Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Automated Tuning for Diffusion Inverse Problem Solvers without Generative Prior Retraining

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read ZADS tunes per-step diffusion fidelity weights from held-out k-space, improving MRI without retraining.

desk verdict A clean, honest extension of the authors' ZAPS idea to DDS-style CG updates with SSDU weight tuning; the gains are real but the evidence is too thin for the general claims. read the letter →

arxiv 2509.09880 v1 pith:ZXTASMMM submitted 2025-09-11 eess.IV cs.AIcs.CVcs.LGphysics.med-ph

Yaşar Utku Alçalar , Junno Yun , Mehmet Akçakaya This is my paper · ORCID
classification eess.IVcs.AIcs.CVcs.LGphysics.med-ph
keywords diffusionmodelsMRIreconstructionself-supervisedlearningdatafidelitytuninginverseproblemszero-shotadaptationfastalgorithmunrolling
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 introduces ZADS, a test-time method that automatically tunes the per-timestep data fidelity weights of a diffusion-based inverse problem solver—specifically the decomposed diffusion sampler (DDS)—using only the undersampled measurements already available. ZADS treats the denoising procedure as a fixed unrolled network, splits the measured k-space into a data-consistency set and a held-out supervision set, and optimizes the fidelity weights with a self-supervised loss on the held-out set. The authors show on fastMRI knee data that this adaptive tuning outperforms fixed-weight diffusion solvers and compressed sensing, especially with fast irregular sampling schedules, and that it works without any retraining of the pretrained diffusion prior. If correct, ZADS removes the need for hand-tuning fidelity weights across noise schedules and measurement conditions, making diffusion reconstruction more practical.

What carries the argument

The unrolled DDS sampler: each diffusion timestep produces a Tweedie denoised estimate, refines it with M conjugate-gradient steps using a timestep-dependent fidelity weight ζ_i, and advances a DDIM step. Around this fixed chain, ZADS wraps a self-supervised loop: acquired k-space is split into Θ (used in the CG data-consistency term) and Λ (held out), and the fidelity weights are optimized by minimizing the normalized ℓ1+ℓ2 error between the final reconstruction and the held-out measurements. The key mechanism is treating ζ_i as learnable parameters of an unrolled network, so the same pretrained diffusion prior is reused unchanged while the weights adapt to the noise schedule and measuremen

What would settle it

Measure ZADS's final SSIM/PSNR against a fully sampled reference while monitoring the held-out loss: if the held-out loss keeps decreasing across epochs but reconstruction quality starts to degrade (or if using a random Λ split of the same size yields the same final quality), then the self-supervised objective is not the driver of the reported gains.

Watch

Extended reading notes

Core claim

ZADS's central claim is that the data fidelity weights in diffusion-based inverse problem solvers need not be hand-tuned or fixed; they can be learned at inference time by minimizing a self-supervised loss on a held-out portion of k-space. Treating the DDS denoising loop as a fixed unrolled sampler, the method assigns a learnable weight ζ_i to each timestep in an arbitrary noise schedule, performs conjugate-gradient data consistency updates with the current weights, and then back-propagates through the entire unrolled sampling chain to update those weights against the held-out measurements. The experiments report consistent gains over ℓ1-wavelet compressed sensing, DPS, and DDS on coronal PD

Load-bearing premise

The load-bearing premise is that minimizing the normalized ℓ1/ℓ2 error on the held-out k-space subset Λ is a reliable proxy for true reconstruction quality, so that optimizing this self-supervised loss for 10 epochs improves the final image rather than just memorizing the held-out points.

Editorial extensions

If this is right

  • If ZADS works as claimed, diffusion-based MRI reconstruction can be deployed across different acquisition settings without retraining or per-dataset hand-tuning of fidelity weights.
  • Irregular, fast sampling schedules that preserve fine details become practical, because the per-timestep weights are learned rather than heuristically chosen.
  • The same self-supervised unrolling principle could be applied to other diffusion inverse problem solvers beyond DDS, since the weight-tuning loop is solver-agnostic.
  • The method reduces the computational overhead of diffusion solvers by using few (25) sampling steps while outperforming 1000-step DPS and 250-step DDS.
  • Because the supervision comes from the acquired data itself, the approach works even when no fully-sampled reference is available, which is the standard clinical setting.

