Pith. sign in

REVIEW 3 major objections 4 minor 36 references

Physics informed guided diffusion for accelerated multi-parametric MRI reconstruction

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read MRF-DiPh, a diffusion reconstruction that interleaves k-space consistency and Bloch dictionary matching into every sampling step, produces more accurate T1 and T2 maps from fivefold-accelerated MRF scans than the tested baselines.

desk verdict A genuinely new combination of diffusion priors with k-space and Bloch consistency for MRF reconstruction, under-supported by a small dataset and an ambiguous train/test split. read the letter →

arxiv 2506.23311 v1 pith:ZWZ2F2EW submitted 2025-06-29 eess.IV cs.LGphysics.med-ph

classification eess.IVcs.LGphysics.med-ph
keywords magneticresonancefingerprintingquantitativeMRIdenoisingdiffusionmodelsphysics-informedreconstructionBlochdictionarymatchingk-spaceconsistencyproximalsplittingmulti-parametrictissuemapping
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 claims that MRF-DiPh, a diffusion-based reconstruction algorithm, improves magnetic resonance fingerprinting (MRF) by enforcing two physical constraints at every sampling step: agreement with the measured k-space data and agreement with the Bloch signal model through dictionary matching. On retrospectively shortened in-vivo brain scans, the authors report lower error than every compared baseline on every metric, with $T_1$ MAPE about 2 percentage points and $T_2$ MAPE about 4.2 percentage points better than the best competing deep-learning method, and TSMI NRMSE about 8.6 percentage points lower. The physics-guided loop also keeps the reconstruction closer to the scanner measurements than a purely data-driven diffusion baseline, reducing k-space NRMSE by about 13 percent. If the result holds, quantitative MRI can be accelerated by a factor of five without changing the scanner protocol, only the reconstruction software.

What carries the argument

The load-bearing mechanism is the MRF-DiPh sampling loop (Algorithm 1), which alternates a diffusion denoising step, a k-space consistency step, and a Bloch dictionary projection at every reverse-diffusion iteration. The denoising diffusion model acts as a learned proximal operator for the image prior; the k-space step is a conjugate-gradient solve of a linear least-squares data-fidelity term; and the Bloch step is a dictionary match that projects the current estimate onto the nearest of roughly 95,000 precomputed $T_1$-$T_2$ fingerprint responses generated by the extended phase graph model. An ADMM dual variable carries the accumulated discrepancy between the data-consistent and Bloch-consistent estimates, and the deterministic noise injected at the next diffusion step is recomputed from the physics-consistent image rather than from the raw denoiser output.

What would settle it

Re-run the Table 1 evaluation under a strictly subject-disjoint split, for example leave-one-subject-out on the eight subjects, and compare $T_1$ and $T_2$ MAPE and k-space NRMSE against the purely data-driven diffusion baseline; if the reported margins vanish or reverse, the claimed advantage is not robust.

Watch

Extended reading notes

Core claim

The paper's central claim is that a pretrained denoising diffusion model can serve as the image prior in a half-quadratic splitting solution of the MRF inverse problem, provided each reverse-diffusion step is followed by an ADMM update that projects the denoised estimate onto two physical constraint sets: k-space measurement consistency and Bloch-model consistency via dictionary matching. The reported experiments on eight healthy volunteers with fivefold temporal acceleration show that this interleaving, called MRF-DiPh, improves $T_1$ MAPE from 8.45 to 6.75 percent and $T_2$ MAPE from 22.54 to 18.40 percent relative to the strongest data-driven baseline, reduces k-space NRMSE from 36.06 to 22.82 percent, and exceeds the classical iterative and CNN baselines on all reported metrics. The paper attributes the gain specifically to the simultaneous enforcement of both physical constraints: ablations that keep only k-space consistency (Mode C) or only the diffusion prior (MRF-IDDPM) perform worse.

Load-bearing premise

The reported gains assume that no subject contributes slices to both training and test sets; the paper describes a 75-25 split without stating that the split is subject-disjoint, so the diffusion prior could memorize a subject's anatomy if its slices appear in both splits.

