Pith. sign in

REVIEW 3 major objections 5 minor 37 references

K-space Gaussian Representation for Parallel MRI

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read KGR reconstructs accelerated parallel MRI by modeling k-space as an explicit continuous Gabor-Gaussian signal rather than discrete samples, and pairs it with Hankel low-rank projection.

desk verdict KGR is a genuinely new k-space continuous representation with consistent empirical gains, but Proposition 1's noiseless premise contradicts Eq. (1) and needs fixing before the theory is credible. read the letter →

arxiv 2608.00075 v2 pith:DOXMP3ER submitted 2026-07-29 cs.CV

Yu Guan , Mingyu Hu , Jiale Hu , Zhuoxu Cui , Dong Liang , Qiegen Liu This is my paper
classification cs.CV
keywords parallelMRIk-spacereconstructionGaussianrepresentationstructuredlow-rankGaborprimitivescalibrationlessfrequency-adaptivefittingcontinuoussignal
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes K-space Gaussian Representation (KGR), a way to reconstruct accelerated parallel MRI that treats the measured multi-coil k-space as a continuous signal built from Gabor-Gaussian primitives, not as discrete Fourier samples to be filled in. It claims that placing an explicit continuous representation directly in native k-space is new, and that pairing it with Hankel structured low-rank projection makes a scan-specific, calibration-free, network-free reconstruction pipeline. On eight dataset-by-mask combinations from two brain MRI benchmarks, it reports the best PSNR, SSIM, and NMSE among compared methods, with gains up to about 0.9 dB over SAKE and about 2.4 dB over an image-domain Gaussian baseline. The paper also proves that if the true k-space lies in a low-rank Hankel set and the Gaussian fit is within epsilon, then the final projection error is at most 2 epsilon, with all error confined to unacquired samples. A sympathetic reader would care because it opens a representation axis—explicit continuous k-space modeling—that existing interpolation, low-rank, and learned approaches do not occupy.

What carries the argument

The central object is the k-space Gabor-Gaussian primitive phi_j(k) = exp(-1/2 d_j^T Sigma_j^{-1} d_j) exp(i xi_j^T d_j), with d_j = k - nu_j, whose center nu_j, covariance Sigma_j, and carrier xi_j are shared across coils while complex amplitudes remain coil-specific. It does the work of representing local frequency variation continuously, including an anisotropic radial-tangential form in the high-frequency band. The second load-bearing mechanism is the block-Hankel low-rank projection Pi_H, which enforces the algebraic redundancy of multi-coil k-space by lifting local neighborhoods into matrices of low rank; KGR alternates between fitting the Gaussians with a frequency-adaptive, band-RMS-

What would settle it

Compute the block-Hankel singular value spectrum of fully sampled k-space slices from the same datasets under 6x6 and 10x10 windows. If the energy captured by the top-q singular values, at the truncation thresholds used, is much less than the reported PSNR gains imply, then the true k-space is not in the low-rank Hankel set and Proposition 1's bound is vacuous. A second check: run KGR under uniform Cartesian R=8; the paper concedes robustness there is limited, so a strong performance collapse would mark the method's coverage boundary.

Watch

Extended reading notes

Core claim

On its own terms, the central claim is that k-space itself is a continuous complex signal and can be parameterized directly by shared-geometry Gabor-Gaussian primitives: for each coil c, K_theta,c(k) is a sum over radial frequency bands of coil-specific complex amplitudes times a common Gaussian envelope modulated by a sinusoidal carrier. All coils share centers, covariances, and carriers, while each coil keeps its own complex amplitudes, so inter-coil correlations are preserved without estimating sensitivity maps. The fitted continuous prediction is not trusted everywhere: a per-band relative-MSE gate accepts it only where it agrees with acquired data, and the accepted k-space is returned t

Load-bearing premise

Everything in the pipeline leans on the true multi-coil k-space being well approximated by a low-rank block-Hankel matrix; if that rank model is wrong for the data, the initialization and final projection both push the reconstruction toward the wrong structure and the paper's 2-epsilon bound has no premise.

