Diffusion-Assisted Frequency Attention Model for Whole-body Low-field MRI Reconstruction

Dong Liang; Qiegen Liu; Xin Xie; Yu Guan; Zhuoxu Cui

REVIEW 4 major objections 6 minor 31 references

Diffusion-Assisted Frequency Attention Model for Whole-body Low-field MRI Reconstruction

T0 review · 4 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 This paper argues that DFAM, a diffusion-assisted frequency attention model whose diffusion prior is compressed into a one-dimensional feature vector, reconstructs whole-body low-field MRI more accurately and much faster than conventional…

desk verdict A credible incremental method for low-field MRI reconstruction whose efficiency story is real, but whose 'consistently outperforms' claim is contradicted by its own Table S1 and needs qualifying before this is publishable. read the letter →

arxiv 2507.17764 v1 pith:4XK4DFMV submitted 2025-07-09 physics.med-ph cs.CV

Xin Xie , Yu Guan , Zhuoxu Cui , Dong Liang , Qiegen Liu This is my paper · ORCID

classification physics.med-phcs.CV PACS 87.61.-c

keywords low-fieldMRIreconstructiondiffusionmodelfrequencyattentionwavelettransformwhole-bodyimagingunder-samplingefficiency-qualityratio

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 sets out to establish that a hybrid reconstruction model, pairing a diffusion model with wavelet-based frequency attention, can reconstruct whole-body low-field MRI from heavily under-sampled k-space data both more accurately and much faster than existing alternatives. The authors' argument is that importing high-field reconstruction methods into low-field imaging fails because signal-to-noise ratio is intrinsically low and relaxation contrasts shift, so the prior should be learned in a compact one-dimensional feature space while attention operates in the frequency domain. If the claim holds, low-field scanners in resource-constrained and underdeveloped settings could produce usable clinical images with reconstruction times on the order of seconds, removing a major obstacle to broader adoption.

What carries the argument

The machinery has two coupled parts. A squeeze-and-excitation feature extraction (SEFE) module fuses the ground-truth and low-quality images, uses channel excitation and convolutional feature extraction, and compresses the result into a compact one-dimensional feature vector $A$. A frequency attention (FA) module takes that vector as a guide and replaces standard spatial attention with wavelet-transform attention, separating low-frequency structure from high-frequency texture. Between them sits a vector diffusion model that adds Gaussian noise to $A$ during training and removes it during inference; because the diffusion target is a 1D vector rather than a 2D image, the reverse process runs in only four iterations. A data consistency step then forces the reconstructed image to agree with the measured k-space samples.

What would settle it

Retrain DFAM and the same comparison methods on paired data produced by under-sampling plus physically realistic low-field degradation, namely Rician noise at 0.3T signal levels, T1/T2 contrast shifts, and realistic coil sensitivity maps, and evaluate on real 0.3T scans; if the reported PSNR and SSIM margins over the baselines shrink to within run-to-run variance, the claim that the learned prior is robust to low-SNR conditions is unsupported.

Watch

Extended reading notes

Core claim

The central discovery is that a diffusion prior for MRI reconstruction can be compressed into a one-dimensional feature vector without losing the generative advantage: the SEFE module fuses the ground-truth and low-quality image into a compact vector $A$, and a vector diffusion model learns and reverses the distribution of $A$, so only four reverse iterations are needed. In parallel, the FA module embeds a wavelet transform into Transformer attention, using low-frequency components for overall structure and high-frequency components for textures and edges. The paper reports that this design consistently yields higher PSNR and SSIM, lower MSE, and faster runs than conventional parallel-imaging, energy-based, unrolled, score-based, and spatial-domain Transformer baselines across head, cervical spine, lumbar spine, knee, and abdomen slices at acceleration factors from 3 to 12.

Load-bearing premise

The load-bearing premise is stated in the dataset description: high-quality 3T images are under-sampled by factors of 3 to 12 to simulate low-field MRI data, with no added low-SNR noise, no shifted relaxation contrasts, and no coil-sensitivity effects; if real low-field images are not well approximated by that proxy, the claimed robustness under low-SNR acquisitions is not actually supported by the training setup.

