Pith. sign in

REVIEW 4 major objections 5 minor 53 references

Self-supervised feature learning for cardiac Cine MR image reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A self-supervised MRI reconstruction method matches supervised learning up to 16x undersampling, using only undersampled training data—no fully-sampled references needed.

desk verdict The performance claim holds up on the reported data, but the key ablation conflates extra capacity with feature learning, so the mechanism is underdetermined. read the letter →

arxiv 2505.23408 v1 pith:SU33LSKQ submitted 2025-05-29 eess.IV

Siying Xu , Marcel Früh , Kerstin Hammernik , Andreas Lingg , Jens Kübler , Patrick Krumm , Daniel Rueckert , Sergios Gatidis
show 1 more author
Thomas Küstner
This is my paper · ORCID
classification eess.IV
keywords self-supervisedlearningMRIreconstructioncardiacCinefeaturecontrastiveVICRegundersampledk-spaceunrollednetwork
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

The paper proposes SSFL-Recon, a two-step self-supervised framework for reconstructing cardiac Cine MRI from undersampled data alone. In the first step, a feature extractor is trained on undersampled images to learn representations that are insensitive to the undersampling pattern; in the second, those features are fed into a self-supervised reconstruction network. The paper claims that this approach outperforms existing self-supervised reconstruction methods and matches or beats supervised learning up to 16× retrospective undersampling, while using no fully-sampled references during training. If true, this would let clinical sites exploit the abundant undersampled acquisitions that are normally discarded, and would reduce the bias introduced by conventionally reconstructed 'fully-sampled' training targets.

What carries the argument

The load-bearing mechanism is the feature extraction network (FE-Net), an unrolled network whose encoder features are trained so that two randomly re-undersampled views of the same subject map to nearby embeddings, either by a contrastive InfoNCE loss or by a variance-invariance-covariance (VICReg) loss. The pre-trained features are then concatenated into the bottleneck of each iterative block of a separate unrolled reconstruction network, giving it anatomical guidance that is shared across sampling patterns. The reconstruction is driven by a self-supervised loss made of an image-consistency mean squared error between the two reconstructed views and a cross k-space mean absolute error that checks each reconstruction against the other view's sampled points.

What would settle it

Train FE-Net on positive pairs generated from a fully-sampled k-space so the two views share no common 2× accelerated baseline, and compare 16× reconstruction quality against the same network without features; if the gain disappears, the features were encoding the shared accelerated baseline rather than sampling-insensitive anatomy.

Watch

Extended reading notes

Core claim

The central discovery is that pre-learning sampling-insensitive features from undersampled cardiac Cine images, and concatenating those features into each iteration of an unrolled self-supervised reconstruction network, yields reconstructions that are artifact-free up to 18× and usable at 21× acceleration. On the test set at 16× acceleration, SSFL-Recon(c) reaches a mean PSNR of 31.344 dB and SSIM of 0.884, slightly above the supervised baseline at 30.955 dB and 0.881; at 8× the margin over supervised learning is larger (36.325 dB vs 35.326 dB PSNR, SSIM 0.954 vs 0.947). Ablation without the feature-learning step (SSL-Recon) falls to 25.371 dB PSNR at 16×, so the feature step carries most of the gain. The same pre-trained networks generalize to a prospectively undersampled, free-breathing dataset from a different scanner without fine-tuning.

Load-bearing premise

The method assumes that two differently re-undersampled views of the same subject share a latent representation that is about the subject's anatomy, not about the common accelerated reconstruction or its artifact pattern.

Editorial extensions

If this is right

  • Training on undersampled data alone becomes sufficient for high-quality cardiac Cine reconstruction, removing the need for fully-sampled references that are often impractical under respiratory or cardiac motion.
  • The feature-learning step nearly doubles 16× PSNR relative to the same reconstruction network without features, so the margin over earlier self-supervised methods is attributable to the features, not the unrolled backbone alone.
  • The method supports a wide range of acceleration factors in a single model, from 2× to 21×, because the training masks and re-undersampling rates are randomized.
  • The information-maximization variant (VICReg) matches contrastive learning without negative pairs, reducing memory and computational cost.
  • Pre-trained features transfer to prospectively undersampled data from a different scanner without fine-tuning, indicating generalization across subjects and acquisitions.