Editorial extensions

If this is right

  • KGR can be used as a drop-in continuous representation module: it augments structured low-rank reconstruction rather than replacing it, so existing Hankel-based solvers can adopt it without retraining.
  • Because the pipeline is scan-specific and needs no coil sensitivity maps or pretrained networks, it is applicable where training data or calibration signal are scarce.
  • The ablation indicates that frequency-adaptive band decomposition is the largest single contributor, so handling k-space's dynamic range is a key mechanism, not just the primitive form.
  • If the reported ranking holds, an explicit continuous parameterization is a viable alternative to learned k-space interpolation and self-supervised networks for accelerated MRI reconstruction.
  • The error analysis provides a design criterion: improve either the Gaussian approximation error or the low-rank compatibility, and the final reconstruction bound improves.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to feed KGR's continuous predictions into calibration-based interpolation under uniform Cartesian undersampling, since the paper concedes that setting remains a weakness and its reliability gate would likely reject predictions in large uniform gaps.
  • The reliability threshold and per-band primitive counts are hand-set; a scan-specific search over them could shift the reported margins, and the ablation suggests the final result is sensitive to at least the band decomposition.
  • The shared-geometry assumption is untested for coils with strongly different sensitivity profiles; a direct check is to compare shared-geometry against per-coil geometry on a dataset with a heterogeneous coil array.
  • Because the continuous representation can be evaluated off-grid, it could extend to non-Cartesian trajectories without separate gridding, but only if the low-rank operator is replaced by a trajectory-aware version—the authors flag this as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes K-space Gaussian Representation (KGR), a scan-specific, network-free method that models multi-coil k-space as a continuous signal using Gabor-Gaussian primitives with geometry shared across coils. The pipeline comprises a Hankel low-rank initialization, frequency-adaptive fitting of band-specific primitives, a reliability-gated augmentation, and a final Hankel low-rank compatibility projection. Experiments on fastMRI and CC359 under Poisson and radial undersampling at R=6 and R=10 report that KGR achieves the best PSNR, SSIM, and NMSE among all compared methods, with gains up to about 0.9 dB over SAKE and about 2.4 dB over an image-domain Gaussian baseline. The authors also provide a theoretical Proposition 1 intended to characterize how Gaussian approximation error and low-rank inconsistency affect the final reconstruction error.

Significance. If the empirical results hold, KGR is a useful contribution to calibrationless, network-free parallel MRI reconstruction: it connects explicit continuous k-space modeling with established structured low-rank priors, and it ships with code, paired t-tests, and an ablation study. The frequency-adaptive, cross-coil shared-geometry formulation is a reasonable and somewhat novel design. However, the theoretical analysis is a stated contribution and is currently not sound under the paper's own noisy forward model; additionally, the empirical comparison would be stronger with sensitivity analysis for per-dataset hyperparameters and with measures of variability for the reported means. The core reconstruction idea is plausible, but the paper needs revision before the claims can be accepted at face value.