Editorial extensions

If this is right

The method is presented as a whole-body tool rather than a single-anatomy solution, with experiments on head, cervical spine, abdomen, lumbar spine, and knee slices.
Reported results at acceleration factors up to 12x suggest the quality advantage holds as under-sampling becomes more aggressive.
Reconstruction times in seconds, from a four-iteration vector diffusion process plus a one-stage frequency attention module, make the approach practical for low-resource settings.
Frequency-domain attention is argued to suppress spatial-domain noise and preserve edges, which is the specific property low-field images need most.
The new efficiency-quality metric, EQRatio, jointly accounting for PSNR, SSIM, and log-transformed time, positions DFAM as better on the speed-accuracy frontier than iterative and generative baselines.

Reading between the lines

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

Beyond the paper: because the vector diffusion prior is decoupled from image space, the same SEFE-plus-frequency-attention recipe could be tried on other paired reconstruction tasks, such as multi-contrast MRI where the low- and high-quality pair differs by contrast weighting rather than by sampling.
Beyond the paper: the reported efficiency ranking is tied to the EQRatio's chosen weights and logarithmic time penalty, so a robustness check that varies those weights would show how stable the speed-accuracy conclusion is.
Beyond the paper: a natural next experiment is to inject realistic low-field noise and coil effects into the simulated training pairs; if the model keeps its margin under that stress, the frequency-domain mechanism is genuinely carrying the low-SNR claim, and if not, the training proxy is the bottleneck.

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.

Desk Editor's Note

read the letter

Two things to know about this paper: the method is a credible, incremental adaptation of DiffIR to low-field MRI, and it ships code plus tests on real 0.3T data from five body regions—better than many papers in this space. But the headline claim that DFAM 'consistently outperforms' all baselines is not true on the paper's own numbers, and the training proxy for low-field is weak.

What's actually new: the SEFE module compresses the image prior into a 1D feature vector, so the diffusion runs in a compressed latent and needs only four reverse iterations; the frequency-attention module uses wavelet-domain attention to guide reconstruction. Both components are known ideas, and DiffIR is the direct base, so this is a combination rather than a new paradigm. But the combination is sensible and the efficiency story is plausible. The ablations against DiffIR, SEFE-only, and FA-only are a genuine plus. Testing on real 0.3T scanner data across head, spine, abdomen, and knee is exactly the right kind of external validation for this subfield.

The soft spots, in proportion. First, the abstract and conclusion say 'consistently outperforms,' but Supporting Table S1 shows Restormer beating DFAM on SSIM at R=8 (0.7794 vs 0.7724). The Discussion acknowledges this and justifies it by speed, which is fine—but the abstract doesn't get to keep the unqualified claim. That is an internal inconsistency, and it is the load-bearing word. The evidence supports 'higher PSNR, comparable or slightly lower SSIM in one setting, faster.' Second, training simulates low-field by undersampling clean 3T images without added noise. The paper's own premise is low-SNR robustness, so the training distribution does not match the stated challenge; the real 0.3T test results help, but they are single slices without error bars or statistical testing. I would not call this disqualifying—the method still wins on most metrics—but it means the strong claims outrun the evidence. Third, the EQRatio uses hand-chosen weights (0.1/0.9) that favor fast methods; it is self-referential and should not carry the efficiency argument alone, even though the raw times in Table S1 are informative. Minor: Algorithm 1 tells you to update A_{t-1} by Eq. (14), which is the loss, not the sampling update (Eq. 15).

Bottom line: this is a serious incremental contribution that deserves referee time, not a desk reject. The fixes are honest ones—qualify the claims, add error bars, address the training proxy, fix the algorithm reference. I'd send it to review with the expectation of major revision.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes DFAM (Diffusion-Assisted Frequency Attention Model) for whole-body low-field MRI reconstruction. The method trains a frequency-attention Transformer (FA) to reconstruct images from undersampled low-quality data, guided by a diffusion model that operates on a one-dimensional feature vector produced by a squeeze-and-excitation feature extraction (SEFE) module; the diffusion prior is compressed to 1D so that only four reverse iterations are needed. Training uses paired high-quality and undersampled high-field images from the SIAT and TotalSegmentator datasets, and testing is performed on real 0.3 T low-field images from five body regions. The authors compare against P-LORAKS, ESPIRiT, EBMRec, MoDL, Score-MRI, HGGDP, SwinIR, and Restormer, and they introduce a composite EQRatio metric that combines PSNR/SSIM improvement with log reconstruction time. The claimed central result is that DFAM consistently outperforms conventional and learning-based baselines while being computationally efficient.

