pith. sign in

arxiv: 2508.09179 · v3 · submitted 2025-08-07 · 📡 eess.IV · cs.CV

HiFi-Mamba: Dual-Stream W-Laplacian Enhanced Mamba for High-Fidelity MRI Reconstruction

Hongli Chen , Pengcheng Fang , Yuxia Chen , Yingxuan Ren , Jing Hao , Fangfang Tang , Xiaohao Cai , Shanshan Shan
show 1 more author
Feng Liu
This is my paper

Pith reviewed 2026-05-19 00:41 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords MRI reconstructionMambadual-stream architectureW-Laplacianhigh-frequency detailsundersampled k-spacestate-space modelspectral decoupling
0
0 comments X

The pith

Dual-stream Mamba with W-Laplacian splitting preserves high-frequency details for superior MRI reconstruction from undersampled data.

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

The paper proposes HiFi-Mamba to fix two problems when using Mamba models on MRI: they miss fine anatomical details and waste computation on repeated scanning directions. It stacks W-Laplacian blocks that split input features into separate low-frequency and high-frequency streams while keeping the original information intact. The HiFi-Mamba blocks then handle global structure on the low-frequency path and selectively add back the high-frequency path using adaptive modulation. A single-direction scan replaces the usual multi-direction scans to cut redundancy without losing long-range context. Tests on standard MRI benchmarks show the model beats prior CNN, Transformer, and Mamba methods in accuracy while using a smaller footprint.

Core claim

The HiFi-Mamba architecture comprises stacked W-Laplacian blocks that perform fidelity-preserving spectral decoupling into complementary low- and high-frequency streams, followed by HiFi-Mamba blocks that apply unidirectional traversal and adaptive state-space modulation to focus low-frequency modeling while selectively integrating high-frequency features, thereby overcoming insensitivity to anatomical details and scanning redundancy in standard Mamba applications to undersampled k-space MRI reconstruction.

What carries the argument

The W-Laplacian block, which performs fidelity-preserving spectral decoupling to produce complementary low- and high-frequency streams that enable focused global modeling and selective detail integration.

If this is right

  • Higher reconstruction accuracy than CNN, Transformer, and existing Mamba models on common MRI benchmarks.
  • Better preservation of high-frequency anatomical structures through the dual-stream separation.
  • Reduced computational cost from the unidirectional traversal while retaining long-range dependency capture.
  • A compact overall model size that still delivers state-of-the-art fidelity.
  • Direct applicability to clinical MRI workflows that require fast, high-quality images from limited k-space samples.

Where Pith is reading between the lines

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

  • The same frequency-splitting idea could be tested on other inverse problems such as CT or ultrasound reconstruction.
  • Unidirectional scanning might generalize to video or 3D medical volumes where multi-directional passes become even costlier.
  • If the spectral streams remain complementary, the architecture could be adapted for tasks that need explicit frequency control like denoising or super-resolution.

Load-bearing premise

The W-Laplacian block actually separates frequencies into truly complementary streams without fidelity loss, and switching to unidirectional traversal keeps the full long-range modeling power.

What would settle it

Run the model on a held-out MRI dataset with a new undersampling mask and measure whether PSNR and SSIM scores drop below those of a standard bidirectional Mamba baseline.

Figures

Figures reproduced from arXiv: 2508.09179 by Fangfang Tang, Feng Liu, Hongli Chen, Jing Hao, Pengcheng Fang, Shanshan Shan, Xiaohao Cai, Yingxuan Ren, Yuxia Chen.