Reading between the lines

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

  • ZADS effectively converts the choice of fidelity schedule into a per-example optimization problem; one natural extension is to also learn the noise schedule itself, which the authors flag as future work.
  • The held-out k-space split (ρ=0.4 fixed) is a sensitive hyperparameter: the same self-supervised objective could be used to validate the split ratio or to adapt it per scan, which the paper does not explore.
  • Because the tuning loss is defined on k-space, the idea should transfer to non-Cartesian or 3D MRI, and to other linear inverse problems such as CT or super-resolution, provided a differentiable forward operator exists.
  • The 10-epoch gradient tuning on Λ could overfit the held-out set in low-SNR regimes; a validation-based early stopping or a regularization of the weights would be a testable safeguard.
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 manuscript introduces Zero-shot Adaptive Diffusion Sampling (ZADS), a test-time optimization procedure that tunes per-timestep data-fidelity weights ζ_i in a fixed DDS-style unrolled diffusion sampler. For each test slice, the acquired k-space is split into Θ (used in the CG data-consistency step) and Λ (held out for supervision), and the 25 weights are optimized for 10 epochs by minimizing the normalized ℓ1/ℓ2 k-space residual in Eq. (11). Experiments on fastMRI knee data (coronal PD and PD-FS, R=4 equispaced, 24 central lines) report PSNR/SSIM gains over ℓ1-wavelet, DPS, and DDS, and a qualitative ablation favoring an irregular '17,5,3' schedule. The authors claim the method generalizes across noise schedules and acquisition settings without retraining the diffusion prior.

Significance. The idea of treating a diffusion sampler as a fixed unrolled network and tuning only the fidelity weights via a self-supervised held-out k-space loss is elegant and practically relevant. If the reported gains are robust, ZADS would provide a simple way to make diffusion-based MRI reconstruction adaptive to SNR and schedule without prior retraining, and the explicit separation of the CG data-consistency set Θ from the tuning set Λ is a sensible safeguard against using the same measurements for both inference and supervision. The algorithm is specified in sufficient detail to be implemented, and the authors use the official DPS/DDS repositories with the same pretrained diffusion model. However, the current evidence base is too narrow to support the advertised consistency: a single undersampling pattern, two contrasts, 200 slices, no significance tests, and no sensitivity analyses for the main hyperparameters (ρ, epochs, CG iterations, initialization) or comparison to ZAPS/grid search.

major comments (3)
  1. [Section III-B, Table I] The central claim of consistent gains rests on one table with two contrasts, one equispaced R=4 mask, and 200 slices. There are no significance tests or confidence intervals for the differences, and no ablations over ρ, the number of fine-tuning epochs, CG iterations M, or initialization of ζ_i. Because these hyperparameters directly control the optimization in Algorithm 1, the reported gains over DDS cannot be separated from favorable choices. Please add sensitivity analyses and per-subject statistical testing.
  2. [Section II-D, Eq. (11); Algorithm 1, line 19] All improvements come from minimizing the held-out k-space loss L(y_Λ, E_Λ x_0), but no evidence is given that this loss is a faithful proxy for full-image reconstruction quality. With ρ fixed at 0.4 and only 10 epochs of per-slice optimization, the 25 scalar weights may overfit the specific Λ locations, especially in equispaced sampling where Λ is structured. The paper itself defers a systematic analysis of ρ to future work. Please report reconstruction quality on the unacquired k-space, vary ρ and epoch count, and compare with optimizing the same loss on the full Ω (or with early stopping/regularization) to demonstrate that the held-out loss transfers.
  3. [Section III-A, Table I] No comparison is made to the authors' prior ZAPS method [11] or to a grid search over per-timestep weights. Since ZADS is motivated as removing heuristic tuning, the appropriate baseline is the best fixed or grid-searched DDS weight(s) at the same NFE. Without such a comparison, the reported advantage could reflect favorable tuning of the loss rather than the adaptive weighting itself. Please include ZAPS and a grid-searched DDS baseline at comparable compute.
minor comments (5)
  1. [Algorithm 1, line 19] The optimized objects are scalar fidelity weights, not 'network parameters'; please adjust the wording to avoid confusion.
  2. [Section III-A] 'Equidistant' and 'equispaced' are used interchangeably; define the '17,5,3' schedule explicitly, including the corresponding timestep indices.
  3. [Figure 4] The uniform-vs-irregular schedule ablation is qualitative only; provide quantitative PSNR/SSIM numbers for that comparison.
  4. [Eq. (11)] Clarify whether y_Λ and E_Λ x_0 are complex vectors and how the ℓ1/ℓ2 norms are computed (e.g., real/imaginary components vs magnitude).
  5. [Table I] Indicate whether the reported standard deviations are across slices or subjects, and report the number of subjects per cell.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ZADS is an explicit test-time fitting method evaluated on external full-image metrics; self-citations are contextual, not load-bearing.