Editorial extensions

If this is right

  • At fivefold temporal acceleration, the reported $T_1$ and $T_2$ maps have lower MAPE than the compared deep-learning and compressed-sensing methods, so quantitative maps could be obtained from shorter scans.
  • The physics-constrained loop keeps reconstructions closer to raw measurements, with k-space NRMSE about 13 percent lower than the purely data-driven diffusion baseline, which reduces the risk of hallucinated anatomy.
  • Competitive accuracy is reached earlier in training when physics guidance is active, suggesting the method needs fewer training iterations for a given error target.
  • Reconstruction time can be traded against accuracy: reducing diffusion steps or conjugate-gradient iterations drops runtime from about 44 seconds into the 9-16 second range with modest metric changes.

Reading between the lines

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

  • An untested direction is spatial rather than temporal k-space subsampling; the paper notes the unconditional variant underperformed in a scenario with further subsampling, so applying the physics-guided loop there is a natural next experiment.
  • Dictionary granularity bounds precision: with about 95,000 atoms, the projection step cannot represent $T_2$ values between entries, so replacing dictionary matching with a continuous Bloch inversion could improve grey-matter and CSF estimates.
  • The checkpoint analysis hints at data efficiency, but the paper does not quantify how much paired training data the physics-guided model saves at a fixed error target; a learning-curve study would settle it.
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 / 4 minor

Summary. The paper proposes MRF-DiPh, a physics-informed denoising diffusion reconstruction method for accelerated magnetic resonance fingerprinting (MRF). The method alternates a diffusion-model denoising step with two physical-consistency steps: k-space measurement consistency via a linear least-squares prox operator, and Bloch-model consistency via dictionary matching. The algorithm is derived from a half-quadratic splitting (HQS) formulation in which the diffusion denoiser replaces the proximal operator of an image prior. Experiments on retrospectively shortened in-vivo brain scans (8 subjects, 15 slices each) at R=5 acceleration report improved T1/T2 MAPE and TSMI NRMSE over compressed sensing, CNN, and diffusion baselines, with ablations showing the contribution of each physical constraint.

Significance. If the reported gains are robust, the paper makes a useful contribution by demonstrating that explicit k-space and Bloch-consistency enforcement can be combined with a pretrained diffusion prior for quantitative MRI. The core idea of integrating dictionary-based physical constraints into diffusion sampling is timely and potentially applicable to other quantitative imaging problems. The paper includes informative ablations (Modes A-D) and an analysis of training-time efficiency. However, the empirical evidence is currently under-supported: the dataset is small, the train/test split is not stated as subject-level, and the reported metrics are single averaged values without error bars or significance tests. The theoretical derivation from HQS to the diffusion-step update is heuristic and is not rigorously justified. These issues are central to the manuscript's claims and require attention.