Figure 1
Figure 1. Figure 1: Illustration of scanning and decoupling. (a) Scanning [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed HiFi-Mamba architecture. (a) The HiFi-Mamba Unit splits the input into high- and low [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on the fastMRI and CC359 datasets under single-coil settings. (a) Reconstruction results on the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of HiFi-Mamba Block TABLE VI: Ablation study of HiFi-Mamba with different depth-wise convolution configurations on the CC359 dataset under 8× AF. Left: Current DConv1D in Mamba block. Right: Pre-Dconv1D before split. Mechanism PSNR SSIM NMSE HiFi-Mamba DConv1D(3 × 3) 27.81 0.796 0.030 HiFi-Mamba DConv1D(5 × 5) 28.05 0.805 0.028 HiFi-Mamba DConv1D(7 × 7)* 28.49 0.810 0.026 Mechanism PSNR SSIM NMSE … view at source ↗
Figure 5
Figure 5. Figure 5: Overview of HiFi-Mamba Block 2) Ablation on Gate Placement.: We further investigate the effect of different gating strategies applied to the modulation branches within the HiFi-Mamba block. As shown in [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Data Processing PipeLine A.3 More Results More results are shown in [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison on the fastMRI and CC359 datasets. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
read the original abstract

Reconstructing high-fidelity MR images from undersampled k-space data remains a challenging problem in MRI. While Mamba variants for vision tasks offer promising long-range modeling capabilities with linear-time complexity, their direct application to MRI reconstruction inherits two key limitations: (1) insensitivity to high-frequency anatomical details; and (2) reliance on redundant multi-directional scanning. To address these limitations, we introduce High-Fidelity Mamba (HiFi-Mamba), a novel dual-stream Mamba-based architecture comprising stacked W-Laplacian (WL) and HiFi-Mamba blocks. Specifically, the WL block performs fidelity-preserving spectral decoupling, producing complementary low- and high-frequency streams. This separation enables the HiFi-Mamba block to focus on low-frequency structures, enhancing global feature modeling. Concurrently, the HiFi-Mamba block selectively integrates high-frequency features through adaptive state-space modulation, preserving comprehensive spectral details. To eliminate the scanning redundancy, the HiFi-Mamba block adopts a streamlined unidirectional traversal strategy that preserves long-range modeling capability with improved computational efficiency. Extensive experiments on standard MRI reconstruction benchmarks demonstrate that HiFi-Mamba consistently outperforms state-of-the-art CNN-based, Transformer-based, and other Mamba-based models in reconstruction accuracy while maintaining a compact and efficient model design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces HiFi-Mamba, a dual-stream Mamba-based architecture for high-fidelity MRI reconstruction from undersampled k-space data. It comprises stacked W-Laplacian (WL) blocks that perform fidelity-preserving spectral decoupling into complementary low- and high-frequency streams, and HiFi-Mamba blocks that focus on low-frequency global modeling while adaptively integrating high-frequency details via state-space modulation. A unidirectional traversal strategy replaces redundant multi-directional scanning to improve efficiency without sacrificing long-range modeling. The central claim is that this design yields consistent outperformance over CNN-, Transformer-, and other Mamba-based baselines on standard MRI reconstruction benchmarks while maintaining a compact model.

Significance. If the performance gains are substantiated and the spectral-decoupling mechanism is shown to be near-lossless, the work could advance efficient long-range modeling in medical image reconstruction by addressing Mamba's documented weaknesses in high-frequency sensitivity. The emphasis on complementary streams and reduced scanning redundancy offers a concrete path toward more accurate yet computationally lighter alternatives to Transformers, with potential clinical relevance for accelerated MRI.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (WL block description): the claim that the WL block 'performs fidelity-preserving spectral decoupling, producing complementary low- and high-frequency streams' is load-bearing for attributing accuracy gains to the dual-stream design rather than parameter count or training details, yet no equation defining the W-Laplacian operator, invertibility proof, or quantitative forward-inverse reconstruction metric (e.g., PSNR/SSIM between input and recombined streams) is supplied.
  2. [§4] §4 (HiFi-Mamba block and traversal strategy): the assertion that the unidirectional traversal 'preserves long-range modeling capability with improved computational efficiency' lacks a direct ablation or comparison (e.g., feature correlation or dependency-range metrics) against multi-directional scanning, which is required to confirm that the efficiency gain does not trade off the core long-range advantage of Mamba.
minor comments (2)
  1. [Results] Results section: the abstract states outperformance but the provided text supplies no quantitative tables, error bars, dataset splits, or ablation studies; these must be presented with explicit numerical comparisons and statistical tests to support the 'consistently outperforms' claim.
  2. [Notation and figures] Notation and figures: define the W-Laplacian operator and any learned parameters explicitly; add error-map visualizations in reconstruction figures to allow readers to assess high-frequency detail preservation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We have addressed each major comment point by point below, providing clarifications and indicating the revisions made to strengthen the presentation of the W-Laplacian operator and the traversal strategy analysis.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (WL block description): the claim that the WL block 'performs fidelity-preserving spectral decoupling, producing complementary low- and high-frequency streams' is load-bearing for attributing accuracy gains to the dual-stream design rather than parameter count or training details, yet no equation defining the W-Laplacian operator, invertibility proof, or quantitative forward-inverse reconstruction metric (e.g., PSNR/SSIM between input and recombined streams) is supplied.

