Pith. sign in

REVIEW 4 major objections 5 minor 32 references

Unconditional MRI priors can be steered at inference time by a secondary scan, reducing hallucinations without retraining the generative model.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 19:01 UTC pith:WBOITVMM

load-bearing objection A genuinely novel and clinically sensible idea undercut by an uncontrolled comparison; the hallucination-reduction claim needs matched-quality baselines and a BraTS ablation before it holds. the 4 major comments →

arxiv 2603.03710 v3 pith:WBOITVMM submitted 2026-03-04 cs.CV cs.AI

MPFlow: Multi-modal Posterior-Guided Flow Matching for Zero-Shot MRI Reconstruction

Seunghoi Kim , Chen Jin , Henry F. J. Tregidgo , Matteo Figini , Daniel C. Alexander This is my paper
classification cs.CV cs.AI
keywords MRI reconstructionzero-shotflow matchinghallucinationmulti-modalself-supervised learningcontrastive pretraininginverse problems
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that a generative prior for MRI reconstruction, trained without any auxiliary information, can be guided during sampling by a second MRI contrast that is already routinely acquired. The steering is done by a self-supervised encoder pair, PAMRI, that aligns image patches across modalities; at inference, the flow model's velocity is adjusted by a term that pulls the current reconstruction toward the auxiliary image in that shared latent space. The authors argue this jointly suppresses intrinsic hallucinations (violating the measurements) and extrinsic hallucinations (measurement-consistent but anatomically wrong), while matching diffusion-quality images in 20% of the sampling steps. A sympathetic reader would care because zero-shot reconstruction currently has no simple way to exploit the multi-sequence scans that are standard in clinical practice.

Core claim

A rectified flow prior, left completely unconditional, can be given multi-modal conditioning at inference time through a self-supervised patch-level alignment network. The paper introduces PAMRI, which learns a shared latent space between target and auxiliary MRI contrasts using an adaptive contrastive loss and a reconstruction regularizer, and then adds a cross-modal feature-matching term to the flow ODE alongside data consistency. The authors show that this reduces both intrinsic and extrinsic hallucinations, with tumor segmentation Dice improving by more than 15 percentage points over a strong diffusion baseline and image quality matching that baseline at one-fifth of the sampling steps.

What carries the argument

PAMRI (Patch-level Multi-modal MR Image Pretraining) is the mechanism that carries the argument: a pair of independent encoders maps target and auxiliary image patches into a shared space, trained by an InfoNCE-style loss with an adaptive temperature set by normalized mutual information, plus a patch-reconstruction loss to preserve spatial detail. At inference, the rectified flow update becomes the prior velocity plus a data-consistency gradient and a gradient of feature-matching loss between the current clean-image estimate and the auxiliary image; a noise-optimization step selects the best initial latent. This combination makes the unconditional prior condition on the auxiliary modality wi

Load-bearing premise

The load-bearing premise is that PAMRI encoders, trained on clean paired patches, provide a feature space in which minimizing the alignment loss on partially reconstructed, out-of-distribution intermediate images pulls the reconstruction toward the true target rather than toward auxiliary-specific appearance.

What would settle it

A direct falsifier would be an experiment where the auxiliary image is replaced by its own ground-truth-inconsistent variant (e.g., T1 with the tumor region removed or swapped from another subject): if the method still reduces hallucinations as measured by Dice or SHAFE, then the cross-modal alignment is not actually using the auxiliary anatomical information. Alternatively, measuring whether the alignment loss gradient decreases the distance to the ground truth at early flow times would test the core assumption.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, any existing unconditional flow or diffusion prior for MRI can leverage available multi-contrast scans without retraining, making multi-modal zero-shot reconstruction practical in clinical pipelines.
  • The method's efficiency at T=100 steps suggests that near-linear flow trajectories combined with lightweight cross-modal guidance could enable near-real-time reconstruction.
  • The reported reduction in extrinsic hallucinations, measured by tumor segmentation Dice and SHAFE, implies that measurement-consistent but anatomically incorrect structures can be suppressed by purely latent-space alignment.
  • The benefit scaling with task severity (larger gains at 8x versus 4x super-resolution) indicates that cross-modal guidance is most valuable exactly when the null space is largest.
  • The framework is task-agnostic within the degradation model, covering both k-space subsampling and spatial super-resolution, so it may generalize to other inverse problems in medical imaging.

