Pith. sign in

REVIEW 3 major objections 6 minor 73 references

Accelerating multiparametric quantitative MRI using self-supervised scan-specific implicit neural representation with model reinforcement

T0 review · 3 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 REFINE-MORE, a self-supervised scan-specific network, reconstructs 4x–5x accelerated multiparametric qMRI without fully sampled training data by coupling an implicit neural representation with MR-physics model reinforcement.

desk verdict A credible scan-specific qMT reconstruction method whose accuracy claims are real but only as strong as the BTS signal model they inherit. read the letter →

arxiv 2508.00891 v1 pith:4ZGE3EIH submitted 2025-07-27 physics.med-ph cs.AI

Ruimin Feng , Albert Jang , Xingxin He , Fang Liu This is my paper
classification physics.med-phcs.AI
keywords self-superviseddeeplearningimplicitneuralrepresentationmodelreinforcementmultiparametricquantitativeMRImagnetizationtransferhashencodinglow-rankadaptationundersampledk-spacereconstruction
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

REFINE-MORE claims that a scan-specific, self-supervised network can reconstruct accelerated multiparametric quantitative MRI without any fully sampled training data by combining an implicit neural representation with explicit MR physics constraints. The target application is simultaneous estimation of free-water relaxation T1F, macromolecular proton fraction f, and magnetization exchange rate kF from a quantitative magnetization transfer sequence (BTS). On in vivo brain data at 4x and 5x undersampling, the paper reports the lowest nRMSE and highest SSIM among the compared methods, and phantom regressions against fully sampled references give slopes near one with R2 values of 1.000, 0.999, and 0.999. A low-rank model-adaptation strategy cuts reconstruction time by roughly fivefold. If the results hold, this would make high-dimensional parameter mapping practical without the large training datasets that supervised qMRI methods need.

What carries the argument

The machinery is a two-stage scan-specific optimization. Stage one represents the whole 4D weighted image series W(x,y,z,t) as a function of spatial and temporal coordinates through separate hash encodings and small MLPs, and represents the parameter maps Q(x,y,z) as a function of spatial coordinates; it is trained by a data-consistency loss plus total variation for the images (L1) and a loss that fits the BTS signal model to those images (L2). Stage two unrolls a proximal gradient descent algorithm into four phases in which each parameter map is updated by a gradient step on the data-consistency term of Eq. (1) and then passed through a U-Net acting as an implicit proximal operator, with a self-supervised loss L3 summing data consistency over all phases. The BTS binary spin-bath signal model (Eq. 11) is the physical constraint that connects the weighted images to the target parameters T1F, f, and kF. A LoRA low-rank adaptation of the U-Net weights, plus reuse of frozen temporal hash encodings and MLP weights after the first subject, provides the approximate fivefold speedup.

What would settle it

Run REFINE-MORE at 4x acceleration on a phantom whose T1F, f, and kF values are known from an independent method (for example, NMR relaxometry or a different quantitative MT model), and compare its output to those independent values rather than to the fully sampled BTS fit; if the reported nRMSE advantage does not persist against this external reference, the claim that the method recovers true tissue parameters would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that REFINE-MORE, by coupling a coordinate-based implicit neural representation with an unrolled proximal gradient descent that enforces the BTS magnetization-transfer signal model, reconstructs multiparametric quantitative maps from undersampled k-space accurately and without any fully sampled reference data at training time. In the comparison on five in vivo brains, the reported nRMSE and SSIM at both 4x and 5x acceleration are the best among Zero Filling, locally low-rank reconstruction, RELAX, MANTIS, and SUMMIT, with the single exception that the SSIM of the kF map at AF=4 trails the best method by 0.3 percentage points. Phantom experiments show vial-average estimates agreeing with fully sampled fits with R2 values of 1.000, 0.999, and 0.999 for T1F, f, and kF, and the LoRA-based adaptation strategy cuts reconstruction time by about fivefold. In the authors' framing, the method works because the INR stage supplies a stable initialization of the parameter maps and the reinforcement stage corrects deviations through explicit physical data consistency.

Load-bearing premise

The load-bearing premise is that the BTS two-pool magnetization-transfer signal model (Eq. 11) correctly describes the measured signal in every imaged tissue; both the fully sampled nonlinear fit used as the reference and REFINE-MORE's data-consistency losses are built from this same model, so if the model is misspecified both share the same systematic bias.

Editorial extensions

If this is right

  • At 4x and 5x undersampling, the method produces T1F, f, and kF maps with lower nRMSE and generally higher SSIM than the compared model-based and supervised baselines on the reported in vivo dataset, meaning scan-specific reconstruction can beat methods that need external training data.
  • The INR's separate temporal encoding supports interpolation to unmeasured contrast time points: weighted images synthesized at t=2.5, 3.5, and 4.5 show plausible intermediate contrast, which could be used for synthetic contrasts or retrospective re-sampling.
  • Ablation shows the physics reinforcement module is what recovers fine detail: without it T1F is over-smoothed and f maps are inaccurate, and four unrolled phases are enough because gains saturate after three.
  • LoRA adaptation cuts trainable U-Net parameters from 7,763,879 to 517,600, and the adapted model reaches the non-adapted model's quality in about 200 iterations instead of 1500, reducing total reconstruction from roughly 71.8 minutes to 14.2 minutes.

