Pith. sign in

REVIEW 4 major objections 9 minor 99 references

PhyMRI-SR: Toward Physics-Aware MRI Image Super-Resolution

T0 review · 4 major / 9 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Lower resolution can yield sharper MRI scans, physics-aware model shows

desk verdict Adapting 2D Gaussian Splatting to MRI with physics-constrained signal modeling is a genuine methodological contribution. The 'optimal resolution isn't the highest' headline claim is not supported by the paper's own real data. read the letter →

arxiv 2607.06238 v1 pith:OX5W2KRA submitted 2026-07-07 cs.CV

Lihua Wei , Huatong Gao , Jia Gong , Zhiyu Tan , Hao Li , Jun Liu , Zhihua Ren This is my paper
classification cs.CV PACS 87.61.-c87.57.N
keywords MRIsuper-resolutionGaussianSplattingphysics-awarereconstructionresolution-SNRtrade-offultra-low-fieldmeta-learningdynamicresolution
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 argues that the starting image fed into an MRI super-resolution pipeline need not be the highest-resolution scan available. Because MRI physics couples spatial resolution to signal-to-noise ratio (SNR) — smaller voxels mean proportionally less signal per voxel while noise stays fixed — pushing for maximum resolution at acquisition time can actually produce a worse input for downstream enhancement. The authors demonstrate this with non-monotonic quality curves across resolution scales on both simulated and real data, showing that an intermediate resolution (around 0.7× on simulated data, 0.76× on real data) yields the best final reconstruction. To operationalize this insight, they adapt 2D Gaussian Splatting — a technique that represents images as collections of parametric geometric primitives rather than pixel grids — to MRI, making it resolution-agnostic so it can accept inputs at any scale. They augment it with three MRI-specific components: anatomically-guided placement of Gaussian primitives (denser in structurally complex tissue like cortical gray matter), a covariance dictionary that constrains primitive shapes to MRI-realistic configurations, and a physics-constrained intensity model that predicts proton density and relaxation parameters rather than directly regressing pixel values. A meta-learning pipeline bridges the gap between simulated training data and real ultra-low-field (64 mT) scanner acquisitions. On the FastMRI benchmark the method achieves 34.26 dB PSNR at 4× upscaling, exceeding the prior best by 1.45 dB, and on real 64 mT–3T paired data it reaches 26.85 dB versus under 20.3 dB for baselines.

What carries the argument

2D Gaussian Splatting adapted to MRI with three innovations: (1) segmentation-guided primitive initialization that distributes Gaussian kernels according to tissue complexity (gray matter gets more primitives than CSF), (2) an MRI-specific covariance dictionary of 1,001 kernel shapes learned from real high-field MRI data rather than natural images, and (3) a physics-constrained signal model that computes pixel intensity as c_i = ρ_i · e^{-R2,i} + δ_i from predicted proton density ρ and effective relaxation rate R2, rather than directly regressing intensities. A first-order MAML meta-learning pipeline trains on simulated data with episodic real-data injection for domain adaptation.

What would settle it

On real 64 mT–3T paired data, the method achieves 26.85 dB PSNR and 0.8856 SSIM, outperforming all baselines (LIIF, LTE, Pixel-to-Gaussian) by over 6.5 dB in PSNR. On the FastMRI benchmark, it achieves 34.26 dB PSNR at 4× upscaling, exceeding the prior best method (MS-PRDDiff) by 1.45 dB. The non-monotonic resolution hypothesis is supported by non-monotonic PSNR/SSIM curves across input resolution scales on both simulated (best at ×0.7) and real (competitive at ×0.76) datasets.

Watch

Extended reading notes

Core claim

The central discovery is empirical: the relationship between input resolution and super-resolution output quality is non-monotonic, with an intermediate resolution-SNR operating point producing the best reconstructions. This finding rests on the well-established MRI physics that SNR scales linearly with voxel volume while noise is resolution-independent, creating a regime where excessively fine resolution at acquisition time destroys more signal information than the extra spatial samples provide. The paper's methodological contribution is a 2D Gaussian Splatting framework adapted to MRI that can accept inputs at arbitrary resolutions and exploits this non-monotonicity, using physics-constrae

Load-bearing premise

The physics-constrained signal model includes an unconstrained learnable residual term that is added to the physics-based intensity formula. This residual can in principle absorb any deviation from the idealized signal equation, meaning the network could learn to bypass the physics constraint entirely and regress target intensities through the residual, making the biophysical-plausibility guarantee depend on whether the network actually uses the physics pathway rather than a兜

Editorial extensions

If this is right

  • If the non-monotonic resolution-SNR finding generalizes, MRI acquisition protocols could be redesigned to capture at an intermediate resolution that is cheaper and faster to acquire, rather than maximally resolved scans, trusting post-processing to recover the detail.
  • The resolution-agnostic Gaussian Splatting representation could be extended to 3D volumetric MRI, where the SNR penalty for isotropic resolution doubling is a factor of 8 (64× more averages to recover), making the optimal-resolution argument even stronger.
  • The physics-constrained intensity decomposition into proton density and relaxation parameters could enable the super-resolution model to output quantitative tissue parameter maps as a byproduct, not just enhanced images.
  • If the intermediate-resolution optimum is confirmed across more pulse sequences and field strengths, scanner manufacturers could build acquisition protocols that intentionally target the SNR-optimal point rather than the finest nominal resolution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 9 minor

Summary. This paper proposes PhyMRI-SR, a physics-aware MRI super-resolution framework based on 2D Gaussian Splatting. The central thesis is that the optimal input resolution for MRI super-resolution is not necessarily the highest achievable resolution, due to the resolution-SNR trade-off inherent in MRI physics. The framework introduces three innovations: (1) a prior-aware Gaussian representation with anatomical structure-guided initialization and an MRI-specific covariance dictionary, (2) physics-constrained signal modeling that predicts proton density and relaxation parameters rather than directly regressing intensities, and (3) a meta-learning framework for domain adaptation from simulated to real data. Experiments are conducted on simulated IXI data, a real 3T-5T dynamic-resolution dataset, the FastMRI benchmark, and a real 64mT-3T paired dataset. The method reports state-of-the-art performance on FastMRI (34.26 dB PSNR at 4x) and substantial improvements on real 64mT-3T data (26.85 dB vs. <20.3 dB for baselines).