Where Pith is reading between the lines

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

  • A testable extension would be feeding a deliberately mismatched auxiliary image (e.g., from a different patient) to see whether the guidance still reduces hallucinations; if it does, the effect could be due to generic feature smoothing rather than true anatomical correspondence.
  • The information-theoretic justification suggests the same conditional-mutual-information argument could apply to other paired observations (e.g., PET, CT, or functional MRI), and one could expect larger gains when the auxiliary modality carries more independent anatomical information.
  • Because PAMRI is self-supervised on patches, the approach might transfer to modalities where paired training labels are scarce, but the encoders' sensitivity to distribution shift at intermediate flow times is untested and could be probed by measuring alignment loss versus ground-truth distance during sampling.

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 / 5 minor

Summary. The paper proposes MPFlow, a zero-shot multi-modal MRI reconstruction framework built on rectified flow. A PAMRI module is pre-trained on paired target/auxiliary patches with an adaptive InfoNCE loss, then used at inference time to guide the flow prior jointly with data consistency. The model is evaluated on HCP 4× super-resolution and BraTS 8× k-space reconstruction, with T1 as the auxiliary modality. The authors report that MPFlow matches or exceeds diffusion baselines in image quality while using substantially fewer sampling steps, and reduces tumor hallucinations as measured by segmentation Dice and SHAFE scores. A theoretical justification is given via a conditional-entropy identity, and ablations/numerical analyses investigate the contribution of each component.

Significance. If the reported benefits hold, MPFlow would be a practically valuable framework because it exploits routinely acquired auxiliary MRI scans without retraining the unconditional generative prior. The core idea—inference-time cross-modal guidance through a self-supervised feature-alignment module—is novel and well motivated by the clinical protocol. The paper is honest in reporting error bars and p-values, and the component ablations in Table 3a isolate individual contributions. However, the headline hallucination-reduction claim is currently confounded by the comparison settings, and the theoretical/experimental support for the PAMRI guidance mechanism is incomplete. The paper does not provide code, data, or a matched-quality baseline comparison, which limits immediate reproducibility and verification.

