Pith. sign in

REVIEW 2 major objections 5 minor 43 references

A dual-branch image-domain INR with feature embeddings improves unsupervised cardiac cine MRI reconstruction without fully sampled references.

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 · grok-4.5

2026-07-11 21:53 UTC pith:E7NW5JZ5

load-bearing objection Solid, carefully ablated image-domain dual-branch INR that reliably beats its single-branch and k-space cousins on cine data; incremental but useful. the 2 major comments →

arxiv 2607.04069 v1 pith:E7NW5JZ5 submitted 2026-07-05 cs.CV cs.AI

Enhancing Implicit Neural Representations with Image Feature Embedding for Unsupervised Cardiac Cine MRI Reconstruction

Donghang Lyu , Marius Staring , Yiming Dong , Keupp Jochen , Hildo J. Lamb , Mariya Doneva This is my paper
classification cs.CV cs.AI
keywords implicit neural representationsdual-branch structureimage feature embeddingunsupervised cardiac cine MRI reconstructionself-supervised MRImulti-coil reconstructioncoil-combined reconstruction
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.

Cardiac cine MRI is routinely accelerated by undersampling k-space, which leaves an ill-posed inverse problem that must be solved without fully sampled ground-truth images. Implicit neural representations already offer a compact, self-supervised route by mapping coordinates to image intensities, yet they usually ignore neighborhood context. This paper claims that adding a second INR branch fed by features extracted from a complex U-Net, letting the two branches exchange information, and periodically refining the image input with conjugate-gradient data consistency systematically improves reconstruction quality. The resulting dual-branch model, I-FP-INR, is shown to outperform classical low-rank-plus-sparse methods and several prior INR baselines on both a large public multi-view cine dataset and private 0.6 T acquisitions across multiple Cartesian sampling patterns and acceleration factors. A sympathetic reader cares because the method removes the need for expensive fully sampled training data while still recovering sharper cardiac structures and cleaner temporal profiles than single-branch or k-space-only INRs.

Core claim

The authors establish that an image-domain dual-branch INR (I-FP-INR) whose second branch consumes multi-scale features extracted from a lightweight complex U-Net, interacts with the coordinate branch, and is alternately refined by conjugate-gradient data consistency, yields consistently higher PSNR, SSIM, DISTS and HaarPSI than L+S, pure coordinate INRs, k-space feature INRs and hash-encoded INRs on cardiac cine data at 4 imes and 8 imes acceleration under three Cartesian patterns, and remains effective on lower-field in-house scans.

What carries the argument

I-FP-INR: two interactive MLP branches (one coordinate-driven with WIRE activations and NeRF/Fourier positional encodings, one feature-driven with Leaky ReLU) whose embeddings are fused twice, together with an alternating inference stage that updates the image input via CG data consistency so that the feature extractor receives progressively cleaner cine volumes.

Load-bearing premise

That features pulled from initially aliased zero-filled images still supply genuinely complementary neighborhood information after a few conjugate-gradient refinements, rather than simply re-encoding the same residual that the coordinate branch already sees.

What would settle it

Run the identical dual-branch architecture with the feature branch permanently disabled (or fed pure noise) on the same CMRxRecon2024 VISTA and PDS subsets; if the quantitative gap to the full I-FP-INR vanishes, the claim that the extracted features are complementary collapses.

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

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

2 major / 5 minor

Summary. The manuscript proposes I-FP-INR, an unsupervised image-domain dual-branch INR for cardiac cine MRI reconstruction. One branch maps positional encodings of (i,j,t) coordinates to complex image intensities (I-P-INR); the second extracts multi-scale features from (initially zero-filled, later CG-refined) images via a lightweight complex U-Net and maps them through a second INR (I-F-INR). The branches interact via linear cross-injection and mid-network fusion, and an alternating inference stage periodically freezes the network, averages the two outputs, and applies conjugate-gradient data consistency to refine the image input for the feature branch (Algorithms 1–2, Eqs. 4–13). Two reconstruction strategies (multi-coil and coil-combined) are examined. Optimization uses only measured k-space via a combination of NIK and relative-L2 losses (Eqs. 15–19). On CMRxRecon2024 (16 subjects, 4 imes/8 imes, three Cartesian patterns) and in-house 0.6 T data the method reports consistent gains over L+S, P-INR, KP-INR, Hash-INR and single-branch I-P-INR in PSNR/SSIM/DISTS/HaarPSI, with Wilcoxon tests and ablations of activation, positional encoding, loss and inference stage.