Significance. The paper addresses a clinically relevant problem and proposes a well-motivated framework. The resolution-SNR derivation (Eqs. 1-4, S17-S36) is standard but correctly applied and clearly presented. The adaptation of 2D Gaussian Splatting to MRI with domain-specific priors is a reasonable architectural choice for continuous-scale super-resolution. The FastMRI benchmark results (Table II) are strong, with a 1.45 dB PSNR improvement over the best diffusion-based baseline at 4x. The real 64mT-3T results (Table III) show a substantial 6.5+ dB improvement over baselines, which is noteworthy. The meta-learning domain adaptation strategy is a practical contribution for the paired-data scarcity problem. However, the significance of the central 'optimal resolution' claim is weakened by inconsistencies in the supporting evidence and by the unconstrained residual in the physics model.

major comments (4)
  1. §IV-B, Eq. (16): The physics-constrained signal model c_i = rho_i * e^{-R2,i} + delta_i includes an unconstrained learnable residual delta_i. Since delta_i can absorb any deviation from the idealized signal equation, the claim of 'biophysically plausible contrast' is at risk of being vacuous — the network can learn delta_i ≈ c_i - rho_i * e^{-R2,i} for any target intensity, reducing the physics constraint to a reparameterization of direct intensity regression. The ablation in Table V shows the physics module improves SSIM from 0.55 to 0.85, but without constraining or analyzing delta_i (e.g., reporting its magnitude relative to the physics term, or adding a regularization term), it is unclear whether the improvement comes from the physics constraint or from the reparameterization acting as a different optimization landscape. The authors should either constrain delta_i (e.g., norm-bounded
  2. §V-C.2 and Table I (Real Multi-Resolution Dataset): The headline claim that 'the optimal resolution for MRI super-resolution is not necessarily the highest achievable resolution' is contradicted by the paper's own real data. On the 3T-5T dataset, the highest input resolution (x1.04) achieves the best PSNR (26.78 dB) and tied-best SSIM (0.8775). The paper acknowledges this but reframes it by pointing to secondary metrics (HFEN best at x0.76, DISTS best at x0.83). However, the claim is stated in terms of overall quality, and the primary metrics favor the highest resolution on real data. On the simulated IXI data, the evidence is marginal: x0.7 gives PSNR 28.10 vs. 28.06 at x1.0, a 0.04 dB difference within typical run-to-run variance. The authors should either soften the headline claim to reflect that different quality metrics favor different resolutions, or provide statistical evidence (e
  3. §S2.11 (Eqs. S37-S41) and Table I: The degradation model for simulating low-field MRI uses task-dependent blur sigma and noise alpha parameters, but the paper does not specify how these parameters are calibrated to the resolution-SNR trade-off in Eq. (4) for each resolution scale. If the noise level at each resolution scale is not set proportional to voxel volume (as Eq. 4 dictates), the 'optimal' intermediate resolution observed on simulated data could be an artifact of the chosen sigma and alpha values rather than a fundamental physics property. The authors should explicitly state the sigma and alpha values used for each resolution scale in Table I and verify that they are consistent with the SNR scaling predicted by Eq. (4).
  4. Table VI: The ablation table for meta-learning labels the second variant 'w/o Physics' in the Method column, but the text describes it as 'w/o Meta' (a model without meta-learning). This appears to be a labeling error. If the variant truly removes meta-learning, the label should read 'w/o Meta'. If it removes physics, then the ablation does not test the meta-learning contribution. This needs correction and clarification.
minor comments (9)
  1. Abstract: 'futher' should be 'further' (appears twice in abstract and introduction).
  2. §IV-A: 'domian-specific' should be 'domain-specific'.
  3. §IV-A.2: The covariance dictionary is built from 4,241 5T MRI slices (§S3.A). The paper should clarify whether these slices overlap with the test set used in Table I (Real Multi-Resolution Dataset), and if so, whether this constitutes a form of train-test leakage. If the dictionary is built from training data only, this should be stated explicitly.
  4. §V-A: 'acquision' should be 'acquisition' (appears in Eq. 3 caption area).
  5. Table I: The simulated IXI dataset shows DISTS values of 0.1231 for both x1.0 and x0.9, which is suspicious. The authors should verify this is not a copy error.
  6. §V-C.2: The phrase 'the optimal resolution lies in an intermediate range rather than at the extreme ends' overstates the real-data evidence, where x1.04 is best or tied-best on two of four metrics. Consider softening this language.
  7. Figure 2 caption: 'segmentator' is non-standard terminology; consider 'segmentation network' or 'segmentation model' for consistency with standard terminology.
  8. §S3.C.3, Eq. (S42): The reference 'Method Section ??' is an unresolved LaTeX cross-reference.
  9. The paper would benefit from reporting standard deviations or confidence intervals for the dynamic-resolution experiments (Table I), especially given that the key PSNR differences on simulated data (28.10 vs. 28.06) are small.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee raises four major points: (1) the unconstrained residual delta_i in the physics-constrained signal model may render the physics constraint vacuous; (2) the headline claim that optimal resolution is not the highest is contradicted by the real 3T-5T data where the highest resolution achieves best PSNR; (3) the degradation model's sigma and alpha parameters may not be calibrated to the SNR scaling of Eq. (4), potentially making the simulated optimal-resolution finding an artifact; and (4) a labeling error in Table VI. We address each point below and describe the revisions we will make.

read point-by-point responses

  1. Referee: §IV-B, Eq. (16): The physics-constrained signal model c_i = rho_i * e^{-R2,i} + delta_i includes an unconstrained learnable residual delta_i. Since delta_i can absorb any deviation from the idealized signal equation, the claim of 'biophysically plausible contrast' is at risk of being vacuous. The ablation in Table V shows the physics module improves SSIM from 0.55 to 0.85, but without constraining or analyzing delta_i, it is unclear whether the improvement comes from the physics constraint or from the reparameterization acting as a different optimization landscape.

    Authors: The referee raises a valid and important concern. We agree that without any constraint or analysis on delta_i, the physics-constrained formulation could in principle degenerate into a reparameterization of direct intensity regression, which would weaken our claim of biophysically plausible contrast. In the revised manuscript, we will make two changes. First, we will add an L2 regularization term on delta_i (lambda * ||delta||_2) to the training objective, penalizing large deviations from the physics-predicted signal and encouraging the network to rely on the physical model rather than absorbing all intensity information into the residual. Second, we will report the ratio ||delta_i|| / ||rho_i * e^{-R2,i}|| measured on the test set, demonstrating that the physics term dominates the residual and that delta_i accounts for a small fraction of the total intensity. This will directly address whether the improvement in SSIM (0.55 to 0.85) comes from the physics constraint rather than from reparameterization alone. We will also add an ablation variant with delta_i = 0 (fully constrained) to isolate the contribution of the physics term from the residual. We note that the qualitative results in Figure S8 already provide indirect evidence that the physics module restores correct grayscale fidelity consistent with the ground truth, which pure reparameterization would not necessarily achieve. However, we agree that quantitative analysis of delta_i is needed to substantiate this claim. revision: yes

  2. Referee: §V-C.2 and Table I: The headline claim that 'the optimal resolution for MRI super-resolution is not necessarily the highest achievable resolution' is contradicted by the paper's own real data. On the 3T-5T dataset, the highest input resolution (x1.04) achieves the best PSNR (26.78 dB) and tied-best SSIM (0.8775). The paper acknowledges this but reframes it by pointing to secondary metrics (HFEN best at x0.76, DISTS best at x0.83). On the simulated IXI data, the evidence is marginal: x0.7 gives PSNR 28.10 vs. 28.06 at x1.0, a 0.04 dB difference within typical run-to-run variance.