Reading between the lines

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

  • Inference: the success of the feature-learning step suggests that enforcing feature-level consistency is the active ingredient; a direct test would be to add the feature-consistency term to the reconstruction loss itself and measure the gain.
  • Inference: because the shared-latent hypothesis is about anatomy rather than sampling, the framework should extend to non-Cartesian trajectories and other organs, but the paper only demonstrates Cartesian VISTA sampling for cardiac Cine.
  • Inference: the observed transfer to a different scanner raises the possibility that a single pre-trained FE-Net could serve as a reusable representation for several downstream tasks, such as segmentation or functional analysis, beyond reconstruction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes SSFL-Recon, a two-stage self-supervised framework for dynamic cardiac Cine MRI reconstruction that requires only undersampled k-space data. In the first stage, a feature extractor (FE-Net) is trained on pairs of re-undersampled views of the same subject using either contrastive (InfoNCE) or information-maximization (VICReg) losses. In the second stage, the pretrained FE-Net features are concatenated into a physics-based unrolled reconstruction network trained with a self-supervised loss combining image-domain consistency and cross k-space fidelity. The authors evaluate on an in-house 129-subject cardiac Cine dataset with retrospective VISTA undersampling at 8x and 16x, and qualitatively on a prospectively undersampled OCMR slice. They report that SSFL-Recon outperforms SSDU and PARCEL and is comparable to or better than a supervised baseline, with an ablation (SSL-Recon) showing a large gain attributed to the feature-learning step.

Significance. If the mechanism is as claimed, the work is a useful contribution: it provides a way to exploit abundant undersampled clinical data, is evaluated on a sizable in-house cardiac cohort, includes ablations and statistical tests, and releases source code. The two feature-learning variants (contrastive and VICReg) are sensible, and the cross k-space loss is a reasonable self-supervised target. However, the paper's central mechanistic attribution—that the gain comes from learning sampling-insensitive features—is not cleanly isolated by the current ablation, and several broad claims (outperforming 'existing' SSL methods, transferability) are not fully supported by the presented experiments.

major comments (4)
  1. [IV.C, Table II] The ablation SSL-Recon removes the entire FE-Net, so the large performance gap at 16x (PSNR 25.371 dB vs 31.344 dB in Table II) simultaneously removes the added 320,432-parameter feature pathway, the fixed auxiliary features, and the specific self-supervised pretraining objective. This confound does not establish the central hypothesis of Section II.A that sampling-insensitive feature learning is the cause of the improvement. A control with a randomly initialized and frozen FE-Net, or an FE-Net pretrained with an objective that does not enforce invariance across re-undersampling views, is needed before the gain can be attributed to the proposed feature-learning mechanism rather than to additional capacity or a generic low-pass regularization.
  2. [Abstract, IV.B] The claim that SSFL-Recon 'outperforms existing self-supervised MRI reconstruction methods' is supported only against two SSL baselines (SSDU and PARCEL). Figure 4 and Table II establish superiority over those two methods on the in-house test set, but not over 'existing' methods generally. Please either restrict the claim to the evaluated baselines or add additional recent self-supervised reconstruction baselines (e.g., multi-mask SSDU, Noise2Recon, DDSS, or a public benchmark) to support the stronger statement.
  3. [V.A] The transferability experiment described in Section V.A—training FE-Net and the reconstruction network on disjoint subsets of the dataset—is asserted to yield 'comparable performance,' but no quantitative or qualitative results for this experiment appear in the manuscript. Since this is presented as evidence for the general applicability of the feature extractor, the claim should be substantiated with a table or figure (or removed).
  4. [IV.B.2, Table I] The prospective OCMR evaluation is qualitative only (Fig. 6), and the comparison to supervised learning there is not quantified. In addition, Table I reports 25 h of training for the supervised baseline versus 96 h for the SSFL-Recon reconstruction step; if both are trained for the same number of epochs, this discrepancy needs explanation, and if the optimization budgets differ, the fairness of the comparison should be addressed.
minor comments (5)
  1. [II.B.1, Eq. (16)] The text states that the cross k-space loss is MAE, but Eq. (16) is a smoothed L2 norm; please align the equation and the description.
  2. [II.A.1, Eq. (4)] The denominator in Eq. (4) should presumably contain exponentiated similarity terms, exp(sim(...)/τ), rather than raw similarities; please clarify the notation or define sim accordingly.
  3. [Figs. 7 and 8] Both figures show duplicated subfigure labels: each has an '(a) Ablation Study' and an '(a) Feature Map Visualization'; please correct the labels and captions.
  4. [Table II] The footnote indicates asterisks mark significant differences with the best method, but bold entries are not visible in the text; please clarify which method is best per metric at each acceleration.
  5. [III.B] There is a typo: 'initilized' should be 'initialized'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is an empirical comparison against held-out fully-sampled references and external baselines, and the self-supervised losses do not encode the evaluation metric.

