Pith. sign in

REVIEW 5 major objections 6 minor 58 references

Adaptive Mask-guided K-space Diffusion for Accelerated MRI Reconstruction

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an adaptive mask-guided diffusion model, AMDM, reconstructs accelerated MRI from under-sampled k-space by generating hybrid masks from each scan's frequency content, reporting consistently higher PSNR and SSIM than…

desk verdict Plausible adaptive-mask diffusion idea for MRI, but the reported gains sit on a two-image test set and baselines explicitly adjusted to favor the method; evaluation needs to be redone. read the letter →

arxiv 2506.18270 v1 pith:2SI4ZGFP submitted 2025-06-23 eess.IV cs.CV

Qinrong Cai , Yu Guan , Zhibo Chen , Dong Liang , Qiuyun Fan , Qiegen Liu This is my paper
classification eess.IVcs.CV
keywords acceleratedMRIk-spacereconstructionscore-baseddiffusionadaptivemaskshybridmaskedimagemodelingmulti-channelembeddingdataconsistency
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

The paper sets out to show that masking is not something to apply blindly: in accelerated MRI, the mask should be derived from the k-space signal itself. It proposes a diffusion model that separates each scan's k-space into high- and low-frequency parts, turns those parts into adaptive binary masks, and stacks the masked components into multiple channels for score learning. Because the mask is recomputed from the evolving reconstruction at each iteration, the model forms a closed loop in which frequency content guides diffusion and diffusion refines frequency content. If the reported numbers hold, the approach would give large, consistent quality gains over established deep and classical reconstruction methods, especially at 8x to 15x acceleration, and would generalize to data distributions not seen in training.

What carries the argument

The load-bearing object is the adaptive mask pair, defined by thresholding the difference between the input k-space and its high-pass or low-pass filtered version, producing one mask for high frequencies and one for low frequencies. A hybrid-masks strategy stacks the masked components, with real and imaginary parts split as separate channels, into a high-dimensional tensor that the score network learns to denoise. The closed-loop part is the iteration in which reconstructed k-space from one diffusion step is used to recompute the masks before the next predictor-corrector step, followed by a data-consistency projection. The mechanism is designed to let the model learn frequency-specific priors and to reduce score-estimation error by sampling in a higher-dimensional space.

What would settle it

Reconstruct a held-out cohort of several dozen volumes with the mask thresholds fixed in advance and report per-volume PSNR and SSIM distributions; if the average gain over the best-tuned baseline falls below roughly 1 dB or is not significant under a paired test, the claimed consistent superiority is not supported.

Watch

Extended reading notes

Core claim

The central claim is that a score-based diffusion model trained on k-space components produced by adaptive hybrid masks can reconstruct undersampled MRI more accurately than models trained on the full k-space. The masks are generated by applying high-pass and low-pass filters to the input k-space, taking the absolute difference from the full data, and thresholding those differences; this produces separate high-frequency and low-frequency components that are embedded together as multiple channels in a high-dimensional tensor. Two score models are trained on different channel orderings, and during iterative reconstruction the predictor-corrector sampling alternates between them, maps the multi-channel state back to k-space, enforces data consistency, and regenerates the masks from the updated k-space. The paper reports that AMDM surpasses all compared methods on PSNR, SSIM, and MSE across Poisson, random, and radial sampling patterns, with the largest margins at high acceleration.

Load-bearing premise

The load-bearing premise is that the comparison methods were tuned to their best settings and that the small number of test images is enough to support the reported multi-decibel gains.

Editorial extensions

If this is right

  • If the central claim is right, accelerated MRI reconstruction at 8x to 15x undersampling can gain roughly 4 to 5 dB PSNR over a strong end-to-end variational baseline on T1-weighted brain data.
  • The method's advantage persists across Poisson, random, and radial sampling patterns, which suggests the mask mechanism is not tied to one undersampling geometry.
  • A model trained only on brain k-space transfers to a different anatomy with higher PSNR and SSIM than baseline diffusion models, indicating improved out-of-distribution robustness.
  • Convergence is reached in fewer sampling iterations than a k-space generative baseline, and total reconstruction time is lower, so the quality gain does not come from extra diffusion steps.