full rationale

The paper's central claim is that test-time optimization of the timestep-dependent fidelity weights ζ_i in Eq. (10) under the self-supervised held-out loss in Eq. (11) improves reconstruction quality. This is not circular: Eq. (11) is evaluated on the Λ subset of acquired k-space, while the reported quantitative results in Table I are PSNR/SSIM computed against fully sampled fastMRI references, an external metric not identical to the training loss. The weights are indeed fitted values by explicit design ('Update network parameters {\zeta_i} via L(y_Λ, E_Λ x_0)'), but they are not disguised as predictions; the paper's contribution is an automated tuning procedure benchmarked against DPS, DDS, and ℓ_1-wavelet using the same pretrained diffusion prior. The DDS sampler and prior are taken from external work [17], and the SSDU split is a standard self-supervised technique [25]; citations to the authors' prior ZAPS [11] and SSDU work are contextual and not load-bearing. The paper itself flags the main limitation: 'the stronger CG-based updates in DDS may lead to overfitting if the same measurements are used for both inference and supervision, necessitating the use of hold-out masking,' and notes that a 'more systematic analysis of the SSDU split ratio ρ' is future work. That is a generalization/proxy-validity concern, not a circular derivation: the reported gains do not reduce to Eq. (11) by construction.

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

No new physical or model entities are introduced; the only new objects are scalar weights, which are free parameters rather than invented entities.

free parameters (5)
  • Per-timestep data fidelity weights ζ_i = learned at test time (values not reported)
    One scalar per sampled timestep, updated by gradient descent on the self-supervised loss (Eq. 11); these are the core fitted variables.
  • CG iterations M = 15
    Set by hand for both DDS and ZADS; no ablation shown.
  • SSDU holdout ratio ρ = 0.4
    Fraction of acquired k-space held out for supervision; fixed following [25], authors flag systematic analysis as future work.
  • Fine-tuning epochs = 10
    Number of complete unrolled sampler passes; chosen by hand, no ablation.
  • Irregular schedule = '17,5,3' (25 steps)
    The only schedule tested; no study of schedule robustness.
assumptions (5)
  • standard math DDIM/Tweedie update equations (4)-(5) hold for the pretrained score model.
    Invoked in Section II-B as the base sampler.
  • domain assumption The data manifold is locally affine, so DDS's tangent-space CG correction is valid.
    Inherited from DDS [17], used in Eq. (7)-(10).
  • domain assumption The self-supervised loss on held-out k-space (Eq. 11) is a valid proxy for reconstruction quality.
    Core supervisory signal for ZADS; paper's results depend entirely on this.
  • domain assumption Retrospectively undersampled noisy measurements with an equispaced R=4 mask simulate clinical acquisition.
    Used for all experiments in Section III-A.
  • domain assumption The pretrained diffusion model from [17] is a suitable prior for the fastMRI knee slices.
    Used without retraining, as stated in Section III-A.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automated Tuning for Diffusion Inverse Problem Solvers without Generative Prior Retraining." pith.science (2026). https://pith.science/paper/ZXTASMMM