major comments (3)
  1. [Section III-D, Eq. (13), Proposition 1] The proposition assumes K* ∈ C_H, where C_H = {K : M_Ω⊙K = \hat K, rank(H(K)) ≤ q}. Under the paper's own forward model Eq. (1), \hat K = M_Ω⊙K* + ε_Ω, so M_Ω⊙K* = \hat K − ε_Ω; hence K* ∉ C_H whenever ε_Ω ≠ 0. This invalidates both conclusions: part (i) is false on Ω, and part (ii)'s bound dist_F(Kaug,C_H) ≤ ε_G relies on K*∈C_H and is unsupported. The final projection in Eq. (11) enforces M_Ω⊙K_final = \hat K, so the reconstructed error on acquired samples is exactly the measurement noise. Please state Proposition 1 under a noiseless model and analyze the noisy case explicitly, or modify the compatibility set to a noise-aware constraint and re-derive the bound.
  2. [Section IV-A, implementation details] Hyperparameters are tuned separately for each dataset: fastMRI uses band counts 50/600/2000/10000, reliability gate δ_max=0.40, 10×10 final-projection window and threshold 2.0, while CC359 uses 10/80/300/1100, δ_max=0.70, 8×8, and threshold 1.8. No sensitivity analysis is provided for δ_max, band counts, λ_Ω, low-rank thresholds, or window sizes. Since the reliability gate in Eq. (10) and the final projection in Eq. (12) are central to the claimed gains, the 'consistent improvements' claim is confounded with per-dataset tuning. A sensitivity study, or at least a single fixed setting across both datasets, is needed to support the general claim.
  3. [Section III-D, Proposition 1 (proof and interpretation)] Proposition 1 does not analyze the Gaussian fitting error ε_G; it merely assumes ∥Kaug−K*∥ ≤ ε_G and then applies the triangle inequality. No bound on ε_G is derived from the KGR fitting procedure, so the theorem amounts to a conditional error-inheritance statement rather than a characterization of when the Gaussian stage benefits reconstruction. The contributions section says 'the derived error analysis characterizes how representation accuracy and low-rank consistency affect reconstruction fidelity,' which overstates what Proposition 1 establishes. Please either derive a nontrivial bound on ε_G or rephrase the theoretical contribution to match the conditional nature of the result.
minor comments (5)
  1. [Section III-D, explanation after Proposition 1] The sentence 'The factor of two comes from the nonconvex rank constraint: Best approximation onto C_H is not firmly nonexpansive' is inaccurate: the factor 2 in Eq. (17) follows directly from the triangle inequality together with dist_F(Kaug,C_H) ≤ ε_G, and does not depend on nonconvexity. Please correct or remove this explanation.
  2. [Section IV-B, Tables I and II] The tables report only mean PSNR/SSIM/NMSE. Paired t-test p-values show that the means differ, but no standard deviations or confidence intervals are given. Since the central claim is that KGR is best 'across all eight settings,' reporting effect-size uncertainty would strengthen the comparison.
  3. [Section IV-A, p-value description] The p-value column is described as a 'two-sided paired t-test against KGR on PSNR averaged over the selected slices and the four sampling settings.' Please clarify whether the test is performed on per-case PSNR values or on setting-level means, and how the four masks are pooled.
  4. [Section III-C, Eq. (10)] In Eq. (10), δ_r is defined as a band-level scalar in Eq. (9), but the condition is written as δ_r(k) ≤ δ_max with r(k) the band containing k. Clarify that the gate applies to all unacquired points in a band based on that band's scalar δ_r.
  5. [Abstract and Section I] The claim of being 'the first explicit continuous representation formulated directly in the native k-space domain' should be carefully qualified relative to neural implicit k-space methods cited as [22], [23]. The authors distinguish explicit from implicit representations, but a reader may see the abstract claim as overly broad; a sentence clarifying the distinction would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KGR is an empirical reconstruction pipeline validated on external benchmarks; Proposition 1 is a conditional inequality and not used to fit or predict the evaluated metrics.

full rationale

KGR's reconstruction is produced by Algorithm 1 from the undersampled measurements alone. The Gaussian parameters are optimized against the low-rank initialization K_init and the acquired samples, and the final output is evaluated against fully sampled references on fastMRI and CC359. No fitted parameter is renamed as a prediction; the missing k-space values are not the training targets, and the reported PSNR/SSIM/NMSE are computed against ground-truth RSS images. The low-rank initialization and final projection use the same Hankel operator, but that is an algorithmic choice, not a statistical circularity: the final reconstruction is not forced to equal the initialization, and the ablation shows the Gaussian fitting contributes beyond the projection. Proposition 1 is explicitly conditional ('Assume...'), and its conclusions follow from those assumptions by the definition of C_H and the triangle inequality; it does not assert that KGR achieves epsilon_G, nor is it used to generate the empirical numbers. One correctness concern should be flagged separately from circularity: with measurement noise in Eq. (1), the true k-space K* does not satisfy M_Omega ⊙ K* = K_hat, so the premise K* ∈ C_H is incompatible with the forward model; but this is a validity gap in the analysis, not a circular derivation. There are no load-bearing self-citations and no imported uniqueness claims. Therefore circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on hyperparameters chosen per dataset (delta_max, band counts, thresholds), on the low-rank Hankel model (a standard domain assumption for multi-coil k-space), and on the inductive bias that Gabor-Gaussian primitives with shared geometry can approximate the underlying signal in each band. No new physical entities are introduced.