Reading between the lines

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

  • A natural extension the paper leaves implicit is to replace the manually tuned thresholds in the mask equation with a learned threshold network; if the frequency-separation principle is correct, learned thresholds should match or improve the reported gains without per-dataset hand-tuning.
  • The same adaptive-mask-in-high-dimensional-space recipe could be transplanted to other inverse problems with structured spectral content, such as CT reconstruction or image super-resolution, where the mask would be derived from the measured transform rather than from k-space.
  • Because the reported comparisons lack variance estimates, a fair test of the claim would be a paired re-run on many volumes with fixed thresholds and identical compute budgets; the paper's own logic would predict that the mask mechanism, not tuning, drives the improvement.
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

5 major / 6 minor

Summary. The paper proposes AMDM, a score-based diffusion model for accelerated MRI reconstruction in which binary adaptive masks are generated from high-pass and low-pass filtered k-space data, a multi-channel high-dimensional tensor is formed by stacking masked frequency components, and a closed-loop predictor-corrector sampling procedure alternates between score-model updates and mask re-computation. The method is evaluated against SAKE, ESPIRiT, EBMRec, E2E-VarNet, score-MRI, and WKGM on T1-GE Brain, T1-weighted Brain, fastMRI+ test images, and an out-of-distribution knee dataset, with reported PSNR/SSIM/MSE improvements of several dB over baselines at acceleration factors R=8 to R=15.

Significance. If the reported gains are real, the adaptive mask-guided frequency separation and multi-channel stacking could be a useful contribution to k-space diffusion reconstruction. The paper provides an open-source code link and compares against several established methods, which is commendable. However, the central claim that AMDM 'consistently surpasses existing methods across all metrics' is not currently supported by the experimental evidence: the baseline comparison protocol is explicitly favorable to AMDM, the external test sets consist of two fastMRI+ images, no error bars or statistical tests are reported, core hyperparameters are missing, and no ablation isolates the adaptive mask mechanism. These issues are load-bearing for the main claim and need to be addressed before the results can be accepted as reliable.

major comments (5)
  1. [Section IV-A] The statement 'The results of each method are adjusted to the best aiming to illustrate the advantages of AMDM' is a direct admission that the baseline results were selected in a way that favors AMDM. As written, Table I cannot support the claim that AMDM 'consistently surpasses existing methods across all metrics.' The authors must specify exactly which hyperparameters and settings were used for each baseline, confirm that all methods share the same undersampling masks, coil handling, and test data, and follow a standard, pre-specified evaluation protocol without tuning baselines to make AMDM look better.
  2. [Tables I, II, III, and V] The test-set sizes are not reported for the in-vivo datasets, and the external fastMRI+ validation is based on only two images, Test1 and Test2. No standard deviations, confidence intervals, or statistical significance tests are provided. On a sample of two images, the reported 4–5 dB PSNR gaps in Table I could easily be an artifact of favorable baseline selection or of adaptive masks overfit to these particular examples. The paper should report the number of volumes/slices used for each table, per-case results, and appropriate statistical measures, and should validate on a substantially larger and more diverse test set.
  3. [Eqs. (7), (15), (18), and Algorithm 1] Several parameters that are essential to reproduce the method are missing: the threshold ranges [tau_min, tau_max] and [phi_min, phi_max] in Eq. (7), the corrector step size epsilon_t in Eq. (15), the data-consistency weight mu in Eq. (18), the total number of diffusion steps T, and the exact channel configuration N used for each experiment. Without these values, the manuscript is not reproducible, and it is impossible to assess whether the thresholds were chosen on the test data, which would constitute a form of test-set overfitting.
  4. [Section IV-B and Table VI] No ablation isolates the adaptive mask mechanism. The comparisons against score-MRI and WKGM change multiple components at once: the adaptive mask, the hybrid multi-channel embedding, the closed-loop mask updating, and the cascaded two-model reconstruction in Eq. (16). The paper needs ablations that separately remove or replace each component, e.g., fixed masks versus adaptive masks, single-channel versus multi-channel input, and with versus without the closed-loop update. In addition, Table VI shows that the 8-channel configuration performs worse than the 6-channel configuration, which complicates the claim in Section III-B that increasing dimensionality improves score estimation; this should be explained or the theoretical discussion should be moderated.
  5. [Eq. (10) and Section III-B] The theoretical justification for the multi-channel strategy is borrowed from a prior bound on score estimation with denoising autoencoders, but the paper does not derive or verify that the bound applies to the specific construction k-tilde formed by stacking masked frequency components. The notation in Eq. (9) is also unclear (e.g., the use of s_theta tilde). Since the claimed benefit of high-dimensional stacking is central to the method, the authors should either provide a derivation for their construction or explicitly present Eq. (10) as an intuitive motivation rather than a rigorous justification.
minor comments (6)
  1. [Abstract] There is a typo in the abstract: 'k-space This work introduces' should be 'k-space. This work introduces'.
  2. [Section III-B] The phrase 'real and virtual components' should be 'real and imaginary components' when describing complex-valued k-space data.
  3. [Eq. (9)] The optimization variable and the notation for the estimated score function are not clearly defined; please rewrite Eq. (9) so that the objective being minimized over s_theta is unambiguous.
  4. [Table II and Fig. 7] The text in Section IV-B says Table II was obtained under Poisson sampling, while Fig. 7 is described as Radial sampling; please clarify which sampling pattern applies to which result.
  5. [Reference [12]] Reference [12] is listed with year 1999, but the cited journal volume and title suggest a much later publication date; please verify and correct the bibliographic details.
  6. [Table IV] The computational cost comparison reports total time, iteration steps, and per-iteration time, but the hardware and implementation details for each method are not specified; please state the GPU, software versions, and any implementation optimizations to make the timing comparison meaningful.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the AMDM pipeline is a constructive neural-network architecture whose reported gains rest on empirical comparisons; the main threats are benchmark-tuning and test-set model selection, not equation-level circularity.