Reading between the lines

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

  • Because the fully sampled reference and REFINE-MORE both use the same BTS signal model, the reported nRMSE and SSIM measure agreement with that model's parameterization rather than with independently known tissue properties; separating reconstruction error from model error would require an external standard.
  • The two-stage design is not tied to magnetization transfer: swapping the signal model S in Eqs. (1), (6), and (9) should extend REFINE-MORE to other quantitative contrasts such as T1/T2 mapping or quantitative susceptibility imaging, though the paper's limitation section notes that hyperparameters may need adjustment for new anatomies and signal models.
  • The model-adaptation result suggests a continual-learning workflow in which temporal and MLP weights learned on earlier subjects are reused for later ones, so per-scan cost could keep shrinking as more subjects are reconstructed; the paper demonstrates the benefit only across the five in vivo subjects.
  • An external calibrated phantom with values measured by independent relaxometry or NMR, rather than by the BTS fit, would give a cleaner test of the reconstruction claim, because the current phantom regression compares the method to the same model used to generate its reference.
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 / 6 minor

Summary. The paper proposes REFINE-MORE, a self-supervised, scan-specific framework for accelerated multiparametric quantitative MRI. The method has two stages: first, an implicit neural representation with hash encodings and MLPs reconstructs undersampled weighted images and initializes quantitative maps (T1F, f, kF) through losses L1 and L2; second, an unrolled proximal gradient descent with a U-Net as a learned proximal operator enforces data consistency with the MR physics model through loss L3. The method is evaluated on BTS quantitative magnetization transfer imaging at 4x and 5x accelerations using five healthy subjects and a phantom, reporting the lowest nRMSE and highest SSIM among compared methods in Table 1, and a fivefold speedup via model adaptation with low-rank adaptation (LoRA) and weight transfer.

Significance. If the accuracy claim holds, the work is a useful engineering contribution: it combines scan-specific INR initialization with model-based unrolled refinement for high-dimensional qMRI, ships public code, and includes ablations of the key components (number of unrolled phases, regularization weight, LoRA adaptation). The method is internally consistent and the optimization losses are correctly formulated. However, the significance is conditional because the evaluation establishes agreement with the BTS signal model rather than accuracy of tissue properties: the reference maps, the reconstruction losses, and the phantom ground truth all derive from the same forward model. The comparison with supervised baselines is also weakened by the very small training set used for those baselines.

major comments (3)
  1. [Section 3.1, 3.4.1, 4.2 (Eq. 11)] The accuracy claim is conditional on the validity of the BTS signal model, which is not independently validated. The fully sampled reference maps are obtained by nonlinear voxel-wise fitting of Eq. (11) (Section 3.4.1), and the reconstruction losses L1, L2, and L3 all enforce consistency with the same forward model S. The phantom experiment (Sections 3.4.2 and 4.2) compares REFINE-MORE against fully sampled BTS fits; the expected T1F, f, and kF values for the agar and egg-white vials are never stated, and no independent measurement (e.g., standard T1 mapping or a reference MT phantom) is used. Consequently, the reported nRMSE/SSIM and the 'strong agreement with reference values' measure agreement with the BTS parameterization rather than true tissue properties. I request an independent validation of at least the phantom (known values or an independent reference method) and a discussion of BTS model validity.
  2. [Section 3.4.1 and Table 1] The comparison with RELAX and MANTIS is not a fair test of state-of-the-art performance. RELAX and MANTIS are trained on four subjects and tested on the remaining one via leave-one-out, while REFINE-MORE is scan-specific and optimized on the test subject's own data; the paper itself attributes the baselines' poor performance to 'limited training data' (Section 4.1). With only five subjects, this design does not support the claim that REFINE-MORE is superior to state-of-the-art methods. Please retrain the baselines on a larger dataset or report externally pretrained models, and/or reframe the claim as feasibility under data-limited conditions.
  3. [Section 3.3 and 4.4] The hyperparameters lambda (Eq. 4) and the number of unrolled phases K were selected by ablations performed on the same five-subject dataset that produced the headline results in Table 1, with no held-out split. This selection on the evaluation set can make the reported nRMSE/SSIM optimistically biased. Because the dataset is small, please report an internal cross-validation for these choices or state clearly that the reported numbers are the result of tuning on the evaluation set.
minor comments (6)
  1. [Section 5.3 vs Section 2.4] The claim that REFINE-MORE 'requires no external training datasets beyond the target scan itself' is inconsistent with the model adaptation strategy, which transfers and freezes weights from previously reconstructed subjects; subsequent subjects are not strictly scan-specific in their initialization.
  2. [Eq. (9) and Section 3.3] The notation 'CNN(q_i_hat^k)' in Eq. (9) is ambiguous: Section 3.3 states that the U-Net takes N=4 quantitative maps as input channels, so please clarify whether the CNN operates jointly on all maps or separately on each map.
  3. [Section 3.3] The hash encoding hyperparameters are described as 'fixed as the default values'; for reproducibility, please report N_min, b, L, T, and F explicitly.
  4. [Section 4.3] The timing comparison ('with adaptation: (6.6+7.6) min vs without adaptation: (13.3+58.5) min') should clarify whether the time for the first unadapted subject is amortized, since the fivefold speedup applies to subsequent subjects only.
  5. [Abstract and Section 4.1] The statement that REFINE-MORE achieves the highest SSIM 'across various acceleration factors and quantitative parameter maps' should be qualified by the kF map at AF=4 exception in Table 1, as the text already notes.
  6. [Table 1] The bold-face indication of the best performer should be checked for the kF SSIM entry at AF=4, where SUMMIT outperforms REFINE-MORE.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: REFINE-MORE's reconstruction is trained against undersampled k-space alone, and the fully sampled reference is an independent, data-derived comparison target.