full rationale

The derivation chain is self-contained rather than circular. The feature extractor is trained with Eq. (4) (InfoNCE) or Eq. (5) (VICReg) to make embeddings of two re-undersampled views of the same 2x-accelerated image similar; this objective does not by construction produce the reported PSNR/SSIM values, which are computed against fully-sampled references on a held-out test set. The reconstruction loss in Eqs. (13)-(16) is a self-supervised consistency loss over two re-undersampled masks, enforcing agreement between the two reconstructed images and k-space fidelity at sampled locations, but it is not the evaluation metric and is the same class of loss used by the external SSL baselines (SSDU, PARCEL). The claim that feature learning helps is supported by an ablation against SSL-Recon, but that ablation removes the entire FE-Net (320,432 parameters), so it conflates the feature-learning objective with added capacity and a fixed auxiliary feature path; this is a validity/confound limitation rather than circularity, because the ablated gap is not forced by the training objective. The authors' own prior work (CINENet [19], MERLIN [49]) appears as architecture and implementation support only and is not invoked as a uniqueness theorem or as proof of the reported gains. Hence no step in the derivation reduces to its own inputs.

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

The framework has no analytic derivation; it is an empirical neural network method. The listed assumptions are the modeling choices on which the reported performance rests: accurate sensitivity maps, valid positive pairs, distributional match between training and test masks, and trustworthy references. No new physical entities are introduced.

free parameters (4)
  • VICReg loss weights lambda, mu, nu = lambda=25, mu=25, nu=1
    Hand-selected hyperparameters in Eq. (5); no ablation or sensitivity analysis is provided, but they follow common VICReg practice and do not encode the target result.
  • InfoNCE temperature tau = not stated
    Temperature in Eq. (4) affecting contrastive loss sharpness; no value or tuning procedure reported.
  • Unrolled iterations n = 3
    Network depth chosen for both FE-Net and reconstruction; no ablation on n is shown.
  • Initial data consistency step size lambda = 1
    Lambda in Eq. (17) is trainable but initialized to 1; the choice is not justified by experiments.
assumptions (4)
  • domain assumption The forward model y = A x with coil sensitivities S estimated by ESPiRIT is an accurate description of the acquisition.
    Used throughout Sections II-A and II-B for re-undersampling, data consistency, and loss computation; inaccurate sensitivities would bias all reconstructions.
  • ad hoc to paper Two random re-undersampled views of the same subject share a latent feature representation that is invariant to aliasing artifacts.
    Section II-A states this as the core hypothesis motivating FE-Net; if false, the pretrained features would not transfer or would encode artifacts.
  • domain assumption Undersampling masks used in training (VISTA distribution, random rates) are representative of inference masks.
    Section III-A keeps VISTA distribution for training and testing, following Noiser2Noise findings in [28]; the generalization claim depends on this distributional match.
  • domain assumption The fully-sampled reference images used for evaluation and supervised training are reliable ground truths.
    Evaluation and the supervised baseline depend on the accuracy of the reference; the paper notes many fully-sampled datasets are actually conventional reconstructions of mildly accelerated data, and this dataset's reference provenance is not fully detailed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-supervised feature learning for cardiac Cine MR image reconstruction." pith.science (2026). https://pith.science/paper/SU33LSKQ

@misc{pith2026250523408,
  author       = {Pith},
  title        = {Pith review of: Self-supervised feature learning for cardiac Cine MR image reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SU33LSKQ}},
  note         = {Machine review of arXiv:2505.23408}
}
abstract