    Authors: We agree that the original manuscript did not include an explicit equation for the W-Laplacian operator or supporting quantitative verification of fidelity preservation. In the revised version, we have expanded Section 3 to include the mathematical definition of the W-Laplacian operator as a wavelet-domain spectral filter that decomposes the input into complementary low- and high-frequency streams. We have also added a concise invertibility argument based on the perfect reconstruction property of the underlying wavelet transform. To directly address the concern, we now report forward-inverse reconstruction metrics (PSNR > 48 dB and SSIM > 0.995) on the benchmark datasets in both the main text and supplementary material, confirming that recombination recovers the original signal with negligible loss. These additions clarify that the performance improvements can be attributed to the dual-stream design rather than incidental factors. revision: yes

  2. Referee: [§4] §4 (HiFi-Mamba block and traversal strategy): the assertion that the unidirectional traversal 'preserves long-range modeling capability with improved computational efficiency' lacks a direct ablation or comparison (e.g., feature correlation or dependency-range metrics) against multi-directional scanning, which is required to confirm that the efficiency gain does not trade off the core long-range advantage of Mamba.

    Authors: The referee is correct that the original submission relied primarily on overall runtime and FLOPs comparisons without targeted metrics for long-range dependency preservation. We have revised Section 4 to include a dedicated ablation study comparing unidirectional versus multi-directional scanning. The new analysis reports inter-patch feature correlation for spatially distant regions (difference < 4%) and effective dependency range measurements, showing that the unidirectional strategy maintains nearly equivalent long-range modeling capacity while reducing scanning overhead by 28–32%. These quantitative results are now integrated into the text and figures to substantiate the efficiency claim without compromising the core Mamba advantage. revision: yes

Circularity Check

0 steps flagged

No circularity: architecture is an independent design validated on external benchmarks

full rationale

The paper introduces HiFi-Mamba as a novel dual-stream architecture with W-Laplacian blocks for spectral decoupling and unidirectional traversal. Claims rest on empirical outperformance on standard MRI benchmarks rather than any derivation chain. No equations, fitted parameters renamed as predictions, or self-citations are referenced in the abstract or described claims. The design choices (fidelity-preserving decoupling, adaptive modulation) are presented as independent innovations, not reductions to prior inputs by construction. This is a standard self-contained empirical contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The paper rests on the domain assumption that standard Mamba models suffer from high-frequency insensitivity and redundant scanning, plus two newly introduced architectural components whose independent evidence is limited to the experiments described in the abstract.

axioms (1)
  • domain assumption Mamba variants for vision tasks are insensitive to high-frequency anatomical details and rely on redundant multi-directional scanning.
    Explicitly listed as the two key limitations the new architecture is designed to address.
invented entities (2)
  • W-Laplacian (WL) block no independent evidence
    purpose: Performs fidelity-preserving spectral decoupling to produce complementary low- and high-frequency streams.
    New component introduced to enable the dual-stream design.
  • HiFi-Mamba block no independent evidence
    purpose: Focuses on low-frequency global modeling while selectively integrating high-frequency features via adaptive state-space modulation and uses unidirectional traversal.
    Core novel processing unit of the proposed architecture.

pith-pipeline@v0.9.0 · 5790 in / 1444 out tokens · 47467 ms · 2026-05-19T00:41:17.509858+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CHASM: Cross-frequency Harmonized Axis-Separable Mixing for Spectral Token Operators

    cs.CV 2026-05 unverdicted novelty 7.0

    CHASM introduces a cross-frequency harmonized axis-separable spectral mixer using a shared channel eigenbasis plus per-frequency positive gains, yielding consistent gains over same-backbone baselines in medical and na...

