscript and consolidated responses facilitate your reassessment. We would be grateful if you could reconsider your evaluations in light of the additional evidence, and we warmly welcome any further questions or discussion.

Thank you again for your time, careful consideration, and constructive feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.