Significance. If the reported gains hold, the work supplies a practical, fully unsupervised alternative for dynamic cardiac MRI that does not require fully-sampled training sets and improves on both classical low-rank+sparse and prior INR baselines. The systematic comparison of multi-coil versus coil-combined strategies under different sampling patterns, the multi-metric evaluation (including perceptual indices), the public-plus-private validation, and the component-wise ablations (Tables II–V) constitute a solid empirical contribution to the growing INR-for-MRI literature. The dual-branch design with alternating refinement is a concrete architectural idea that other self-supervised reconstruction methods can build on.

major comments (2)
  1. §V-A and Table I: reconstruction times are reported as ~44–45 min per slice for both I-FP-INR variants versus ~12–14 min for single-branch image-domain INRs and <1 min for L+S. Given that the clinical value of unsupervised cine reconstruction is tightly coupled to turnaround time, the manuscript should either (i) quantify how many iterations can be removed while preserving the statistically significant gains over I-P-INR, or (ii) discuss concrete acceleration routes (e.g., reduced feature-extractor capacity, early stopping, or warm-start across slices). Without this, the practical significance of the dual-branch improvement remains incompletely established.
  2. §IV-A and §V-C: only 16 subjects are used from CMRxRecon2024 and the in-house evaluation is purely qualitative (Fig. 5). While the multi-pattern, multi-acceleration design and Bonferroni-corrected Wilcoxon tests are appropriate, the limited cohort size and absence of quantitative metrics on the private data leave open the possibility that the observed ranking is dataset-specific. Adding at least summary statistics (mean±std of the same four metrics) for the in-house cases, or expanding the public subset, would strengthen the robustness claim that currently rests heavily on a modest sample.
minor comments (5)
  1. Eq. (3) vs. Eq. (4): the regularization term λ∥xθ−fθ(P)∥2 is written as an explicit soft constraint; clarify whether this term is actually optimized or is merely notational, since the subsequent loss (Eq. 19) contains only k-space data-consistency terms.
  2. Fig. 1 caption and Algorithms 1–2: the precise schedule at which the inference stage begins (i>t) and the value of the CG regularization parameter λ are not stated numerically; adding them would improve reproducibility.
  3. Table I header: “VD w/ ACS” is defined only in the caption; a short expansion in the table note would help readers who start from the results section.
  4. §III-A: the learnable fusion weight α is introduced but never reported (final value or range); a brief statement of its typical magnitude would confirm that both feature scales are used.
  5. References [13] and [31] are the authors’ own prior works; ensure the distinction between KP-INR (k-space) and the present image-domain extension is stated once more in the introduction for readers unfamiliar with the earlier papers.

Circularity Check

0 steps flagged

No significant circularity: empirical dual-branch INR method with standard prior-work self-citation; claims rest on data-consistency optimization and external benchmarks, not definitional reductions.

full rationale

The paper is a self-supervised reconstruction methods contribution. The core optimization (Eqs. 3–4, Algorithms 1–2) minimizes data-consistency loss solely on measured multi-coil k-space samples (NIK + relative L2, Eq. 19); reconstructed images are free parameters of the dual-branch INR, not algebraically forced by any fitted constant or self-defined quantity. Ablations (Tables II–V) and multi-pattern comparisons (Table I, Figs. 3–5) against independent baselines (L+S, Hash-INR, P-INR) and the single-branch ablation I-P-INR supply external falsifiability. The sole self-citation of note is the authors’ prior KP-INR [31], invoked only as the k-space predecessor being extended to the image domain; it does not supply a uniqueness theorem, ansatz, or uniqueness claim that forces the reported metrics or the dual-branch gains. No fitted-input-as-prediction, self-definitional loop, or renaming of a known result appears. Circularity burden is therefore negligible (score 1 for the ordinary self-citation of prior method work).