  2. SO-Mamba: State-Ownership Mamba for Unrolled MRI Reconstruction

    cs.CV 2026-05 unverdicted novelty 6.0

    SO-Mamba introduces state-ownership routing in Mamba regularizers for unrolled MRI reconstruction to separate resident carrier content from non-resident evidence across stages.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages · cited by 2 Pith papers · 3 internal anchors

  1. [1]

    Magnetic resonance imaging (mri) studies of knee joint under mechanical loading,

    S. Jerban, E. Y . Chang, and J. Du, “Magnetic resonance imaging (mri) studies of knee joint under mechanical loading,” Magnetic Resonance Imaging, vol. 65, pp. 27–36, 2020. I

    work page 2020

  2. [2]

    Motion artifact reduction techniques in mri: A review,

    L. Varela-Mattatall and R. C. N. D’Arcy, “Motion artifact reduction techniques in mri: A review,” Journal of Magnetic Resonance Imaging , vol. 45, no. 6, pp. 1779–1790, 2017. I

    work page 2017

  3. [3]

    Learning a variational network for reconstruction of accelerated mri data,

    K. Hammernik, T. Klatzer, E. Kobler, M. P. Recht, D. K. Sodickson, T. Pock, and F. Knoll, “Learning a variational network for reconstruction of accelerated mri data,”Magnetic Resonance in Medicine, vol. 79, no. 6, pp. 3055–3071, 2018. I, I

    work page 2018

  4. [4]

    Deep-learning methods for parallel magnetic resonance imaging reconstruction: A survey of the current approaches, trends, and issues,

    F. Knoll, K. Hammernik, C. Zhang, S. Moeller, T. Pock, and D. K. Sodickson, “Deep-learning methods for parallel magnetic resonance imaging reconstruction: A survey of the current approaches, trends, and issues,” IEEE Signal Processing Magazine , vol. 37, no. 1, pp. 128–140,

    work page

  5. [5]

    Compressed sensing mri: a review from signal processing perspective,

    J. C. Ye, “Compressed sensing mri: a review from signal processing perspective,” BMC Biomedical Engineering , vol. 1, no. 1, p. 8, 2019. I

    work page 2019

  6. [6]

    fastMRI: An Open Dataset and Benchmarks for Accelerated MRI

    J. Zbontar, F. Knoll, A. Sriram, M. J. Muckley, M. Bruno, A. Defazio, M. Parente, C. L. Zitnick, D. K. Sodickson, N. Yakubova et al. , “fastmri: An open dataset and benchmarks for accelerated mri,” arXiv preprint arXiv:1811.08839 , 2018. [Online]. Available: https://arxiv.org/abs/1811.08839 I, IV-A1, IV-A1

    work page internal anchor Pith review arXiv 2018

  7. [7]

    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, vol. 58, no. 6, pp. 1182–1195, 2007. I

    work page 2007

  8. [8]

    Unet++: A nested u-net architecture for medical image segmentation,

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: A nested u-net architecture for medical image segmentation,” in International workshop on deep learning in medical image analysis . Springer, 2018, pp. 3–11. I, I

    work page 2018

  9. [9]

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

    J. Schlemper, J. Caballero, J. V . Hajnal, A. N. Price, and D. Rueckert, “A deep cascade of convolutional neural networks for dynamic mr image reconstruction,” IEEE Transactions on Medical Imaging, vol. 37, no. 2, pp. 491–503, 2018. [Online]. Available: https://ieeexplore.ieee.org/document/8067520 I, II-0a

    work page arXiv 2018

  10. [10]

    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, 2019. [Online]. Available: https://ieeexplore.ieee.org/document/8425639 I

    work page arXiv 2019

  11. [11]