free parameters (5)
  • band primitive counts = central/low/mid/high: 50/600/2000/10000 (fastMRI); 10/80/300/1100 (CC359)
    Chosen per dataset with no sensitivity analysis; controls representation capacity in each radial band.
  • reliability gate threshold delta_max = 0.40 (fastMRI), 0.70 (CC359)
    Tuned per dataset; determines which bands' KGR predictions are accepted in Eq. (10).
  • anisotropic covariance widths = sigma_rad = 3.0, sigma_tan = 12.0
    Manual choice for high-frequency band primitive shape.
  • data anchor weight lambda_Omega = 0.1
    Manual choice controlling the strength of the acquired-sample anchoring term in Eq. (8).
  • low-rank thresholds and window sizes = initialization: 6x6 window, threshold 1.5, 100 iters; final: 10x10 window, threshold 2.0 (8x8, 1.8 on CC359)
    Manual tuning; these define the Hankel projection in Eqs. (7), (11), and (12).
assumptions (4)
  • standard math Fourier encoding and coil sensitivity model in Eq. (1) holds.
    Standard data model for parallel MRI; the reconstruction objective in Eq. (2) relies on it.
  • domain assumption True multi-coil k-space K* lies in the low-rank Hankel set C_H (Eq. 13).
    Same assumption as SAKE/LORAKS; underlies the initialization, final projection, and Proposition 1.
  • domain assumption Gabor-Gaussian primitives with shared geometry can approximate the k-space signal within a small band-wise relative error epsilon_G.
    The core inductive bias of the method; the paper does not provide a constructive bound on epsilon_G.
  • ad hoc to paper Bands whose fitted representation matches acquired data are also accurate at unacquired locations (reliability gate).
    Used in Eqs. (9)-(10) to decide where KGR predictions replace the initialization; this is assumed transfer, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of K-space Gaussian Representation for Parallel MRI." pith.science (2026). https://pith.science/paper/DOXMP3ER

@misc{pith2026260800075,
  author       = {Pith},
  title        = {Pith review of: K-space Gaussian Representation for Parallel MRI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DOXMP3ER}},
  note         = {Machine review of arXiv:2608.00075}
}
read the original abstract

Accelerated magnetic resonance imaging (MRI) aims to recover the k-space signal from acquired measurements, where accurate estimation of missing samples is essential for high-fidelity reconstruction. Existing k-space reconstruction methods estimate missing samples through interpolation operators or structure priors defined on discrete sampling grids. Although these formulations effectively exploit local interpolation relationships and global k-space redundancy, they reconstruct only discrete frequency coefficients and therefore do not explicitly model the underlying continuous signal. To overcome this limitation, we propose K-space Gaussian Representation (KGR), the first explicit continuous representation formulated directly in the native k-space domain. Rather than estimating unknown samples on discrete grids, KGR parameterizes the continuous signal using Gabor-Gaussian primitives with shared spatial geometry, yielding a compact representation that naturally preserves inter-coil correlations. Because unconstrained continuous fitting does not necessarily satisfy the intrinsic structural properties of multi-coil signal, the estimated representation is projected onto a low-rank manifold to enforce the algebraic constraints arising from smoothly varying phase and coil redundancy. A frequency-adaptive fitting strategy accommodates the heterogeneous characteristics of different k-space regions. Comprehensive validation across multiple datasets and sampling schemes shows consistent improvements over representative reconstruction baselines in both quantitative metrics and visual quality. These results suggest that explicit continuous parameterization of native k-space provides a principled framework for integrating continuous signal modeling with structured low-rank reconstruction.

Figures

Figures reproduced from arXiv: 2608.00075 by the authors.

