pith. sign in

arxiv: 2512.23726 · v2 · submitted 2025-12-19 · ⚛️ physics.med-ph · cs.AI· cs.CV

q3-MuPa: Quick, Quiet, Quantitative Multi-Parametric MRI using Physics-Informed Diffusion Models

Shishuai Wang , Florian Wiesinger , Noemi Sgambelluri , Carolin Pirkl , Stefan Klein , Juan A. Hernandez-Tamames , Dirk H.J. Poot This is my paper

Pith reviewed 2026-05-16 21:09 UTC · model grok-4.3

classification ⚛️ physics.med-ph cs.AIcs.CV
keywords quantitative MRIdiffusion modelsphysics-informedMuPa-ZTEmulti-parametric mappingsynthetic datasilent MRIaccelerated acquisition
0
0 comments X

The pith

Physics-informed diffusion model turns fourfold-accelerated silent MRI scans into accurate 3D quantitative maps of T1, T2 and proton density.

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

The paper presents q3-MuPa, which pairs a nearly silent 3D MuPa-ZTE acquisition that uses phyllotaxis sampling with a denoising diffusion probabilistic model to produce quantitative maps. The model is trained exclusively on synthetic image series generated from digital brain phantoms and uses the known MuPa-ZTE forward signal equation as an explicit data-consistency constraint during inference. This combination supports fourfold acceleration, shortening scans to roughly one minute while delivering maps that show lower noise and better structural detail than either dictionary matching or a purely data-driven diffusion baseline. The approach generalizes from synthetic training data to real scans of healthy volunteers and a patient with brain metastases. A sympathetic reader would care because quantitative MRI could become far more practical in clinics once scan time, acoustic noise, and motion sensitivity are reduced without the need for large real-world training sets.

Core claim

By training a denoising diffusion probabilistic model on synthetic MuPa-ZTE image series from digital brain phantoms and enforcing the MuPa-ZTE physics forward model as a data-consistency term at inference, the method recovers high-accuracy 3D maps of T1, T2, and proton density from fourfold-accelerated acquisitions, with reduced noise and improved detail preservation relative to dictionary matching and data-only diffusion models, while generalizing directly to real patient data.

What carries the argument

A denoising diffusion probabilistic model (DDPM) whose reverse sampling process is constrained by the MuPa-ZTE forward signal model acting as an explicit data-consistency term.

If this is right

  • High-quality quantitative maps become feasible in approximately one-minute scans.
  • Noise is reduced and fine anatomical detail is better preserved than with dictionary matching or purely data-driven diffusion models.
  • The method works on real clinical scans even though it was trained only on synthetic phantoms.
  • Patient comfort improves through near-silent acquisition and greater tolerance to motion.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same physics-constrained diffusion approach could be ported to other quantitative MRI sequences whose forward models are known, enabling similar acceleration without new real-data collections.
  • Reliance on synthetic phantoms lowers the data barrier for developing quantitative mapping tools in rare-disease or pediatric settings where large patient cohorts are unavailable.
  • The framework might be extended to dynamic or contrast-enhanced studies where motion and short acquisition windows are limiting factors.

Load-bearing premise

Digital brain phantoms used to generate the synthetic training data already contain enough variability in tissue properties, noise levels, and acquisition imperfections that the model will perform well on real patient scans.

What would settle it

Acquire reference T1, T2, and proton-density maps on the same patients using established slow, non-accelerated quantitative sequences and compare them voxel-wise with the maps produced by the accelerated q3-MuPa method; systematic large errors or loss of structural fidelity on real data would falsify the generalization claim.

Figures

Figures reproduced from arXiv: 2512.23726 by Carolin Pirkl, Dirk H.J. Poot, Florian Wiesinger, Juan A. Hernandez-Tamames, Noemi Sgambelluri, Shishuai Wang, Stefan Klein.