full rationale

The paper does not derive a quantitative prediction from a formula that is preset to equal its own output. The adaptive masks (Eqs. 7-8), hybrid-channel priors (Eqs. 11-12), score objective (Eq. 13), and closed-loop updates (Eqs. 14-18) are constructive transformations of the input k-space; the reconstruction is obtained by running an iterative sampler, not by unpacking a definition. The high-dimensional score bound of Eq. (10) is quoted from external work [49] and used as motivation; even if the interpretation is questionable, that is a correctness or citation issue, not a circular reduction. The self-references ([15], [50]) motivate cascaded score networks and complex-valued channel handling, but the current paper implements and evaluates those choices, so the central comparison is not defined by the citations. Section IV-A's statement that results are 'adjusted to the best' and Section V's selection of the 6-channel configuration from the reported metrics are legitimate concerns about benchmark fairness and optimistic model selection, but they are evaluation-protocol problems rather than instances where a predicted quantity is equal by construction to a fitted parameter or to the paper's own definition. No circular step satisfying the quoted-evidence standard was found.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The paper's improvements depend on several hand-tuned hyperparameters (mask thresholds, channel count) that are not fully reported, and the theoretical justification is borrowed from a cited bound. No external validation of the mask construction is provided.

free parameters (5)
  • Mask threshold ranges [tau_min, tau_max] and [phi_min, phi_max] = Not reported
    Eq. (7) defines masks using threshold ranges described as 'manually adjusted' (Section III-B). Their values directly determine which k-space regions are masked, and they are not reported, so the method's performance depends on hidden choices.
  • Number of high-frequency channels N = 2
    Section III-B states 'the proposed AMDM set N to be 2 during the training phase.' This controls the size of the stacked tensor and is an architectural choice.
  • Channel count in high-dimensional tensor = 6-ch selected as best
    Table VI compares 4, 6, and 8 channels under R=8 random sampling; 6-ch gives the best PSNR/SSIM, indicating the channel number was tuned on test data.
  • Data consistency weight mu = Not reported
    Eq. (18) requires a weight mu balancing data fidelity and prior; no value is given in the paper.
  • Corrector step size epsilon_t = Not reported
    Eq. (15) uses epsilon_t in the Langevin corrector; the schedule is not specified.
assumptions (5)
  • standard math Score-based diffusion SDE (Eq. 3-5) correctly models the k-space data distribution and the score can be learned by the network.
    The method relies on the standard score-based diffusion framework from Song et al. [17,32]; no proof is given here.
  • domain assumption The error bound from Block et al. [49] (Eq. 10) implies that stacking masked k-space components into a higher-dimensional tensor reduces score estimation error for their specific construction.
    Eq. (9)-(10) are cited from prior work and applied heuristically; the paper does not prove that the bound applies to the hybrid-mask stacked tensor.
  • domain assumption High-pass and low-pass filtering of k-space, followed by thresholding, separates high- and low-frequency information without losing the information needed for reconstruction.
    The mask generation in Eq. (7) assumes wavelet-based filtering is a faithful decomposition; the paper does not analyze the loss of information.
