Abstract

Accelerated cardiac cine MRI requires reconstructing spatiotemporal images from highly undersampled k-space data. Implicit neural representations (INRs) enable scan-specific reconstruction without large training datasets, but encode content implicitly in network weights without physically interpretable parameters. Gaussian primitives provide an explicit and geometrically interpretable alternative, but their spectra are confined near the k-space origin, limiting high-frequency representation. We propose Gabor primitives for MRI reconstruction, modulating each Gaussian envelope with a complex exponential to place its spectral support at an arbitrary k-space location, enabling efficient representation of both smooth structures and sharp boundaries. To exploit spatiotemporal redundancy in cardiac cine, we decompose per-primitive temporal variation into a low-rank geometry basis capturing cardiac motion and a signal-intensity basis modeling contrast changes. Experiments on cardiac cine data with Cartesian and radial trajectories show that Gabor primitives consistently outperform compressed sensing, Gaussian primitives, and hash-grid INR baselines, while providing a compact, continuous-resolution representation with physically meaningful parameters.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0972_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/wenqihuang/FastGaborMRI

Link to the Dataset(s)

OCMR dataset: https://www.ocmr.info Radial Cardiac MRI cine dataset: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/E9FFZK

BibTex

@InProceedings{HuaWen_Gabor_MICCAI2026,
        author = { Huang, Wenqi AND Spieker, Veronika AND Stolt-Ansó, Nil AND Niessen, Natascha AND Dannecker, Maik AND Kafali, Sevgi Gokce AND Kurugol, Sila AND Schnabel, Julia A. AND Rueckert, Daniel},
        title = { { Gabor Primitives for Accelerated Cardiac Cine MRI Reconstruction } },
        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

    The authors propose Gabor primitives for accelerated cardiac cine MRI reconstruction, extending Gaussian primitives by modulating each Gaussian envelope with a complex exponential to enable arbitrary k-space spectral positioning. This allows efficient representation of both smooth structures and sharp boundaries. To exploit spatiotemporal redundancy in cardiac cine imaging, the method decomposes per-primitive temporal variation into a low-rank geometry basis (capturing cardiac motion) and a signal-intensity basis (modeling contrast changes). Experiments on cardiac cine data with Cartesian (R=12, R=16) and radial (R≈23) trajectories demonstrate consistent improvements over compressed sensing, Gaussian primitives, and hash-grid INR baselines across multiple 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.

    Strong Technical Innovation: The introduction of Gabor primitives for MRI reconstruction is novel and well-motivated. By modulating Gaussian primitives with complex exponentials, the method elegantly addresses a fundamental limitation of standard Gaussian primitives—their spectral support being confined to the k-space origin. The mathematical formulation (Eq. 1-2) is rigorous and the frequency-domain interpretation is clear and physically meaningful. Elegant Temporal Modeling: The two-component low-rank decomposition (Eq. 4-5) that separates geometry dynamics from signal-intensity variations is conceptually sound and well-suited to cardiac cine imaging. The coupling between geometry and intensity bases is particularly thoughtful, as it models motion-induced intensity changes such as partial-volume and through-plane flow effects that are clinically relevant. Comprehensive Experimental Validation: The evaluation is thorough, covering both Cartesian and non-Cartesian (radial) trajectories at clinically relevant acceleration factors (R=12, 16, 23). The comparison against four diverse baselines (conventional CS methods, implicit neural representations, and Gaussian primitives) provides strong evidence of the method’s effectiveness. The consistent improvements across all settings demonstrate robustness. Compact and Interpretable Representation: The method achieves the best reconstruction quality while maintaining the most compact representation (ρ<0.5, significantly lower than Hash-INR’s ρ=2.60). The explicit parameterization offers unique capabilities such as frequency decomposition by modulation frequency (Fig. 3c) and natural super-resolution (Fig. 3d), which are not available in voxel-based or purely implicit methods. Strong Ablation Design: Using Gaussian primitives as a baseline is an excellent ablation choice that isolates the contribution of frequency modulation. The consistent gains over this baseline clearly validate the core innovation.