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The work is an empirical deep-learning method paper. Load-bearing content is architectural choices and hyper-parameters rather than physical axioms. The MRI forward model and INR universal-approximation premise are standard domain assumptions; the dual-branch interaction and loss weights are free design choices fitted by ablation.

free parameters (6)
  • loss weights λ1–λ4 = 0.5 each
    All set to 0.5 by hand with no sensitivity study beyond the NIK vs RelL2 ablation; equal weighting is a free design choice that affects the reported metrics.
  • NIK regularization weight λr = 0.5
    Fixed at 0.5 inside LNIK; controls high-frequency emphasis.
  • positional encoding dimensions (spatial 480, temporal 96) = 480 / 96
    Chosen following FMLP; total embedding length 576 is a free architectural hyper-parameter.
  • optimization schedule (300 iters, inference every 50, lr 1e-4, decay 0.95/50) = 300 / 50 / 1e-4
    Hand-tuned for convergence; directly determines both quality and the 44–45 min runtime.
  • feature fusion scalar α = learnable in [0,1]
    Learnable but initialized and constrained to [0,1]; its optimum is data-dependent.
  • CG data-consistency λ
    Regularization strength inside the conjugate-gradient step of the inference stage; value not explicitly reported.
axioms (4)
  • domain assumption Multi-coil MRI forward model yt_c = M_u F S_c x_t holds with known or pre-estimated coil sensitivity maps S_c.
    Eq. (1)–(2); standard parallel-imaging model assumed throughout optimization and CG steps.
  • domain assumption A compact MLP with suitable positional encoding can represent the continuous cine image sequence well enough that data-consistency on measured k-space yields a clinically useful reconstruction.
    Core INR premise restated in §II–III; inherited from NeRF/WIRE literature.
  • ad hoc to paper Features extracted from (refined) image-domain inputs supply information complementary to pure coordinate embeddings.
    Central design hypothesis of I-FP-INR (§III-A); validated only by ablation against I-P-INR, not independently proven.
  • domain assumption AdamW with the stated schedule and gradient accumulation over frames converges to a stable minimum of the non-convex dual-branch loss.
    Implementation §IV-B; standard deep-learning optimization assumption.
invented entities (1)
  • I-FP-INR dual-branch architecture (I-P-INR + I-F-INR with mid-network fusion and alternating inference) no independent evidence
    purpose: To inject neighborhood image context into coordinate-based INR reconstruction and iteratively clean the feature input.
    New composite architecture; no independent physical existence outside the optimization procedure. independent_evidence is false because the only evidence is the paper’s own reconstruction metrics.

pith-pipeline@v1.1.0-grok45 · 24679 in / 3311 out tokens · 33269 ms · 2026-07-11T21:53:30.958505+00:00 · methodology

0 comments
read the original abstract

Cardiac cine Magnetic Resonance Imaging (MRI) is a critical diagnostic tool that provides dynamic insights for radiologists. To accelerate acquisition, under-sampled k-space data is often used, requiring reconstruction methods that combine coil sensitivity encoding with prior information to recover missing data. Deep learning approaches have gained more attention for leveraging data-adaptive priors. While supervised learning approaches are a common choice, they depend on fully sampled reference data, which is not always available. Unsupervised methods eliminate the need for fully sampled reference data, which can be advantageous in cardiac cine MRI reconstruction. Among them, implicit neural representations (INRs) have shown great potential due to their simple architecture and good quality reconstructions. In this work, we propose an image-domain dual-branch INR framework, termed I-FP-INR, which extends the original INR design by introducing an additional feature-processing branch. This design aims to extract complementary feature embeddings to enhance the overall representation, thereby benefiting reconstruction. Extensive evaluations on both public datasets and in-house data show consistent improvements over baseline methods in reconstruction quality, with strong robustness across varied scenarios.