Significance. If the claims were fully supported, the paper would make a useful contribution: compressing the diffusion prior into a 1D vector to cut reverse sampling to four iterations is an interesting design that addresses the well-known inference cost of diffusion-based MRI reconstruction; the wavelet-based frequency attention is a plausible mechanism for preserving edges under low SNR; and testing on real 0.3 T data from multiple body parts gives the paper practical relevance. The authors also provide source code, which aids reproducibility. However, the strength of the current evidence is limited by single-slice quantitative evaluations without error bars, a training simulation that undersamples high-field images without modeling low-field SNR, and an unqualified 'consistently outperforms' claim that is contradicted by the authors' own Table S1 for SSIM at R=8. The significance of the paper would increase substantially if the quantitative comparisons were reported over multiple slices with statistical measures and if the central claim were appropriately qualified.

major comments (4)

[Abstract / Conclusion; Supporting Information Table S1] The abstract and conclusion state that DFAM 'consistently outperforms' all compared methods, but Supporting Information Table S1 shows that under radial sampling at R=8, Restormer achieves SSIM 0.7794 while DFAM achieves 0.7724, i.e., DFAM is 0.007 lower. The Discussion acknowledges this outcome, yet the unqualified claim is retained. Please qualify the central claim (e.g., 'often outperforms on PSNR with comparable SSIM') or define an explicit aggregate criterion, such as winning on a majority of metrics or passing a statistical significance test across slices, and apply it consistently to the abstract and conclusion.
[Results, Tables 1-3] The quantitative comparisons in Tables 1-3 report single-slice results: Table 1 uses one lumbar and one cervical slice, Table 2 uses two head slices, and Table 3 uses one knee and one abdominal slice. No standard deviations, confidence intervals, or statistical significance tests are provided. With a single slice per test condition, 'consistently outperforms' is not supported as a general statement; the authors should report summary statistics over multiple test slices or at minimum clearly state that the tables are illustrative single-slice examples.
[Description of Datasets] The training data are generated by undersampling high-quality 3T images (3x-12x) to 'simulate low-field MRI data,' with no stated addition of noise, no altered T1/T2 contrasts, and no coil-related noise. Because real low-field images have substantially lower SNR, this simulation does not reproduce the low-SNR statistics that the abstract claims DFAM is robust to. The authors should either add noise and relaxation-contrast simulation to the training pipeline or temper the low-SNR robustness claim to what is demonstrated by the real 0.3 T test data.
[Discussion B, Eq. (17)] The efficiency comparison relies on the proposed EQRatio metric with hand-selected weights omega_1=0.1 and omega_2=0.9 and a logarithmic time term. Because the weights are ad hoc and the metric is not validated against any external or clinical criterion, conclusions such as 'the comprehensive efficiency is better' are metric-dependent. At a minimum, the paper should report raw PSNR, SSIM, and runtime with error bars, and present the EQRatio as a secondary, sensitivity-checked metric rather than the basis for its efficiency claim.

minor comments (6)