full rationale

The derivation chain is self-contained. REFINE-MORE optimizes scan-specific INR and unrolled proximal-gradient parameters against undersampled k-space through losses L1 (Eq. 4), L2 (Eq. 6), and L3 (Eq. 9), none of which uses the fully sampled reference maps as training input. The reference is produced independently by nonlinear voxel-wise fitting of fully sampled k-space (Section 3.4.1), so the reported nRMSE and SSIM compare reconstruction outputs against an external data-derived target rather than against the method's own fitted values. The BTS forward model S (Eq. 11) is shared by the reconstruction losses and the reference fit; this makes the evaluation conditional on the validity of the BTS signal model, and the paper's statement in Section 5.2 that final outputs remain consistent 'with the true underlying tissue properties' is stronger than the experiment can establish. However, this is a modeling and validation limitation, not a circular reduction: the shared model does not by construction force the undersampled reconstruction to equal the fully sampled fit. Prior self-citations (RELAX, MANTIS, SUMMIT, MORE) appear as baselines or component precedents, not as unverified load-bearing justifications for the central claim. No step was found in which a prediction is equivalent to its input by definition.

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

The central claim rests on the BTS signal model, on fixed coil sensitivity estimation, and on a set of hyperparameters (lambda, K, LoRA rank/beta, iteration counts) selected either by ablation or default. The model adaptation strategy adds a paper-specific transfer assumption about temporal encoding stability. No new physical entities are introduced. The main cost is that the forward model and the reference share the same assumptions, so the evaluation is internal to the BTS model.

free parameters (4)
  • regularization weight lambda = 0.01
    Total variation weight in Eq. 4; studied in Fig. 9 and fixed at 0.01 across experiments, effectively tuned on the same evaluation data (Sec. 3.3, 4.4).
  • number of unrolled phases K = 4
    Chosen after ablation (Fig. 8) showing diminishing returns beyond 3 phases; selected on the same in vivo data used for evaluation (Sec. 3.3, 4.4).
  • LoRA rank r and scaling factor beta = not reported
    Rank r and beta in Eq. 10 control the low-rank adaptation of the U-Net. The paper reports resulting parameter counts (7,763,879 to 517,600) but not the exact rank or scaling factor (Sec. 2.4).
  • training iteration count for each stage = 1500 + 1500
    INR is trained for 1500 iterations, then both modules jointly for 1500 more (Sec. 3.3). This schedule is a design choice not varied in the paper.
assumptions (5)
  • domain assumption BTS binary spin-bath magnetization transfer signal model (Eq. 11) accurately describes the measured signal.
    The forward model S is used both to generate the reference via nonlinear fitting and inside REFINE-MORE's data consistency losses (Eqs. 6 and 9). Its accuracy is inherited from ref 53, not revalidated here (Sec. 3.1).
  • domain assumption Coil sensitivity maps estimated by ESPIRiT from a 24x24 center k-space region are correct and fixed.
    The forward operator U F C S assumes known sensitivities (Eq. 1); errors in C propagate into all data consistency terms (Sec. 3.4.1).
  • standard math The neural networks (hash-encoded MLPs and U-Net) have sufficient capacity and trainability to represent the 4D image series and the implicit proximal operator.
    Relies on universal approximation arguments and on the empirical success of hash encoding INR (refs 36-41, 67); no convergence guarantee is provided for the unrolled optimization.
  • ad hoc to paper The temporal hash encoding H1 and MLPs learned from earlier subjects transfer to new subjects under the same BTS sequence.
    This is the paper's own hypothesis in Sec. 2.4, supported by one demonstration on 5 healthy volunteers (Fig. 6A), not by a multi-site or pathology evaluation.
  • domain assumption Total variation and CNN-based proximal regularization are sufficient to disambiguate aliasing artifacts from true signal in the self-supervised setting.
    The method enforces data consistency only against the undersampled measurements; the regularization is what prevents trivial overfitting. The paper provides empirical but not theoretical support (Sec. 2.2, 2.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accelerating multiparametric quantitative MRI using self-supervised scan-specific implicit neural representation with model reinforcement." pith.science (2026). https://pith.science/paper/4ZGE3EIH

@misc{pith2026250800891,
  author       = {Pith},
  title        = {Pith review of: Accelerating multiparametric quantitative MRI using self-supervised scan-specific implicit neural representation with model reinforcement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ZGE3EIH}},
  note         = {Machine review of arXiv:2508.00891}
}
read the original abstract