    Authors: The referee is correct that the real 3T-5T data does not unambiguously support the headline claim when PSNR and SSIM are the primary metrics. On the real data, the highest resolution (x1.04) achieves the best PSNR (26.78 dB) and tied-best SSIM (0.8775), while intermediate resolutions are favored only by HFEN (x0.76) and DISTS (x0.83). On the simulated IXI data, the PSNR difference between x0.7 (28.10 dB) and x1.0 (28.06 dB) is indeed marginal at 0.04 dB. We will revise the manuscript in two ways. First, we will soften the headline claim from 'the optimal resolution for MRI super-resolution is not necessarily the highest achievable resolution' to a more precise statement: 'different quality metrics favor different resolution settings, and intermediate resolutions can achieve competitive or superior performance on perceptual and high-frequency metrics, suggesting that the highest input resolution is not universally optimal across all aspects of image quality.' Second, we will run multiple random seeds (at least 3) on the simulated IXI experiment and report mean and standard deviation for each resolution scale, so that the reader can assess whether the 0.04 dB difference is within run-to-run variance. If the difference is not statistically significant, we will state this explicitly and frame the simulated result as 'no significant degradation at intermediate resolution' rather than 'intermediate resolution is superior.' We acknowledge that the current evidence does not support a strong universal claim, and the revised framing will reflect this honestly. revision: yes

  3. Referee: §S2.11 (Eqs. S37-S41) and Table I: The degradation model for simulating low-field MRI uses task-dependent blur sigma and noise alpha parameters, but the paper does not specify how these parameters are calibrated to the resolution-SNR trade-off in Eq. (4) for each resolution scale. If the noise level at each resolution scale is not set proportional to voxel volume (as Eq. 4 dictates), the 'optimal' intermediate resolution observed on simulated data could be an artifact of the chosen sigma and alpha values rather than a fundamental physics property.

    Authors: This is a fair concern. The referee is right that if the noise level alpha at each resolution scale is not set consistently with the SNR scaling predicted by Eq. (4) (i.e., SNR proportional to voxel volume), then the observed non-monotonic relationship between input resolution and SR performance on simulated data could be an artifact of the degradation parameter choices rather than a physics-driven phenomenon. In the revised manuscript, we will add a table specifying the exact sigma and alpha values used for each resolution scale in the simulated IXI experiment. We will also verify and explicitly state whether these values are calibrated to be proportional to voxel volume as dictated by Eq. (4). If the current values are not calibrated this way, we will re-run the simulated experiment with properly calibrated noise levels and update Table I accordingly. If the non-monotonic trend disappears under properly calibrated noise, we will report this honestly and revise our claims. We agree that this calibration is essential for the simulated-data evidence to be credible. revision: yes

  4. Referee: Table VI: The ablation table for meta-learning labels the second variant 'w/o Physics' in the Method column, but the text describes it as 'w/o Meta' (a model without meta-learning). This appears to be a labeling error. If the variant truly removes meta-learning, the label should read 'w/o Meta'. If it removes physics, then the ablation does not test the meta-learning contribution.

    Authors: The referee is correct. This is a labeling error in Table VI. The text in the accompanying paragraph clearly describes the variant as 'a model without meta-learning (w/o Meta),' and the purpose of the ablation is to test the contribution of meta-learning. The label in the Method column should read 'w/o Meta,' not 'w/o Physics.' We will correct this in the revised manuscript. We apologize for the confusion. revision: yes

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity found; one minor concern with covariance dictionary encoding target-domain statistics, but it is a learned prior, not a circular derivation.

full rationale

The paper's derivation chain does not exhibit circularity. The resolution-SNR trade-off (Eqs. 1–4, S17–S36) is standard MRI physics properly cited from external textbooks [33, 34, 73] and derived from the spin-echo signal equation without self-reference. The physics-constrained signal model (Eqs. 13–16) uses the well-known T2-weighted signal equation S ≈ ρ·e^{-R2}; the paper does not claim to derive this equation but adopts it as a known physical relationship. The unconstrained residual δ_i (Eq. 16) raises a correctness concern (it could absorb any deviation, making the physics constraint vacuous), but this is a validity issue, not circularity — the paper does not present δ_i as a prediction or first-principles result. The covariance dictionary (Section IV-A.2) is built by fitting 2D GS to 4,241 high-resolution 5T MRI slices and then constraining predicted covariances to linear combinations of dictionary entries. While this encodes target-domain spatial statistics into the model, it is presented as a learned prior (a dictionary), not as a prediction or derivation. The paper does not claim to 'predict' MRI system characteristics from first principles; it explicitly states the dictionary is constructed empirically from data. No self-citation chain is load-bearing: the 2D GS framework is cited from [39] (Pixel-to-Gaussian, external authors), meta-learning from MAML [74] (external), and segmentation from SynthSeg [95] (external). The 'optimal resolution' hypothesis is an empirical finding from experiments, not a derivation that reduces to its inputs. The minor score of 2 reflects the covariance dictionary's encoding of target-domain statistics, which could limit generalization claims, but this is a methodological design choice rather than a circular derivation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The paper introduces several hand-set or learned parameters (complexity weights, offset scale, dictionary size, residual term) and relies on domain assumptions about T2-weighted signal simplification, tissue property invariance, and degradation model fidelity. The covariance dictionary is a new entity without independent falsifiable handles.

free parameters (6)
  • Tissue complexity weights w_k (w_GM, w_WM, w_CSF) = w_GM > w_WM > w_CSF (exact values not stated)
    Hand-set weights controlling primitive density per tissue type in Eq. 9. Not fitted to data but chosen by anatomical intuition.
  • Position offset scale δ = Not specified
    Controls maximum displacement range in Eq. 11. Stated as a scaling factor but exact value not given.
  • Covariance dictionary size M = 1001
    Number of dictionary entries built from 4,241 5T MRI slices (Section S3.A). 10 values per parameter × 3 parameters + zero kernel.
  • Learnable residual δ_i = Learned per-primitive
    Eq. 16. Unconstrained residual absorbing physics model deviations. Effectively a free parameter per Gaussian primitive.
  • Meta-learning inner-loop rate α = Not specified
    Eq. S42. Inner-loop learning rate for task adaptation.
  • Meta-learning outer-loop rate β = Not specified
    Eq. S43. Outer-loop meta-update learning rate.
assumptions (4)
  • domain assumption MRI signal for T2-weighted spin echo simplifies to S ≈ ρ·e^{-TE/T2} when TR >> T1
    Eq. 14. Standard MRI physics assumption for T2-weighted imaging. Used as the basis for physics-constrained signal modeling.
  • domain assumption Tissue properties ρ and R2 are resolution-invariant
    Section IV-B.2: 'since ρ and R2 reflect intrinsic tissue properties that are independent of imaging resolution.' This motivates the decomposition but is not independently verified in the paper.
  • domain assumption The empirical covariance distribution from 5T MRI generalizes to other field strengths
    Section IV-A.2. The covariance dictionary is built from 5T data but applied to 64mT and 3T super-resolution. No validation that 5T covariance statistics transfer.
  • domain assumption Simulated degradation model (Eq. S37) accurately represents real low-field MRI
    Section S2.11. The degradation model combines PSF blurring, Rician noise, and bias field. Used to generate all simulated training data and to support the dynamic-resolution hypothesis.
invented entities (1)
  • MRI-Specific Covariance Dictionary
    purpose: Constrains Gaussian primitive covariance to MRI-realistic configurations
    Built from 5T MRI data (Section S3.A). No falsifiable prediction outside the paper — it is a precomputed lookup table from training-domain statistics. Whether it generalizes to unseen field strengths or pathologies is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PhyMRI-SR: Toward Physics-Aware MRI Image Super-Resolution." pith.science (2026). https://pith.science/paper/OX5W2KRA