    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, 2019. I

    work page 2019

  12. [12]

    Image reconstruction with b 0 inhomogeneity using a deep unrolled network on an open-bore mri- linac,

    S. Shan, Y . Gao, D. Waddington, H. Chen, B. Whelan, P. Liu, Y . Wang, C. Liu, H. Gan, M. Gao et al. , “Image reconstruction with b 0 inhomogeneity using a deep unrolled network on an open-bore mri- linac,” IEEE Transactions on Instrumentation and Measurement , 2024. I

    work page 2024

  13. [13]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , vol. 30,

    work page

  14. [14]

    Available: https://proceedings.neurips.cc/paper files/ paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf I

    [Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf I

    work page 2017

  15. [15]

    Reconformer: Accelerated mri reconstruction using recurrent transformer,

    P. Guo, Y . Mei, J. Zhou, S. Jiang, and V . M. Patel, “Reconformer: Accelerated mri reconstruction using recurrent transformer,”IEEE trans- actions on medical imaging , vol. 43, no. 1, pp. 582–593, 2023. I, II-0b, I

    work page 2023

  16. [16]

    Deep learning based mri reconstruction with transformer,

    Z. Wu, W. Liao, C. Yan, M. Zhao, G. Liu, and N. Ma, “Deep learning based mri reconstruction with transformer,” Computer Methods and Programs in Biomedicine , vol. 234, p. 107602, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0169260723001189 I

    work page 2023

  17. [17]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2021, pp. 10 012–10 022. I

    work page 2021

  18. [18]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, T. Dao et al. , “Efficiently modeling long sequences with structured state spaces,” in Advances in Neural Information Processing Systems , vol. 35, 2022, pp. 21 915–21 929. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/ 2022/file/a8d1c416cfa3ef548e23f9fef3f65c41-Paper-Conference.pdf I

    work page 2022

  19. [19]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    A. Gu, T. Dao et al. , “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023. [Online]. Available: https://arxiv.org/abs/2312.00752 I

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  20. [20]

    Mambair: A simple baseline for image restoration with state-space model,

    H. Guo, J. Li, T. Dai, Z. Ouyang, X. Ren, and S.-T. Xia, “Mambair: A simple baseline for image restoration with state-space model,” in European conference on computer vision . Springer, 2024, pp. 222–

    work page 2024

  21. [21]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,” inAdvances in Neural In- formation Processing Systems, vol. 37, 2024, pp. 103 031–103 063. [On- line]. Available: https://proceedings.neurips.cc/paper files/paper/2024/ file/baa2da9ae4bfed26520bb61d259a3653-Paper-Conference.pdf I, II-0c

    work page 2024

  22. [22]

    Mamba in vision: A comprehensive survey of techniques and applications,

    M. M. Rahman, A. A. Tutul, A. Nath, L. Laishram, S. K. Jung, and T. Hammond, “Mamba in vision: A comprehensive survey of techniques and applications,” arXiv preprint arXiv:2410.03105 , 2024. [Online]. Available: https://arxiv.org/abs/2410.03105 I

    work page arXiv 2024

  23. [23]

    Computation-efficient era: A comprehensive survey of state space models in medical image analysis,

    X. Zhang, R. He, F. Wang, and Q. Liu, “Computation-efficient era: A comprehensive survey of state space models in medical image analysis,” arXiv preprint arXiv:2405.07639 , 2024. [Online]. Available: https://arxiv.org/abs/2405.07639 I

    work page arXiv 2024

  24. [24]

    Tinyvim: Frequency decoupling for tiny hybrid vision mamba,

    X. Ma, Z. Ni, and X. Chen, “Tinyvim: Frequency decoupling for tiny hybrid vision mamba,” arXiv preprint arXiv:2411.17473 , 2024. I

    work page arXiv 2024

  25. [25]

    Ista-net: Interpretable optimization-inspired deep network for image compressive sensing,

    J. Zhang and B. Ghanem, “Ista-net: Interpretable optimization-inspired deep network for image compressive sensing,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 1828–1837. II-0a, I

    work page 2018

  26. [26]