Purpose: To develop a self-supervised scan-specific deep learning framework for reconstructing accelerated multiparametric quantitative MRI (qMRI). Methods: We propose REFINE-MORE (REference-Free Implicit NEural representation with MOdel REinforcement), combining an implicit neural representation (INR) architecture with a model reinforcement module that incorporates MR physics constraints. The INR component enables informative learning of spatiotemporal correlations to initialize multiparametric quantitative maps, which are then further refined through an unrolled optimization scheme enforcing data consistency. To improve computational efficiency, REFINE-MORE integrates a low-rank adaptation strategy that promotes rapid model convergence. We evaluated REFINE-MORE on accelerated multiparametric quantitative magnetization transfer imaging for simultaneous estimation of free water spin-lattice relaxation, tissue macromolecular proton fraction, and magnetization exchange rate, using both phantom and in vivo brain data. Results: Under 4x and 5x accelerations on in vivo data, REFINE-MORE achieved superior reconstruction quality, demonstrating the lowest normalized root-mean-square error and highest structural similarity index compared to baseline methods and other state-of-the-art model-based and deep learning approaches. Phantom experiments further showed strong agreement with reference values, underscoring the robustness and generalizability of the proposed framework. Additionally, the model adaptation strategy improved reconstruction efficiency by approximately fivefold. Conclusion: REFINE-MORE enables accurate and efficient scan-specific multiparametric qMRI reconstruction, providing a flexible solution for high-dimensional, accelerated qMRI applications.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

73 extracted references · 43 canonical work pages

  1. [1]

    Quantitative magnetic resonance imaging of brain anatomy and in vivo histology

    Weiskopf N, Edwards LJ, Helms G, Mohammadi S, Kirilina E. Quantitative magnetic resonance imaging of brain anatomy and in vivo histology. Nat Rev Phys. 2021;3(8):570-

    work page 2021

  2. [2]

    Primary Multiparametric Quantitative Brain MRI: State-of- the-Art Relaxometric and Proton Density Mapping Techniques

    Jara H, Sakai O, Farrher E, et al. Primary Multiparametric Quantitative Brain MRI: State-of- the-Art Relaxometric and Proton Density Mapping Techniques. Radiology. 2022;305(1):5-

    work page 2022

  3. [3]

    Multiparametric Quantitative MRI in Neurological Diseases

    Seiler A, Nö th U, Hok P, et al. Multiparametric Quantitative MRI in Neurological Diseases. Front Neurol. 2021;12:640239. doi:10.3389/fneur.2021.640239

    work page arXiv 2021

  4. [4]

    Generalized autocalibrating partially parallel acquisitions (GRAPPA)

    Griswold MA, Jakob PM, Heidemann RM, et al. Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn Reson Med. 2002;47(6):1202-1210. doi:10.1002/mrm.10171

    doi:10.1002/mrm.10171 2002

  5. [5]

    SENSE: sensitivity encoding for fast MRI

    Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE: sensitivity encoding for fast MRI. Magn Reson Med. 1999;42(5):952-962

    work page 1999

  6. [6]

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

    Lustig M, Donoho D, Pauly JM. Sparse MRI: The application of compressed sensing for rapid MR imaging. Magn Reson Med. 2007;58(6):1182-1195. doi:10.1002/mrm.21391

    doi:10.1002/mrm.21391 2007

  7. [7]

    SPIRiT: Iterative Self-consistent Parallel Imaging Reconstruction from Arbitrary k-Space

    Lustig M, Pauly JM. SPIRiT: Iterative Self-consistent Parallel Imaging Reconstruction from Arbitrary k-Space. Magn Reson Med. 2010;64(2):457-471. doi:10.1002/mrm.22428

    doi:10.1002/mrm.22428 2010

  8. [8]

    Compressed sensing reconstruction for magnetic resonance parameter mapping

    Doneva M, Bö rnert P, Eggers H, Stehning C, Sé né gas J, Mertins A. Compressed sensing reconstruction for magnetic resonance parameter mapping. Magn Reson Med. 2010;64(4):1114-1120. doi:10.1002/mrm.22483

    work page doi:10.1002/mrm.22483 2010

Show all 73 references
  1. [9]

    T2 mapping from highly undersampled data by reconstruction of principal component coefficient maps using compressed sensing

    Huang C, Graff CG, Clarkson EW, Bilgin A, Altbach MI. T2 mapping from highly undersampled data by reconstruction of principal component coefficient maps using compressed sensing. Magn Reson Med. 2012;67(5):1355-1366. doi:10.1002/mrm.23128

    2012 doi

  2. [10]

    Fast MR parameter mapping using k-t principal component analysis

    Petzschner FH, Ponce IP, Blaimer M, Jakob PM, Breuer FA. Fast MR parameter mapping using k-t principal component analysis. Magn Reson Med. 2011;66(3):706-716. doi:10.1002/mrm.22826

    2011 doi

  3. [11]

    Accelerating MR parameter mapping using sparsity-promoting regularization in parametric dimension

    Velikina JV, Alexander AL, Samsonov A. Accelerating MR parameter mapping using sparsity-promoting regularization in parametric dimension. Magn Reson Med. 2013;70(5):1263-1273. doi:10.1002/mrm.24577