major comments (4)
  1. [Tables 1 and 2, §3.2–3.3] The hallucination metrics in Table 2 are reported only at T=100. In Table 1, DynamicDPS at T=100 has SSIM 0.68 (BraTS) and 0.69 (HCP) with high variance, whereas its T=500 run reaches 0.89/0.79. MPFlow at T=100 reaches 0.90/0.78, i.e., it matches or beats the T=500 diffusion baseline on image quality. Therefore the Dice improvement of 0.740 vs 0.639 (claimed >15%) could stem largely from the rectified-flow prior's robustness to large step sizes rather than from PAMRI's cross-modal guidance. Please report hallucination metrics at a matched sampling/computation budget (e.g., all methods at T=500, or MPFlow at a comparable computational cost) and include T=500 hallucination numbers for MPFlow and DynamicDPS. Without this, the claim that auxiliary guidance reduces hallucinations is not separated from the flow model's efficiency advantage.
  2. [Table 3a, §3.3] The component ablation isolating PAMRI is shown only for HCP. The BraTS tumor Dice claim in the abstract and §3.3 is central to the paper's clinical motivation, but Table 2b compares full MPFlow to vanilla MPFlow without a BraTS ablation table. Add a BraTS ablation (base vs +PAMRI vs +Noise Optimization) reporting Dice and measurement-space loss. This is necessary to verify that the auxiliary modality, rather than the noise-optimization heuristic alone, drives the tumor-hallucination reduction.
  3. [§2.2, Eq. (2); §2.5, Eq. (8)] The theoretical justification in Eq. (2) is the standard conditional-entropy identity H(x|y,x_aux)=H(x|y)−I(x;x_aux|y). It applies to the true posterior and does not by itself justify the particular gradient update in Eq. (8). No derivation is given showing that λP L_P(φ(P(\hat{x}_{1|t})), ψ(P(x_aux))) approximates the conditional posterior p(x|y,x_aux), nor that the MSE in feature space is calibrated to the mutual information term. Please clarify the status of Eq. (8) as a heuristic and provide either a derivation of the posterior approximation or an empirical validation that the update behaves as intended (e.g., by comparing against an oracle posterior-sampling variant on a small task).
  4. [§2.4 and §2.5] PAMRI encoders are trained on clean paired patches, but at inference they are applied to intermediate estimates \hat{x}_{1|t}, which are partially reconstructed and out-of-distribution. The paper does not test whether the learned alignment transfers to such inputs. If the feature-space MSE responds to auxiliary-specific appearance rather than to shared anatomy, the large measurement-space decreases in Table 2 could be misleading. Please add an experiment that either evaluates PAMRI's feature alignment on intermediate ODE states, or compares guidance using features from clean and degraded reconstructions, to support the transfer assumption.
minor comments (5)
  1. [Table 1] The column header 'Acc.×8' should be expanded to 'Acceleration ×8' and the k-space sampling pattern (e.g., Cartesian with center fraction) should be specified. The p-value row is ambiguous; clarify which column it refers to and how it is computed for each metric.
  2. [Tables 2a/2b] Unlike Table 1, Table 2 does not report standard deviations or error bars. Please include them, especially since the differences in SHAFE and Dice may be within variability.
  3. [§3.1] The baselines do not include FlowDPS [13], despite the reference being present. Since MPFlow is flow-based, a comparison with FlowDPS would help isolate the contribution of the flow prior from the cross-modal guidance. Please add this baseline or explain its omission.
  4. [§2.4] PAMRI training details are sparse: no number of training epochs, batch size, optimizer, learning rate, or augmentation strategy for patch extraction. These details are necessary for reproducibility.
  5. [General] The paper states 'Code will be released upon acceptance' but no code or data is provided. For a zero-shot method whose value depends on transfer, releasing the pretrained PAMRI checkpoints and sampling code would substantially improve verifiability.

Circularity Check

0 steps flagged

No significant circularity; self-citations (DynamicDPS, SHAFE) are present but not load-bearing.

full rationale

The paper's central claims are empirically tested against external or independent metrics rather than reducing to fitted inputs or self-referential definitions. The entropy identity in Eq. (2) is a standard information-theoretic identity; the assumption I(x;x_aux|y)>0 is a plausible empirical premise, not constructed to force the result. The cross-modal guidance in Eq. (8) minimizes a PAMRI feature-space loss, but the reported hallucination metrics are not defined in terms of that loss: Dice uses a Swin-UNet trained on FLAIR, SHAFE uses a pre-trained DINO encoder, and measurement-space loss is directly computed from the forward model. Thus the evaluation is not circular with respect to the method's own objective. The ablations in Tables 3a and 3b (removing PAMRI or noise optimization, and scaling of benefit with task severity) provide genuine falsifiable tests. The self-citations to the authors' prior work (DynamicDPS [15], SHAFE [16]) are used as a baseline and a metric, respectively, but they are not the load-bearing premises of the derivation; the claim would stand on the Dice and measurement-space results even without SHAFE. The concern that hallucination metrics are only reported at T=100 where diffusion baselines are degraded is an experimental fairness issue, not a circularity. Overall, no equation reduces to its input, and no fitted parameter is renamed as a prediction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The free parameters are standard hyperparameters, but alpha_t and tau bounds are unreported, which hampers reproduction. The main load-bearing axioms are the domain assumption of complementary registered modalities and the unverified OOD transfer of PAMRI features.