We propose a self-supervised feature learning assisted reconstruction (SSFL-Recon) framework for MRI reconstruction to address the limitation of existing supervised learning methods. Although recent deep learning-based methods have shown promising performance in MRI reconstruction, most require fully-sampled images for supervised learning, which is challenging in practice considering long acquisition times under respiratory or organ motion. Moreover, nearly all fully-sampled datasets are obtained from conventional reconstruction of mildly accelerated datasets, thus potentially biasing the achievable performance. The numerous undersampled datasets with different accelerations in clinical practice, hence, remain underutilized. To address these issues, we first train a self-supervised feature extractor on undersampled images to learn sampling-insensitive features. The pre-learned features are subsequently embedded in the self-supervised reconstruction network to assist in removing artifacts. Experiments were conducted retrospectively on an in-house 2D cardiac Cine dataset, including 91 cardiovascular patients and 38 healthy subjects. The results demonstrate that the proposed SSFL-Recon framework outperforms existing self-supervised MRI reconstruction methods and even exhibits comparable or better performance to supervised learning up to $16\times$ retrospective undersampling. The feature learning strategy can effectively extract global representations, which have proven beneficial in removing artifacts and increasing generalization ability during reconstruction.

Figures

Figures reproduced from arXiv: 2505.23408 by the authors.

Figure 1
Figure 1. The first step of the proposed SSFL-Recon framework: Self-supervised feature learning. Input images x1,u and x2,u are generated from the same subject (R = 2 conventional parallel imaging undersampled image x1) with two different sampling masks (contrastive learning: same acceleration rate but random generation seeds, information maximization method: random acceleration rate). The input image x3,u is additionally nee… view at source ↗
Figure 2
Figure 2. The second step of the proposed SSFL-Recon framework: Self-supervised reconstruction. (a) Inputs generation (training): two undersampled images, x1 and x2, are obtained by re-undersampling the accelerated k-space y. Dashed arrows indicate the correspondence between the k-space and the image domain. (b) Masks illustration of training inputs: the initially undersampled k-space y corresponds to the undersampling mask M… view at source ↗
Figure 3
Figure 3. Reconstructions in spatial (x-y) and spatio-temporal (y-t) plane of the proposed (a) SSFL-Recon(c) and (b) SSFL-Recon(v) using a retro￾spectively VISTA undersampling in a patient with transposition of the great vessels. Results for the acceleration factors R = 3, 6, 9, 12, 15, 18, 21 are shown in each column. The dynamic performance in the y-t plane corresponds to the blue line in the reference x-y plane image. The … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Quantitative comparison in terms of NRMSE, PSNR, and SSIM between PARCEL [43], SSDU [26], supervised learning, and the proposed SSFL-Recon frameworks. Each metric was computed for accelerations R = 8 and R = 16. Results are calculated for all subjects in the test datas…
Figure 5
Figure 5. Figure 5: Reconstructions in spatial (x-y) and spatio-temporal (y-t) plane of the proposed SSFL-Recon frameworks in comparison to zero-filled, kt-SLR [51], PARCEL [43], SSDU [26], supervised learning, and the ablation study SSL-Recon for a healthy subject who was retrospectively…
Figure 6
Figure 6. Figure 6: Reconstructions in spatial (x-y) plane of the proposed SSFL￾Recon frameworks in comparison to zero-filled, PARCEL [43], SSDU [26], and supervised learning for a healthy subject of the OCMR dataset, which was prospectively undersampled with VISTA sampling (R=10). End-sy…
Figure 7
Figure 7. Figure 7: (a) The learned features in the first step of SSFL-Recon(v). Each column presents the features at different iterations in the unrolled network under R = 5, 10, 15. (b) Ablation study: the reconstruction re￾sults of SSL-Recon compared to SSFL-Recon(v) under R = 5, 10, 1…
Figure 8
Figure 8. Figure 8: (a) The learned features in the first step of SSFL-Recon(c). Each column presents the features at different iterations in the unrolled network under R = 5, 10, 15. (b) Ablation study: the reconstruction re￾sults of SSL-Recon compared to SSFL-Recon(c) under R = 5, 10, 1…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 47 canonical work pages

  1. [1]

    SENSE: sensitivity encoding for fast MRI,

    K. P . Pruessmann, M. Weiger, M. B. Scheidegger, and P . Boe siger, “SENSE: sensitivity encoding for fast MRI,” Magn. Reson. Med. , vol. 42, no. 5, pp. 952–962, Nov. 1999

    work page 1999

  2. [2]

    Generalized autocalibrating partially parallel acquisitions (GRAPPA),

    M. A. Griswold et al., “ Generalized autocalibrating partially parallel acquisitions (GRAPPA),” Magn. Reson. Med. , vol. 47, no. 6, pp. 1202– 1210, Jun. 2002

    work page 2002

  3. [3]