    2013 doi

  4. [12]

    Model-based iterative reconstruction for radial fast spin-echo MRI

    Block KT, Uecker M, Frahm J. Model-based iterative reconstruction for radial fast spin-echo MRI. IEEE Trans Med Imaging. 2009;28(11):1759-1769. doi:10.1109/TMI.2009.2023119

    2009

  5. [13]

    Accelerating parameter mapping with a locally low rank constraint

    Zhang T, Pauly JM, Levesque IR. Accelerating parameter mapping with a locally low rank constraint. Magn Reson Med. 2015;73(2):655-661. doi:10.1002/mrm.25161 21

    2015 doi

  6. [14]

    Accelerated High-Dimensional MR Imaging With Sparse Sampling Using Low-Rank Tensors

    He J, Liu Q, Christodoulou AG, Ma C, Lam F, Liang ZP. Accelerated High-Dimensional MR Imaging With Sparse Sampling Using Low-Rank Tensors. IEEE Trans Med Imaging. 2016;35(9):2119-2129. doi:10.1109/TMI.2016.2550204

    2016

  7. [15]

    Three‐dimensional simultaneous brain mapping of T1, T2, and magnetic susceptibility with MR Multitasking

    Cao T, Ma S, Wang N, et al. Three‐dimensional simultaneous brain mapping of T1, T2, and magnetic susceptibility with MR Multitasking. Magn Reson Med. 2022;87(3):1375-1389. doi:10.1002/mrm.29059

    2022 doi

  8. [16]

    Echo planar time‐resolved imaging with subspace reconstruction and optimized spatiotemporal encoding

    Dong Z, Wang F, Reese TG, Bilgic B, Setsompop K. Echo planar time‐resolved imaging with subspace reconstruction and optimized spatiotemporal encoding. Magn Reson Med. 2020;84(5):2442-2455. doi:10.1002/mrm.28295

    2020 doi

  9. [17]

    Accelerated MR Parameter Mapping with Low-Rank and Sparsity Constraints

    Zhao B, Lu W, Hitchens TK, Lam F, Ho C, Liang ZP. Accelerated MR Parameter Mapping with Low-Rank and Sparsity Constraints. Magn Reson Med. 2015;74(2):489-498. doi:10.1002/mrm.25421

    2015 doi

  10. [18]

    doi:10.1148/radiol.211519

    doi

  11. [19]

    Deep model-based magnetic resonance parameter mapping network (DOPAMINE) for fast T1 mapping using variable flip angle method

    Jun Y, Shin H, Eo T, Kim T, Hwang D. Deep model-based magnetic resonance parameter mapping network (DOPAMINE) for fast T1 mapping using variable flip angle method. Med Image Anal. 2021;70:102017. doi:10.1016/j.media.2021.102017

    2021

  12. [20]

    Magnetic resonance fingerprinting

    Ma D, Gulani V, Seiberlich N, et al. Magnetic resonance fingerprinting. Nature. 2013;495(7440):187-192. doi:10.1038/nature11971

    2013 doi

  13. [21]

    Single-shot T2 mapping using overlapping-echo detachment planar imaging and a deep convolutional neural network

    Cai C, Wang C, Zeng Y, et al. Single-shot T2 mapping using overlapping-echo detachment planar imaging and a deep convolutional neural network. Magn Reson Med. 2018;80(5):2202-2214. doi:10.1002/mrm.27205

    2018 doi

  14. [22]

    MANTIS: Model-Augmented Neural neTwork with Incoherent k-space Sampling for efficient MR parameter mapping

    Liu F, Feng L, Kijowski R. MANTIS: Model-Augmented Neural neTwork with Incoherent k-space Sampling for efficient MR parameter mapping. Magn Reson Med. 2019;82(1):174-

    2019

  15. [23]

    A simultaneous multi-slice T2 mapping framework based on overlapping-echo detachment planar imaging and deep learning reconstruction

    Li S, Wu J, Ma L, Cai S, Cai C. A simultaneous multi-slice T2 mapping framework based on overlapping-echo detachment planar imaging and deep learning reconstruction. Magn Reson Med. 2022;87(5):2239-2253. doi:10.1002/mrm.29128

    2022 doi

  16. [24]

    MOdel-Based SyntheTic Data-Driven Learning (MOST-DL): Application in Single-Shot T2 Mapping With Severe Head Motion Using Overlapping-Echo Acquisition

    Yang Q, Lin Y, Wang J, et al. MOdel-Based SyntheTic Data-Driven Learning (MOST-DL): Application in Single-Shot T2 Mapping With Severe Head Motion Using Overlapping-Echo Acquisition. IEEE Trans Med Imaging. 2022;41(11):3167-3181. doi:10.1109/TMI.2022.3179981

    2022

  17. [25]

    Robust Single-Shot T2 Mapping via Multiple Overlapping- Echo Acquisition and Deep Neural Network

    Zhang J, Wu J, Chen S, et al. Robust Single-Shot T2 Mapping via Multiple Overlapping- Echo Acquisition and Deep Neural Network. IEEE Trans Med Imaging. 2019;38(8):1801-