@misc{pith2026260706238,
  author       = {Pith},
  title        = {Pith review of: PhyMRI-SR: Toward Physics-Aware MRI Image Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OX5W2KRA}},
  note         = {Machine review of arXiv:2607.06238}
}
read the original abstract

Magnetic resonance imaging (MRI) super-resolution is vital for improving diagnostic accessibility, yet most methods treat it as a deterministic mapping from a fixed low-resolution input to a high-resolution target. This overlooks a key property of MRI acquisition physics: spatial resolution and signal-to-noise ratio (SNR) are inherently coupled, making any given low-resolution scan merely one of many possible realizations under varying acquisition trade-offs. We rethink MRI super-resolution as a physics-aware reconstruction problem, in which the goal is to identify the optimal resolution-SNR configuration and then super-resolve it to obtain high-quality MRI results. A key implication of this formulation is that MRI resolution becomes dynamic rather than fixed. To handle such resolution-heterogeneous inputs, we adapt 2D Gaussian Splatting (2D GS) to MRI by formulating reconstruction as a coordinate-based, resolution-agnostic rendering problem. To further enhance fidelity, we introduce three innovations: (1) a prior-aware Gaussian representation that combines an Anatomical Structure Prior for tissue-specific kernel initialization with an Imaging System Prior that captures hardware characteristics via a covariance dictionary; (2) a physics-constrained signal modeling scheme that predicts intrinsic tissue parameters (proton density rho and effective relaxation rate R2) and synthesizes intensities through governing physical equations, ensuring biophysically plausible contrast; and (3) a meta-learning framework that alleviates paired-data scarcity by pretraining on simulated data and adapting to real-world conditions. Extensive experiments on dynamic-resolution datasets and standard benchmarks demonstrate that our method achieves state-of-the-art performance, highlighting its strong potential for clinical deployment.

Figures

Figures reproduced from arXiv: 2607.06238 by the authors.

Figure 1
Figure 1. Illustration of the trade-off between spatial resolution and signal [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. This underscores that the given low-resolution image [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed physics-aware 2D GS framework. An arbitrary-resolution input is processed through two parallel pathways: a segmentator [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Effect of input resolution on super-resolution quality on the simulated IXI dataset. Two representative cases are shown with varying scales from [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 4
Figure 4. Figure 4: Visual comparison on the real 3T–5T dynamic-resolution dataset. Our method produces reconstructions most consistent with the 5T Ground Truth, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The qualitative results on three simulated datasets under the optimal input resolution. Our method reconstructs fine details of the Ground Truth on [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison on real 64mT-3T dataset. The SR results of our method can recover fine details from the Ground Truth. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

99 extracted references · 99 canonical work pages

  1. [1]

    Magnetic resonance imaging of musculoskeletal tumors,

    M. J. Kransdorf, “Magnetic resonance imaging of musculoskeletal tumors,”Orthopedics, vol. 17, no. 11, pp. 1003–1016, 1994

    work page 1994

  2. [2]

    Noninvasive mri evaluation of cerebral blood flow in cerebrovascular disease,

    J. A. Detre, D. C. Alsop, L. Vives, L. Maccotta, J. Teener, and E. Raps, “Noninvasive mri evaluation of cerebral blood flow in cerebrovascular disease,”Neurology, vol. 50, no. 3, pp. 633–641, 1998

    work page 1998

  3. [3]

    Brain mri analysis for alzheimer’s disease diag- nosis using an ensemble system of deep convolutional neural networks,

    J. Islam and Y . Zhang, “Brain mri analysis for alzheimer’s disease diag- nosis using an ensemble system of deep convolutional neural networks,” Brain informatics, vol. 5, no. 2, p. 2, 2018

    work page 2018

  4. [4]

    Deep learning for accelerated and robust mri reconstruction,

    R. Heckel, M. Jacob, A. Chaudhari, O. Perlman, and E. Shimron, “Deep learning for accelerated and robust mri reconstruction,”Magnetic Resonance Materials in Physics, Biology and Medicine, vol. 37, no. 3, pp. 335–368, 2024

    work page 2024

  5. [5]

    A review on deep learning mri reconstruction without fully sampled k-space,

    G. Zeng, Y . Guo, J. Zhan, Z. Wang, Z. Lai, X. Du, X. Qu, and D. Guo, “A review on deep learning mri reconstruction without fully sampled k-space,”BMC Medical Imaging, vol. 21, no. 1, p. 195, 2021

    work page 2021

  6. [6]

    Super-resolution in magnetic resonance imaging: a review,

    E. Van Reeth, I. W. Tham, C. H. Tan, and C. L. Poh, “Super-resolution in magnetic resonance imaging: a review,”Concepts in Magnetic Reso- nance Part A, vol. 40, no. 6, pp. 306–325, 2012

    work page 2012

  7. [7]

    Super- resolution reconstruction of mr image with a novel residual learning network algorithm,

    J. Shi, Q. Liu, C. Wang, Q. Zhang, S. Ying, and H. Xu, “Super- resolution reconstruction of mr image with a novel residual learning network algorithm,”Physics in Medicine & Biology, vol. 63, no. 8, p. 085011, 2018

    work page 2018

  8. [8]

    Medical image super-resolution reconstruction algorithms based on deep learning: A survey,

    D. Qiu, Y . Cheng, and X. Wang, “Medical image super-resolution reconstruction algorithms based on deep learning: A survey,”Computer Methods and Programs in Biomedicine, vol. 238, p. 107590, 2023

    work page 2023

Show all 99 references
  1. [9]

    Brain mri image super-resolution reconstruction: a systematic review,

    A. Muhammad, S. Aramvith, K. Duangchaemkarn, and M.-T. Sun, “Brain mri image super-resolution reconstruction: a systematic review,” IEEE Access, vol. 12, pp. 156 347–156 362, 2024

    2024

  2. [10]

    Sparse reconstruction techniques in magnetic resonance imaging: meth- ods, applications, and challenges to clinical adoption,

    A. C. Yang, M. Kretzler, S. Sudarski, V . Gulani, and N. Seiberlich, “Sparse reconstruction techniques in magnetic resonance imaging: meth- ods, applications, and challenges to clinical adoption,”Investigative radiology, vol. 51, no. 6, pp. 349–364, 2016

    2016

  3. [11]

    Super- resolution methods in mri: can they improve the trade-off between reso- lution, signal-to-noise ratio, and acquisition time?