major comments (3)
  1. [Section 4 (Dataset)] The train/test split is not stated to be subject-level. The text says 'The dataset includes 8 subjects with 15 axial slices each, split 75%-25% for training and testing.' With only 8 subjects, a random slice-level split would place slices from the same subject in both training and test sets. Because the conditional diffusion denoiser is trained on x_c and reference TSMIs of the same subjects, slice-level leakage could allow the network to memorize subject anatomy and would inflate the reported improvements over baselines. Please clarify explicitly whether all slices of each subject are kept in one split. If the split is not subject-level, the experiments must be rerun with a subject-wise split, and per-subject variance (e.g., boxplots or error bars) should be reported. Given that the reported advantages over MRF-IDDPM are modest (T1 MAPE 6.75 vs 8.45; T2 MAPE 18.40 vs 22.54), this issue is load-bearing for the central claim of generalization.
  2. [Section 5, Table 1] Table 1 reports a single averaged metric per method with no error bars, no per-subject statistics, and no significance tests. With only two test subjects under a subject-level split (25% of 8 subjects), the reported differences may not be statistically meaningful. The claim in Section 5 that MRF-DiPh 'outperforms the tested baselines across all reconstruction metrics' is also contradicted by Table 1 for k-space NRMSE: MRF-DiPh (22.82) is worse than MRF-ADMM (18.69) and LRTV (11.72). The abstract's statement about 'better preserving measurement fidelity' is only true relative to MRF-IDDPM and SVDMRF, not relative to all baselines. Please qualify the claims accordingly and add measures of variability.
  3. [Section 3, Eqs. (5)-(10)] The derivation from HQS to the DDM update is heuristic and should be presented as such or rigorously justified. In the HQS splitting, equation (6) requires the proximal operator of the prior h(x). Equation (8) replaces this with a Tweedie-style denoising estimate from the diffusion model, but the paper does not show that the learned denoiser corresponds to the proximal operator of any h(x), nor that the resulting iteration converges. Additionally, the statement that 'a single ADMM iteration is sufficient for updating (9)' is an empirical observation without supporting analysis. These are not necessarily fatal if the method is positioned as a heuristic algorithm inspired by HQS, but the current wording ('derived from a proximal splitting formulation') overstates the theoretical foundation. Please either provide a convergence or consistency argument, or explicitly describe the derivation as a heuristic approximation.
minor comments (4)
  1. [Algorithm 1] In Algorithm 1, lines 7-8 use the index t in 'vt-1 = vt + ...' and 'ˆϵk = ... (xk - √ ¯αk z_{k-1})' while the surrounding code uses k; the indices should be made consistent (e.g., v_{k-1} = v_k + ...).
  2. [Section 4] There are typos and spacing issues, e.g., 'dimentionality reduction' and 'pefermed best' in Section 4, and 'T able 1' / 'T able 2' at the start of the table captions. The captions of Figure 1 also appear incomplete ('zoomed-in absolute percentage error maps' is cut off).
  3. [Table 2] Table 2's header layout is confusing: the row 'MRF-DiPh Base K CG' mixes the base-model column with the K and CG columns. Please restructure the table so that the base model and the varied parameters are clearly separated, and define what each column represents.
  4. [Section 3.1] The text says 'range-normalized to [-1,1]' for complex-valued data processed by concatenating real and imaginary parts. It would be clearer to specify whether the normalization is applied separately to the real and imaginary channels or jointly to the magnitude.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the physics-constrained diffusion reconstruction is derived from an explicit HQS/ADMM formulation with independently trained priors, and the reported gains are evaluated against external baselines.

full rationale

The derivation chain is self-contained. MRF-DiPh is obtained by applying HQS to Eq. (5), replacing the proximal operator of the prior with a pretrained DDM denoiser (Eqs. 6-10), and solving the constrained subproblem via ADMM with a dictionary-matching projection (Eqs. 11-13). The Bloch consistency projection and k-space consistency steps are constraints imposed by the method, not fitted outputs, and the diffusion prior is trained independently on matched xc/xref pairs. The reference TSMIs are generated from the Bloch model, which is the same physical forward model used by the method; this is standard model-based reconstruction rather than a circular reduction, because the method's qmaps are not defined as the reference values. The closest method, MRF-IDDPM, is the authors' prior work, but it is used as an ablation baseline with the same trained network and does not supply any load-bearing premise. The k-space NRMSE improvement is expected because the method explicitly enforces measurement consistency, but the paper presents it as a constraint-satisfaction property, not as an independent prediction. The train/test split ambiguity noted in Section 4 is a data-leakage concern about experimental validity, not a circularity of the derivation; it does not make the predicted quantities equal to the inputs by construction. Overall, no step in the claimed derivation reduces to its own input.

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

The paper introduces no new physical entities. The method rests on standard MRF physics, a pretrained diffusion prior, and several tuned hyperparameters. The main contributions are algorithmic and empirical, not new physics.