  • ad hoc to paper The mask recomputed from the evolving reconstruction in the closed-loop process improves, rather than degrades, subsequent diffusion steps.
    Section III-C describes iterative mask updating using the current estimate; no analysis shows this feedback is stable or beneficial.
  • domain assumption Score models trained on SIAT brain generalize to T1-GE Brain, T1-weighted Brain, and fastMRI+ knee data.
    Section IV-A trains on SIAT brain; Section V tests on different datasets to claim out-of-distribution performance.
invented entities (1)
  • High-dimensional hybrid-mask tensor k-tilde
    purpose: To provide multiple frequency-specific views of the k-space so the score model can estimate the distribution in a higher-dimensional space.
    This is a computational construction in Eq. (11)-(12); it has no falsifiable handle outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Mask-guided K-space Diffusion for Accelerated MRI Reconstruction." pith.science (2026). https://pith.science/paper/2SI4ZGFP

@misc{pith2026250618270,
  author       = {Pith},
  title        = {Pith review of: Adaptive Mask-guided K-space Diffusion for Accelerated MRI Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2SI4ZGFP}},
  note         = {Machine review of arXiv:2506.18270}
}
read the original abstract

As the deep learning revolution marches on, masked modeling has emerged as a distinctive approach that involves predicting parts of the original data that are proportionally masked during training, and has demonstrated exceptional performance in multiple fields. Magnetic Resonance Imaging (MRI) reconstruction is a critical task in medical imaging that seeks to recover high-quality images from under-sampled k-space data. However, previous MRI reconstruction strategies usually optimized the entire image domain or k-space, without considering the importance of different frequency regions in the k-space This work introduces a diffusion model based on adaptive masks (AMDM), which utilizes the adaptive adjustment of frequency distribution based on k-space data to develop a hybrid masks mechanism that adapts to different k-space inputs. This enables the effective separation of high-frequency and low-frequency components, producing diverse frequency-specific representations. Additionally, the k-space frequency distribution informs the generation of adaptive masks, which, in turn, guide a closed-loop diffusion process. Experimental results verified the ability of this method to learn specific frequency information and thereby improved the quality of MRI reconstruction, providing a flexible framework for optimizing k-space data using masks in the future.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 47 canonical work pages

  1. [1]

    SENSE: sensi- tivity encoding for fast MRI,

    K. P. Pruessmann, M. Weiger, M. B. Scheidegger, et al., “SENSE: sensi- tivity encoding for fast MRI,” Magnetic Resonance in Medicine: An Offi- cial Journal of the International Society for Magnetic Resonance in Med- icine, vol. 42, no. 5, pp. 952–962, 1999

    work page 1999

  2. [2]

    Generalized au- tocalibrating partially parallel acquisitions (GRAPPA),

    M. A. Griswold, P . M. Jakob, R. M. Heidemann, et al., “Generalized au- tocalibrating partially parallel acquisitions (GRAPPA),” Magn. Reson. Med., vol. 47, no. 6, pp. 1202–1210, 2002

    work page 2002

  3. [3]

    Parallel imaging with nonlinear reconstruction using variational penalties,

    F. Knoll, C. Clason, K. Bredies, et al., “Parallel imaging with nonlinear reconstruction using variational penalties,” Magn. Reson. Med., vol. 67, no. 1, pp. 34–41, 2011

    work page 2011

  4. [4]

    Deep learning-based diffusion tensor cardiac magnetic resonance reconstruction: a comparison study ,