    E. Plenge, D. H. Poot, M. Bernsen, G. Kotek, G. Houston, P. Wielopol- ski, L. Van Der Weerd, W. J. Niessen, and E. Meijering, “Super- resolution methods in mri: can they improve the trade-off between reso- lution, signal-to-noise ratio, and acquisition time?”Magnetic resonance...

    1983

  4. [12]

    Measuring information gain for frequency-encoded super-resolution mri,

    G. S. Mayer and E. R. Vrscay, “Measuring information gain for frequency-encoded super-resolution mri,”Magnetic resonance imaging, vol. 25, no. 7, pp. 1058–1069, 2007

    2007

  5. [13]

    Mri resolution enhancement: How useful are shifted images obtained by changing the demodulation frequency?

    Q. M. Tieng, G. J. Cowin, D. C. Reutens, G. J. Galloway, and V . Vegh, “Mri resolution enhancement: How useful are shifted images obtained by changing the demodulation frequency?”Magnetic resonance in medicine, vol. 65, no. 3, pp. 664–672, 2011

    2011

  6. [14]

    Fast single image super-resolution using estimated low-frequency k-space data in mri,

    J. Luo, Z. Mou, B. Qin, W. Li, F. Yang, M. Robini, and Y . Zhu, “Fast single image super-resolution using estimated low-frequency k-space data in mri,”Magnetic resonance imaging, vol. 40, pp. 1–11, 2017

    2017

  7. [15]

    Mri superresolution using self-similarity and image priors,

    J. V . Manj ´on, P. Coup´e, A. Buades, D. L. Collins, and M. Robles, “Mri superresolution using self-similarity and image priors,”International journal of biomedical imaging, vol. 2010, no. 1, p. 425891, 2010

    2010

  8. [16]

    Resolution enhancement in mri,

    E. Carmi, S. Liu, N. Alon, A. Fiat, and D. Fiat, “Resolution enhancement in mri,”Magnetic resonance imaging, vol. 24, no. 2, pp. 133–154, 2006

    2006

  9. [17]

    Mri inter-slice recon- struction using super-resolution,

    H. Greenspan, G. Oz, N. Kiryati, and S. Peled, “Mri inter-slice recon- struction using super-resolution,”Magnetic resonance imaging, vol. 20, no. 5, pp. 437–446, 2002

    2002

  10. [18]

    Example-based super- resolution,

    W. T. Freeman, T. R. Jones, and E. C. Pasztor, “Example-based super- resolution,”IEEE Computer graphics and Applications, vol. 22, no. 2, pp. 56–65, 2002

    2002

  11. [19]

    Multiscale brain mri super-resolution using deep 3d convolutional networks,

    C.-H. Pham, C. Tor-D ´ıez, H. Meunier, N. Bednarek, R. Fablet, N. Passat, and F. Rousseau, “Multiscale brain mri super-resolution using deep 3d convolutional networks,”Computerized Medical Imaging and Graphics, vol. 77, p. 101647, 2019

    2019

  12. [20]

    Super-resolution of brain tumor mri images based on deep learning,

    Z. Zhou, A. Ma, Q. Feng, R. Wang, L. Cheng, X. Chen, X. Yang, K. Liao, Y . Miao, and Y . Qiu, “Super-resolution of brain tumor mri images based on deep learning,”Journal of applied clinical medical physics, vol. 23, no. 11, p. e13758, 2022

    2022

  13. [21]

    Residual dense network for medical magnetic resonance images super-resolution,

    D. Zhu and D. Qiu, “Residual dense network for medical magnetic resonance images super-resolution,”Computer Methods and Programs in Biomedicine, vol. 209, p. 106330, 2021

    2021

  14. [22]

    Spatio-angular convolutions for super-resolution in diffusion mri,

    M. Lyon, P. Armitage, and M. A. ´Alvarez, “Spatio-angular convolutions for super-resolution in diffusion mri,”Advances in Neural Information Processing Systems, vol. 36, pp. 12 457–12 475, 2023

    2023

  15. [23]

    Mul- ticontrast mri super-resolution via transformer-empowered multiscale contextual matching and aggregation,

    J. Lyu, G. Li, C. Wang, Q. Cai, Q. Dou, D. Zhang, and J. Qin, “Mul- ticontrast mri super-resolution via transformer-empowered multiscale contextual matching and aggregation,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 9, pp. 12 004–12 014, 2023

    2023

  16. [24]

    Accurate multi-contrast mri super-resolution via a dual cross-attention transformer network,

    S. Huang, J. Li, L. Mei, T. Zhang, Z. Chen, Y . Dong, L. Dong, S. Liu, and M. Lyu, “Accurate multi-contrast mri super-resolution via a dual cross-attention transformer network,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer,...

    2023

  17. [25]

    Rethinking multi-contrast mri super-resolution: Rectangle- window cross-attention transformer and arbitrary-scale upsampling,

    G. Li, L. Zhao, J. Sun, Z. Lan, Z. Zhang, J. Chen, Z. Lin, H. Lin, and W. Xing, “Rethinking multi-contrast mri super-resolution: Rectangle- window cross-attention transformer and arbitrary-scale upsampling,” in Proceedings of the IEEE/CVF international conference on computer v...

    2023

  18. [28]

    Inversesr: 3d brain mri super-resolution using a latent diffusion model,

    J. Wang, J. Levman, W. H. L. Pinaya, P.-D. Tudosiu, M. J. Cardoso, and R. Marinescu, “Inversesr: 3d brain mri super-resolution using a latent diffusion model,” inInternational conference on medical image computing and computer-assisted intervention. Springer, 2023, pp. 438–447

    2023

  19. [29]

    Mri super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting,

    M. Safari, S. Wang, Z. Eidex, Q. Li, R. L. Qiu, E. H. Middlebrooks, D. S. Yu, and X. Yang, “Mri super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting,”Physics in Medicine & Biology, vol. 70, no. 12, p. 125008, 2025

    2025

  20. [30]

    Super- resolution musculoskeletal mri using deep learning,

    A. S. Chaudhari, Z. Fang, F. Kogan, J. Wood, K. J. Stevens, E. K. Gibbons, J. H. Lee, G. E. Gold, and B. A. Hargreaves, “Super- resolution musculoskeletal mri using deep learning,”Magnetic reso- nance in medicine, vol. 80, no. 5, pp. 2139–2154, 2018

    2018

  21. [31]

    Brain mri super resolution using 3d deep densely connected neural networks,

    Y . Chen, Y . Xie, Z. Zhou, F. Shi, A. G. Christodoulou, and D. Li, “Brain mri super resolution using 3d deep densely connected neural networks,” in2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018). IEEE, 2018, pp. 739–742

    2018

  22. [32]

    Multi-contrast super-resolution mri through a progressive network,

    Q. Lyu, H. Shan, C. Steber, C. Helis, C. Whitlow, M. Chan, and G. Wang, “Multi-contrast super-resolution mri through a progressive network,”IEEE transactions on medical imaging, vol. 39, no. 9, pp. 2738–2749, 2020