Figure 1
Figure 1. Figure 1: Schematic of a MuPa-ZTE acquisition. The PD-weighted [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A schematic of the inference workflow of the proposed method. With a trained noise-prediction model, the qMRI maps corresponding to [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results of experiments on synthetic test datasets. The inferences were conducted on 30 patches of size 40 [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Quantitative evaluation of T1 and T2 mapping using the [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example of mapping results (PD in a.u., T1 and T2 in ms) from all three methods: DictMatch, DL-Di [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example uncertainty maps comparing DL-Di [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example of mapping results (PD in a.u., T1 and T2 in ms) from all three methods: DictMatch, DL-Di [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Conventional MR images of the patient. lowed to recover the fine structures from highly under￾sampled data, and the data consistency step ensured that these structures were consistent with the acquired sig￾nals. This capability could facilitate the broader inte￾gration of MuPa-ZTE into routine clinical practice. In terms of limitations, the T2 mapping exhibits a rel￾atively larger systematic error compared… view at source ↗
read the original abstract

The 3D fast silent multi-parametric mapping sequence with zero echo time (MuPa-ZTE) is a novel quantitative MRI (qMRI) acquisition that enables nearly silent scanning by using a 3D phyllotaxis sampling scheme. MuPa-ZTE improves patient comfort and motion robustness, and generates quantitative maps of T1, T2, and proton density using the acquired weighted image series. In this work, we propose a diffusion model-based qMRI mapping method that leverages both a deep generative model and physics-based data consistency to further improve the mapping performance. Furthermore, our method enables additional acquisition acceleration, allowing high-quality qMRI mapping from a fourfold-accelerated MuPa-ZTE scan (approximately 1 minute). Specifically, we trained a denoising diffusion probabilistic model (DDPM) to map MuPa-ZTE image series to qMRI maps, and we incorporated the MuPa-ZTE forward signal model as an explicit data consistency (DC) constraint during inference. We compared our mapping method against a baseline dictionary matching approach and a purely data-driven diffusion model. The diffusion models were trained entirely on synthetic data generated from digital brain phantoms, eliminating the need for large real-scan datasets. We evaluated on synthetic data, a NISM/ISMRM phantom, healthy volunteers, and a patient with brain metastases. The results demonstrated that our method produces 3D qMRI maps with high accuracy, reduced noise and better preservation of structural details. Notably, it generalised well to real scans despite training on synthetic data alone. The combination of the MuPa-ZTE acquisition and our physics-informed diffusion model is termed q3-MuPa, a quick, quiet, and quantitative multi-parametric mapping framework, and our findings highlight its strong clinical potential.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces q3-MuPa, a framework that combines the MuPa-ZTE acquisition sequence with a physics-informed denoising diffusion probabilistic model (DDPM) for 3D quantitative mapping of T1, T2, and proton density. The DDPM is trained exclusively on synthetic data generated from digital brain phantoms and incorporates the MuPa-ZTE forward signal model as an explicit data-consistency constraint during inference. The method is evaluated on synthetic data, an ISMRM phantom, healthy volunteers, and one patient with brain metastases, with claims of high accuracy, reduced noise, improved structural detail preservation, and successful generalization from synthetic training to real scans, enabling fourfold acceleration to approximately 1-minute acquisitions.

Significance. If the generalization from synthetic phantoms to real data holds under quantitative scrutiny, the work could meaningfully advance clinical qMRI by enabling silent, motion-robust, and accelerated multi-parametric mapping without large real-world training datasets, potentially improving patient comfort and accessibility in settings where conventional qMRI is limited by scan time or noise.

major comments (3)
  1. [Results] Results section: claims of 'high accuracy' and successful generalization to real volunteer and patient scans are supported only by qualitative visual comparisons; no RMSE, bias, correlation, or SSIM values are reported for T1/T2/PD maps on the volunteer or metastases-patient data relative to dictionary matching or the pure diffusion baseline, preventing assessment of the magnitude of improvement or domain-shift magnitude.
  2. [Methods] Methods (synthetic data generation): the ranges, distributions, and joint statistics of tissue parameters, B1 inhomogeneity, noise, and phyllotaxis-sampling artifacts simulated in the digital brain phantoms are not specified, leaving the weakest assumption—that these phantoms capture sufficient real-world variability—unverifiable and load-bearing for the generalization claim.
  3. [Results] Results (ablation): no quantitative ablation isolating the physics-based data-consistency term on real data (e.g., performance of the DDPM with versus without the DC constraint on volunteer scans) is presented, so the contribution of the physics-informed component to the reported improvements on real acquisitions remains unquantified.
minor comments (2)
  1. [Abstract] Abstract: the statement that the method 'generalised well' would be strengthened by inclusion of at least one key quantitative metric (e.g., mean T1 error on the phantom) if space allows.
  2. [Figures] Figure captions: several result figures lack scale bars or color-bar ranges for the quantitative maps, reducing interpretability of the visual comparisons.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback on our manuscript. We address each of the major comments below and have made revisions to strengthen the paper.

read point-by-point responses

  1. Referee: [Results] Results section: claims of 'high accuracy' and successful generalization to real volunteer and patient scans are supported only by qualitative visual comparisons; no RMSE, bias, correlation, or SSIM values are reported for T1/T2/PD maps on the volunteer or metastases-patient data relative to dictionary matching or the pure diffusion baseline, preventing assessment of the magnitude of improvement or domain-shift magnitude.