    2019

  18. [26]

    qMTNet: Accelerated quantitative 22 magnetization transfer imaging with artificial neural networks

    Luu HM, Kim DH, Kim JW, Choi SH, Park SH. qMTNet: Accelerated quantitative 22 magnetization transfer imaging with artificial neural networks. Magn Reson Med. 2021;85(1):298-308. doi:10.1002/mrm.28411

    2021 doi

  19. [27]

    One for multiple: Physics-informed synthetic data boosts generalizable deep learning for fast MRI reconstruction

    Wang Z, Yu X, Wang C, et al. One for multiple: Physics-informed synthetic data boosts generalizable deep learning for fast MRI reconstruction. Med Image Anal. 2025;103:103616. doi:10.1016/j.media.2025.103616

    2025

  20. [28]

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

    Yaman B, Hosseini SAH, Moeller S, Ellermann J, Uğurbil K, Akçakaya M. Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data. Magn Reson Med. 2020;84(6):3172-3191. doi:10.1002/mrm.28378

    2020 doi

  21. [29]

    MR fingerprinting Deep RecOnstruction NEtwork (DRONE)

    Cohen O, Zhu B, Rosen MS. MR fingerprinting Deep RecOnstruction NEtwork (DRONE). Magn Reson Med. 2018;80(3):885-894. doi:10.1002/mrm.27198

    2018 doi

  22. [30]

    SSL-QALAS: Self-Supervised Learning for rapid multiparameter estimation in quantitative MRI using 3D-QALAS

    Jun Y, Cho J, Wang X, et al. SSL-QALAS: Self-Supervised Learning for rapid multiparameter estimation in quantitative MRI using 3D-QALAS. Magn Reson Med. 2023;90(5):2019-2032. doi:10.1002/mrm.29786

    2023 doi

  23. [31]

    Zero-DeepSub: Zero-shot deep subspace reconstruction for rapid multiparametric quantitative MRI using 3D-QALAS

    Jun Y, Arefeen Y, Cho J, et al. Zero-DeepSub: Zero-shot deep subspace reconstruction for rapid multiparametric quantitative MRI using 3D-QALAS. Magn Reson Med. 2024;91(6):2459-2482. doi:10.1002/mrm.30018

    2024 doi

  24. [32]

    Joint MAPLE: Accelerated joint T1 and T2* mapping with scan‐specific self‐supervised networks

    Heydari A, Ahmadi A, Kim TH, Bilgic B. Joint MAPLE: Accelerated joint T1 and T2* mapping with scan‐specific self‐supervised networks. Magn Reson Med. 2024;91(6):2294-

    2024

  25. [33]

    Magnetic resonance parameter mapping using model-guided self-supervised deep learning

    Liu F, Kijowski R, El Fakhri G, Feng L. Magnetic resonance parameter mapping using model-guided self-supervised deep learning. Magn Reson Med. 2021;85(6):3211-3226. doi:10.1002/mrm.28659

    2021 doi

  26. [34]

    Coordinate-based neural representation enabling zero-shot learning for fast 3D multiparametric quantitative MRI

    Lao G, Feng R, Qi H, et al. Coordinate-based neural representation enabling zero-shot learning for fast 3D multiparametric quantitative MRI. Med Image Anal. 2025;102:103530. doi:10.1016/j.media.2025.103530

    2025

  27. [35]

    High-Dimensional MR Reconstruction Integrating Subspace and Adaptive Generative Models

    Zhao R, Peng X, Kelkar VA, Anastasio MA, Lam F. High-Dimensional MR Reconstruction Integrating Subspace and Adaptive Generative Models. IEEE Trans Biomed Eng. 2024;71(6):1969-1979. doi:10.1109/TBME.2024.3358223

    2024

  28. [36]

    NeRF: representing scenes as neural radiance fields for view synthesis

    Mildenhall B, Srinivasan PP, Tancik M, Barron JT, Ramamoorthi R, Ng R. NeRF: representing scenes as neural radiance fields for view synthesis. Commun ACM. 2021;65(1):99-106. doi:10.1145/3503250

    2021 doi

  29. [37]

    NeRF-SR: High Quality Neural Radiance Fields using Supersampling

    Wang C, Wu X, Guo YC, Zhang SH, Tai YW, Hu SM. NeRF-SR: High Quality Neural Radiance Fields using Supersampling. In: Proceedings of the 30th ACM International Conference on Multimedia. MM ’22. New York, NY, USA: Association for Computing Machinery; 2022:6445-6454. doi:10.1145/...

    2022

  30. [38]

    Improving quantitative MRI using self-supervised deep learning with model reinforcement: Demonstration for rapid T1 mapping

    Bian W, Jang A, Liu F. Improving quantitative MRI using self-supervised deep learning with model reinforcement: Demonstration for rapid T1 mapping. Magn Reson Med. 2024;92(1):98-111. doi:10.1002/mrm.30045

    2024 doi

  31. [40]

    NeX: Real-time View Synthesis with Neural Basis Expansion

    Wizadwongsa S, Phongthawee P, Yenphraphai J, Suwajanakorn S. NeX: Real-time View Synthesis with Neural Basis Expansion. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, TN, USA: IEEE; 2021:8530-8539. doi:10.1109/CVPR46437.2021.00843

    2021

  32. [42]

    Dynamic CT Reconstruction from Limited Views with Implicit Neural Representations and Parametric Motion Fields

    Reed AW, Kim H, Anirudh R, et al. Dynamic CT Reconstruction from Limited Views with Implicit Neural Representations and Parametric Motion Fields. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, QC, Canada: IEEE; 2021:2238-2248. doi:10.1109/ICCV4...