[Algorithm 1] Algorithm 1 says 'Update A_{t-1} via Eq. (14)', but Eq. (14) is the joint loss function; the reverse-diffusion update is given by Eq. (15). Please correct the cross-reference.
[Eq. (16)] Equation (16) is typeset obscurely, so the two branches of the data-consistency update are hard to read. Please rewrite with explicit cases for j in Omega and j not in Omega.
[Introduction, Ref. [15]] The sentence citing Ref. [15] describes WKGM as 'trained a DM based on unsupervised scoring for MRI reconstruction,' but Ref. [15] is a weighted k-space generative model; this description needs clarification.
[Table 1, Test 2] The first row of Test 2 contains a stray slash in the MSE column ('11.91/') and an extra slash after '0.727'; please check all table entries for formatting consistency.
[Supporting Information Figure S1] Figure S1 is described as a bar chart without error bars; the ablation would be more informative with numeric values, error bars, or a corresponding table.
[Title and Datasets] The term 'whole-body' is used for a method tested on five body-part datasets; please either specify the exact body coverage or use a more precise term such as 'multi-region' in the title and abstract.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: DFAM's reconstruction accuracy is benchmarked against external methods with standard PSNR/SSIM/MSE, though the efficiency claim leans on an author-defined EQRatio metric.

full rationale

The central reconstruction claim is not circular. DFAM is compared against external baselines (P-LORAKS, ESPIRiT, EBMRec, MoDL, Score-MRI, HGGDP, SwinIR, Restormer) using standard PSNR, SSIM, and MSE, which are independent of the model's training targets and fitted parameters. The method's design—SEFE extracting a 1D feature vector, DM denoising that vector, and FA performing wavelet-domain attention—is a standard supervised generative pipeline; no reported equation reduces by construction to its own input. The paper's 'consistently outperforms' wording is internally inconsistent with Supporting Information Table S1, where Restormer achieves higher SSIM (0.7794) than DFAM (0.7724) at R=8, but that is an accuracy/consistency defect, not a circularity. The author-defined EQRatio uses hand-chosen weights (0.1 and 0.9) and a logarithmic time transform, which favors fast methods and makes the efficiency comparison partly self-referential; however, EQRatio is an evaluation metric rather than a training target or a predicted quantity, so it does not make the derivation circular. The training-set limitation (simulating low-field MRI only by under-sampling 3T images, without added noise or low-field contrast changes) is a generalization risk, not a circular step. Self-citations to the authors' earlier works appear as baselines or related methods, but none is load-bearing as a uniqueness theorem or as the sole justification for DFAM's architecture.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the standard MRI model and on the unvalidated proxy that undersampled 3T images emulate low-field acquisition. Efficiency claims rely on a new metric with hand-picked weights, and the small, non-public test set limits generality.

free parameters (3)

EQRatio weights = omega1=0.1, omega2=0.9
Hand-chosen weights in the new Efficiency-Quality Ratio metric used to demonstrate efficiency advantages; no validation or sensitivity analysis.
Diffusion reverse iterations = 4
The number of reverse denoising iterations is set to 4 and presented as a breakthrough, but it is a hyperparameter with no ablation.
Data consistency weight lambda = not stated
Eq. (16) includes a weight coefficient lambda balancing measurement noise; the paper does not specify its value or tuning.

assumptions (4)

domain assumption Standard linear MRI forward model y = Px + n with P = MFS
Used in Eqs. (1)-(3) as the basis for the inverse problem.
ad hoc to paper Undersampled high-field images are a faithful proxy for real low-field MRI data
Training simulates low-field by undersampling 3T images; ignores true low-field SNR, contrast and coil characteristics.
domain assumption Diffusion model can learn a useful prior over compressed one-dimensional feature vectors
The vector DM is trained on SEFE feature vectors; no theoretical guarantee is provided.
domain assumption Wavelet-based frequency attention improves reconstruction under low SNR
Central design premise of the FA module, supported only by ablations on selected slices.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Diffusion-Assisted Frequency Attention Model for Whole-body Low-field MRI Reconstruction." pith.science (2026). https://pith.science/paper/4XK4DFMV

@misc{pith2026250717764,
  author       = {Pith},
  title        = {Pith review of: Diffusion-Assisted Frequency Attention Model for Whole-body Low-field MRI Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4XK4DFMV}},
  note         = {Machine review of arXiv:2507.17764}
}

read the original abstract

By integrating the generative strengths of diffusion models with the representation capabilities of frequency-domain attention, DFAM effectively enhances reconstruction performance under low-SNR condi-tions. Experimental results demonstrate that DFAM consistently outperforms both conventional reconstruction algorithms and recent learning-based approaches. These findings highlight the potential of DFAM as a promising solution to advance low-field MRI reconstruction, particularly in resource-constrained or underdeveloped clinical settings.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 30 canonical work pages