    SPIRiT: iterative self-cons istent parallel imaging reconstruction from arbitrary k-space,

    M. Lustig, and J. M. Pauly, “SPIRiT: iterative self-cons istent parallel imaging reconstruction from arbitrary k-space,” Magn. Reson. Med., vol. 64, no. 2, pp. 457–471, Aug. 2010

    work page 2010

  4. [4]

    ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA.,

    M. Uecker et al., “ ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA.,” Magn. Reson. Med. , vol. 71, no. 3, pp. 990–1001, Mar. 2014

    work page 2014

  5. [5]

    A wavelet-based regularized reconstruction algorithm for S ENSE parallel MRI with applications to neuroimaging,

    L. Chaˆ ari, J. C. Pesquet, A. Benazza-Benyahia, and P . Ci uciu, “A wavelet-based regularized reconstruction algorithm for S ENSE parallel MRI with applications to neuroimaging,” Med. Image Anal. , vol. 15, no. 2, pp. 185–201, Apr. 2011

    work page 2011

  6. [6]

    An ite rative regu- larization method for total variation-based image restora tion,

    S. Osher, M. Burger, D. Goldfarb, J. Xu, and W. Yin, “An ite rative regu- larization method for total variation-based image restora tion,” Multiscale Model. Simul. , vol. 4, no. 2, pp. 460–489, 2005

    work page 2005

  7. [7]

    Undersampled radia l MRI with multiple coils. Iterative image reconstruction using a total variation constraint,

    K. T. Block, M. Uecker, and J. Frahm, “Undersampled radia l MRI with multiple coils. Iterative image reconstruction using a total variation constraint,” Magn. Reson. Med. , vol. 57, no. 6, pp. 1086–1098, Jun. 2007

    work page 2007

  8. [8]

    Ada ptive dictionary learning in sparse gradient domain for image rec overy,

    Q. Liu, S. Wang, L. Ying, X. Peng, Y . Zhu, and D. Liang, “Ada ptive dictionary learning in sparse gradient domain for image rec overy,” IEEE Trans. Image Process. , vol. 22, no. 12, pp. 4652–4663, Aug. 2013

    work page 2013

Show all 53 references
  1. [9]

    Dictionary learning and time sparsity for dynamic MR data reconstructi on,

    J. Caballero, A. N. Price, D. Rueckert, and J. V . Hajnal, “ Dictionary learning and time sparsity for dynamic MR data reconstructi on,” IEEE Trans. Med. Imaging , vol. 33, no. 4, pp. 979–994, Jan. 2014

    2014

  2. [10]

    Reconstruction with dictionary learning f or accelerated parallel magnetic resonance imaging,

    D. Weller, “Reconstruction with dictionary learning f or accelerated parallel magnetic resonance imaging,” IEEE SSIAI, Santa Fe, NM, USA, pp. 105–108, Mar. 2016

    2016

  3. [11]

    SparseSENSE: applicatio n of compressed sensing in parallel MRI,

    B. Liu, Y . M. Zou, and L. Ying, “SparseSENSE: applicatio n of compressed sensing in parallel MRI,” IEEE EMBS Int. Conf. ITAB , Shenzhen, China, pp. 127–130, May. 2008

    2008

  4. [12]

    Accel eration of MR parameter mapping using annihilating filter-based low ra nk hankel matrix (ALOHA),

    D. Lee, K. H. Jin, E. Y . Kim, S. H. Park, and J. C. Y e, “Accel eration of MR parameter mapping using annihilating filter-based low ra nk hankel matrix (ALOHA),” Magn. Reson. Med. , vol. 76, no. 6, pp. 1848–1864, Dec. 2016

    2016

  5. [13]

    Scan-specific robust artificial-neural-networks for k-space interpolat ion (RAKI) recon- struction: Database-free deep learning for fast imaging,

    M. Akc ¸akaya, S. Moeller, S. Weing¨ artner, and K. U˘ gurbil, “Scan-specific robust artificial-neural-networks for k-space interpolat ion (RAKI) recon- struction: Database-free deep learning for fast imaging,” Magn. Reson. Med., vol. 81, no. 1, pp. 439–453, Jan. 2019

    2019

  6. [14]

    Deep residual learni ng for accelerated MRI using magnitude and phase networks,

    D. Lee, J. Y oo, S. Tak, and J. C. Y e, “Deep residual learni ng for accelerated MRI using magnitude and phase networks,” IEEE Trans. Biomed. Eng. , vol. 65, no. 9, pp. 1985–1995, Apr. 2018