free parameters (5)
  • lambda (regularization weight) = 1e-4
    Balances k-space consistency against the diffusion prior; reported as experimentally tuned in Section 3.1.
  • tau (ADMM balance parameter) = 0.01
    Sets gamma_t = tau * mu_t to balance k-space and Bloch consistency; tuned in the range [0.01, 0.1].
  • xi (sampling stochasticity) = 1.0 in base mode
    Controls the mix of deterministic and stochastic noise in the diffusion update; ablated at 0.5 and 0.
  • K (number of diffusion sampling steps) = 30
    Sub-sequence length for accelerated sampling; chosen as a runtime-accuracy trade-off.
  • CG maximum iterations = 5
    Number of conjugate gradient iterations in the k-space consistency subproblem; selected for speed with controlled accuracy loss.
assumptions (5)
  • domain assumption The Bloch response dictionary D and the EPG simulation accurately represent the MRF tissue signal model.
    The method enforces Bloch consistency by projecting onto this dictionary, and the reference TSMIs are generated using the same model, so dictionary accuracy is load-bearing.
  • ad hoc to paper DDM denoising in equation (8) approximates the proximal operator of an image prior h(x) in the HQS splitting.
    The paper replaces the proximal step (6) with a diffusion denoising step without identifying h(x) or proving equivalence; this is a plug-and-play assumption.
  • ad hoc to paper A single ADMM iteration per diffusion step is sufficient for the constrained subproblem (9).
    Justified only by 'Early experiments showed' in Section 3, with no convergence analysis or systematic study.
  • domain assumption Reference TSMIs estimated from LRTV qmaps via the Bloch model are adequate ground truth for training and evaluation.
    Section 4 describes building reference TSMIs from LRTV qmaps; errors in LRTV or in the Bloch inversion propagate into training targets and evaluation.
  • domain assumption The time-domain SVD compression with s = 5 preserves enough signal information for accurate T1/T2 mapping.
    All methods except SCQ use the linear dimensionality reduction from [27]; the reconstruction operates in a 5-dimensional subspace, which may discard discriminative signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics informed guided diffusion for accelerated multi-parametric MRI reconstruction." pith.science (2026). https://pith.science/paper/ZWZ2F2EW

@misc{pith2026250623311,
  author       = {Pith},
  title        = {Pith review of: Physics informed guided diffusion for accelerated multi-parametric MRI reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZWZ2F2EW}},
  note         = {Machine review of arXiv:2506.23311}
}
read the original abstract

We introduce MRF-DiPh, a novel physics informed denoising diffusion approach for multiparametric tissue mapping from highly accelerated, transient-state quantitative MRI acquisitions like Magnetic Resonance Fingerprinting (MRF). Our method is derived from a proximal splitting formulation, incorporating a pretrained denoising diffusion model as an effective image prior to regularize the MRF inverse problem. Further, during reconstruction it simultaneously enforces two key physical constraints: (1) k-space measurement consistency and (2) adherence to the Bloch response model. Numerical experiments on in-vivo brain scans data show that MRF-DiPh outperforms deep learning and compressed sensing MRF baselines, providing more accurate parameter maps while better preserving measurement fidelity and physical model consistency-critical for solving reliably inverse problems in medical imaging.

Figures

Figures reproduced from arXiv: 2506.23311 by the authors.

Figure 1
Figure 1. Reconstructed T1-T2 maps and their zoomed-in absolute percentage error maps (rows), using different methods (columns) for a representative brain slice in test set (electronic zoom recommended) [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 34 canonical work pages

  1. [1]

    https://github.com/openai/guided-diffusion

  2. [2]

    IEEE Signal Processing Magazine37(1), 105–116 (2020)

    Ahmad, R., Bouman, C.A., Buzzard, G.T., Chan, S., Liu, S., Reehorst, E.T., Schniter, P.: Plug-and-play methods for magnetic resonance imaging: Using denois- ers for image recovery. IEEE Signal Processing Magazine37(1), 105–116 (2020)

  3. [3]

    Magnetic resonance in medicine79(1), 83–96 (2018)

    Assländer, J., et al.: Low rank alternating direction method of multipliers re- construction for MR fingerprinting. Magnetic resonance in medicine79(1), 83–96 (2018)

  4. [4]

    In: Machine Learning for Medical Image Re- construction

    Balsiger, F., et al.: Magnetic resonance fingerprinting reconstruction via spatiotem- poral convolutional neural networks. In: Machine Learning for Medical Image Re- construction. pp. 39–46. Springer (2018)

  5. [5]

    IEEE Transactions on Medical Imaging pp

    Bian, W., Jang, A., Zhang, L., Yang, X., Stewart, Z., Liu, F.: Diffusion Modeling with Domain-conditioned Prior Guidance for Accelerated MRI and qMRI Recon- struction. IEEE Transactions on Medical Imaging pp. 1–1 (2024)