Figures

Figures reproduced from arXiv: 2607.04069 by Donghang Lyu, Hildo J. Lamb, Keupp Jochen, Marius Staring, Mariya Doneva, Yiming Dong.

Figure 1
Figure 1. Figure 1: The overall optimization pipeline of I-FP-INR under two reconstruction strategies is illustrated: (a) the multi-coil [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization examples of three sampling mask types [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of the proposed I-FP-INR with baseline methods at R=4 under three sampling patterns. “GT” [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of the proposed I-FP-INR with baseline methods at R=8 under three sampling patterns. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative visualizations of the proposed I-FP-INR on in-house data at different acceleration factors under a variable [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] 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

43 extracted references · 1 linked inside Pith

  1. [1]

    Interventions to alleviate patient anxiety during magnetic resonance imaging: a review,

    S. Phillips and I. J. Deary, “Interventions to alleviate patient anxiety during magnetic resonance imaging: a review,”Radiography, vol. 1, no. 1, pp. 29–34, 1995

    1995

  2. [2]

    Sense: sensitivity encoding for fast mri,

    K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boesiger, “Sense: sensitivity encoding for fast mri,”Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 42, no. 5, pp. 952–962, 1999

    1999

  3. [3]

    Generalized autocalibrating partially parallel acquisitions (grappa),

    M. A. Griswold, P. M. Jakob, R. M. Heidemann, M. Nittka, V . Jellus, J. Wang, B. Kiefer, and A. Haase, “Generalized autocalibrating partially parallel acquisitions (grappa),”Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 47, no. 6, pp. 1202–1210, 2002

    2002

  4. [4]

    Sparse mri: The application of compressed sensing for rapid mr imaging,

    M. Lustig, D. Donoho, and J. M. Pauly, “Sparse mri: The application of compressed sensing for rapid mr imaging,”Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 58, no. 6, pp. 1182–1195, 2007. 2026 13

    2007

  5. [5]

    A review on accelerated magnetic resonance imaging techniques: Parallel imaging, compressed sensing, and machine learning,

    M. Tavakkoli and M. D. Noseworthy, “A review on accelerated magnetic resonance imaging techniques: Parallel imaging, compressed sensing, and machine learning,”Critical Reviews™ in Biomedical Engineering, vol. 53, no. 5, 2025

    2025

  6. [6]

    Electrodynamics and ultimate snr in parallel mr imaging,

    F. Wiesinger, P. Boesiger, and K. P. Pruessmann, “Electrodynamics and ultimate snr in parallel mr imaging,”Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 52, no. 2, pp. 376–390, 2004

    2004

  7. [7]

    Modl: Model-based deep learning architecture for inverse problems,

    H. K. Aggarwal, M. P. Mani, and M. Jacob, “Modl: Model-based deep learning architecture for inverse problems,”IEEE transactions on medical imaging, vol. 38, no. 2, pp. 394–405, 2018

    2018

  8. [8]

    Admm-csnet: A deep learning approach for image compressive sensing,

    Y . Yang, J. Sun, H. Li, and Z. Xu, “Admm-csnet: A deep learning approach for image compressive sensing,”IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 3, pp. 521–538, 2018

    2018

  9. [9]

    Idpcnn: Iterative denoising and projecting cnn for mri reconstruction,

    R. Hou and F. Li, “Idpcnn: Iterative denoising and projecting cnn for mri reconstruction,”Journal of Computational and Applied Mathematics, vol. 406, p. 113973, 2022

    2022

  10. [10]