  • 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 Novelty in Core Concept: While the application to MRI is novel, Gabor primitives themselves are well-established in signal processing and have recently been explored in computer vision (Gabor splatting [24, 25, 28]). The authors acknowledge this but do not sufficiently differentiate their complex-valued MRI formulation from existing real-valued Gabor splatting methods. A more detailed comparison of the mathematical differences and why direct adaptation is non-trivial would strengthen the contribution. Insufficient Comparison with State-of-the-Art Deep Learning Methods: The baselines are limited to classical methods (L+S, PICS) and one scan-specific INR (Hash-INR). Modern supervised deep learning methods for dynamic MRI reconstruction (e. g. , [9, 17] cited in the paper, or more recent transformer-based or diffusion-based methods) are not included. While the authors mention that supervised methods require large datasets, a comparison would contextualize the performance gap and clarify when scan-specific methods are preferable. Even a comparison on a subset of data would be valuable. Computational Cost Not Fully Addressed: Gabor primitives are 2–3× slower than Gaussian primitives (Table 1: 2.5 vs 1.2 min for Cartesian R=12) due to wider spatial support and increased per-pixel overlap during rasterization. While the authors explain this trade-off, they do not explore potential optimizations (e. g. , spatial culling, hierarchical rasterization, or GPU kernel optimizations). For clinical translation, reconstruction time is critical, and a discussion of scalability or potential acceleration strategies is needed. Lack of Statistical Significance Testing: The quantitative results (Table 1) report mean PSNR/SSIM/FSIM across all subjects, but no standard deviations, confidence intervals, or paired statistical tests are provided. Given the relatively modest improvements in some settings (e. g. , +0.4 dB PSNR over PICS at Cartesian R=12), statistical validation is essential to confirm that the gains are significant and not due to random variation. Limited Discussion of Failure Cases and Generalizability: The paper does not discuss cases where the method might struggle or fail. For example: How does performance degrade with extreme undersampling (R>25)? How sensitive is the method to initialization of parameters? What happens when cardiac motion is highly irregular (e. g. , arrhythmia)? Are the hyperparameters (Rg=6, Rc=4, λs, λt) robust across different anatomies, field strengths, or pathologies, or do they require per-dataset tuning? 2D Limitation and Lack of 3D Extension Discussion: The method operates on 2D+time data. While the authors mention 3D extension as future work, cardiac MRI is increasingly acquired as 3D volumes. A more substantive discussion of the challenges and potential strategies for 3D+time extension (e. g. , computational complexity, primitive density, memory requirements) would strengthen the paper. Reproducibility Concerns: While the authors promise to release code upon acceptance, critical implementation details are missing: (1) How are coil sensitivities estimated for the radial data? (2) What are the specific learning rates for each parameter group? (3) How sensitive is the method to the density control schedule? (4) What is the convergence criterion? Providing these details or pseudocode would improve reproducibility.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

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

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

    Only mentioned using two cardiac cine MRI datasets, but not describe the dataset at all.

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

    N/A

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

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    This paper presents a solid technical contribution with a novel and well-motivated approach to MRI reconstruction. The introduction of Gabor primitives addresses a real limitation of Gaussian primitives, and the two-component temporal model is elegant and well-suited to cardiac imaging. The experimental validation is comprehensive, covering multiple trajectories and acceleration factors, and the results consistently outperform relevant baselines. However, several issues prevent a stronger acceptance recommendation: (1) the lack of comparison with state-of-the-art supervised deep learning methods limits the contextualization of performance, (2) the absence of statistical significance testing raises questions about the robustness of the reported improvements, (3) computational cost is a concern for clinical translation and is not adequately addressed, and (4) reproducibility is hindered by missing implementation details. Additionally, the novelty is somewhat incremental given existing Gabor splatting work in computer vision, and the 2D limitation reduces immediate clinical impact. If the authors can address these concerns in the rebuttal—particularly by providing statistical validation, clarifying implementation details, and discussing computational optimization strategies—the paper would be suitable for acceptance. The method shows promise and offers unique interpretability advantages, but the evaluation needs to be strengthened to fully support the claims.

  • 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



Review #2

  • Please describe the contribution of the paper

    The authors propose Gabor primitives for scan-specific cardiac cine MRI reconstruction. A standard Gaussian splatting primitive is multiplied by a complex exponential exp(i2πξₙ·(r−µₙ)), which by the Fourier shift theorem moves the primitive’s k-space support from the origin to a learnable center frequency ξₙ. This generalizes Gaussian splatting (recovered when ξₙ=0) and is meant to represent sharp tissue boundaries more efficiently. A second contribution is a temporal model that decomposes per-primitive dynamics into two shared low-rank bases with a coupling term that lets motion drive intensity changes. Validation is on Cartesian cine (R=12, 16) and radial cine (R≈23) against L+S, PICS, Hash-INR, and a Gaussian-primitive ablation.