[1]

LaPierre, Najat Salameh, David EJ Waddington, Thomas Witzel, and Matthew S

Sarracanie, Mathieu, Cristen D. LaPierre, Najat Salameh, David EJ Waddington, Thomas Witzel, and Matthew S. Rosen. Low-cost high-performance MRI. Sci. Rep. 2015; 5(1): 15177

work page 2015
[2]

and Temel, Y., Ultra high field MRI -guided deep brain stimulation

Forstmann, B.U., Isaacs, B.R. and Temel, Y., Ultra high field MRI -guided deep brain stimulation. Trends Bio- technol. 2017;35(10):904-907

work page 2017
[3]

and Tang, L., Insight into RF power requirements and B1 field homogeneity for human MRI via rigorous FDTD approach

Ibrahim, T.S. and Tang, L., Insight into RF power requirements and B1 field homogeneity for human MRI via rigorous FDTD approach. J. Magn. Reson. Imaging. 2007; 25(6):1235-1247

work page 2007
[4]

and Sarracanie, M., MRI at low field: A review of software solutions for improving SNR

Ayde, R., Vornehm, M., Zhao, Y., Knoll, F., Wu, E.X. and Sarracanie, M., MRI at low field: A review of software solutions for improving SNR. NMR Biomed. 2025; 38(1): e5268

work page 2025
[5]

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

Guan, Y., Yu, C., Cui, Z., Zhou, H. and Liu, Q., Correlated and multi-frequency diffusion modeling for highly under-sampled MRI reconstruction. IEEE Trans. Med. Imaging. 2024; 43(10): 3490-3502

work page 2024
[6]

Calibrationless parallel imaging reconstruction by structured low -rank matrix completion

Lustig, M., Elad, M., Pauly, J M. Calibrationless parallel imaging reconstruction by structured low -rank matrix completion. In Proc. 18th Annu. Meet. ISMRM, Stockholm, 2010: 2870

work page 2010
[7]

and Liu, Q., Partition-based K-space synthesis for multi-contrast parallel imaging

Huang, Y., Wu, Z., Xu, X., Zhang, M., Wang, S. and Liu, Q., Partition-based K-space synthesis for multi-contrast parallel imaging. Magn. Reson. Imaging. 2024: 110297

work page 2024
[9]

High -frequency space diffusion model for accelerated mri

Cao C, Cui Z X, Wang Y, et al. High -frequency space diffusion model for accelerated mri. IEEE Trans. Med. Imaging, 2024;43(5):1853-1865

work page 2024

Show all 31 references

[10]

IEEE Trans

Han, Yoseo, Leonard Sunwoo, Jong Chul Ye , K-space deep learning for accelerated MRI. IEEE Trans. Med. Imaging. 2019; 39(2): 377-386

2019
[11]

, Scan‐specific robust artificial‐neural‐networks for k‐space interpolation (RAKI) reconstruction: database‐free deep learning for fast imaging

Akçakaya, M., Moeller, S., Weingärtner, S., & Uğurbil, K. , Scan‐specific robust artificial‐neural‐networks for k‐space interpolation (RAKI) reconstruction: database‐free deep learning for fast imaging. Magn. Reson. Med . 2019; 81(1): 439-453

2019
[12]

B., Tumwiine, J., & Obungoloch, J

Ahishakiye, E., Van Gijzen, M. B., Tumwiine, J., & Obungoloch, J. Adaptive -size dictionary learning using in- formation theoretic criteria for image reconstruction from undersampled k -space data in low field magnetic res- onance imaging. BMC medical imaging, 2020; 20, 1-12

2020
[13]

and McNab, J.A

Dai, E., Mani, M. and McNab, J.A. Multi‐band multi‐shot diffusion MRI reconstruction with joint usage of structured low‐rank constraints and explicit phase map. Magn. Reson. Med. 2023; 89(1):95-111