Figure 1
Figure 1. Conceptual comparison of representations for parallel MRI reconstruction. (a) Discrete k-space reconstruction estimates missing samples on a discrete grid. (b) Image-domain Gaussian representation places primitives in the spatial domain. (c) K-space Gaussian Represen￾tation (KGR) places the Gaussian representation in native k-space and models missing samples through frequency-adaptive decomposition. Existing methods… view at source ↗
Figure 3
Figure 3. Cross-coil sharing of Gaussian geometry. (a) Per-coil geometry uses an independent set of primitive centers, covariances, and carriers for each receiver. (b) KGR shares one geometry set across C coils while retaining coil-specific complex amplitudes. For the same number of primitives available to each coil, the normalized geometry parameter count decreases from 1 to 1/C. The predicted value of coil c is: Kθ,c(k) = X… view at source ↗
Figure 2
Figure 2. Frequency-adaptive Gaussian fitting. (a) A full-grid loss can be biased toward the high-energy central k-space. (b) KGR computes a mean-squared residual separately within the central (C), low (L), mid (M), and high (H) bands and normalizes by squared band RMS. (a) Full-grid fitting (b) Frequency-adaptive fitting (a) Per-coil geometry (b) Shared geometry R ela tiv e C o n t rib u tio n R ela tiv e C o n t rib u tio n… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of the proposed KGR reconstruction pipeline. The undersampled multi-coil k-space first undergoes a Hankel projection with acquired-sample replacement to obtain a low-rank initialization. Frequency-adaptive KGR then fits band-specific Gaussian and Gabor-Gaussia…
Figure 5
Figure 5. Figure 5: The primitive is centered at νj and uses a Gabor carrier ξj inside an anisotropic Gaussian envelope. In the high-frequency band, the covariance frame is aligned with the local radial and tangential directions ˆerad,j and ˆetan,j , with widths σrad,j and σtan,j . Here η…
Figure 6
Figure 6. Figure 6: Qualitative comparison on fastMRI and CC359 under Poisson R6 and radial R6 undersampling. Columns show the fully sampled reference, zero-fill, CS-TV, SAKE, P-LORAKS, Image-GR, ZS-SSL, and KGR. The top two examples are from CC359, and the bottom two are from fastMRI. Ea…
Figure 7
Figure 7. Figure 7: compares the full-grid and frequency-adaptive fitting on a representative fastMRI T1 slice under Poisson R6. Both variants use identical initialization, low-rank projection, and output processing, thus isolating the effect of the fitting strategy. Frequency-adaptive fi…

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

  1. [16]

    Gabor primitives for accelerated cardiac cine MRI reconstruction,

    W. Huang, V . Spieker, N. Stolt-Ans ´o, N. Niessen, M. Dannecker, S. Gokce Kafali, S. Kurugol, J. A. Schnabel, and D. Rueckert, “Gabor primitives for accelerated cardiac cine MRI reconstruction,” arXiv:2603.05681, 2026

    work page arXiv 2026

  2. [1]

    SENSE: sensitivity encoding for fast MRI,

    K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boesiger, “SENSE: sensitivity encoding for fast MRI,”Magnetic Resonance in Medicine, vol. 42, no. 5, pp. 952–962, 1999

    work page 1999

  3. [2]

    Generalized autocalibrating partially parallel acquisitions (GRAPPA),

    M. A. Griswold, P. M. Jakob, R. M. Heidemann, M. Nittka, V . Jellus, J. Wang, B. Kiefer, and A. Haase, “Generalized autocalibrating partially parallel acquisitions (GRAPPA),”Magnetic Resonance in Medicine, vol. 47, no. 6, pp. 1202–1210, 2002

    work page 2002

  4. [3]

    Calibrationless parallel imaging reconstruction based on structured low-rank matrix completion,

    P. J. Shin, P. E. Z. Larson, M. A. Ohliger, M. Elad, J. M. Pauly, D. B. Vi- gneron, and M. Lustig, “Calibrationless parallel imaging reconstruction based on structured low-rank matrix completion,”Magnetic Resonance in Medicine, vol. 72, no. 4, pp. 959–970, 2014

    work page 2014

  5. [4]

    Scan-specific robust artificial-neural-networks for k-space interpolation (RAKI) re- construction: Database-free deep learning for fast imaging,

    M. Akc ¸akaya, S. Moeller, S. Weing¨artner, and K. U˘gurbil, “Scan-specific robust artificial-neural-networks for k-space interpolation (RAKI) re- construction: Database-free deep learning for fast imaging,”Magnetic Resonance in Medicine, vol. 81, no. 1, pp. 439–453, 2019

    work page 2019

  6. [5]