  • 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 Gaussian-spectrum-stuck-at-origin observation is a real limitation and the proposed fix is mathematically clean. Eq. 2 gives a closed-form FT, and the fact that ξₙ=0 recovers the Gaussian case makes the Gaussian baseline a fair architectural ablation rather than an apples-to-oranges comparison. The complex-valued extension is the right move for MRI. Existing graphics Gabor splatting work (refs 24, 25, 28) targets real non-negative signals, and complex-valued primitive weights are needed to cleanly couple to multi-coil k-space. The temporal model is sensible. Splitting geometry and intensity into separate low-rank bases plus an explicit coupling term to capture motion-induced intensity (partial volume, flow) is physically reasonable and applies the low-rank prior in primitive parameter space, not pixel space. Figure 3 is the part of the paper I liked most. Rather than just showing PSNR tables, the authors visualize where learned ξₙ end up in k-space (the mechanism is doing what the theory predicts), confirm via band-limited PSNR that gains concentrate in mid/high frequency bands, and use the |ξₙ| partition for a frequency decomposition that voxel- or Gaussian-based methods cannot offer. These feel like mechanism checks, not decoration. Evaluation spans Cartesian and radial trajectories on two distinct public datasets, with parameter counts carefully matched between Gaussian and Gabor. The compact representation (ρ<0.5) compared to Hash-INR (ρ=2.60) is a real practical advantage.

  • 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.

    Novelty relative to graphics Gabor splatting is incremental. Refs 24, 25, 28 already propose Gabor primitives for frequency-adaptive scene representation. The MRI-specific delta is essentially the complex-valued formulation plus the multi-coil forward model coupling. The paper would be stronger with an explicit short paragraph stating what is new versus 3DGabSplat. Cartesian gains are small. Gabor beats Gaussian by +1.11 dB at R=12 and +0.72 dB at R=16, with SSIM deltas of 0.002 and 0.001.PICS actually wins or ties on FSIM (0.886/0.875 vs 0.883/0.871). The “consistently outperform” framing in the abstract is really driven by radial; on Cartesian the picture is more nuanced. No paired significance test is reported despite N=99/102.The second claimed contribution is not ablated. The two-component temporal model gets no dedicated study. I would want to see: geometry basis only, intensity basis only, removing the c_n^[w] coupling, and sensitivity to the ranks Rg=6, Rc=4.As it stands one cannot tell whether the low-rank decomposition contributes meaningfully or whether per-frame parameters would do equally well. Rasterization of high-|ξₙ| primitives. The forward model rasterizes Eq. 3 on a Cartesian grid before the FFT. For ξₙ approaching the grid Nyquist frequency, the modulated envelope oscillates near the sampling rate, which can introduce pre-FFT aliasing. The paper says ξₙ “evolves freely” without mentioning any cap or regularization. A clarification on whether/how this is handled is important for correctness. Hash-INR may be under-tuned. It uses 6× more parameters than the primitive methods and ranks last among learned baselines, which is inconsistent with prior INR results on dynamic MRI (e.g., ref 6). Worth either reporting Hash-INR at a parameter-matched budget or describing the tuning protocol. Super-resolution claim has no ground truth. Fig 3d compares 4× Gabor against bicubic upsampling of the reference rather than against an actual high-resolution acquisition. “Sharper than Gaussian” is then qualitative only and could equally reflect plausible hallucination. A downsample-reconstruct-upsample-evaluate experiment with a true high-res reference would settle this. Fig. 2 looks cherry-picked. The displayed Cartesian R=12 frame shows a ≈4 dB Gabor-vs-PICS gap (46.37 vs 42.43) while Table 1 reports a mean gap of 0.40 dB. A near-mean slice plus a worst case would be more honest. Runtime trade-off should be foregrounded. Gabor is ~2× slower than Gaussian (2.5 vs 1.2 min Cartesian; 4.2 vs 3.4 min radial). For a scan-specific method where reconstruction time is inference time, this matters and should be discussed alongside the modest Cartesian gains.

  • Please rate the clarity and organization of this paper

    Good

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

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

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

    N/A

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

    N/A

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

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The paper identifies a precise and real limitation of Gaussian-primitive MRI reconstruction (spectra anchored at the k-space origin) and proposes a clean fix that strictly generalizes the Gaussian case. The complex-valued extension is a non-trivial but well-handled step, and the auxiliary experiments in Fig. 3 go beyond headline numbers and convincingly demonstrate that the proposed mechanism behaves as the theory predicts. The radial R≈23 result is the strongest part: a +2.34 dB gain over PICS at a compact parameter budget is meaningful for non-Cartesian high-acceleration cine. What holds me back from a stronger score: novelty over graphics Gabor splatting (refs 24, 25, 28) is incremental and not clearly delineated; Cartesian gains are small and unsupported by significance testing; the temporal model is entirely un-ablated; and the high-|ξₙ| rasterization question is unaddressed. None of these are fatal, and all are within reach in a short rebuttal: an explicit delta-vs-3DGabSplat paragraph, paired significance tests, a compact temporal ablation, and a sentence on ξₙ bounds. If the authors handle these I would move to 5.