    1985

  7. [15]

    Real-time cardiovascular MR with spatio-temporal artifa ct suppression using deep learning–proof of concept in congenital heart di sease,

    A. Hauptmann, S. Arridge, F. Lucka, V . Muthurangu, and J . A. Steeden, “Real-time cardiovascular MR with spatio-temporal artifa ct suppression using deep learning–proof of concept in congenital heart di sease,” Magn. Reson. Med. , vol. 81, no. 2, pp. 1143–1156, Feb. 2019

    2019

  8. [16]

    Spatio- temporal deep learning-based undersampling artefact redu ction for 2D radial cine MRI with limited training data,

    A. Kofler, M. Dewey, T. Schaeffter, C. Wald, and C. Kolbit sch, “Spatio- temporal deep learning-based undersampling artefact redu ction for 2D radial cine MRI with limited training data,” IEEE Trans. Med. Imaging , vol. 39, no. 3, pp. 703–717, Aug. 2019

    2019

  9. [17]

    A deep cascade of convolutional neural networks for dynamic M R image reconstruction,

    J. Schlemper, J. Caballero, J. V . Hajnal, A. N. Price, an d D. Rueckert, “A deep cascade of convolutional neural networks for dynamic M R image reconstruction,” IEEE Trans. Med. Imaging , vol. 37, no. 2, pp. 491–503, Oct. 2017

    2017

  10. [18]

    Learning a variational network for reconstruction of accelerated MRI data,

    K. Hammernik et al., “ Learning a variational network for reconstruction of accelerated MRI data,” Magn. Reson. Med. , vol. 79, no. 6, pp. 3055– 3071, Jun. 2018

    2018

  11. [19]

    CINENet: deep learning-based 3D cardiac CINE MRI reconstruction with multi-coil complex-valued 4D spat io-temporal convolutions,

    T. K¨ ustner et al., “ CINENet: deep learning-based 3D cardiac CINE MRI reconstruction with multi-coil complex-valued 4D spat io-temporal convolutions,” Sci. Rep. , vol. 10, no. 1, pp. 13710, Aug. 2020

    2020

  12. [20]

    KIK I-net: cross-domain convolutional neural networks for reconstru cting under- sampled magnetic resonance images,

    T. Eo, Y . Jun, T. Kim, J. Jang, H. J. Lee, and D. Hwang, “KIK I-net: cross-domain convolutional neural networks for reconstru cting under- sampled magnetic resonance images,” Magn. Reson. Med. , vol. 80, no. 5, pp. 2188–2201, Nov. 2018

    2018

  13. [21]

    Multi-domain convolutional neural network (MD- CNN) for radial reconstruction of dynamic cardiac MRI,

    H. El-Rewaidy et al., “Multi-domain convolutional neural network (MD- CNN) for radial reconstruction of dynamic cardiac MRI,” Magn. Reson. Med., vol. 85, no. 3, pp. 1195–1208, Mar. 2021

    2021

  14. [22]

    Knowledge-driven deep learning for fast MR imaging: Under sampled MR image reconstruction from supervised to un-supervised l earning

    S. Wang, R. Wu, S. Jia, A. Diakite, C. Li, Q. Liu, H. Zheng, and L. Ying, “Knowledge-driven deep learning for fast MR imaging: Under sampled MR image reconstruction from supervised to un-supervised l earning.” Magn. Reson. Med. , vol. 92, no. 2, pp. 496–518, Aug. 2024

    2024

  15. [23]

    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 Med. Imaging , vol. 21, no. 1, pp. 195, Dec. 2021

    2021

  16. [24]

    Deep learning for accelerated and robust MRI reconstruction

    R. Heckel, M. Jacob, A. Chaudhari, O. Perlman, and E. Shi mron, “Deep learning for accelerated and robust MRI reconstruction.” Magn. Reson. Mater . Phy., pp. 1–34, Jul. 2024

    2024

  17. [25]

    On instabilities of deep learning in image reconstruction and the potential 12 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. XX, NO. XX, XXX X 2020 costs of AI

    V . Antun, F. Renna, C. Poon, B. Adcock, and A. C. Hansen, “ On instabilities of deep learning in image reconstruction and the potential 12 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. XX, NO. XX, XXX X 2020 costs of AI.” Proc. Natl. Acad. Sci. , vol. 117, no. 48, pp. 30088–3009...