    Authors: We agree that quantitative metrics on real data would better support our claims of high accuracy and generalization. Although quantitative evaluations (RMSE, etc.) are provided for synthetic and phantom data, the volunteer and patient results are presented qualitatively. In the revised manuscript, we will add quantitative comparisons for the healthy volunteer data against dictionary matching and the pure diffusion baseline. For the patient scan, where ground truth is unavailable, we will include supplementary quantitative measures such as noise reduction estimates and structural fidelity metrics to better quantify the improvements. revision: yes

  2. Referee: [Methods] Methods (synthetic data generation): the ranges, distributions, and joint statistics of tissue parameters, B1 inhomogeneity, noise, and phyllotaxis-sampling artifacts simulated in the digital brain phantoms are not specified, leaving the weakest assumption—that these phantoms capture sufficient real-world variability—unverifiable and load-bearing for the generalization claim.

    Authors: We acknowledge that the lack of detailed specifications for the synthetic phantom generation limits the verifiability of our generalization claims. In the revised Methods section, we will provide comprehensive details on the ranges, distributions, and joint statistics of the simulated tissue parameters (T1, T2, PD), B1 inhomogeneity, noise levels, and the incorporation of phyllotaxis-sampling artifacts. revision: yes

  3. Referee: [Results] Results (ablation): no quantitative ablation isolating the physics-based data-consistency term on real data (e.g., performance of the DDPM with versus without the DC constraint on volunteer scans) is presented, so the contribution of the physics-informed component to the reported improvements on real acquisitions remains unquantified.

    Authors: We agree that a quantitative ablation study on real data is necessary to isolate the effect of the physics-based data consistency constraint. The manuscript currently demonstrates the benefit of the DC term through ablation on synthetic data. We will extend this analysis to the volunteer scans in the revised Results section, reporting performance metrics (e.g., RMSE, SSIM) for the model with and without the DC constraint relative to dictionary matching. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; derivation remains self-contained

full rationale

The paper trains a DDPM exclusively on synthetic data generated from external digital brain phantoms and enforces the MuPa-ZTE forward signal model as an explicit data-consistency constraint only at inference time. This forward model is an independent physical description of the acquisition sequence, not a parameter fitted to the output maps. The central performance claims are supported by evaluation on held-out synthetic data, an ISMRM phantom, volunteers, and a patient scan, none of which are used in training. No self-citations, uniqueness theorems, or ansatzes are invoked to justify load-bearing steps, and no fitted input is relabeled as a prediction. The generalization step from synthetic to real data is an empirical claim rather than a definitional reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the accuracy of the MuPa-ZTE forward signal model used as a data-consistency constraint and on the representativeness of the digital brain phantoms for real tissue and noise distributions; no new physical entities are postulated.

free parameters (1)
  • DDPM training hyperparameters
    Specific learning rate, number of diffusion steps, and network architecture details are not provided in the abstract and must be chosen to achieve the reported performance.
axioms (1)
  • domain assumption The MuPa-ZTE forward signal model accurately describes the relationship between acquired weighted images and the target T1, T2, and proton density maps
    Invoked explicitly as the data consistency constraint during inference.

pith-pipeline@v0.9.0 · 5658 in / 1354 out tokens · 31375 ms · 2026-05-16T21:09:28.363181+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    Blystad, J

    I. Blystad, J. B. M. Warntjes, O. Smedby, A.-M. Landtblom, P. Lundberg, E.-M. Larsson, Synthetic MRI of the brain in a clinical setting, Acta Radiol. 53 (10) (2012) 1158–1163

    work page 2012

  2. [2]

    D. Ma, V . Gulani, N. Seiberlich, K. Liu, J. L. Sun- shine, J. L. Duerk, M. A. Griswold, Magnetic res- onance fingerprinting, Nature 495 (7440) (2013) 187–192

    work page 2013

  3. [3]

    Kvernby, M

    S. Kvernby, M. J. B. Warntjes, H. Haraldsson, C.- J. Carlhäll, J. Engvall, T. Ebbers, Simultaneous three-dimensional myocardial T1 and T2 mapping in one breath hold with 3D-QALAS, J. Cardio- vasc. Magn. Reson. 16 (1) (2014) 102

    work page 2014

  4. [4]