  • 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



Review #3

  • Please describe the contribution of the paper

    This paper introduces Gabor primitives for accelerated cardiac cine MRI reconstruction. The core idea is to modulate each Gaussian primitive with a complex exponential, breaking the constraint that Gaussian spectra are anchored at the k-space origin and allowing each primitive to freely position its spectral support at an arbitrary k-space location. This idea is exceptionally well-matched to the MRI setting: MRI signals are inherently complex-valued and data acquisition occurs directly in k-space, so the analytic Fourier transform of a Gabor primitive (Eq. 2) participates directly in the forward model without any additional approximation. The paper further contributes a two-component low-rank temporal model that separates geometry dynamics from signal-intensity variations. Experiments cover Cartesian and radial trajectories at multiple acceleration factors, with informative analysis experiments beyond standard metric tables.

  • 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.Problem–method alignment is stronger. In MRI, the signal is already complex-valued, and data are acquired in k-space. Each Gabor primitive’s Fourier transform (Eq. 2) is a Gaussian blob centered at the modulation frequency $\boldsymbol{\xi}n$ with a position-dependent phase. Standard Gaussian primitives, regardless of spatial position, always have spectra centered at the k-space origin (spatial translation introduces phase shift, not frequency shift), which is a particularly severe limitation in MRI where high-frequency k-space regions encode tissue boundaries and other clinically important structures. The modulation frequency $\boldsymbol{\xi}_n$ precisely fills this gap. 2.The two-component low-rank temporal model is physically motivated and well-designed. Equations (4)–(5) decompose temporal variation into a geometry bases (capturing in-plane cardiac motion via position, scale, rotation, and frequency perturbations) and an signal-intensity bases (capturing through-plane motion-induced signal changes). The coupling term $c_n^{[w]} \cdot \mathbf{v}{g,t}$ in Eq. (5), which links primitive weights to the geometry basis, models motion-induced intensity effects such as partial-volume and flow. This reflects genuine understanding of cardiac cine physics rather than being a purely mathematical convenience.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The low-rank temporal model lacks component-level evaluation and visualization. In classical low-rank dynamic MRI reconstruction (e.g., L+S decomposition), it is standard to display the static/low-rank component and the dynamic/sparse component separately, allowing direct visual assessment of what each subspace captures. This paper’s two-component temporal model (geometry and signal-intensity bases) is presented as a key contribution, yet no analogous decomposition visualization is provided. Specifically, the paper does not show: (a) images reconstructed using only the the static component geometric parameter or a low-rank perturbation geometric parameter; (b) the effect of removing the coupling term $c_n^{[w]} \cdot \mathbf{v}_{g,t}$ from Eq. (5), which links primitive weights to the geometry subspace. That the geometry basis captures in-plane cardiac motion while the intensity basis captures through-plane signal variations remains unverified. It is unclear whether the two bases learn distinct, interpretable dynamics or merely provide redundant capacity that improves reconstruction metrics without meaningful decomposition. This omission is particularly notable given that the two-component temporal model is listed as one of the paper’s main contributions.

  • Please rate the clarity and organization of this paper

    Good

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

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

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

    N/A

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

    N/A

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

    (5) Accept — should be accepted, independent of rebuttal

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

    This paper introduces Gabor primitives for accelerated cardiac cine MRI reconstruction. The core idea is to modulate each Gaussian primitive with a complex exponential, breaking the constraint that Gaussian spectra are anchored at the k-space origin and allowing each primitive to freely position its spectral support at an arbitrary k-space location. This idea is exceptionally well-matched to the MRI setting: MRI signals are inherently complex-valued and data acquisition occurs directly in k-space, so the analytic Fourier transform of a Gabor primitive (Eq. 2) participates directly in the forward model without any additional approximation.

  • 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

N/A




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 are overall positive, with scores of 4, 4, and 5, and I agree with the accept recommendation. The paper presents a clean and well-motivated extension of Gaussian primitives to Gabor primitives for accelerated cardiac cine MRI reconstruction, allowing learnable frequency positioning in k-space. This is technically appropriate for complex-valued MRI data and is supported by a sensible low-rank temporal model for cardiac dynamics.

    The main concerns are moderate rather than fatal: the novelty over prior Gabor splatting work should be clarified, Cartesian gains are relatively modest, statistical testing and temporal-model ablations are missing, and some implementation details should be added for reproducibility.



back to top