    2020

  18. [26]

    Self-supervised learning of physics-gu ided recon- struction neural networks without fully sampled reference data,

    B. Y aman, S. A. H. Hosseini, S. Moeller, J. Ellermann, K. U˘ gurbil, and M. Akc ¸akaya, “Self-supervised learning of physics-gu ided recon- struction neural networks without fully sampled reference data,” Magn. Reson. Med. , vol. 84, no. 6, pp. 3172–3191, Dec. 2020

    2020

  19. [27]

    Multi-mask self-supervised learning for physics- guided neural networks in highly accelerated magnetic reso nance imag- ing,

    B. Y aman et al., “ Multi-mask self-supervised learning for physics- guided neural networks in highly accelerated magnetic reso nance imag- ing,” NMR Biomed. , vol. 35, no. 12, pp. e4798, Dec. 2022

    2022

  20. [28]

    A theoretical framework for s elf-supervised MR image reconstruction using sub-sampling via variable de nsity Nois- ier2Noise

    C. Millard, and M. Chiew, “A theoretical framework for s elf-supervised MR image reconstruction using sub-sampling via variable de nsity Nois- ier2Noise.” IEEE Trans. Comput. Imaging , vol. 9, Jul. 2023

    2023

  21. [29]

    Noisier2n oise: Learn- ing to denoise from unpaired noisy data

    N. Moran, D. Schmidt, Y . Zhong, and P . Coady, “Noisier2n oise: Learn- ing to denoise from unpaired noisy data.” IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 12064–12072, 2020

    2020

  22. [30]

    Dual-domain self-supervised learning for accelerated non-cartesian mri reconstruction,

    B. Zhou et al., “ Dual-domain self-supervised learning for accelerated non-cartesian mri reconstruction,” Med. Image Anal. , vol. 81, pp. 102538, Oct. 2022

    2022

  23. [31]

    Noise2Recon: Enabling SNR-robust MRI recon- struction with semi-supervised and self-supervised learn ing,

    A. D. Desai et al., “ Noise2Recon: Enabling SNR-robust MRI recon- struction with semi-supervised and self-supervised learn ing,” Magn. Reson. Med. , vol. 90, no. 5, pp. 2052–2070, Nov. 2023

    2023

  24. [32]

    Solving inverse problems in medical imaging with score-based generative mo dels,

    Y . Song, L. Shen, L. Xing, and S. Ermon, “Solving inverse problems in medical imaging with score-based generative mo dels,” arXiv:2111.08005, Nov. 2021

    2021 arXiv

  25. [33]

    Self-score: self-supervised learning on score-based models for MRI reconstruction,

    Z. X. Cui et al., “ Self-score: self-supervised learning on score-based models for MRI reconstruction,” arXiv:2209.00835, Sep. 2022

    2022 arXiv

  26. [34]

    Solving i nverse problems with score-based generative priors learned from n oisy data,

    A. Aali, M. Arvinte, S. Kumar, and J. I. Tamir, “Solving i nverse problems with score-based generative priors learned from n oisy data,” arXiv:2305.01166, May. 2023

    2023 arXiv

  27. [35]

    Learning re presentations by maximizing mutual information across views,

    P . Bachman, R. D. Hjelm, and W. Buchwalter, “Learning re presentations by maximizing mutual information across views,” NeurIPS, V ancouver, Canada, vol. 32, Dec. 2019

    2019

  28. [36]

    Momentum co ntrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum co ntrast for unsupervised visual representation learning,” Proc. CVPR , Seattle, W A, USA, pp. 9729–9738, Jun. 2020

    2020

  29. [37]

    Exploring simple siamese represent ation learning,

    X. Chen, and K. He, “Exploring simple siamese represent ation learning,” Proc. CVPR , pp. 15750–15758, Jun. 2021

    2021

  30. [38]

    VICReg: variance-in variance- covariance regularization for self-supervised learning,

    A. Bardes, J. Ponce, and Y . LeCun, “VICReg: variance-in variance- covariance regularization for self-supervised learning,” ICLR, Apr. 2022

    2022

  31. [39]

    Signature verification using a

    J. Bromley, I. Guyon, Y . LeCun, E. S¨ ackinger, and R. Sha h, “Signature verification using a” siamese” time delay neural network,” Adv. Neural Inf. Processing Syst. , vol. 6, 1993