    Kiki- net: Cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images,

    T. Eo, Y . Jun, T. Kim, J. Jang, H. Lee, D. Hwang, and J. C. Ye, “Kiki- net: Cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images,” Magnetic Resonance in Medicine, vol. 80, no. 5, pp. 2188–2201, 2018. [Online]. Available: https://onlinelibrary.wiley.com/doi/10.1002/mrm.27178 II-0a

    work page doi:10.1002/mrm.27178 2018

  27. [27]

    Dudornet: Learning a dual-domain recurrent network for fast mri reconstruction with deep t1 prior,

    B. Zhou, S. Zhou, L. Wang, Y . Xing, Q. Wang, S. Zhang, C. Liu, and H. Lu, “Dudornet: Learning a dual-domain recurrent network for fast mri reconstruction with deep t1 prior,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020, pp. 4273–4282. [Online]. Available: http://openaccess.thecvf.com/content CVPR 202...

    work page 2020

  28. [28]

    Unsupervised mri reconstruction via zero-shot learned adversarial transformers,

    Y . Korkmaz, S. U. Dar, M. Yurt, M. ¨Ozbey, and T. Cukur, “Unsupervised mri reconstruction via zero-shot learned adversarial transformers,” IEEE Transactions on Medical Imaging , vol. 41, no. 7, pp. 1747–1763, 2022. II-0b

    work page 2022

  29. [29]

    Reference-based magnetic resonance image reconstruction using texture transformer,

    Y . Gu, Y . Lu, H. You, Y . Zhan, S. Zhou, and D. Shen, “Reference-based magnetic resonance image reconstruction using texture transformer,” arXiv preprint arXiv:2111.09492 , 2021. [Online]. Available: https://arxiv.org/pdf/2111.09492 II-0b

    work page arXiv 2021

  30. [30]

    Dual-domain accelerated mri reconstruction using transformers with learning-based undersampling,

    J. Wang, S. Wu, Z. Xu, R. Shi, Y . Qian, J. Cai, Y . Huang et al. , “Dual-domain accelerated mri reconstruction using transformers with learning-based undersampling,” Computerized Medical Imaging and Graphics, vol. 106, p. 102179, 2023. [Online]. Available: https://www. sciencedirect.com/science/article/abs/pii/S0895611123000241 II-0b

    work page 2023

  31. [31]

    Swin transformer for fast mri,

    J. Huang, Y . Fang, Y . Wu, H. Wu, Z. Gao, Y . Li, J. Del Ser, J. Xia, and G. Yang, “Swin transformer for fast mri,” Neurocomputing, vol. 493, pp. 281–304, 2022. II-0b

    work page 2022

  32. [32]

    A survey on visual mamba,

    H. Li, Y . Wang, Y . Xu, Z. Ding, C. Xu, Y . Lu, X. Ye, and S. Bai, “A survey on visual mamba,” arXiv preprint arXiv:2404.15956 , 2024. [Online]. Available: https://arxiv.org/abs/2404.15956 II-0c

    work page arXiv 2024

  33. [33]

    Enhancing global sensitiv- ity and uncertainty quantification in medical image reconstruction with monte carlo arbitrary-masked mamba,

    J. Huang, L. Yang, F. Wang, Y . Wu, Y . Nan, W. Wu, C. Wang, K. Shi, A. I. Aviles-Rivero, C.-B. Schoenlieb et al., “Enhancing global sensitiv- ity and uncertainty quantification in medical image reconstruction with monte carlo arbitrary-masked mamba,” Medical Image Analysis, vol. 99, p. 103334, 2025. II-0c

    work page 2025

  34. [34]

    Lmo: Linear mamba operator for mri reconstruction,

    J. Li, C. Wang, Y . Xu, Y . Qian, Y . Yang, and D. Shen, “Lmo: Linear mamba operator for mri reconstruction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2025, pp. 5112–5122. [Online]. Available: https: //openaccess.thecvf.com/content/CVPR2025/papers/Li LMO Linear Mamba Operator for MRI Reconstruction CVPR...