    2020

  23. [33]

    Physics of mri: a primer,

    D. B. Plewes and W. Kucharczyk, “Physics of mri: a primer,”Journal of magnetic resonance imaging, vol. 35, no. 5, pp. 1038–1054, 2012

    2012

  24. [34]

    Understanding mri: basic mr physics for physicians,

    S. Currie, N. Hoggard, I. J. Craven, M. Hadjivassiliou, and I. D. Wilkin- son, “Understanding mri: basic mr physics for physicians,”Postgraduate medical journal, vol. 89, no. 1050, pp. 209–223, 2013

    2013

  25. [35]

    Information content of snr/resolution trade-offs in three-dimensional magnetic resonance imaging,

    S. Portnoy, S. Kale, A. Feintuch, C. Tardif, G. Pike, and R. Henkelman, “Information content of snr/resolution trade-offs in three-dimensional magnetic resonance imaging,”Medical physics, vol. 36, no. 4, pp. 1442– 1451, 2009

    2009

  26. [36]

    Rethinking dif- fusion model for multi-contrast mri super-resolution,

    G. Li, C. Rao, J. Mo, Z. Zhang, W. Xing, and L. Zhao, “Rethinking dif- fusion model for multi-contrast mri super-resolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 11 365–11 374

    2024

  27. [37]

    Deep attention super-resolution of brain magnetic resonance images acquired under clinical protocols,

    B. M. Li, L. V . Castorina, M. d. C. Vald ´es Hern ´andez, U. Clancy, S. J. Wiseman, E. Sakka, A. J. Storkey, D. Jaime Garcia, Y . Cheng, F. Doubal et al., “Deep attention super-resolution of brain magnetic resonance images acquired under clinical protocols,”Frontiers in Compu...

    2022

  28. [38]

    Smore: a self-supervised anti-aliasing and super-resolution algorithm for mri using deep learning,

    C. Zhao, B. E. Dewey, D. L. Pham, P. A. Calabresi, D. S. Reich, and J. L. Prince, “Smore: a self-supervised anti-aliasing and super-resolution algorithm for mri using deep learning,”IEEE transactions on medical imaging, vol. 40, no. 3, pp. 805–817, 2020

    2020

  29. [39]

    Pixel to gaussian: Ultra-fast continuous super-resolution with 2d gaussian modeling,

    L. Peng, A. Wu, W. Li, P. Xia, X. Dai, X. Zhang, X. Di, H. Sun, R. Pei, Y . Wanget al., “Pixel to gaussian: Ultra-fast continuous super-resolution with 2d gaussian modeling,”arXiv preprint arXiv:2503.06617, 2025

    2025 arXiv

  30. [40]

    Mri super-resolution with deep learning: A comprehensive survey,

    M. Khateri, S. Vasylechko, M. Ghahremani, L. Timms, D. Ko- canaogullari, S. K. Warfield, C. Jaimes, D. Karimi, A. Sierra, J. Tohka et al., “Mri super-resolution with deep learning: A comprehensive survey,”arXiv preprint arXiv:2511.16854, 2025

    2025

  31. [41]

    Application of tikhonov regularization to super-resolution reconstruction of brain mri images,

    X. Zhang, E. Y . Lam, E. X. Wu, and K. K. Wong, “Application of tikhonov regularization to super-resolution reconstruction of brain mri images,” inMedical Imaging and Informatics: 2nd International Conference, MIMI 2007, Beijing, China, August 14–16, 2007, Revised Selected Pap...

    2007

  32. [42]

    A tool for super-resolving multimodal clinical mri,

    M. Brudfors, Y . Balbastre, P. Nachev, and J. Ashburner, “A tool for super-resolving multimodal clinical mri,”arXiv preprint arXiv:1909.01140, 2019. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    1909 arXiv

  33. [43]

    Nonlinear total variation based noise removal algorithms,

    L. I. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise removal algorithms,”Physica D: Nonlinear Phenomena, vol. 60, no. 1–4, pp. 259–268, 1992

    1992

  34. [44]

    An efficient total variation algorithm for super-resolution in fetal brain mri with adaptive regularization,

    S. Tourbier, X. Bresson, P. Hagmann, J.-P. Thiran, R. Meuli, and M. B. Cuadra, “An efficient total variation algorithm for super-resolution in fetal brain mri with adaptive regularization,”NeuroImage, vol. 118, pp. 584–597, 2015

    2015

  35. [45]

    Lrtv: Mr image super-resolution with low-rank and total variation regularizations,

    F. Shi, J. Cheng, L. Wang, P.-T. Yap, and D. Shen, “Lrtv: Mr image super-resolution with low-rank and total variation regularizations,”IEEE Transactions on Medical Imaging, vol. 34, no. 12, pp. 2459–2466, 2015

    2015

  36. [46]

    Super-resolution from a single image,

    D. Glasner, S. Bagon, and M. Irani, “Super-resolution from a single image,” in2009 IEEE 12th International Conference on Computer Vision. IEEE, 2009, pp. 349–356

    2009

  37. [47]

    Isotropic reconstruction of mr images using 3d patch-based self-similarity learning,

    A. Bustin, D. V oilliot, A. Menini, J. Felblinger, C. de Chillou, D. Burschka, L. Bonnemains, and F. Odille, “Isotropic reconstruction of mr images using 3d patch-based self-similarity learning,”IEEE Transactions on Medical Imaging, vol. 37, no. 8, pp. 1932–1942, 2018

    1932

  38. [48]

    Deep mr brain image super-resolution using spatio-structural priors,

    V . Cherukuri, T. Guo, S. J. Schiff, and V . Monga, “Deep mr brain image super-resolution using spatio-structural priors,”IEEE Transactions on Image Processing, vol. 29, pp. 1368–1383, 2019

    2019

  39. [49]

    Motion-aligned 4dmri reconstruction using higher degree total variation and locally low-rank regularization,

    P. Li, J. Chen, D. Nan, J. Zou, D. Lin, and Y . Hu, “Motion-aligned 4dmri reconstruction using higher degree total variation and locally low-rank regularization,”Magnetic Resonance Imaging, vol. 93, pp. 97–107, 2022

    2022

  40. [50]

    Accelerated mri reconstruction with separable and enhanced low-rank hankel regularization,

    X. Zhang, H. Lu, D. Guo, Z. Lai, H. Ye, X. Peng, B. Zhao, and X. Qu, “Accelerated mri reconstruction with separable and enhanced low-rank hankel regularization,”IEEE Transactions on Medical Imaging, vol. 41, no. 9, pp. 2486–2498, 2022

    2022

  41. [51]

    Image super-resolution based on structure-modulated sparse representation,

    Y . Zhang, J. Liu, W. Yang, and Z. Guo, “Image super-resolution based on structure-modulated sparse representation,”IEEE Transactions on Image Processing, vol. 24, no. 9, pp. 2797–2810, 2015