    1993

  32. [40]

    A simp le frame- work for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simp le frame- work for contrastive learning of visual representations,” ICML, Vienna, Austria, Nov. 2020

    2020

  33. [41]

    Bar low twins: Self-supervised learning via redundancy reduction,

    J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny, “Bar low twins: Self-supervised learning via redundancy reduction,” ICML, Jul. 2021

    2021

  34. [42]

    A review of self-supervised, generative, and fe w-shot deep learning methods for data-limited magnetic resonance imaging segmentation,

    Z. Liu, K. Kainth, A. Zhou, T. W. Deyer, Z. A. Fayad, H. Gre enspan, and X. Mei, “A review of self-supervised, generative, and fe w-shot deep learning methods for data-limited magnetic resonance imaging segmentation,” NMR Biomed. , pp. e5143, Mar. 2024

    2024

  35. [43]

    PARCEL: physics-based unsupervised contrastive representation learning for multi-coil MR imaging,

    S. Wang et al., “ PARCEL: physics-based unsupervised contrastive representation learning for multi-coil MR imaging,” IEEE/ACM Trans. Comput. Biol. Bioinform. , Oct. 2022

    2022

  36. [44]

    Contrastive l earning for local and global learning mri reconstruction,

    Q. Yi, J. Liu, L. Hu, F. Fang, and G. Zhang, “Contrastive l earning for local and global learning mri reconstruction,” arXiv:2111.15200, Nov. 2021

    2021 arXiv

  37. [45]

    Representation lea rning with contrastive predictive coding,

    A. V . D. Oord, Y . Li, and O. Vinyals, “Representation lea rning with contrastive predictive coding,” arXiv:1807.03748, Jul. 2018

    2018 arXiv

  38. [46]

    Unitary evolution recurrent neural networks,

    M. Arjovsky, A. Shah, and Y . Bengio, “Unitary evolution recurrent neural networks,” ICML, New Y ork City, NY , USA, Jun. 2016

    2016

  39. [47]

    V ariable density incoherent spatiotemporal acquisition (VISTA) for highly accelerated cardiac MRI,

    R. Ahmad, H. Xue, S. Giri, Y . Ding, J. Craft, and O. P . Simo netti, “V ariable density incoherent spatiotemporal acquisition (VISTA) for highly accelerated cardiac MRI,” Magn. Reson. Med. , vol. 74, no. 5, pp. 1266–1278, Nov. 2015

    2015

  40. [48]

    OCMR (v1. 0)–open-access multi-coil k-space dataset for cardiovascular magnetic resonance imaging

    C. Chen et al., “ OCMR (v1. 0)–open-access multi-coil k-space dataset for cardiovascular magnetic resonance imaging.” ArXiv Preprint ArXiv:2008.03410, 2020

    2008 arXiv

  41. [49]

    Machine enhanced recon struction learning and interpretation networks (MERLIN),

    K. Hammernik, and T. K¨ ustner, “Machine enhanced recon struction learning and interpretation networks (MERLIN),” ISMRM, London, UK, May. 2022

    2022

  42. [50]

    Adam: A method for stochastic op timization,

    D. P . Kingma, and J. Ba, “Adam: A method for stochastic op timization,” arXiv:1412.6980, Dec. 2014

    2014 arXiv

  43. [51]

    Accelera ted dynamic MRI exploiting sparsity and low-rank structure: kt SLR,

    S. G. Lingala, Y . Hu, E. DiBella, and M. Jacob, “Accelera ted dynamic MRI exploiting sparsity and low-rank structure: kt SLR,” IEEE Trans. Med. Imaging , vol. 30, no. 5, pp. 1042–1054, Jan. 2011

    2011

  44. [52]

    Learning- based optimization of the under-sampling pattern in MRI,

    C. D. Bahadir, A. V . Dalca, and M. R. Sabuncu, “Learning- based optimization of the under-sampling pattern in MRI,” IPMI, Proceedings 26, pp. 780–792, Jun. 2019

    2019

  45. [53]

    Fast data -driven learning of parallel MRI sampling patterns for large scale p roblems,

    M. V . Zibetti, G. T. Herman, and R. R. Regatte, “Fast data -driven learning of parallel MRI sampling patterns for large scale p roblems,” Sci. Rep. , vol. 11, no. 1, pp. 19312, Sep. 2021

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.