    Self-supervised learning of physics-guided reconstruc- tion neural networks without fully sampled reference data,

    B. Yaman, S. A. H. Hosseini, S. Moeller, J. Ellermann, K. U ˘gurbil, and M. Akc ¸akaya, “Self-supervised learning of physics-guided reconstruc- tion neural networks without fully sampled reference data,”Magnetic Resonance in Medicine, vol. 84, no. 6, pp. 3172–3191, 2020

    work page 2020

  7. [6]

    Deep magnetic resonance image reconstruction: Inverse problems meet neural networks,

    D. Liang, J. Cheng, Z. Ke, and L. Ying, “Deep magnetic resonance image reconstruction: Inverse problems meet neural networks,”IEEE Signal Processing Magazine, vol. 37, no. 1, pp. 141–151, 2020

    work page 2020

  8. [7]

    A general framework for compressed sensing and parallel MRI using annihilating filter based low-rank Hankel matrix,

    K. H. Jin, D. Lee, and J. C. Ye, “A general framework for compressed sensing and parallel MRI using annihilating filter based low-rank Hankel matrix,”IEEE Transactions on Computational Imaging, vol. 2, no. 4, pp. 480–495, 2016

    work page 2016

Show all 37 references
  1. [8]

    Implicit neural representations with periodic activation functions,

    V . Sitzmann, J. N. P. Martel, A. W. Bergman, D. B. Lindell, and G. Wetzstein, “Implicit neural representations with periodic activation functions,” inAdvances in Neural Information Processing Systems, vol. 33, pp. 7462–7473, 2020

    2020

  2. [9]

    Low-rank modeling of local k-space neighborhoods (LO- RAKS) for constrained MRI,

    J. P. Haldar, “Low-rank modeling of local k-space neighborhoods (LO- RAKS) for constrained MRI,”IEEE Transactions on Medical Imaging, vol. 33, no. 3, pp. 668–681, 2014

    2014

  3. [10]

    P-LORAKS: Low-rank modeling of local k- space neighborhoods with parallel imaging data,

    J. P. Haldar and J. Zhuo, “P-LORAKS: Low-rank modeling of local k- space neighborhoods with parallel imaging data,”Magnetic Resonance in Medicine, vol. 75, no. 4, pp. 1499–1514, 2016

    2016

  4. [11]

    Universal approximation using radial-basis- function networks,

    J. Park and I. W. Sandberg, “Universal approximation using radial-basis- function networks,”Neural Computation, vol. 3, no. 2, pp. 246–257, 1991

    1991

  5. [12]

    3D Gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3D Gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, pp. 1–14, 2023

    2023

  6. [13]

    R2-Gaussian: Rectifying radiative Gaussian splatting for tomographic reconstruction,

    R. Zha, T. J. Lin, Y . Cai, J. Cao, Y . Zhang, and H. Li, “R2-Gaussian: Rectifying radiative Gaussian splatting for tomographic reconstruction,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  7. [14]

    Fast undersampled dynamic MRI reconstruction using explicit representation learning with Gaussian splatting,

    M. L. Terpstra and C. A. T. van den Berg, “Fast undersampled dynamic MRI reconstruction using explicit representation learning with Gaussian splatting,”arXiv:2603.21980, 2026

    2026

  8. [15]

    SPIRiT: Iterative self-consistent parallel imaging reconstruction from arbitrary k-space,

    M. Lustig and J. M. Pauly, “SPIRiT: Iterative self-consistent parallel imaging reconstruction from arbitrary k-space,”Magnetic Resonance in Medicine, vol. 64, no. 2, pp. 457–471, 2010

    2010

  9. [17]

    Sparse MRI: The application of compressed sensing for rapid MR imaging,

    M. Lustig, D. Donoho, and J. M. Pauly, “Sparse MRI: The application of compressed sensing for rapid MR imaging,”Magnetic Resonance in Medicine, vol. 58, no. 6, pp. 1182–1195, 2007

    2007

  10. [18]

    An annihilating filter-based low-rank hankel matrix approach for image inpainting,