2023
[14]

and Stein, J.M., Low‐field MRI: clinical promise and challenges

Arnold, T.C., Freeman, C.W., Litt, B. and Stein, J.M., Low‐field MRI: clinical promise and challenges. J. Magn. Reson. Imaging. 2023;57(1):25-44

2023
[15]

and Liu, Q., WKGM: weighted k‐space generative model for parallel imaging reconstruction

Tu, Z., Liu, D., Wang, X., Jiang, C., Zhu, P., Zhang, M., Wang, S., Liang, D. and Liu, Q., WKGM: weighted k‐space generative model for parallel imaging reconstruction. NMR Biomed. 2023;36(11):5005

2023
[16]

Self -supervised MRI reconstruction with unrolled diffusion models

Korkmaz, Y, Cukur, T, Patel, V M. Self -supervised MRI reconstruction with unrolled diffusion models . In MICCAI. Cham: Springer Nature Switzerland, 2023: 491-501

2023
[17]

DiffIR: Efficient diffusion model for image restoration

Zhang, B, Wang, S, Wang, Y, et al. DiffIR: Efficient diffusion model for image restoration. In Proc. IEEE/CVF Int. Conf. Comput. Vis, 2023: 13095-13105. 16

2023
[18]

and Lin, S., WaveletFormerNet: A Transformer -based wavelet network for real -world non-homogeneous and dense fog removal

Zhang, S., Tao, Z. and Lin, S., WaveletFormerNet: A Transformer -based wavelet network for real -world non-homogeneous and dense fog removal. Image Vis. Comput. 2024; 146:105014

2024
[19]

C., Freeman, C

Arnold, T. C., Freeman, C. W., Litt, B., & Stein, J. M. Low‐field MRI: clinical promise and challenges. Magn. Reson. Imaging, 2023; 57(1): 25-44

2023
[20]

M., Truong, M

Coffey, A. M., Truong, M. L., & Chekmenev, E. Y. Low -field MRI can be more sensitive than high -field MRI. Magn. Reson. Imaging, 2013; 237, 169-174

2013
[21]

B., Brinkmann, I

Khodarahmi, I., Keerthivasan, M. B., Brinkmann, I. M., Grodzki, D., & Fritz, J. Modern low -field MRI of the musculoskeletal system: practice considerations, opportunities, and challenges. Investigative radiology, 2023; 58(1), 76-87

2023
[22]

Totalsegmentator mri: Sequence -independent segmentation of 59 anatomical structures in mr images

D’Antonoli TA, Berger LK, Indrakanti AK, et al. Totalsegmentator mri: Sequence -independent segmentation of 59 anatomical structures in mr images. arXiv preprint arXiv:2405.19492. 2024

2024 arXiv
[23]

L., van Gijzen, M

de Leeuw den Bouter, M. L., van Gijzen, M. B., & Remis, R. F. Conjugate gradient variants for ℓ p -regularized image reconstruction in low-field MRI. SN Applied Sciences, 2019; 1(12), 1736

2019
[24]

and Zhuo, J., P‐LORAKS: low‐rank modeling of local k‐space neighborhoods with parallel imaging data

Haldar, J.P. and Zhuo, J., P‐LORAKS: low‐rank modeling of local k‐space neighborhoods with parallel imaging data. Magn. Reson. Med. 2016; 75(4):1499-1514

2016
[25]

and Lustig, M., ESPIRiT-an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA

Uecker, M., Lai, P., Murphy, M.J., Virtue, P., Elad, M., Pauly, J.M., Vasanawala, S.S. and Lustig, M., ESPIRiT-an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA. Magn. Reson. Med. 2014; 71(3):990-1001

2014
[26]

and Liang, D., Magnetic resonance imaging reconstruction using a deep energy‐based model

Guan, Y., Tu, Z., Wang, S., Wang, Y., Liu, Q. and Liang, D., Magnetic resonance imaging reconstruction using a deep energy‐based model. NMR Biomed. 2023; 36(3): e4848

2023
[27]

and Jacob, M., MoDL: Model-based deep learning architecture for inverse problems

Aggarwal, H.K., Mani, M.P. and Jacob, M., MoDL: Model-based deep learning architecture for inverse problems. IEEE Trans. Med. Imaging. 2018; 38(2):394-405