    P. A. Gómez, M. Cencini, M. Golbabaee, R. F. Schulte, C. Pirkl, I. Horvath, G. Fallo, L. Peretti, M. Tosetti, B. H. Menze, G. Buonincontri, Rapid three-dimensional multiparametric MRI with quantitative transient-state imaging, Sci. Rep. 10 (1) (2020) 13769

    work page 2020

  5. [5]

    Ljungberg, T

    E. Ljungberg, T. C. Wood, A. B. Solana, S. Kolind, S. C. Williams, F. Wiesinger, G. J. Barker, Silent T1 mapping using the variable flip angle method with B1 correction, Magn. Reson. Med. 84 (2) (2020) 813–824

    work page 2020

  6. [6]

    Ljungberg, N

    E. Ljungberg, N. L. Damestani, T. C. Wood, D. J. Lythgoe, F. Zelaya, S. C. Williams, A. B. Solana, G. J. Barker, F. Wiesinger, Silent zero TE MR neu- roimaging: current state-of-the-art and future di- rections, Prog. Nucl. Magn. Reson. Spectrosc. 123 (2021) 73–93

    work page 2021

  7. [7]

    Wiesinger, G

    F. Wiesinger, G. McKinnon, S. Kaushik, A. B. Solana, E. Ljungberg, M. V ogel, N. Takei, R. F. Schulte, C. Pirkl, C. Cozzini, L. Nuñez-Gonzalez, J. A. Hernandez-Tamames, M. Engström, 3D silent parameter mapping: Further refinements & quantitative assessment, in: Proceedings of the International Society Magnetic Resonance in Medicine, V ol. 29, 2021, p. 1828

    work page 2021

  8. [8]

    Wiesinger, E

    F. Wiesinger, E. Ljungberg, M. Engström, S. Kaushik, T. C. Wood, S. Williams, G. Barker, A. B. Solana, PSST. . . Parameter mapping Swift and SilenT, in: Proceedings of the International Society Magnetic Resonance in Medicine, V ol. 882, 2020

    work page 2020

  9. [9]

    C. Cai, C. Wang, Y . Zeng, S. Cai, D. Liang, Y . Wu, Z. Chen, X. Ding, J. Zhong, Single-shot T2 map- ping using overlapping-echo detachment planar imaging and a deep convolutional neural network, Magn. Reson. Med. 80 (5) (2018) 2202–2214. 14 Figure B.10: Example of mapping results (PD in a.u., T1 and T2 in ms) from all three methods: DictMatch, DL-Diffusi...

    work page 2018

  10. [10]

    Jeelani, Y

    H. Jeelani, Y . Yang, R. Zhou, C. M. Kramer, M. Salerno, D. S. Weller, A myocardial T1- mapping framework with recurrent and U-Net convolutional neural networks, in: 2020 IEEE 17th international symposium on biomedical imaging (ISBI), IEEE, 2020, pp. 1941–1944

    work page 2020

  11. [11]

    J. Shao, V . Ghodrati, K.-L. Nguyen, P. Hu, Fast and accurate calculation of myocardial T1 and T2 values using deep learning Bloch equation simu- lations (DeepBLESS), Magn. Reson. Med. 84 (5) (2020) 2831–2845

    work page 2020

  12. [12]

    E. R. Sabidussi, S. Klein, M. W. Caan, 15 Figure C.11: Supplementary visualisation examples to Fig. 5, presenting the sagittal and coronal views. Units: PD in a.u., T1 and T2 in ms. Acquisition settings: (a)longMuPa-ZTE and (b)shortMuPa-ZTE. S. Bazrafkan, A. J. den Dekker, J. Sijbers, W. J. Niessen, D. H. Poot, Recurrent inference ma- chines as inverse pr...

    work page 2021

  13. [13]

    Zhang, D

    C. Zhang, D. Karkalousos, P.-L. Bazin, B. F. Coolen, H. Vrenken, J.-J. Sonke, B. U. Forstmann, D. H. Poot, M. W. Caan, A unified model for re- construction and R2* mapping of accelerated 7T data using the quantitative recurrent inference ma- chine, NeuroImage 264 (2022) 119680

    work page 2022

  14. [14]

    E. R. Sabidussi, S. Klein, B. Jeurissen, D. H. Poot, dtiRIM: A generalisable deep learning method for diffusion tensor imaging, NeuroImage 269 (2023) 119900. 16 Figure C.12: Supplementary visualisation examples to Fig. B.10, presenting the sagittal and coronal views. Units: PD in a.u., T1 and T2 in ms. Acquisition settings: (a)longMuPa-ZTE and (b)shortMuPa-ZTE

    work page 2023

  15. [15]