    J. Huang, P. F. Ferreira, L. Wang, et al., “Deep learning-based diffusion tensor cardiac magnetic resonance reconstruction: a comparison study ,” Scientific Reports, vol. 14, no. 1, pp. 5658, 2024

    work page 2024

  5. [5]

    Improving synthesis and analysis prior blind compressed sensing with low -rank constraints for dynamic MRI reconstruction,

    A. Majumdar, “Improving synthesis and analysis prior blind compressed sensing with low -rank constraints for dynamic MRI reconstruction, ” Magn. Reson. Imag., vol. 33, no. 1, pp. 174–179, 2015

    work page 2015

  6. [6]

    Compressed sensing,

    D. L. Donoho, “Compressed sensing,” IEEE Transactions on information theory, vol. 52, no. 4, pp. 1289–1306, 2006

    2006

  7. [7]

    Application of magnetic resonance imaging techniques to particulate systems,

    A. J. Sederman, L. F. Gladden, M. D. Mantle, “Application of magnetic resonance imaging techniques to particulate systems,” Advanced Powder Technology, vol. 18, no. 1, pp. 23–38, 2007

    work page 2007

  8. [8]

    Magnetic resonance imaging of granular materials,

    R. Stannarius, “Magnetic resonance imaging of granular materials,” Re- view of Scientific Instruments, vol. 88, no. 5, 2017

    work page 2017

Show all 58 references
  1. [9]

    Image reconstruction by domain- transform manifold learning,

    B. Zhu, J. Z. Liu, S. F. Cauley, et al., “Image reconstruction by domain- transform manifold learning,” Nature, vol. 555, no. 7697, pp. 1476–4687, 2018

    2018

  2. [10]

    Diffusion model based on generalized map for accelerated MRI,

    Z. Xiao, Y. Lu, B. He, et al., “Diffusion model based on generalized map for accelerated MRI,” NMR in Biomedicine , vol. 37, no. 12, pp. e5232, 2024

    2024

  3. [11]

    DP-MDM: Detail-preserving mr recon- struction via multiple diffusion models,

    M. Geng, J. Zhu, X. Zhu, et al., “DP-MDM: Detail-preserving mr recon- struction via multiple diffusion models,” Phys. Med. Biol., vol. 70, no. 11 pp. 115004, 2024

    2024

  4. [12]

    Low-rank tensor assisted K-space gen- erative model for parallel imaging reconstruction ,

    W. Zhang, Z. Xiao, H. Tao, et al., “Low-rank tensor assisted K-space gen- erative model for parallel imaging reconstruction ,” Magnetic Resonance Imaging, vol. 103, pp. 198–207, 1999

    1999

  5. [13]

    Universal generative modeling in dual do- mains for dynamic MRI,

    C. Yu, Y. Guan, Z. Ke, et al., “Universal generative modeling in dual do- mains for dynamic MRI,” NMR in Biomedicine, vol. 36, no. 12, pp. e5011, 2023

    2023

  6. [14]

    Sub-DM: Subspace diffusion model with orthogonal decomposition for MRI reconstruction,

    Y. Guan, Q. Cai, W. Li, et al., “Sub-DM: Subspace diffusion model with orthogonal decomposition for MRI reconstruction,”. arXiv preprint arXiv: 2411.03758, 2024

    2024 arXiv

  7. [15]

    Correlated and multi-frequency diffusion modeling for highly under -sampled MRI reconstruction,

    Y . Guan, C. Y u, Z. Cui, et al., “Correlated and multi-frequency diffusion modeling for highly under -sampled MRI reconstruction,” IEEE Trans. Med. Imag., vol. 43, no. 10, pp. 3490-3502, 2024

    2024

  8. [16]

    Robust compressed sensing MRI with deep generative priors,

    A. Jalal, M. Arvinte, G. Daras, et al., “Robust compressed sensing MRI with deep generative priors,” Advances in Neural Information Processing Systems, vol. 34, pp. 14938-14954, 2021. 10

    2021

  9. [17]