2018
[28]

and Ye, J.C., Score-based diffusion models for accelerated MRI

Chung, H. and Ye, J.C., Score-based diffusion models for accelerated MRI. Med. Image Anal. 2022; 80:102479

2022
[29]

and Liu, Q., Homotopic gradients of generative density priors for MR image reconstruction

Quan, C., Zhou, J., Zhu, Y., Chen, Y., Wang, S., Liang, D. and Liu, Q., Homotopic gradients of generative density priors for MR image reconstruction. IEEE Trans. Med. Imaging. 2021; 40(12):3265-3278

2021
[30]

Swinir: Image restoration using swin Transformer

Liang, J, Cao, J, Sun, G, Zhang, K, Van Gool, L, Timofte, R. Swinir: Image restoration using swin Transformer. In Proc. IEEE/CVF ICCV, 2021: 1833-1844

2021
[31]

Restormer: Efficient Transformer for high-resolution image restoration

Zamir, S W, Arora, A, Khan, S, Hayat, M, Khan, F S, Yang, M H. Restormer: Efficient Transformer for high-resolution image restoration. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit, 2022: 5728-5739

2022
[32]

Lustig, M

M. Lustig, M. Elad, and J. M. Pauly, Calibrationless parallel imaging reconstruction by structured low-rank matrix completion. In Annu. Meet. ISMRM, 2010: 2870. 17 Captions Fig. 1 . The training and reconstruction proces s of the DFAM. (a) Training of the FA module. (b) Traini...

2010

Pith tools

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

[1] [1]

LaPierre, Najat Salameh, David EJ Waddington, Thomas Witzel, and Matthew S

Sarracanie, Mathieu, Cristen D. LaPierre, Najat Salameh, David EJ Waddington, Thomas Witzel, and Matthew S. Rosen. Low-cost high-performance MRI. Sci. Rep. 2015; 5(1): 15177

work page 2015

[2] [2]

and Temel, Y., Ultra high field MRI -guided deep brain stimulation

Forstmann, B.U., Isaacs, B.R. and Temel, Y., Ultra high field MRI -guided deep brain stimulation. Trends Bio- technol. 2017;35(10):904-907

work page 2017

[3] [3]

and Tang, L., Insight into RF power requirements and B1 field homogeneity for human MRI via rigorous FDTD approach

Ibrahim, T.S. and Tang, L., Insight into RF power requirements and B1 field homogeneity for human MRI via rigorous FDTD approach. J. Magn. Reson. Imaging. 2007; 25(6):1235-1247

work page 2007

[4] [4]

and Sarracanie, M., MRI at low field: A review of software solutions for improving SNR

Ayde, R., Vornehm, M., Zhao, Y., Knoll, F., Wu, E.X. and Sarracanie, M., MRI at low field: A review of software solutions for improving SNR. NMR Biomed. 2025; 38(1): e5268

work page 2025

[5] [5]

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

Guan, Y., Yu, C., Cui, Z., Zhou, H. and Liu, Q., Correlated and multi-frequency diffusion modeling for highly under-sampled MRI reconstruction. IEEE Trans. Med. Imaging. 2024; 43(10): 3490-3502

work page 2024

[6] [6]

Calibrationless parallel imaging reconstruction by structured low -rank matrix completion

Lustig, M., Elad, M., Pauly, J M. Calibrationless parallel imaging reconstruction by structured low -rank matrix completion. In Proc. 18th Annu. Meet. ISMRM, Stockholm, 2010: 2870

work page 2010

[7] [7]

and Liu, Q., Partition-based K-space synthesis for multi-contrast parallel imaging

Huang, Y., Wu, Z., Xu, X., Zhang, M., Wang, S. and Liu, Q., Partition-based K-space synthesis for multi-contrast parallel imaging. Magn. Reson. Imaging. 2024: 110297

work page 2024

[8] [9]

High -frequency space diffusion model for accelerated mri

Cao C, Cui Z X, Wang Y, et al. High -frequency space diffusion model for accelerated mri. IEEE Trans. Med. Imaging, 2024;43(5):1853-1865

work page 2024