free parameters (6)
  • lambda_P = 0.1
    Weight of PAMRI guidance in Eq (8); chosen by hand; no sensitivity analysis reported.
  • lambda_rec = 0.5
    Weight of reconstruction loss in PAMRI objective (Eq 7).
  • S = 8
    Number of candidate noise seeds in warm-start selection (Eq 9).
  • t_noise = 0.2T
    Fraction of sampling steps used for noise optimization.
  • alpha_t = not reported
    Guidance step-size schedule in Eq (8); never specified, yet central to data-consistency strength.
  • tau_min, tau_max = not reported
    Temperature bounds for adaptive InfoNCE in Sec 2.4; not reported.
axioms (5)
  • domain assumption Degradation model y = F(x) + eta with Gaussian noise holds (Eq 1).
    Standard in inverse problems; forward operator known exactly, no model error or motion. Used throughout experiments.
  • domain assumption Registered MR modalities encode overlapping anatomical information, so I(x; x_aux | y) > 0 (Sec 2.2).
    Theoretical justification for the benefit of the auxiliary modality. True in well-registered scans, but registration is not addressed.
  • standard math Conditional entropy identity H(x|y,x_aux) = H(x|y) - I(x;x_aux|y) (Eq 2).
    Standard information-theoretic identity; valid as stated.
  • ad hoc to paper PAMRI encoders, trained on clean paired patches, provide a distortion-invariant shared space that transfers to partial reconstructions at intermediate flow time steps.
    The guidance term in Eq (8) relies on this; not verified in the paper. This is the weakest assumption.
  • domain assumption The trained rectified flow prior approximates the target data distribution well enough for the posterior guidance to converge to meaningful solutions.
    Standard for generative-prior inverse problems; no convergence guarantees are provided.

pith-pipeline@v1.3.0-alltime-deepseek · 7982 in / 14096 out tokens · 123954 ms · 2026-08-02T19:01:12.947977+00:00 · methodology

0 comments
read the original abstract

Zero-shot MRI reconstruction relies on generative priors, but single-modality unconditional priors produce hallucinations under severe ill-posedness. In many clinical workflows, complementary MRI acquisitions (e.g. high-quality structural scans) are routinely available, yet existing reconstruction methods lack mechanisms to leverage this additional information. We propose MPFlow, a zero-shot multi-modal reconstruction framework built on rectified flow that incorporates auxiliary MRI modalities at inference time without retraining the generative prior to improve anatomical fidelity. Cross-modal guidance is enabled by our proposed self-supervised pretraining strategy, Patch-level Multi-modal MR Image Pretraining (PAMRI), which learns shared representations across modalities. Sampling is jointly guided by data consistency and cross-modal feature alignment using pre-trained PAMRI, systematically suppressing intrinsic and extrinsic hallucinations. Extensive experiments on HCP and BraTS show that MPFlow matches diffusion baselines on image quality using only 20% of sampling steps while reducing tumor hallucinations by more than 15% (segmentation dice score). This demonstrates that cross-modal guidance enables more reliable and efficient zero-shot MRI reconstruction.

Figures

Figures reproduced from arXiv: 2603.03710 by Chen Jin, Daniel C. Alexander, Henry F. J. Tregidgo, Matteo Figini, Seunghoi Kim.

Figure 1
Figure 1. Figure 1: Schematic diagram of the proposed framework. (a) PAMRI: cross￾modal feature alignment by learning a shared representation space across dif￾ferent MRI contrasts, (b) MPFlow: flow-matching prior is guided by both data consistency and PAMRI, to reduce intrinsic and extrinsic hallucinations. 20], where a learned prior is guided by data consistency during sampling, elim￾inating the need for paired supervision. … view at source ↗
Figure 2
Figure 2. Figure 2: presents visual comparisons of MPFlow against baseline methods. As highlighted in the red boxes, baselines introduce hallucinated structures includ￾ing distorted sulci and incorrect tumor morphology, which can directly compro￾mise clinical decisions such as surgical planning and radiotherapy contouring. In contrast, MPFlow preserves anatomically faithful structures with notably sharper tumor boundaries, de… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

32 extracted references · 5 linked inside Pith

  1. [1]

    NeuroImage152, 283–298 (2017)

    Alexander, D.C., Darko, Z., Ghosh, A., Tanno, R., Wottschel, V., et al.: Image quality transfer and applications in diffusion mri. NeuroImage152, 283–298 (2017)