@misc{pith2026250909880,
  author       = {Pith},
  title        = {Pith review of: Automated Tuning for Diffusion Inverse Problem Solvers without Generative Prior Retraining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZXTASMMM}},
  note         = {Machine review of arXiv:2509.09880}
}
read the original abstract

Diffusion/score-based models have recently emerged as powerful generative priors for solving inverse problems, including accelerated MRI reconstruction. While their flexibility allows decoupling the measurement model from the learned prior, their performance heavily depends on carefully tuned data fidelity weights, especially under fast sampling schedules with few denoising steps. Existing approaches often rely on heuristics or fixed weights, which fail to generalize across varying measurement conditions and irregular timestep schedules. In this work, we propose Zero-shot Adaptive Diffusion Sampling (ZADS), a test-time optimization method that adaptively tunes fidelity weights across arbitrary noise schedules without requiring retraining of the diffusion prior. ZADS treats the denoising process as a fixed unrolled sampler and optimizes fidelity weights in a self-supervised manner using only undersampled measurements. Experiments on the fastMRI knee dataset demonstrate that ZADS consistently outperforms both traditional compressed sensing and recent diffusion-based methods, showcasing its ability to deliver high-fidelity reconstructions across varying noise schedules and acquisition settings.

Figures

Figures reproduced from arXiv: 2509.09880 by the authors.

Figure 1
Figure 1. Comparison of fast sampling strategies for diffusion models. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our Zero-shot Adaptive Diffusion Sampling (ZADS) framework treats the diffusion sampling process as an unrolled architecture, where timestep [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Representative reconstructions from the coronal PD and PD-FS datasets ( [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of uniform and irregular noise schedules within the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 1 linked inside Pith

  1. [11]

    Zero-shot adaptation for approximate posterior sampling of diffusion models in inverse problems,

    Y . U. Alc ¸alar and M. Akc ¸akaya, “Zero-shot adaptation for approximate posterior sampling of diffusion models in inverse problems,” inProc. Eur . Conf. Comput. Vis., 2024, pp. 444–460

    2024

  2. [1]

    Generative modeling by estimating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” inProc. Adv. Neural Inf. Process. Syst., 2019, pp. 11918–11930

    2019

  3. [2]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inProc. Adv. Neural Inf. Process. Syst., 2020, pp. 6840–6851

    2020

  4. [3]

    DiffWave: A versatile diffusion model for audio synthesis,

    Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “DiffWave: A versatile diffusion model for audio synthesis,” inProc. Int. Conf. Learn. Represent., 2021

    2021

  5. [4]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inProc. Int. Conf. Learn. Represent., 2021

    2021

  6. [5]

    Video diffusion models,

    J. Ho, et al., “Video diffusion models,” inProc. Adv. Neural Inf. Process. Syst., 2022, pp. 8633–8646

    2022

  7. [6]

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

    Y . Song, L. Shen, L. Xing, and S. Ermon, “Solving inverse problems in medical imaging with score-based generative mod- els,” inProc. Int. Conf. Learn. Represent., 2022

    2022

  8. [7]

    Diffusion posterior sampling for general noisy inverse problems,

    H. Chung, J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Ye, “Diffusion posterior sampling for general noisy inverse problems,” inProc. Int. Conf. Learn. Represent., 2023

    2023

Show all 36 references
  1. [8]

    Zero-shot image restoration using denoising diffusion null-space model,

    Y . Wang, J. Yu, and J. Zhang, “Zero-shot image restoration using denoising diffusion null-space model,” inProc. Int. Conf. Learn. Represent., 2023

    2023

  2. [9]