    2021

  33. [43]

    Recovery of continuous 3D refractive index maps from discrete intensity-only measurements using neural fields

    Liu R, Sun Y, Zhu J, Tian L, Kamilov US. Recovery of continuous 3D refractive index maps from discrete intensity-only measurements using neural fields. Nat Mach Intell. 2022;4(9):781-791. doi:10.1038/s42256-022-00530-3

    2022 doi

  34. [44]

    CoIL: Coordinate-Based Internal Learning for Tomographic Imaging

    Sun Y, Liu J, Xie M, Wohlberg B, Kamilov US. CoIL: Coordinate-Based Internal Learning for Tomographic Imaging. IEEE Trans Comput Imaging. 2021;7:1400-1412. doi:10.1109/TCI.2021.3125564

    2021

  35. [45]

    L. Shen, J. Pauly, L. Xing. NeRP: Implicit Neural Representation Learning With Prior Embedding for Sparsely Sampled Image Reconstruction. IEEE Trans Neural Netw Learn Syst. 2024;35(1):770-782. doi:10.1109/TNNLS.2022.3177134

    2024

  36. [46]

    R. Feng, Q. Wu, J. Feng, et al. IMJENSE: Scan-Specific Implicit Representation for Joint Coil Sensitivity and Image Estimation in Parallel MRI. IEEE Trans Med Imaging. 2024;43(4):1539-1553. doi:10.1109/TMI.2023.3342156

    2024

  37. [47]

    Spatiotemporal Implicit Neural Representation for Unsupervised Dynamic MRI Reconstruction

    Feng J, Feng R, Wu Q, et al. Spatiotemporal Implicit Neural Representation for Unsupervised Dynamic MRI Reconstruction. IEEE Trans Med Imaging. 2025;44(5):2143-

    2025

  38. [48]

    Self-Supervised Coordinate Projection Network for Sparse-View Computed Tomography

    Wu Q, Feng R, Wei H, Yu J, Zhang Y. Self-Supervised Coordinate Projection Network for Sparse-View Computed Tomography. IEEE Trans Comput Imaging. 2023;9:517-529. doi:10.1109/TCI.2023.3281196

    2023

  39. [49]

    IMJ-PLUS: Implicit Representation for Dynamic 24 MRI and Coil Sensitivity Joint Reconstruction Using Low-Rank PLUS Sparse Regularization

    Shen X, Feng J, Feng R, Zhang Y, Wei H. IMJ-PLUS: Implicit Representation for Dynamic 24 MRI and Coil Sensitivity Joint Reconstruction Using Low-Rank PLUS Sparse Regularization. In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI). Houston; 2025:1-5. doi:10....

    2025

  40. [50]

    Neural Implicit k-Space for Binning-Free Non-Cartesian Cardiac MR Imaging

    Huang W, Li HB, Pan J, Cruz G, Rueckert D, Hammernik K. Neural Implicit k-Space for Binning-Free Non-Cartesian Cardiac MR Imaging. In: Information Processing in Medical Imaging (IPMI). Cham: Springer Nature Switzerland; 2023:548-560. doi:10.1007/978-3-031-34048-2_42

    2023 doi

  41. [51]

    Implicit Neural Networks With Fourier-Feature Inputs for Free-Breathing Cardiac MRI Reconstruction

    Kunz JF, Ruschke S, Heckel R. Implicit Neural Networks With Fourier-Feature Inputs for Free-Breathing Cardiac MRI Reconstruction. IEEE Trans Comput Imaging. 2024;10:1280-

    2024

  42. [52]

    Unsupervised reconstruction of accelerated cardiac cine MRI using neural fields

    Catalá n T, Courdurier M, Osses A, et al. Unsupervised reconstruction of accelerated cardiac cine MRI using neural fields. Comput Biol Med. 2025;185:109467. doi:10.1016/j.compbiomed.2024.109467

    2025

  43. [53]

    B1 inhomogeneity-corrected T1 mapping and quantitative magnetization transfer imaging via simultaneously estimating Bloch-Siegert shift and magnetization transfer effects

    Jang A, Han PK, Ma C, et al. B1 inhomogeneity-corrected T1 mapping and quantitative magnetization transfer imaging via simultaneously estimating Bloch-Siegert shift and magnetization transfer effects. Magn Reson Med. 2023;90(5):1859-1873. doi:10.1002/mrm.29778

    2023 doi

  44. [54]

    3D cine-magnetic resonance imaging using spatial and temporal implicit neural representation learning (STINR-MR)

    Shao HC, Mengke T, Deng J, Zhang Y. 3D cine-magnetic resonance imaging using spatial and temporal implicit neural representation learning (STINR-MR). Phys Med Biol. 2024;69(9):095007. doi:10.1088/1361-6560/ad33b7