    2017

  2. [2]

    IEEE Transactions on Medical Imaging40(11), 3249–3260 (2021)

    Bhadra, S., Kelkar, V.A., Brooks, F.J., Anastasio, M.A.: On hallucinations in to- mographic image reconstruction. IEEE Transactions on Medical Imaging40(11), 3249–3260 (2021)

    2021

  3. [3]

    In: Proceedings of the European Conference on Computer Vision Workshops(ECCVW) (2022)

    Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin-unet: Unet-like pure transformer for medical image segmentation. In: Proceedings of the European Conference on Computer Vision Workshops(ECCVW) (2022)

    2022

  4. [4]

    arXiv preprint arXiv:2002.05709 (2020)

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. arXiv preprint arXiv:2002.05709 (2020)

    Pith/arXiv arXiv 2002

  5. [5]

    In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025 (2025)

    Choi, Y., Jung, S., Al-masni, M.A., Yang, M.H., Kim, D.H.: TESLA: Test-time Reference-free Through-plane Super-resolution for Multi-contrast Brain MRI . In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025 (2025)

    2025

  6. [6]

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

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

    2023

  7. [7]

    McNabb, C., Palombo, M., Evans, C.J., Cercignani, M., Jones, D.K., C

    Cicimen, A.G., Tregidgo, H.F.J., Figini, M., Messaritaki, E., B. McNabb, C., Palombo, M., Evans, C.J., Cercignani, M., Jones, D.K., C. Alexander, D.: Im- age quality transfer of diffusion mri guided by high-resolution structural mri. In: Computational Diffusion MRI (2025)

    2025

  8. [8]

    Imaging Neuroscience 2, 1–22 (2024)

    Gopinath, K., Hoopes, A., Alexander, D.C., Arnold, S.E., Balbastre, Y., et al.: Syn- thetic data in generalizable, learning-based neuroimaging. Imaging Neuroscience 2, 1–22 (2024)

    2024

  9. [9]

    arXiv preprint arXiv:1512.03385 (2015)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385 (2015)

    Pith/arXiv arXiv 2015

  10. [10]

    Science Advances9(5) (2023)

    Iglesias, J.E., Billot, B., Balbastre, Y., Magdamo, C., Arnold, S.E., et al.: Synthsr: A public ai tool to turn heterogeneous clinical brain scans into high-resolution t1-weighted images for 3d morphometry. Science Advances9(5) (2023)

    2023

  11. [11]

    Computer Methods and Programs in Biomedicine248, 108110 (2024)

    Kang, L., Tang, B., Huang, J., Li, J.: 3d-mri super-resolution reconstruction using multi-modality based on multi-resolution cnn. Computer Methods and Programs in Biomedicine248, 108110 (2024)

    2024

  12. [12]

    In: Advances in Neural Information Processing Systems (2022)

    Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. In: Advances in Neural Information Processing Systems (2022)

    2022

  13. [13]

    Kim, J., Kim, B.S., Ye, J.C.: Flowdps: Flow-driven posterior sampling for inverse problems (2025)

    2025

  14. [14]

    In: European Conference on Computer Vision (ECCV) (2024) 10 Kim et al

    Kim, S., Jin, C., Diethe, T., Figini, M., Tregidgo, H.F.J., et al.: Tackling structural hallucination in image translation with local diffusion. In: European Conference on Computer Vision (ECCV) (2024) 10 Kim et al

    2024

  15. [15]

    In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025 (2025)

    Kim, S., Tregidgo, H.F.J., Figini, M., Jin, C., Joshi, S., Alexander, D.C.: Tackling Hallucination from Conditional Models for Medical Image Reconstruction with Dy- namicDPS . In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025 (2025)

    2025

  16. [16]

    Kim, S., Tregidgo, H.F.J., Jin, C., Figini, M., Alexander, D.C.: Hallugen: Syn- thesizing realistic and controllable hallucinations for evaluating image restoration (2025), https://arxiv.org/abs/2512.03345

    arXiv 2025

  17. [17]

    arXiv preprint arXiv:2311.06631 (2023)