    Pseudoinverse- guided diffusion models for inverse problems,

    J. Song, A. Vahdat, M. Mardani, and J. Kautz, “Pseudoinverse- guided diffusion models for inverse problems,” inProc. Int. Conf. Learn. Represent., 2023

    2023

  3. [10]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,” inProc. Adv. Neural Inf. Process. Syst., 2021, pp. 8780–8794

    2021

  4. [12]

    Learning a variational network for reconstruction of accelerated MRI data,

    K. Hammernik, et al., “Learning a variational network for reconstruction of accelerated MRI data,”Magn. Reson. Med., vol. 79, no. 6, pp. 3055–3071, 2018

    2018

  5. [13]

    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 Trans. Med. Imag., vol. 38, no. 2, pp. 394–405, 2019

    2019

  6. [14]

    Dense recurrent neural networks for accelerated MRI: history-cognizant unrolling of optimization algorithms,

    S. A. H. Hosseini, B. Yaman, S. Moeller, M. Hong, and M. Akc ¸akaya, “Dense recurrent neural networks for accelerated MRI: history-cognizant unrolling of optimization algorithms,” IEEE J. Sel. Topics Signal Process., vol. 14, no. 6, pp. 1280– 1291, Oct. 2020

    2020

  7. [15]

    Zero-shot self-supervised learning for MRI reconstruction,

    B. Yaman, S. A. H. Hosseini, and M. Akcakaya, “Zero-shot self-supervised learning for MRI reconstruction,” inProc. Int. Conf. Learn. Represent., 2022

    2022

  8. [16]

    Robust compressed sensing MRI with deep generative priors,

    A. Jalal, et al., “Robust compressed sensing MRI with deep generative priors,” inProc. Adv. Neural Inf. Process. Syst., 2021, pp. 14938–14954

    2021

  9. [17]

    Decomposed diffusion sampler for accelerating large-scale inverse problems,

    H. Chung, S. Lee, and J. C. Ye, “Decomposed diffusion sampler for accelerating large-scale inverse problems,” inProc. Int. Conf. Learn. Represent., 2024

    2024

  10. [18]

    Optimization methods for magnetic resonance image reconstruction,

    J. A. Fessler, “Optimization methods for magnetic resonance image reconstruction,”IEEE Signal Process. Mag., vol. 37, no. 1, pp. 33–40, 2020

    2020

  11. [19]

    Self-supervised physics-guided deep learning reconstruction for high-resolution 3D LGE CMR,

    B. Yaman, et al., “Self-supervised physics-guided deep learning reconstruction for high-resolution 3D LGE CMR,” inProc. IEEE Int. Symp. Biomed. Imag., 2021, pp. 100–104

    2021

  12. [20]

    Unsuper- vised deep learning methods for biological image reconstruction and enhancement: An overview from a signal processing per- spective,

    M. Akc ¸akaya, B. Yaman, H. Chung, and J. C. Ye, “Unsuper- vised deep learning methods for biological image reconstruction and enhancement: An overview from a signal processing per- spective,”IEEE Signal Process. Mag., vol. 39, no. 2, pp. 28–44, 2022

    2022

  13. [21]

    Fast MRI for all: Bridging equity gaps via training without raw data access,

    Y . U. Alc ¸alar, M. G ¨ulle, and M. Akc ¸akaya, “Fast MRI for all: Bridging equity gaps via training without raw data access,” 2025, arXiv:2411.13022

    2025

  14. [22]

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

    F. Knoll, et al., “Deep-learning methods for parallel magnetic resonance imaging reconstruction: A survey of the current approaches, trends, and issues,”IEEE Signal Process. Mag., vol. 37, no. 1, pp. 128–140, 2020

    2020

  15. [23]

    Revisitingℓ 1-wavelet compressed-sensing MRI in the era of deep learning,

    H. Gu, et al., “Revisitingℓ 1-wavelet compressed-sensing MRI in the era of deep learning,”Proc. Natl. Acad. Sci., vol. 119, no. 33, 2022, Art. no. e2201062119