    2024 doi

  45. [55]

    Recurrent Generative Adversarial Networks for Proximal Learning and Automated Compressive Image Recovery

    Mardani M, Monajemi H, Papyan V, Vasanawala S, Donoho D, Pauly J. Recurrent Generative Adversarial Networks for Proximal Learning and Automated Compressive Image Recovery. November 2017. http://arxiv.org/abs/1711.10046. Accessed March 22, 2024

    2017 arXiv

  46. [56]

    Learned Proximal Networks for Quantitative Susceptibility Mapping

    Lai KW, Aggarwal M, van Zijl P, Li X, Sulam J. Learned Proximal Networks for Quantitative Susceptibility Mapping. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). Berlin, Heidelberg: Springer-Verlag; 2020:125-135. doi:10.1007/978-3-030-59713-9_13

    2020 doi

  47. [57]

    MoDL-QSM: Model-based deep learning for quantitative susceptibility mapping

    Feng R, Zhao J, Wang H, et al. MoDL-QSM: Model-based deep learning for quantitative susceptibility mapping. NeuroImage. 2021;240:118376. doi:10.1016/j.neuroimage.2021.118376

    2021

  48. [58]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, eds. Medical Image Computing and Computer-Assisted Intervention (MICCAI). Cham: Springer International Publishing; 2015:234-241. do...

    2015 doi

  49. [59]

    LoRA: Low-Rank Adaptation of Large Language Models

    Hu EJ, Shen Y, Wallis P, et al. LoRA: Low-Rank Adaptation of Large Language Models. October 2021. doi:10.48550/arXiv.2106.09685

    2021 doi arXiv

  50. [60]

    Quantitative magnetization transfer imaging in postmortem multiple sclerosis brain

    Schmierer K, Tozer DJ, Scaravilli F, et al. Quantitative magnetization transfer imaging in postmortem multiple sclerosis brain. J Magn Reson Imaging. 2007;26(1):41-51. doi:10.1002/jmri.20984 25

    2007 doi

  51. [61]

    Proximal algorithms

    Parikh N, Boyd S. Proximal algorithms. Found Trends Optim. 2013;1(3):123-231

    2013

  52. [62]

    Quantitative interpretation of magnetization transfer

    Henkelman RM, Huang X, Xiang QS, Stanisz GJ, Swanson SD, Bronskill MJ. Quantitative interpretation of magnetization transfer. Magn Reson Med. 1993;29(6):759-766. doi:10.1002/mrm.1910290607

    1993 doi

  53. [63]

    Magnetization transfer effects on T1-weighted three-dimensional gradient-echo MR images of a phantom simulating enhancing brain lesions

    Finelli DA. Magnetization transfer effects on T1-weighted three-dimensional gradient-echo MR images of a phantom simulating enhancing brain lesions. AJNR Am J Neuroradiol. 1997;18(1):147-159

    1997

  54. [64]

    Adam: A Method for Stochastic Optimization

    Kingma DP, Ba J. Adam: A Method for Stochastic Optimization. January 2017. doi:10.48550/arXiv.1412.6980

    2017 doi arXiv

  55. [65]

    ESPIRiT — An Eigenvalue Approach to Autocalibrating Parallel MRI: Where SENSE meets GRAPPA

    Uecker M, Lai P, Murphy MJ, et al. ESPIRiT — An Eigenvalue Approach to Autocalibrating Parallel MRI: Where SENSE meets GRAPPA. Magn Reson Med. 2014;71(3):990-1001. doi:10.1002/mrm.24751

    2014 doi

  56. [66]

    Statistical methods for assessing agreement between two methods of clinical measurement

    Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet. 1986;1(8476):307-310

    1986

  57. [67]

    Multilayer feedforward networks are universal approximators

    Hornik K, Stinchcombe M, White H. Multilayer feedforward networks are universal approximators. Neural Networks. 1989;2(5):359-366. doi:10.1016/0893-6080(89)90020-8 26 LIST OF TABLES Table 1 . Comparison of nRMSE and SSIM values relative to the fully sampled reference for diffe...

    1989 doi

  58. [68]

    Magnetization transfer in cartilage and its constituent macromolecules

    Gray ML, Burstein D, Lesperance LM, Gehrke L. Magnetization transfer in cartilage and its constituent macromolecules. Magn Reson Med. 1995;34(3):319-325. doi:10.1002/mrm.1910340307

    1995 doi

  59. [75]

    Instant Neural Graphics Primitives with a Multiresolution Hash Encoding

    Mü ller T, Evans A, Schied C, Keller A. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans Graph. 2022;41(4):1 -15. doi:10.1145/3528223.3530127

    2022

  60. [188]

    doi:10.1002/mrm.27707

    doi

  61. [588]

    doi:10.1038/s42254-021-00326-1

    doi

  62. [1289]

    doi:10.1109/TCI.2024.3452008

    2024

  63. [1811]

    doi:10.1109/TMI.2019.2896085

    2019

  64. [2156]

    doi:10.1109/TMI.2025.3526452

    2025

  65. [2309]

    doi:10.1002/mrm.29989

    doi

Pith tools

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