  6. [6]

    Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J.: (2011)

  7. [7]

    Magnetic Resonance in Medicine88(1), 133–150 (2022)

    Cao, X., et al.: Optimized multi-axis spiral projection MR fingerprinting with sub- space reconstruction for rapid whole-brain high-isotropic-resolution quantitative imaging. Magnetic Resonance in Medicine88(1), 133–150 (2022)

  8. [8]

    Magnetic resonance in medicine74(2), 523–528 (2015)

    Cauley, S.F., Setsompop, K., Ma, D., Jiang, Y., Ye, H., Adalsteinsson, E., Gris- wold, M.A., Wald, L.L.: Fast group matching for mr fingerprinting reconstruction. Magnetic resonance in medicine74(2), 523–528 (2015)

Show all 36 references
  1. [9]

    In: The Eleventh International Conference on Learning Representations (2023)

    Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sam- pling for general noisy inverse problems. In: The Eleventh International Conference on Learning Representations (2023)

  2. [10]

    Magnetic resonance in medicine81(6), 3530–3543 (2019)

    Lima da Cruz, G., Bustin, A., Jaubert, O., Schneider, T., Botnar, R.M., Prieto, C.: Sparsity and locally low rank regularization for MR fingerprinting. Magnetic resonance in medicine81(6), 3530–3543 (2019)

  3. [11]

    Siam journal on imaging sciences7(4), 2623–2656 (2014)

    Davies, M., Puy, G., Vandergheynst, P., Wiaux, Y.: A compressed sensing frame- work for magnetic resonance fingerprinting. Siam journal on imaging sciences7(4), 2623–2656 (2014)

  4. [12]

    Dhariwal,P.,Nichol,A.:DiffusionmodelsbeatGANsonimagesynthesis.Advances in neural information processing systems34, 8780–8794 (2021)

  5. [13]

    IEEE transactions on medical imaging38(10), 2364–2374 (2019) 10 Authors Suppressed Due to Excessive Length

    Fang, Z., et al.: Deep learning for fast and spatially constrained tissue quantifi- cation from highly accelerated data in magnetic resonance fingerprinting. IEEE transactions on medical imaging38(10), 2364–2374 (2019) 10 Authors Suppressed Due to Excessive Length

  6. [14]

    In: 2022 IEEE 19th International Symposium on Biomed- ical Imaging (ISBI)

    Fatania, K., Pirkl, C.M., Menzel, M.I., Hall, P., Golbabaee, M.: A plug-and-play approach to multiparametric quantitative MRI: image reconstruction using pre- trained deep denoisers. In: 2022 IEEE 19th International Symposium on Biomed- ical Imaging (ISBI). pp. 1–4. IEEE (2022)

  7. [15]

    Inverse Problems 36(1), 015003 (2019)

    Golbabaee, M., Chen, Z., Wiaux, Y., Davies, M.: CoverBLIP: accelerated and scal- able iterative matched-filtering for magnetic resonance fingerprint reconstruction. Inverse Problems 36(1), 015003 (2019)

  8. [16]

    Medical image analysis69, 101945 (2021)

    Golbabaee, M., et al.: Compressive mri quantification using convex spatiotempo- ral priors and deep encoder-decoder networks. Medical image analysis69, 101945 (2021)

  9. [17]

    Scientific reports10(1), 13769 (2020)

    Gómez, P.A., et al.: Rapid three-dimensional multiparametric mri with quantita- tive transient-state imaging. Scientific reports10(1), 13769 (2020)

  10. [18]

    Medical image analysis88, 102872 (2023)

    Güngör, A., et al.: Adaptive diffusion priors for accelerated MRI reconstruction. Medical image analysis88, 102872 (2023)

  11. [19]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

  12. [20]

    In: Advances in Neural In- formation Processing Systems

    Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A.G., Tamir, J.: Robust compressed sensing mri with deep generative priors. In: Advances in Neural In- formation Processing Systems. vol. 34, pp. 14938–14954. Curran Associates, Inc. (2021)

  13. [21]

    Magnetic resonance in medicine74(6), 1621–1631 (2015)