    An adaptive intelligence algorithm for undersampled knee mri reconstruction,

    N. Pezzotti, S. Yousefi, M. S. Elmahdy, J. H. F. Van Gemert, C. Schuelke, M. Doneva, T. Nielsen, S. Kastryulin, B. P. Lelieveldt, M. J. Van Osch et al., “An adaptive intelligence algorithm for undersampled knee mri reconstruction,”Ieee Access, vol. 8, pp. 204 825–204 838, 2020

    2020

  11. [11]

    Score-based diffusion models for accelerated mri,

    H. Chung and J. C. Ye, “Score-based diffusion models for accelerated mri,”Medical image analysis, vol. 80, p. 102479, 2022

    2022

  12. [12]

    Convolutional recurrent neural networks for dynamic mr image reconstruction,

    C. Qin, J. Schlemper, J. Caballero, A. N. Price, J. V . Hajnal, and D. Rueckert, “Convolutional recurrent neural networks for dynamic mr image reconstruction,”IEEE transactions on medical imaging, vol. 38, no. 1, pp. 280–290, 2018

    2018

  13. [13]

    Convolutional recurrent u-net for cardiac cine mri re- construction via effective spatio-temporal feature exploitation,

    D. Lyu, M. Staring, M. J. van Osch, M. Doneva, H. J. Lamb, and N. Pezzotti, “Convolutional recurrent u-net for cardiac cine mri re- construction via effective spatio-temporal feature exploitation,”Medical Physics, vol. 53, no. 1, p. e70245, 2026

    2026

  14. [14]

    Fill the k-space and refine the image: Prompting for dynamic and multi-contrast mri reconstruction,

    B. Xin, M. Ye, L. Axel, and D. N. Metaxas, “Fill the k-space and refine the image: Prompting for dynamic and multi-contrast mri reconstruction,” inInternational Workshop on Statistical Atlases and Computational Models of the Heart. Springer, 2023, pp. 261–273

    2023

  15. [15]

    Self-supervised learning of physics-guided reconstruc- tion neural networks without fully sampled reference data,

    B. Yaman, S. A. H. Hosseini, S. Moeller, J. Ellermann, K. U ˘gurbil, and M. Akc ¸akaya, “Self-supervised learning of physics-guided reconstruc- tion neural networks without fully sampled reference data,”Magnetic resonance in medicine, vol. 84, no. 6, pp. 3172–3191, 2020

    2020

  16. [16]

    Clean self-supervised mri reconstruction from noisy, sub-sampled training data with robust ssdu,

    C. Millard and M. Chiew, “Clean self-supervised mri reconstruction from noisy, sub-sampled training data with robust ssdu,”Bioengineering, vol. 11, no. 12, p. 1305, 2024

    2024

  17. [17]

    Neural implicit k-space for binning-free non-cartesian cardiac mr imaging,

    W. Huang, H. B. Li, J. Pan, G. Cruz, D. Rueckert, and K. Hammernik, “Neural implicit k-space for binning-free non-cartesian cardiac mr imaging,” inInternational Conference on Information Processing in Medical Imaging. Springer, 2023, pp. 548–560

    2023

  18. [18]

    Multi-dynamic deep image prior for cardiac mri,

    M. V ornehm, C. Chen, M. A. Sultan, S. M. Arshad, Y . Han, F. Knoll, and R. Ahmad, “Multi-dynamic deep image prior for cardiac mri,”Magnetic Resonance in Medicine, vol. 94, no. 6, pp. 2668–2679, 2025

    2025

  19. [19]

    A theoretical framework for self-supervised mr image reconstruction using sub-sampling via variable density nois- ier2noise,

    C. Millard and M. Chiew, “A theoretical framework for self-supervised mr image reconstruction using sub-sampling via variable density nois- ier2noise,”IEEE transactions on computational imaging, vol. 9, pp. 707–720, 2023

    2023

  20. [20]

    Noise2noise: Learning image restoration without clean data,

    J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M. Aittala, and T. Aila, “Noise2noise: Learning image restoration without clean data,”arXiv preprint arXiv:1803.04189, 2018

    Pith/arXiv arXiv 2018

  21. [21]