    D. Lee, K. H. Jin, E. Y . Kim, S.-H. Park, and J. C. Ye, “An annihilating filter-based low-rank hankel matrix approach for image inpainting,” IEEE Transactions on Image Processing, vol. 25, no. 11, pp. 5622– 5636, 2016

    2016

  11. [19]

    Learning a variational network for reconstruction of accelerated MRI data,

    K. Hammernik, T. Klatzer, E. Kobler, M. P. Recht, D. K. Sodickson, T. Pock, and F. Knoll, “Learning a variational network for reconstruction of accelerated MRI data,”Magnetic Resonance in Medicine, vol. 79, no. 6, pp. 3055–3071, 2018

    2018

  12. [20]

    MoDL: Model-based deep learning architecture for inverse problems,

    H. K. Aggarwal, M. P. Mani, and M. Jacob, “MoDL: Model-based deep learning architecture for inverse problems,”IEEE Transactions on Medical Imaging, vol. 38, no. 2, pp. 394–405, 2019

    2019

  13. [21]

    Deep-learning methods for parallel magnetic resonance imaging reconstruction: A survey of the current approaches, trends, and issues,

    F. Knoll, K. Hammernik, C. Zhang, S. Moeller, T. Pock, D. K. Sodick- son, and M. Akc ¸akaya, “Deep-learning methods for parallel magnetic resonance imaging reconstruction: A survey of the current approaches, trends, and issues,”IEEE Signal Processing Magazine, vol. 37, no. 1, ...

    2020

  14. [22]

    Neural implicit k-space for binning-free non-Cartesian cardiac MR imaging,

    W. Huang, H. B. Li, J. Pan, G. Cruz, D. Rueckert, and K. Ham- mernik, “Neural implicit k-space for binning-free non-Cartesian cardiac MR imaging,” inInformation Processing in Medical Imaging (IPMI), vol. 13939 ofLecture Notes in Computer Science, pp. 548–560, Springer, 2023

    2023

  15. [23]

    PISCO: Self-supervised k-space regular- ization for improved neural implicit k-space representations of dynamic MRI,

    V . Spieker, H. Eichhorn, W. Huang, J. K. Stelter, T. Catalan, R. F. Braren, D. Rueckert, F. Sahli Costabal, K. Hammernik, D. C. Karampinos, C. Prieto, and J. A. Schnabel, “PISCO: Self-supervised k-space regular- ization for improved neural implicit k-space representations of ...

    2025 arXiv

  16. [24]

    Free- breathing dynamic MRI reconstruction via joint time-dependent coil sensitivity estimation using implicit neural representation,

    X. Shen, J. Feng, Z. Li, Q. Zou, Y . Zhang, and H. Wei, “Free- breathing dynamic MRI reconstruction via joint time-dependent coil sensitivity estimation using implicit neural representation,”Medical Image Analysis, vol. 108, p. 103847, 2026

    2026

  17. [25]

    DDGS-CT: Direction-disentangled Gaussian splatting for realistic vol- ume rendering,

    Z. Gao, B. Planche, M. Zheng, X. Chen, T. Chen, and Z. Wu, “DDGS-CT: Direction-disentangled Gaussian splatting for realistic vol- ume rendering,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  18. [26]

    Ultra- Gauss: Ultrafast Gaussian reconstruction of 3D ultrasound volumes,

    M. C. Eid, A. I. L. Namburete, and J. F. Henriques, “Ultra- Gauss: Ultrafast Gaussian reconstruction of 3D ultrasound volumes,” arXiv:2505.05643, 2025

    2025 arXiv

  19. [27]

    Zero-shot CT super-resolution using diffusion-based 2D projection priors and signed 3D Gaussians,

    J. Noh, H.-J. Oh, and W.-K. Jeong, “Zero-shot CT super-resolution using diffusion-based 2D projection priors and signed 3D Gaussians,” arXiv:2508.15151, 2025

    2025 arXiv

  20. [28]

    InnerGS: Internal scenes recon- struction and segmentation via factorized 3D Gaussian splatting,

    S. Liang, Y . Xiao, and W. Tang, “InnerGS: Internal scenes recon- struction and segmentation via factorized 3D Gaussian splatting,” arXiv:2508.13287, 2025