    Solving inverse problems in medical imaging with score -based generative models,

    Y. Song, L. Shen, L. Xing, et al., “Solving inverse problems in medical imaging with score -based generative models,” arXiv preprint arXiv: 2111.08005, 2021

    2021 arXiv

  10. [18]

    One-shot generative prior in Hankel- k-space for parallel imaging reconstruction,

    H. Peng, C. Jiang, J. Cheng, et al., “One-shot generative prior in Hankel- k-space for parallel imaging reconstruction,” IEEE Transactions on Med- ical Imaging, vol. 42, no. 11, pp. 3420–3425, 2023

    2023

  11. [19]

    WKGM: Weight-k-space generative model for parallel imaging recon-struction,

    Z. Tu, D. Liu, X. Wang, et al., “WKGM: Weight-k-space generative model for parallel imaging recon-struction,” NMR in Biomed., vol. 36, no.11, pp. e5005, 2023

    2023

  12. [20]

    Masked image modeling with denoising con- trast,

    K Yi, Y Ge, X Li, et al. , “Masked image modeling with denoising con- trast,” arXiv preprint arXiv: 2205.09616, 2022

    2022 arXiv

  13. [21]

    Masked scene contrast: A scalable frame- work for unsupervised 3D representation learning ,

    X Wu, X Wen, X Liu, et al. , “Masked scene contrast: A scalable frame- work for unsupervised 3D representation learning ,”. In: Proceedings of CVPR.: 9415-9424, 2023

    2023

  14. [22]

    Videomac: Video masked autoencoders meet convnets,

    G Pei, T Chen, X Jiang, et al. , “Videomac: Video masked autoencoders meet convnets,” In: Proceedings of CVPR: 22733-22743, 2024

    2024

  15. [23]

    Self-distillation augmented masked auto- encoders for histopathological image classification,

    Y Luo, Z Chen, S Zhou, et al., “Self-distillation augmented masked auto- encoders for histopathological image classification,” arXiv preprint arXiv: 2203.16983, 2022

    2022 arXiv

  16. [24]

    AutoSMIM: Automatic superpixel -based masked image modeling for skin lesion segmentation,

    Z Wang, J Lyu, X Tang. “AutoSMIM: Automatic superpixel -based masked image modeling for skin lesion segmentation,” IEEE Transac- tions on Medical Imaging, 2023

    2023

  17. [25]

    Ensemble and low-Frequency mixing with diffusion models for accelerated MRI reconstruction,

    Y. Shin, G. Son, D. Hwang, et al., “Ensemble and low-Frequency mixing with diffusion models for accelerated MRI reconstruction,” Medical Im- age Analysis, pp. 103477, 2025

    2025

  18. [26]

    Measurement-conditioned denoising diffusion proba- bilistic model for under-sampled medical image reconstruction,

    Y. Xie, and Q. Li, “Measurement-conditioned denoising diffusion proba- bilistic model for under-sampled medical image reconstruction,” Interna- tional Conference on Medical Image Computing and Computer Assisted Intervention-MICCAI, pp. 655-664, 2022

    2022

  19. [27]

    K-space deep learning for accelerated MRI,

    Y. Han, L. Sunwoo, J. C. Ye, et al. “K-space deep learning for accelerated MRI,” IEEE Transactions on Medical Imaging , vol. 39, no. 2, pp. 377– 386, 2019

    2019

  20. [28]

    Evaluation of variable density and data -driven K-space undersampling for compressed sensing magnetic resonance imaging,

    F. Zijlstra, M. A. Viergever, and P. R. Seevinck, “Evaluation of variable density and data -driven K-space undersampling for compressed sensing magnetic resonance imaging,” Investigative Radiol., vol. 51, no. 6, pp. 410–419, 2016

    2016

  21. [29]

    Deep-learning-based op- timization of the under-sampling pattern in MRI,

    C. D. Bahadir, A. Q. Wang, A. V. Dalca, et al., “Deep-learning-based op- timization of the under-sampling pattern in MRI,” IEEE Trans. Comput. Imag., vol. 6, pp. 1139–1152, 2020

    2020

  22. [30]

    J-MoDL: Joint model-based deep learn- ing for optimized sampling and reconstruction,