    Kim,S.,Tregidgo,H.F.,Eldaly,A.K.,Figini,M.,Alexander,D.C.:A3dconditional diffusion model for image quality transfer–an application to low-field mri. arXiv preprint arXiv:2311.06631 (2023)

    Pith/arXiv arXiv 2023

  18. [18]

    In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (2023)

    Li, G., Zhao, L., Sun, J., Lan, Z., Zhang, Z., Chen, J., Lin, Z., Lin, H., Xing, W.: Rethinking multi-contrast mri super-resolution: Rectangle-window cross-attention transformer and arbitrary-scale upsampling. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (2023)

    2023

  19. [19]

    Medical Image Analysis87, 102807 (2023)

    Lin, H., Figini, M., D’Arco, F., Ogbole, G., Tanno, R., et al.: Low-field magnetic resonance image enhancement via stochastic image quality transfer. Medical Image Analysis87, 102807 (2023)

    2023

  20. [20]

    In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024 (2024)

    Lin, X., Du, C., Wu, Q., Tian, X., Yu, J., et al.: Zero-shot Low-field MRI En- hancement via Denoising Diffusion Driven Neural Representation . In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024 (2024)

    2024

  21. [21]

    arXiv preprint arXiv:2209.03003 (2022)

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003 (2022)

    Pith/arXiv arXiv 2022

  22. [22]

    IEEE Transac- tions on Medical Imaging34(10), 1993–2024 (2015)

    Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K.e.a.: The multimodal brain tumor image segmentation benchmark (brats). IEEE Transac- tions on Medical Imaging34(10), 1993–2024 (2015)

    1993

  23. [23]

    arXiv preprint arXiv:2412.00100 (2024)

    Patel, M., Wen, S., Metaxas, D.N., Yang, Y.: Steering rectified flow models in the vector field for controlled image generation. arXiv preprint arXiv:2412.00100 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., Jégou, H., Labatut, P., Bojanowski, P.: DINOv3 (2025)

    2025

  25. [25]

    NeuroImage80, 125–143 (2013)

    Sotiropoulos, S.N., Jbabdi, S., Xu, J., Andersson, J.L.R., Moeller, S., et al.: Advances in diffusion mri acquisition and processing in the human connectome project. NeuroImage80, 125–143 (2013)

    2013

  26. [26]

    ArXiv (2020)

    Sriram, A., Zbontar, J., Murrell, T., Defazio, A., Zitnick, C.L., Yakubova, N., Knoll, F., Johnson, P.M.: End-to-end variational networks for accelerated mri re- construction. ArXiv (2020)

    2020

  27. [27]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2018)

    Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2018)

    2018

  28. [28]

    Wang, J., Chen, Y., Wu, Y., Shi, J., Gee, J.: Enhanced generative adversarial net- workfor3dbrainmrisuper-resolution.In:IEEE/CVFWinterConference(WACV) (2020)

    2020

  29. [29]

    The Eleventh International Conference on Learning Representa- tions (2023)

    Wang, Y., Yu, J., Zhang, J.: Zero-shot image restoration using denoising diffusion null-space model. The Eleventh International Conference on Learning Representa- tions (2023)

    2023

  30. [30]

    In: Proceedings of the 32nd ACM International Conference on Multimedia (2024) Title Suppressed Due to Excessive Length 11

    Wu, H., He, L., Zhang, M., Chen, D., Luo, K., Luo, M., Zhou, J.Z., Chen, H., Lv, J.: Diffusion posterior proximal sampling for image restoration. In: Proceedings of the 32nd ACM International Conference on Multimedia (2024) Title Suppressed Due to Excessive Length 11

    2024

  31. [31]

    Tomography8, 905 – 919 (2021)

    Zhang, K., Hu, H., Philbrick, K.A., Conte, G.M., Sobek, J.D., et al.: Soup-gan: Super-resolution mri using generative adversarial networks. Tomography8, 905 – 919 (2021)

    2021

  32. [32]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effec- tiveness of deep features as a perceptual metric. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    2018