    2015

  42. [52]

    Sparse representation-based mri super-resolution reconstruction,

    Y .-H. Wang, J. Qiao, J.-B. Li, P. Fu, S.-C. Chu, and J. F. Roddick, “Sparse representation-based mri super-resolution reconstruction,”Mea- surement, vol. 47, pp. 946–953, 2014

    2014

  43. [53]

    Hierarchical patch-based sparse representation—a new approach for resolution enhancement of 4d-ct lung data,

    Y . Zhang, G. Wu, P.-T. Yap, Q. Feng, J. Lian, W. Chen, and D. Shen, “Hierarchical patch-based sparse representation—a new approach for resolution enhancement of 4d-ct lung data,”IEEE Transactions on Medical Imaging, vol. 31, no. 11, pp. 1993–2005, 2012

    1993

  44. [54]

    Generalizing the nonlocal-means to super-resolution reconstruction,

    M. Protter, M. Elad, H. Takeda, and P. Milanfar, “Generalizing the nonlocal-means to super-resolution reconstruction,”IEEE Transactions on Image Processing, vol. 18, no. 1, pp. 36–51, 2008

    2008

  45. [55]

    Non-local mri upsampling,

    J. V . Manj´on, P. Coup´e, A. Buades, V . Fonov, D. L. Collins, and M. Rob- les, “Non-local mri upsampling,”Medical Image Analysis, vol. 14, no. 6, pp. 784–792, 2010

    2010

  46. [56]

    Mri upsampling using feature-based nonlocal means approach,

    K. Jafari-Khouzani, “Mri upsampling using feature-based nonlocal means approach,”IEEE Transactions on Medical Imaging, vol. 33, no. 10, pp. 1969–1985, 2014

    1969

  47. [57]

    Fast and robust multiframe super resolution,

    S. Farsiu, M. D. Robinson, M. Elad, and P. Milanfar, “Fast and robust multiframe super resolution,”IEEE Transactions on Image Processing, vol. 13, no. 10, pp. 1327–1344, 2004

    2004

  48. [58]

    Isotropic mri super-resolution reconstruction with multi-scale gradient field prior,

    Y . Sui, O. Afacan, A. Gholipour, and S. K. Warfield, “Isotropic mri super-resolution reconstruction with multi-scale gradient field prior,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2019, pp. 3–11

    2019

  49. [59]

    Learning a gradient guidance for spatially isotropic mri super- resolution reconstruction,

    ——, “Learning a gradient guidance for spatially isotropic mri super- resolution reconstruction,” inInternational Conference on Medical Im- age Computing and Computer-Assisted Intervention. Springer, 2020, pp. 136–146

    2020

  50. [60]

    Photo-realistic single image super-resolution using a generative adversarial network,

    C. Ledig, L. Theis, F. Huszar, J. Caballeroet al., “Photo-realistic single image super-resolution using a generative adversarial network,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4681–4690

    2017

  51. [61]

    Enhanced generative adversarial network for 3d brain mri super-resolution,

    J. Wang, Y . Chen, Y . Wu, J. Shi, and J. Gee, “Enhanced generative adversarial network for 3d brain mri super-resolution,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2020, pp. 3627–3636

    2020

  52. [62]

    Soup-gan: Super-resolution mri using generative adversarial networks,

    K. Zhang, H. Hu, K. Philbrick, G. M. Conte, J. D. Sobek, P. Rouzrokh, and B. J. Erickson, “Soup-gan: Super-resolution mri using generative adversarial networks,”Tomography, vol. 8, no. 2, pp. 905–919, 2022

    2022

  53. [63]

    Disgan: Wavelet-informed discriminator guides gan to mri super-resolution with noise cleaning,

    Q. Wang, L. Mahler, J. Steiglechner, F. Birk, K. Scheffler, and G. Lohmann, “Disgan: Wavelet-informed discriminator guides gan to mri super-resolution with noise cleaning,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 2452–2461

    2023

  54. [64]

    Super-resolution of 3d medical images by generative adversarial networks with long and short-term memory and attention,

    Q. Zhang, Y . Hang, F. Wu, S. Wang, and Y . Hong, “Super-resolution of 3d medical images by generative adversarial networks with long and short-term memory and attention,”Scientific Reports, vol. 15, no. 1, p. 20828, 2025

    2025

  55. [65]

    Nesvor: Implicit neural representation for slice-to- volume reconstruction in mri,

    J. Xu, D. Moyer, B. Gagoski, J. E. Iglesias, P. E. Grant, P. Golland, and E. Adalsteinsson, “Nesvor: Implicit neural representation for slice-to- volume reconstruction in mri,”IEEE Transactions on Medical Imaging, vol. 42, no. 6, pp. 1707–1719, 2023

    2023

  56. [66]

    Cycleinr: Cycle implicit neural representation for arbitrary-scale volumetric super-resolution of medical data,

    W. Fang, Y . Tang, H. Guo, M. Yuan, T. C. Mok, K. Yan, J. Yao, X. Chen, Z. Liu, L. Luet al., “Cycleinr: Cycle implicit neural representation for arbitrary-scale volumetric super-resolution of medical data,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Patte...

    2024

  57. [67]

    Gaussiansr: High fidelity 2d gaussian splatting for arbitrary-scale image super-resolution,

    J. Hu, B. Xia, B. Chen, W. Yang, and L. Zhang, “Gaussiansr: High fidelity 2d gaussian splatting for arbitrary-scale image super-resolution,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025, pp. 3554–3562

    2025

  58. [68]

    Meta-sr: A magnification-arbitrary network for super-resolution,

    X. Hu, H. Mu, X. Zhang, Z. Wang, T. Tan, and J. Sun, “Meta-sr: A magnification-arbitrary network for super-resolution,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 1575–1584

    2019

  59. [69]

    Learning continuous image representa- tion with local implicit image function,

    Y . Chen, S. Liu, and X. Wang, “Learning continuous image representa- tion with local implicit image function,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  60. [70]

    Local texture estimator for implicit representation func- tion,

    Y . Xuet al., “Local texture estimator for implicit representation func- tion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  61. [71]

    Ciaosr: Continuous implicit attention- inattention network for arbitrary-scale image super-resolution,

    J. Cao, Q. Wang, Y . Xian, Y . Li, B. Ni, Z. Pi, K. Zhang, Y . Zhang, R. Timofte, and L. Van Gool, “Ciaosr: Continuous implicit attention- inattention network for arbitrary-scale image super-resolution,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern...