    Jiang, Y., Ma, D., Seiberlich, N., Gulani, V., Griswold, M.A.: MR fingerprinting us- ing fast imaging with steady state precession (FISP) with spiral readout. Magnetic resonance in medicine74(6), 1621–1631 (2015)

  14. [22]

    In: Greenspan, H., Madabhushi, A., Mousavi, P., Salcud- ean, S., Duncan, J., Syeda-Mahmood, T., Taylor, R

    Korkmaz, Y., Cukur, T., Patel, V.M.: Self-supervised mri reconstruction with un- rolled diffusion models. In: Greenspan, H., Madabhushi, A., Mousavi, P., Salcud- ean, S., Duncan, J., Syeda-Mahmood, T., Taylor, R. (eds.) Medical Image Com- puting and Computer Assisted Intervent...

  15. [23]

    In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision

    Liu,J.,etal.:DOLCE:Amodel-basedprobabilisticdiffusionframeworkforlimited- angle ct reconstruction. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision. pp. 10498–10508 (2023)

  16. [24]

    Nature 495(7440), 187–192 (2013)

    Ma, D., et al.: Magnetic resonance fingerprinting. Nature 495(7440), 187–192 (2013)

  17. [25]

    arXiv preprint arXiv:2410.23318 (2024)

    Mayo, P., Pirkl, C.M., Achim, A., Menze, B.H., Golbabaee, M.: Denoising dif- fusion probabilistic models for magnetic resonance fingerprinting. arXiv preprint arXiv:2410.23318 (2024)

  18. [26]

    Medical physics45(9), 4066–4084 (2018)

    Mazor, G., Weizman, L., Tal, A., Eldar, Y.C.: Low-rank magnetic resonance fin- gerprinting. Medical physics45(9), 4066–4084 (2018)

  19. [27]

    IEEE transactions on medical imaging33(12), 2311–2322 (2014)

    McGivney, D., et al.: SVD compression for magnetic resonance fingerprinting in the time domain. IEEE transactions on medical imaging33(12), 2311–2322 (2014)

  20. [28]

    In: International conference on machine learning

    Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International conference on machine learning. pp. 8162–8171. PMLR (2021)

  21. [29]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Peng, C., Guo, P., Zhou, S.K., Patel, V.M., Chellappa, R.: Towards performant and reliable undersampled MR reconstruction via diffusion model sampling. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 623–633. Springer (2022)

  22. [30]

    In: Interna- tional Conference on Learning Representations (2021)

    Song, J., Meng, C., Ermon, S.: Denoising Diffusion Implicit Models. In: Interna- tional Conference on Learning Representations (2021)

  23. [31]

    In: Interna- tional Conference on Learning Representations (2021) Title Suppressed Due to Excessive Length 11

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: Interna- tional Conference on Learning Representations (2021) Title Suppressed Due to Excessive Length 11

  24. [32]

    In: Mukhopadhyay, A., Oksuz, I., Engelhardt, S., Mehrof, D., Yuan, Y

    Wang, S., Ma, H., Hernandez-Tamames, J.A., Klein, S., Poot, D.H.J.: qMRI Dif- fuser: Quantitative T1 Mapping of the Brain Using a Denoising Diffusion Proba- bilistic Model. In: Mukhopadhyay, A., Oksuz, I., Engelhardt, S., Mehrof, D., Yuan, Y. (eds.) Deep Generative Models. pp....

  25. [33]

    Journal of Magnetic Resonance Imaging41(2), 266–295 (2015)

    Weigel, M.: Extended phase graphs: dephasing, RF pulses, and echoes-pure and simple. Journal of Magnetic Resonance Imaging41(2), 266–295 (2015)

  26. [34]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(10), 6360–6376 (2022)

    Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence44(10), 6360–6376 (2022)

  27. [35]

    Magnetic resonance in medicine79(2), 933–942 (2018)

    Zhao, B., et al.: Improved magnetic resonance fingerprinting reconstruction with low-rank and subspace modeling. Magnetic resonance in medicine79(2), 933–942 (2018)

  28. [36]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    Zhu, Y., et al.: Denoising diffusion models for plug-and-play image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 1219–1229 (June 2023)

Pith tools

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