    2022

  16. [24]

    Deep learning for accelerated and robust MRI reconstruc- tion,

    R. Heckel, M. Jacob, A. Chaudhari, O. Perlman, and E. Shim- ron, “Deep learning for accelerated and robust MRI reconstruc- tion,”Magn. Reson. Mater . Phys. Biol. Med., vol. 37, no. 3, pp. 335–368, 2024

    2024

  17. [25]

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

    B. Yaman, et al., “Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data,”Magn. Reson. Med., vol. 84, no. 6, pp. 3172–3191, Dec. 2020

    2020

  18. [26]

    Multi-mask self-supervised learning for physics-guided neural networks in highly accelerated magnetic resonance imaging,

    B. Yaman, et al., “Multi-mask self-supervised learning for physics-guided neural networks in highly accelerated magnetic resonance imaging,”NMR Biomed., vol. 35, no. 12, 2022, Art. no. e4798

    2022

  19. [27]

    Tweedie’s formula and selection bias,

    B. Efron, “Tweedie’s formula and selection bias,”J. Amer . Stat. Assoc., vol. 106, no. 496, pp. 1602–1614, Dec. 2011

    2011

  20. [28]

    Algorithm unrolling: Interpretable, efficient deep learning for signal and image pro- cessing,

    V . Monga, Y . Li, and Y . C. Eldar, “Algorithm unrolling: Interpretable, efficient deep learning for signal and image pro- cessing,”IEEE Signal Process. Mag., vol. 38, no. 2, pp. 18–44, 2021

    2021

  21. [29]

    20-fold accelerated 7T fMRI using referenceless self-supervised deep learning reconstruction,

    ¨O. B. Demirel, et al., “20-fold accelerated 7T fMRI using referenceless self-supervised deep learning reconstruction,” in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), 2021, pp. 3765–3769

    2021

  22. [30]

    Physics-driven deep learning for computational magnetic resonance imaging: Combining physics and machine learning for improved medical imaging,

    K. Hammernik, et al., “Physics-driven deep learning for computational magnetic resonance imaging: Combining physics and machine learning for improved medical imaging,”IEEE Signal Process. Mag., vol. 40, no. 1, pp. 98–114, 2023

    2023

  23. [31]

    A convex compressibility-inspired unsupervised loss function for physics- driven deep learning reconstruction,

    Y . U. Alc ¸alar, M. G ¨ulle, and M. Akc ¸akaya, “A convex compressibility-inspired unsupervised loss function for physics- driven deep learning reconstruction,” inProc. IEEE Int. Symp. Biomed. Imag., 2024, pp. 1–5

    2024

  24. [32]

    Sparsity-driven parallel imaging consistency for improved self-supervised MRI recon- struction,

    Y . U. Alc ¸alar and M. Akc ¸akaya, “Sparsity-driven parallel imaging consistency for improved self-supervised MRI recon- struction,” inProc. IEEE Int. Conf. Image Process., 2025, pp. 851–856

    2025

  25. [33]

    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,” Magn. Reson. Med., vol. 58, no. 6, pp. 1182–1195, Dec. 2007

    2007

  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, et al., “fastMRI: a publicly available raw k-space and DICOM dataset of knee images for accelerated MR image reconstruction using machine learning,”Radiol., Artif. Intell, vol. 2, no. 1, Jan. 2020, Art. no. e190007

    2020

  27. [35]

    Assessment of the generalization of learned image reconstruction and the potential for transfer learning,

    F. Knoll, et al., “Assessment of the generalization of learned image reconstruction and the potential for transfer learning,” Magn. Reson. Med., vol. 81, no. 1, pp. 116–128, 2019

    2019

  28. [36]

    Diff-Unfolding: A model-based score learning framework for inverse problems,

    Y . Wang, S. Shoushtari, and U. S. Kamilov, “Diff-Unfolding: A model-based score learning framework for inverse problems,” 2025, arXiv:2505.11393

    2025 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.