    2023

  62. [72]

    Spin echo magnetic resonance imaging,

    B. A. Jung and M. Weigel, “Spin echo magnetic resonance imaging,” Journal of Magnetic Resonance Imaging, vol. 37, no. 4, pp. 805–817, 2013

    2013

  63. [73]

    R. W. Brown, Y .-C. N. Cheng, E. M. Haacke, M. R. Thompson, and R. Venkatesan,Magnetic resonance imaging: physical principles and sequence design. John Wiley & Sons, 2014

    2014

  64. [74]

    On first-order meta-learning algorithms,

    A. Nichol, J. Achiam, and J. Schulman, “On first-order meta-learning algorithms,”arXiv preprint arXiv:1803.02999, 2018

    2018 arXiv

  65. [75]

    Ixi dataset,

    Information eXtraction from Images (IXI) Consortium, “Ixi dataset,” https://brain-development.org/ixi-dataset/, accessed: 2026

    2026

  66. [76]

    Diffusion-based arbitrary- scale magnetic resonance image super-resolution via progressive k-space reconstruction and denoising,

    J. Wang, Z. Shi, X. Gu, Y . Yang, and J. Sun, “Diffusion-based arbitrary- scale magnetic resonance image super-resolution via progressive k-space reconstruction and denoising,”Medical Image Analysis, p. 103814, 2025

    2025

  67. [77]

    Paired 64mt and 3t brain mri scans of healthy subjects for neuroimaging research,

    R. van den Broek, B. Lena, and A. Webb, “Paired 64mt and 3t brain mri scans of healthy subjects for neuroimaging research,” May 2024. [Online]. Available: https://doi.org/10.5281/zenodo.15862148

    2024 doi

  68. [78]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2021, pp. 1833–1844

    2021

  69. [79]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

    2015

  70. [80]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  71. [81]

    Multicontrast image super-resolution with deformable atten- tion and neighborhood-based feature aggregation (dance): Applications in anatomic and metabolic mri,

    W. Chen, S. Wu, S. Wang, Z. Li, J. Yang, H. Yao, Q. Tian, and X. Song, “Multicontrast image super-resolution with deformable atten- tion and neighborhood-based feature aggregation (dance): Applications in anatomic and metabolic mri,”Medical Image Analysis, vol. 99, p. 103359, 2025

    2025

  72. [82]

    Wavtrans: Synergizing wavelet and cross-attention transformer for multi-contrast mri super- resolution,

    G. Li, J. Lyu, C. Wang, Q. Dou, and J. Qin, “Wavtrans: Synergizing wavelet and cross-attention transformer for multi-contrast mri super- resolution,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2022, pp. 463–473

    2022

  73. [83]

    Spatial attention-based implicit neural representation for arbitrary reduction of mri slice spacing,

    X. Wang, S. Wang, H. Xiong, K. Xuan, Z. Zhuang, M. Liu, Z. Shen, X. Zhao, L. Zhang, and Q. Wang, “Spatial attention-based implicit neural representation for arbitrary reduction of mri slice spacing,”Medical Image Analysis, vol. 94, p. 103158, 2024

    2024

  74. [84]

    Mr image denoising and super- resolution using regularized reverse diffusion,

    H. Chung, E. S. Lee, and J. C. Ye, “Mr image denoising and super- resolution using regularized reverse diffusion,”IEEE Transactions on Medical Imaging, vol. 42, no. 4, pp. 922–934, 2023

    2023

  75. [85]

    Image super-resolution via iterative refinement,

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image super-resolution via iterative refinement,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4713– 4726, 2023

    2023

  76. [86]

    Diracdiffusion: Denoising and incremental reconstruction with assured data-consistency,

    Z. Fabian, B. Tinaz, and M. Soltanolkotabi, “Diracdiffusion: Denoising and incremental reconstruction with assured data-consistency,” inPro- JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14 ceedings of the International Conference on Machine Learning (ICML), 2024

    2021

  77. [87]

    Disc-diff: Disentangled con- ditional diffusion model for multi-contrast mri super-resolution,

    Y . Mao, L. Jiang, X. Chen, and C. Li, “Disc-diff: Disentangled con- ditional diffusion model for multi-contrast mri super-resolution,” in International Conference on Medical Image Computing and Computer- Assisted Intervention (MICCAI), 2023, pp. 387–397

    2023

  78. [88]

    Metags: A meta-learned gaussian- phong model for out-of-distribution 3d scene relighting,

    Y . He, Y . Wang, and X. Yang, “Metags: A meta-learned gaussian- phong model for out-of-distribution 3d scene relighting,”arXiv preprint arXiv:2405.20791, 2024

    2024

  79. [89]

    Local texture estimator for implicit representation function,

    J. Lee and K. H. Jin, “Local texture estimator for implicit representation function,” inCVPR, 2022

    2022

  80. [90]

    Scope of validity of psnr in im- age/video quality assessment,

    Q. Huynh-Thu and M. Ghanbari, “Scope of validity of psnr in im- age/video quality assessment,”Electronics Letters, vol. 44, no. 13, pp. 800–801, 2008

    2008

  81. [91]

    Image quality assessment: From error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,”IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  82. [92]

    Image quality assess- ment: Unifying structure and texture similarity,

    K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Image quality assess- ment: Unifying structure and texture similarity,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 5, pp. 2567– 2581, 2022

    2022

  83. [93]

    Mr image reconstruction from highly undersampled k-space data by dictionary learning,

    S. Ravishankar and Y . Bresler, “Mr image reconstruction from highly undersampled k-space data by dictionary learning,”IEEE Transactions on Medical Imaging, vol. 30, no. 5, pp. 1028–1041, 2011

    2011

  84. [94]

    Factors influencing contrast in fast spin-echo mr imaging,

    R. Constable, A. Anderson, J. Zhong, and J. Gore, “Factors influencing contrast in fast spin-echo mr imaging,”Magnetic resonance imaging, vol. 10, no. 4, pp. 497–511, 1992

    1992

  85. [95]

    Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain mri datasets,

    B. Billot, C. Magdamo, Y . Cheng, S. E. Arnold, S. Das, and J. E. Iglesias, “Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain mri datasets,”Proceedings of the National Academy of Sciences, vol. 120, no. 9, p. e2216399120, 2023. S1. ...

    2023

  86. [96]

    Direct application of models trained on simulated data often leads to suboptimal performance on real data

    Motivation:Despite careful design, simulated low-field MRI inevitably differs from real acquisitions due to unmodeled physical factors and scanner-specific characteristics. Direct application of models trained on simulated data often leads to suboptimal performance on real dat...

  87. [97]

    Learning a robust initialization that can quickly adapt to new degradation domains with minimal samples

  88. [98]

    Leveraging limited real paired data as target-domain guidance during meta-training

  89. [99]

    Modeling domain variability through task-based episodic training across a spectrum of simulated and real degra- dation conditions

  90. [100]

    This enables the model to learn domain-agnostic rep- resentations that generalize across varying imaging conditions

    Task Formulation:We formulate domain adaptation as a meta-learning problem where eachtaskcorresponds to a distinct MRI degradation domain characterized by specific blur, noise, and resolution parameters, rather than a semantic category. This enables the model to learn domain-a...

  91. [101]

    Episode Sampling: In each meta-training episode, we sample a task batchB={T b1 ,

    Episodic Meta-Optimization:We adopt a first-order Model-Agnostic Meta-Learning (MAML) framework, where the meta-objective is to find a model initializationθthat can be rapidly adapted to any task with a small number of gradient steps. Episode Sampling: In each meta-training ep...

    2021

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.