    S. Wang, H. Ma, J. A. Hernandez-Tamames, S. Klein, D. H. Poot, qMRI diffuser: quantitative T1 mapping of the brain using a denoising diffu- sion probabilistic model, in: MICCAI Workshop on Deep Generative Models, Springer, 2024, pp. 129–138

    work page 2024

  16. [16]

    W. Bian, A. Jang, L. Zhang, X. Yang, Z. Stew- art, F. Liu, Diffusion modeling with domain- conditioned prior guidance for accelerated MRI and qMRI reconstruction, IEEE Trans. Med. Imaging (2024)

    work page 2024

  17. [17]

    B. Song, S. M. Kwon, Z. Zhang, X. Hu, Q. Qu, L. Shen, Solving inverse problems with latent dif- fusion models via hard data consistency, arXiv 17 Figure C.13: Supplementary visualisation examples to Fig. 7, presenting the sagittal and coronal views. Units: PD in a.u., T1 and T2 in ms. Acquisition settings: (a)longMuPa-ZTE and (b)shortMuPa-ZTE. preprint ar...

    work page arXiv 2023

  18. [18]

    Wiesinger, M

    F. Wiesinger, M. A. Janich, E. Ljungberg, G. J. Barker, A. B. Solana, Silent, 3D MR parameter mapping using magnetization prepared zero TE, in: Proceedings of the International Society Mag- netic Resonance in Medicine, V ol. 61, 2018

    work page 2018

  19. [19]

    Haase, J

    A. Haase, J. Frahm, D. Matthaei, W. Hanicke, K.- D. Merboldt, FLASH imaging. Rapid NMR imag- ing using low flip-angle pulses, J. Magn. Reson. 67 (2) (1986) 258–266

    work page 1986

  20. [20]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Advances in Neural Infor- mation Processing Systems 33 (2020) 6840–6851. 18

    work page 2020

  21. [21]

    Z. Zhan, D. Chen, J.-P. Mei, Z. Zhao, J. Chen, C. Chen, S. Lyu, C. Wang, Conditional image synthesis with diffusion models: A survey, arXiv preprint arXiv:2409.19365 (2024)

    work page arXiv 2024

  22. [22]

    J. Yu, Y . Wang, C. Zhao, B. Ghanem, J. Zhang, Freedom: Training-free energy-guided condi- tional diffusion model, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 23174–23184

    work page 2023

  23. [23]

    Piccini, A

    D. Piccini, A. Littmann, S. Nielles-Vallespin, M. O. Zenge, Spiral phyllotaxis: the natural way to construct a 3D radial trajectory in MRI, Magn. Reson. Med. 66 (4) (2011) 1049–1056

    work page 2011

  24. [24]

    Ljungberg, T

    E. Ljungberg, T. C. Wood, A. B. Solana, S. C. Williams, G. J. Barker, F. Wiesinger, Motion cor- rected silent ZTE neuroimaging, Magn. Reson. Med. 88 (1) (2022) 195–210

    work page 2022

  25. [25]

    C. A. Cocosco, Brainweb: Online interface to a 3D MRI simulated brain database., NeuroImage (1997) 425

    work page 1997

  26. [26]

    M. J. Muckley, R. Stern, T. Murrell, F. Knoll, TorchKbNufft: A high-level, hardware-agnostic non-uniform fast Fourier transform, in: ISMRM Workshop on Data Sampling & Image Recon- struction, V ol. 22, 2020

    work page 2020

  27. [27]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High-resolution image synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10684–10695

    work page 2022

  28. [28]

    W. H. Pinaya, P.-D. Tudosiu, J. Dafflon, P. F. Da Costa, V . Fernandez, P. Nachev, S. Ourselin, M. J. Cardoso, Brain imaging generation with latent diffusion models, in: MICCAI Workshop on Deep Generative Models, Springer, 2022, pp. 117–126

    work page 2022

  29. [29]

    Z. Ding, M. Zhang, J. Wu, Z. Tu, Patched de- noising diffusion models for high-resolution im- age synthesis, arXiv preprint arXiv:2308.01316 (2023)

    work page arXiv 2023

  30. [30]

    Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simon- celli, Image quality assessment: from error visi- bility to structural similarity, IEEE Trans. Image Process. 13 (4) (2004) 600–612. 19

    work page 2004