    Time- dependent deep image prior for dynamic mri,

    J. Yoo, K. H. Jin, H. Gupta, J. Yerly, M. Stuber, and M. Unser, “Time- dependent deep image prior for dynamic mri,”IEEE Transactions on Medical Imaging, vol. 40, no. 12, pp. 3337–3348, 2021

    2021

  22. [22]

    Deep image prior,

    D. Ulyanov, A. Vedaldi, and V . Lempitsky, “Deep image prior,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 9446–9454

    2018

  23. [24]

    Implicit neural representations for deformable image registration,

    J. M. Wolterink, J. C. Zwienenberg, and C. Brune, “Implicit neural representations for deformable image registration,” inInternational Conference on medical imaging with deep learning. PMLR, 2022, pp. 1349–1359

    2022

  24. [25]

    Cina: Conditional implicit neural atlas for spatio-temporal representation of fetal brains,

    M. Dannecker, V . Kyriakopoulou, L. Cordero-Grande, A. N. Price, J. V . Hajnal, and D. Rueckert, “Cina: Conditional implicit neural atlas for spatio-temporal representation of fetal brains,” inInternational Confer- ence on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 181–191

    2024

  25. [26]

    Vestibular schwannoma growth prediction from longitudinal mri by time-conditioned neural fields,

    Y . Chen, J. M. Wolterink, O. M. Neve, S. R. Romeijn, B. M. Verbist, E. F. Hensen, Q. Tao, and M. Staring, “Vestibular schwannoma growth prediction from longitudinal mri by time-conditioned neural fields,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 508–518

    2024

  26. [27]

    Spatiotemporal implicit neural representation for unsupervised dynamic mri reconstruction,

    J. Feng, R. Feng, Q. Wu, X. Shen, L. Chen, X. Li, L. Feng, J. Chen, Z. Zhang, C. Liuet al., “Spatiotemporal implicit neural representation for unsupervised dynamic mri reconstruction,”IEEE Transactions on Medical Imaging, vol. 44, no. 5, pp. 2143–2156, 2025

    2025

  27. [28]

    Implicit neural networks with fourier-feature inputs for free-breathing cardiac mri reconstruction,

    J. F. Kunz, S. Ruschke, and R. Heckel, “Implicit neural networks with fourier-feature inputs for free-breathing cardiac mri reconstruction,” IEEE Transactions on Computational Imaging, vol. 10, pp. 1280–1289, 2024

    2024

  28. [29]

    Pisco: Self-supervised k-space regularization for improved neural im- plicit k-space representations of dynamic mri,

    V . Spieker, H. Eichhorn, W. Huang, J. K. Stelter, T. Catalan, R. F. Braren, D. Rueckert, F. S. Costabal, K. Hammernik, D. C. Karampinoset al., “Pisco: Self-supervised k-space regularization for improved neural im- plicit k-space representations of dynamic mri,”Medical Image Analysis, p. 103890, 2025

    2025

  29. [30]

    Implicit functions in feature space for 3d shape reconstruction and completion,

    J. Chibane, T. Alldieck, and G. Pons-Moll, “Implicit functions in feature space for 3d shape reconstruction and completion,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 6970–6981

    2020

  30. [31]

    Kp- inr: A dual-branch implicit neural representation model for cardiac cine mri reconstruction,

    D. Lyu, M. Staring, M. Doneva, H. J. Lamb, and N. Pezzotti, “Kp- inr: A dual-branch implicit neural representation model for cardiac cine mri reconstruction,” inInternational Workshop on Statistical Atlases and Computational Models of the Heart. Springer, 2025, pp. 56–66

    2025

  31. [32]

    Cmrxrecon2024: a multimodality, multiview k- space dataset boosting universal machine learning for accelerated cardiac mri,

    Z. Wang, F. Wang, C. Qin, J. Lyu, C. Ouyang, S. Wang, Y . Li, M. Yu, H. Zhang, K. Guoet al., “Cmrxrecon2024: a multimodality, multiview k- space dataset boosting universal machine learning for accelerated cardiac mri,”Radiology: Artificial Intelligence, vol. 7, no. 2, p. e240443, 2025