    H. K. Aggarwal, and M. Jacob, “J-MoDL: Joint model-based deep learn- ing for optimized sampling and reconstruction,” IEEE J. Sel. Topics Sig- nal Process., vol. 14, no. 6, pp. 1151–1162, 2020

    2020

  23. [31]

    Reducing uncertainty in undersampled MRI reconstruction with active acquisition,

    Z. Zhang, A. Romero, M. J. Muckley, et al., “Reducing uncertainty in undersampled MRI reconstruction with active acquisition, ” Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 2049–2058, 2019

    2019

  24. [32]

    Score-based generative modeling through stochastic differential equations ,

    Y. Song, J. S. Dickstein, D. P. Kingma, et al., “Score-based generative modeling through stochastic differential equations ,” arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  25. [33]

    Diffusion models beat GAN on image syn- thesis,

    P. Dhariwal, and A. Nichol, “Diffusion models beat GAN on image syn- thesis,” Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  26. [34]

    Image super-resolution via iterative refinement,

    C. Saharia, J. Ho, W. Chan, et al., “Image super-resolution via iterative refinement,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 4, pp. 4713–4726, 2023

    2023

  27. [35]

    PixelDefend: Leveraging generative models to understand and defend against adversarial examples ,

    Y . Song, T. Kim, S. Nowozin, et al., “PixelDefend: Leveraging generative models to understand and defend against adversarial examples ,” arXiv preprint arXiv:1710.10766, 2017

    2017 arXiv

  28. [36]

    DeScoD-ECG: Deep score-based dif- fusion model for ECG baseline wander and noise removal,

    H. Li, G. Ditzler, J. Roveda, et al., “DeScoD-ECG: Deep score-based dif- fusion model for ECG baseline wander and noise removal,” IEEE J. Bio- med. Health Informat., 2023

    2023

  29. [37]

    Generative modeling in sinogram domain for sparseview CT reconstruction,

    B. Guan, C. Yang, L. Zhang, et al., “Generative modeling in sinogram domain for sparseview CT reconstruction,” IEEE Transactions on Radia- tion and Plasma Medical Sciences, vol. 8, no. 2, pp. 195–207, 2023

    2023

  30. [38]

    WINNet: Wavelet-inspired invertible net- work for image denoising ,

    J. J. Huang, and P. L. Dragotti, “WINNet: Wavelet-inspired invertible net- work for image denoising ,” IEEE Trans. Image Process., vol. 31, pp. 4377–4392, 2022

    2022

  31. [39]

    Reverse-time diffusion equation models,

    B. D. O. Anderson, “Reverse-time diffusion equation models,” Stochastic Processes Appl., vol. 12, no. 3, pp. 313–326, 1982

    1982

  32. [40]

    Masked Image Modeling: A Survey,

    V. Hondru, F. A. Croitoru, Minaee, et al., “Masked Image Modeling: A Survey,” arXiv preprint arXiv:2408.06687, 2024

    2024 arXiv

  33. [41]

    Masked modeling for self -supervised representation learning on vision and beyond ,

    S. Li, L. Zhang, Z. Wang, et al., “Masked modeling for self -supervised representation learning on vision and beyond ,” arXiv preprint arXiv:2401.00897, 2023

    2023 arXiv

  34. [42]

    Fast and reliable score-based generative model for parallel MRI,

    R. Hou, F. Li, and T . Zeng, “Fast and reliable score-based generative model for parallel MRI,” IEEE Transactions on Neural Networks and Learning Systems, 2023

    2023

  35. [43]

    Score-based diffusion models as principled priors for inverse imaging ,

    B. T. Feng, J. Smith, M. Rubinstein, et al., “Score-based diffusion models as principled priors for inverse imaging ,” Proceedings of the IEEE/CVF International Conference on Computer Vision., pp. 313–326, 2023

    2023

  36. [44]

    Diffusion posterior sampling for general noisy inverse problems,

    H. Chung, J. Kim, M. T. Mccann, et al., “Diffusion posterior sampling for general noisy inverse problems,” arXiv preprint arXiv:2209.14687, 2022

    2022 arXiv

  37. [45]

    Subspace diffusion generative models,

    B. Jing, G. Corso, R. Berlinghieri, et al., “Subspace diffusion generative models,” European Conference on Computer Vision. Cham: Springer Na- ture Switzerland, 2022