    2025 arXiv

  21. [29]

    Physics-driven 3D Gaussian rendering for zero-shot MRI super-resolution,

    S. Liu, L. Zhang, W. Huang, Z. Zhang, and Z. Wang, “Physics-driven 3D Gaussian rendering for zero-shot MRI super-resolution,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026

    2026

  22. [30]

    PINGS-X: Physics-informed normalized Gaussian splatting with axes alignment for efficient super-resolution of 4D flow MRI,

    S. Jo, S. Y . Hong, J. Kim, S. Kang, A. Choi, D.-G. An, S. Song, and J. H. Hong, “PINGS-X: Physics-informed normalized Gaussian splatting with axes alignment for efficient super-resolution of 4D flow MRI,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2026

    2026

  23. [31]

    M-Gaussian: An magnetic Gaussian framework for efficient multi-stack MRI reconstruction,

    K. Zheng, X. Cai, J. Wang, G. Fu, Z. Li, Y . Chen, X. Ge, L. Qu, and M. Liu, “M-Gaussian: An magnetic Gaussian framework for efficient multi-stack MRI reconstruction,”arXiv:2603.00145, 2026

    2026

  24. [32]

    Three-dimensional MRI reconstruction with 3D gaussian representations: Tackling the undersam- pling problem,

    T. Peng, R. Zha, Z. Li, X. Liu, and Q. Zou, “Three-dimensional MRI reconstruction with 3D gaussian representations: Tackling the undersam- pling problem,”IEEE Transactions on Medical Imaging, vol. 45, no. 5, pp. 1905–1917, 2026

    1905

  25. [33]

    Neural Gabor splatting: Enhanced Gaussian splatting with neural Gabor for high-frequency surface recon- struction,

    H. Watanabe and N. Umetani, “Neural Gabor splatting: Enhanced Gaussian splatting with neural Gabor for high-frequency surface recon- struction,”arXiv:2604.15941, 2026

    2026 arXiv

  26. [34]

    fastMRI: A publicly available raw k-space and DICOM dataset of knee images for accelerated MR image reconstruction using machine learning,

    F. Knoll, J. Zbontar, A. Sriram, M. J. Muckley, M. Bruno, A. Defazio, M. Parente, K. J. Geras, J. Katsnelson, H. Chandarana, Z. Zhang, M. Drozdzal, A. Romero, M. Rabbat, P. Vincent, J. Pinkerton, D. Wang, N. Yakubova, E. Owens, C. L. Zitnick, M. P. Recht, D. K. Sodickson, and ...

    2020

  27. [35]

    Advancing machine learning for MR image reconstruction with an open competition: Overview of the 2019 fastMRI challenge,

    F. Knoll, T. Murrell, A. Sriram, N. Yakubova, J. Zbontar, M. Rabbat, A. Defazio, M. J. Muckley, D. K. Sodickson, C. L. Zitnick, and M. P. Recht, “Advancing machine learning for MR image reconstruction with an open competition: Overview of the 2019 fastMRI challenge,” Magnetic ...

    2019

  28. [36]

    Results of the 2020 fastMRI challenge for machine learning MR image reconstruc- tion,

    M. J. Muckley, B. Riemenschneider, A. Radmanesh, S. Kim, G. Jeong, J. Ko, Y . Jun, H. Shin, D. Hwang, M. Mostapha, S. Arberet, D. Nickel, Z. Ramzi, P. Ciuciu, J.-L. Starck, J. Teuwen, D. Karkalousos, C. Zhang, A. Sriram, Z. Huang, N. Yakubova, Y . W. Lui, and F. Knoll, “Result...

    2020

  29. [37]

    An open, multi-vendor, multi- field-strength brain MR dataset and analysis of publicly available skull stripping methods agreement,

    R. Souza, O. Lucena, J. Garrafa, D. Gobbi, M. Saluzzi, S. Appenzeller, L. Rittner, R. Frayne, and R. Lotufo, “An open, multi-vendor, multi- field-strength brain MR dataset and analysis of publicly available skull stripping methods agreement,”NeuroImage, vol. 170, pp. 482–494, 2018

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.