    2025

  32. [33]

    Accelerated dynamic mri exploiting sparsity and low-rank structure: kt slr,

    S. G. Lingala, Y . Hu, E. DiBella, and M. Jacob, “Accelerated dynamic mri exploiting sparsity and low-rank structure: kt slr,”IEEE transactions on medical imaging, vol. 30, no. 5, pp. 1042–1054, 2011

    2011

  33. [34]

    Image reconstruction from highly undersampled (k, t)-space data with joint partial separability and sparsity constraints,

    B. Zhao, J. P. Haldar, A. G. Christodoulou, and Z.-P. Liang, “Image reconstruction from highly undersampled (k, t)-space data with joint partial separability and sparsity constraints,”IEEE transactions on medical imaging, vol. 31, no. 9, pp. 1809–1820, 2012

    2012

  34. [35]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  35. [36]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier features let networks learn high frequency functions in low dimensional domains,” Advances in neural information processing systems, vol. 33, pp. 7537– 7547, 2020

    2020

  36. [37]

    Wire: Wavelet implicit neural representations,

    V . Saragadam, D. LeJeune, J. Tan, G. Balakrishnan, A. Veeraraghavan, and R. G. Baraniuk, “Wire: Wavelet implicit neural representations,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 18 507–18 516

    2023

  37. [38]

    Variable density incoherent spatiotemporal acquisition (vista) for highly accelerated cardiac mri,

    R. Ahmad, H. Xue, S. Giri, Y . Ding, J. Craft, and O. P. Simonetti, “Variable density incoherent spatiotemporal acquisition (vista) for highly accelerated cardiac mri,”Magnetic resonance in medicine, vol. 74, no. 5, pp. 1266–1278, 2015

    2015

  38. [39]

    Practical parallel imaging compressed sensing mri: Summary of two years of experience in accelerating body mri of pediatric patients,

    S. S. Vasanawala, M. Murphy, M. T. Alley, P. Lai, K. Keutzer, J. M. Pauly, and M. Lustig, “Practical parallel imaging compressed sensing mri: Summary of two years of experience in accelerating body mri of pediatric patients,” in2011 ieee international symposium on biomedical imaging: From nano to macro. IEEE, 2011, pp. 1039–1043

    2011

  39. [40]

    Image quality assessment: Unifying structure and texture similarity,

    K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Image quality assessment: Unifying structure and texture similarity,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 5, pp. 2567– 2581, 2020

    2020

  40. [41]

    A haar wavelet- based perceptual similarity index for image quality assessment,

    R. Reisenhofer, S. Bosse, G. Kutyniok, and T. Wiegand, “A haar wavelet- based perceptual similarity index for image quality assessment,”Signal Processing: Image Communication, vol. 61, pp. 33–43, 2018

    2018

  41. [42]

    Image quality assessment for magnetic resonance imaging,

    S. Kastryulin, J. Zakirov, N. Pezzotti, and D. V . Dylov, “Image quality assessment for magnetic resonance imaging,”IEEE Access, vol. 11, pp. 14 154–14 168, 2023

    2023

  42. [43]

    Low-rank plus sparse matrix decomposition for accelerated dynamic mri with separation of background and dynamic components,

    R. Otazo, E. Candes, and D. K. Sodickson, “Low-rank plus sparse matrix decomposition for accelerated dynamic mri with separation of background and dynamic components,”Magnetic resonance in medicine, vol. 73, no. 3, pp. 1125–1136, 2015

    2015

  43. [44]

    Implicit neural representations with periodic activation functions,

    V . Sitzmann, J. Martel, A. Bergman, D. Lindell, and G. Wetzstein, “Implicit neural representations with periodic activation functions,” Advances in neural information processing systems, vol. 33, pp. 7462– 7473, 2020

    2020