    2022

  38. [46]

    Structured denoising diffusion models in discrete state-spaces

    J. Austin, D. D. Johnson, J. Ho, et al., “Structured denoising diffusion models in discrete state-spaces”, In Advances in Neural Information Pro- cessing Systems, vol. 34, pp. 17981–17933, 2021

    2021

  39. [47]

    Masked Autoencoders Are Effective Tokenizers for Diffusion Models ,

    H. Chen, Y. Han, F. Chen, et al., “Masked Autoencoders Are Effective Tokenizers for Diffusion Models ,” arXiv preprint arXiv: 2502.03444, 2025

    2025 arXiv

  40. [48]

    Simplified and generalized masked dif- fusion for discrete data,

    J. Shi, K. Han, Z. Wang, et al., “Simplified and generalized masked dif- fusion for discrete data,” arXiv preprint arXiv: 2406.04329 ,2025

    2025 arXiv

  41. [49]

    Generative modeling withde- noising auto -encoders and Langevin sampling,

    A. Block, Y . Mroueh, and A. Rakhlin, “Generative modeling withde- noising auto -encoders and Langevin sampling,” arXiv preprint arXiv: 2002.00107, 2020

    2002 arXiv

  42. [50]

    Highly undersampled magnetic reso- nance imaging reconstruction using autoencoding priors,

    Q. Liu, Q. Yang, H. Cheng, et al., “Highly undersampled magnetic reso- nance imaging reconstruction using autoencoding priors,” Magn. Reson. Med., vol. 83, no. 1, pp.322-336, 2020

    2020

  43. [51]

    Wavelet-inspired multi-channel score- based model for limited -angle CT reconstruction,

    J. Zhang, H. Mao, X. Wang, et al., “Wavelet-inspired multi-channel score- based model for limited -angle CT reconstruction,” IEEE Trans. Med. Imag., 2024

    2024

  44. [52]

    FastMRI+: Clinical pathology an- notations for knee and brain fully sampled multi -coil MRI data,

    R. Zhao., B. Yaman., Y . Zhang, et al., “FastMRI+: Clinical pathology an- notations for knee and brain fully sampled multi -coil MRI data,” arXiv preprint arXiv: 2109.03812, 2021

    2021 arXiv

  45. [53]

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

    P. J. Shin, P. E. Larson, M. A. Ohliger, et al., “Calibrationless parallel im- aging reconstruction based on structured low -rank matrix completion,” Magn. Reason. Med., vol. 72, no. 4, pp. 959-970, 2014

    2014

  46. [54]

    ESPIRiT —An eigenvalue ap- proach to autocalibrating parallel MRI: Where SENSE meets GRAPPA,

    M. Uecker, P. Lai, M. J. Murphy, et al., “ESPIRiT —An eigenvalue ap- proach to autocalibrating parallel MRI: Where SENSE meets GRAPPA,” Magn. Reson. Med., vol. 71, no. 3, pp. 990–1001, 2014

    2014

  47. [55]

    MRI reconstruction using deep energy- based model,

    Y . Guan, Z. Tu, S. Wang, et al., “MRI reconstruction using deep energy- based model,” NMR Biomed., pp. 1-19, 2022

    2022

  48. [56]

    End-to-end variational networks for accelerated MRI reconstruction,

    A. Sriram, J. Zbontar, T. Murrell, et al., “End-to-end variational networks for accelerated MRI reconstruction,” Proc. Int. Conf. Med.Image Comput. Comput.-Assisted Intervention., pp. 64-73, 2020

    2020

  49. [57]

    Score-based diffusion models for accelerated MRI

    H. Chung, and J. C. Ye, “Score-based diffusion models for accelerated MRI.” Medical image analysis, vol. 80, pp. 102479, 2022

    2022

  50. [58]

    Theory of consistency diffusion mod- els: Distribution estimation meets fast sampling,

    Z. Dou, M. Chen, M. Wang, et al., “Theory of consistency diffusion mod- els: Distribution estimation meets fast sampling,” Forty-first Interna- tional Conference on Machine Learning, 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.