    work page 2025

  35. [35]

    The laplacian pyramid as a compact image code,

    P. J. Burt and E. H. Adelson, “The laplacian pyramid as a compact image code,” in Readings in computer vision . Elsevier, 1987, pp. 671–679. III-C

    work page 1987

  36. [36]

    Boosting vit-based mri reconstruction from the perspectives of frequency modulation, spatial purification, and scale diversification,

    Y . Meng, Z. Yang, Y . Shi, and Z. Song, “Boosting vit-based mri reconstruction from the perspectives of frequency modulation, spatial purification, and scale diversification,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 6, 2025, pp. 6135–

    work page 2025

  37. [37]

    I JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 9

    work page 2021

  38. [38]

    Simultaneous truth and performance level estimation (staple): an algorithm for the validation of image segmentation,

    S. K. Warfield, K. H. Zou, and W. M. Wells, “Simultaneous truth and performance level estimation (staple): an algorithm for the validation of image segmentation,” IEEE Transactions on Medical Imaging , vol. 23, no. 7, pp. 903–921, 2004. IV-A1

    work page 2004

  39. [39]

    The scope of psnr in image and video quality assessment,

    Q. Huynh-Thu and M. Ghanbari, “The scope of psnr in image and video quality assessment,” Electronics letters, vol. 44, no. 13, pp. 800–801,

    work page

  40. [40]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Processing , vol. 13, no. 4, pp. 600–612, 2004. IV-A1

    work page 2004

  41. [41]

    Loss Functions for Neural Networks for Image Processing

    H. Zhao, O. Gallo, I. Frosio, and J. Kautz, “Loss functions for neural networks for image processing,” arXiv preprint arXiv:1511.08861, 2016. IV-A1 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 10 APPENDIX A.1 Ablation Study Details Linear DConv Linear DConv1DDConv1DDConv1D S6 NormNorm Linear Linear Linear Linear Linear Linear Linear DConv Lin...

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  42. [42]

    As shown in Figure 4, we compare two architectural variants

    Ablation on Convolution Placement and Kernel Size.: To assess the impact of depth-wise convolution design in the Mamba block, we conduct ablation experiments on both the placement and kernel size of the 1D depth-wise convolution (DConv1D) using the CC359 dataset under an 8× acceleration factor. As shown in Figure 4, we compare two architectural variants. ...

    work page

  43. [43]

    As shown in Figure 5, we compare three designs that vary in the placement and scope of the 1D gating operations

    Ablation on Gate Placement.: We further investigate the effect of different gating strategies applied to the modulation branches within the HiFi-Mamba block. As shown in Figure 5, we compare three designs that vary in the placement and scope of the 1D gating operations. In the baseline HiFi-Mamba design (Figure 5a), 1D gating is applied only to the high-f...

    work page 2021

  44. [44]

    Normalization: Each 2D image is rescaled to the [0, 1] range using min-max normalization to ensure consistent intensity across samples

    work page

  45. [45]

    Fourier Transform: The normalized image is transformed to the frequency domain using a centered 2D Fast Fourier Transform (FFT)

    work page

  46. [46]

    The mask remains fixed across the dataset and corresponds to a predefined acceleration factor

    Undersampling Mask: A 1D Cartesian equispaced binary mask is applied along the column direction of the k-space. The mask remains fixed across the dataset and corresponds to a predefined acceleration factor

    work page

  47. [47]

    Inverse FFT: The masked k-space is converted back to the image domain using inverse FFT to obtain an aliased (undersampled) image

    work page

  48. [48]

    This preprocessing pipeline simulates aliasing artifacts in a controlled and reproducible manner, enabling supervised learning for MRI reconstruction tasks

    Complex Representation: Both the fully-sampled and undersampled images are represented as two-channel tensors, with real and imaginary components stored separately. This preprocessing pipeline simulates aliasing artifacts in a controlled and reproducible manner, enabling supervised learning for MRI reconstruction tasks. Ground truth k-space Under-sampled ...

    work page 2021