pith. sign in

arxiv: 2605.22031 · v1 · pith:NTPQNLBBnew · submitted 2026-05-21 · 💻 cs.CV

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

Pengcheng Fang , Hongli Chen , Fangfang Tang , Feng Liu , Xiaohao Cai , Shanshan Shan This is my paper

Pith reviewed 2026-05-22 06:56 UTC · model grok-4.3

classification 💻 cs.CV
keywords MRI reconstructionMambaunrolled networksstate-space modelsState-Ownership Routeraccelerated imagingimage reconstructionstate ownership
0
0 comments X

The pith

SO-Mamba assigns reconstruction evidence to separate recurrent residency and non-resident streams via a State-Ownership Router in unrolled MRI solvers.

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

The paper introduces SO-Mamba to address mixing of persistent and update-dependent information in Mamba-based regularizers for accelerated MRI reconstruction. In data-consistency-coupled unrolled networks, stages operate on different iterates, so the method routes evidence to recurrent residency for coherent content, state-interface access for adaptation, and non-state output correction to avoid placing both in the same recurrent path. This is realized through the State-Ownership Router that builds a resident carrier for the main Mamba route while sending non-resident signals to modulate B/C interfaces and an output outlet. A two-level outer-band leakage diagnostic is added to separate hidden-state storage from readout expression. Tests on five public benchmarks with varied anatomies, sampling, and coils show gains over CNN, Transformer, and Mamba baselines at similar efficiency.

Core claim

In a data-consistency-coupled unrolled solver, different stages operate on different reconstruction iterates, where the resident carrier should preserve coherent reconstruction content across stages while stage-dependent non-resident evidence is tied to the current update. SO-Mamba implements this ownership rule with a State-Ownership Router, which constructs a resident carrier for recurrent content and routes non-resident evidence to affine modulation of the B/C state interfaces and an output correction outlet. The resident carrier supplies the Mamba content route, while the non-resident evidence stream adapts the state interfaces and contributes through the output outlet without entering 0

What carries the argument

The State-Ownership Router (SOR), which separates reconstruction evidence into a resident carrier for recurrent Mamba content and non-resident streams for affine modulation of state interfaces plus output correction.

If this is right

  • Preservation of anatomically coherent structures across successive unrolled reconstruction stages.
  • Consistent quality gains over CNN, Transformer, and Mamba baselines on diverse MRI benchmarks.
  • Competitive computational efficiency while maintaining the long-range modeling advantages of state-space models.
  • A two-level outer-band leakage diagnostic that isolates hidden-state storage from post-scan readout expression.

Where Pith is reading between the lines

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

  • The ownership routing idea could transfer to unrolled solvers for other inverse problems such as CT reconstruction where stage-wise iterate differences also arise.
  • The outer-band leakage diagnostic offers a general tool for inspecting state trajectories in selective-scan models applied to sequential data outside imaging.
  • If the separation reduces mixing artifacts, it may allow fewer unrolling iterations while retaining reconstruction fidelity in clinical pipelines.

Load-bearing premise

The assumption that separating reconstruction evidence into recurrent residency, state-interface access, and non-state output correction via the State-Ownership Router will preserve coherent structures across unrolled stages without introducing new inconsistencies or artifacts.

What would settle it

If SO-Mamba produces no measurable gains in structural coherence metrics or introduces new artifacts on the five public MRI benchmarks relative to standard Mamba, the separation approach would be shown not to deliver the intended benefit.

Figures

Figures reproduced from arXiv: 2605.22031 by Fangfang Tang, Feng Liu, Hongli Chen, Pengcheng Fang, Shanshan Shan, Xiaohao Cai.

Figure 1
Figure 1. Figure 1: Overview of the SO-Mamba architecture. (a) The full network stacks six SO-Mamba groups; each group [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison on fastMRI and CC359 under single-coil settings. (a) Reconstruction results on [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
read the original abstract

Accelerated MRI reconstruction requires recovering missing details while preserving anatomically coherent structures across large spatial regions. State-space models such as Mamba provide efficient long-range modeling, making them attractive learned regularizers for unrolled reconstruction. However, in a data-consistency-coupled unrolled solver, different stages operate on different reconstruction iterates, where the resident carrier should preserve coherent reconstruction content across stages while stage-dependent non-resident evidence is tied to the current update. Treating these roles uniformly can place persistent resident-carrier evidence and update-dependent non-resident evidence into the same recurrent content route. We therefore propose SO-Mamba, a state-ownership Mamba regularizer that assigns reconstruction evidence within each Mamba stage to recurrent residency, state-interface access, and non-state output correction. SO-Mamba implements this ownership rule with a State-Ownership Router (SOR), which constructs a resident carrier for recurrent content and routes non-resident evidence to affine modulation of the B/C state interfaces and an output correction outlet. The resident carrier supplies the Mamba content route, while the non-resident evidence stream adapts the state interfaces and contributes through the output outlet without entering the recurrent content route. We further introduce a two-level outer-band leakage diagnostic that separates hidden-state storage from readout expression by measuring outer-band energy in the selective-scan state trajectory and the post-scan Mamba readout. Experiments on five public MRI reconstruction benchmarks spanning diverse anatomies, sampling patterns, and coil configurations show that SO-Mamba consistently improves over CNN-, Transformer-, and Mamba-based baselines with competitive computational efficiency.

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 paper proposes SO-Mamba, a state-ownership Mamba regularizer for unrolled MRI reconstruction. It introduces a State-Ownership Router (SOR) that assigns reconstruction evidence to recurrent residency (resident carrier), state-interface access (affine modulation of B/C), and non-state output correction. A two-level outer-band leakage diagnostic is added to monitor hidden-state storage versus readout. Experiments on five public MRI benchmarks with diverse anatomies, sampling patterns, and coil configurations report consistent improvements over CNN-, Transformer-, and Mamba-based baselines while maintaining competitive computational efficiency.

Significance. If the empirical gains hold under rigorous controls, the work offers a targeted architectural fix for mixing persistent and stage-dependent signals in data-consistency-coupled unrolled solvers. This could improve long-range coherence in accelerated MRI without sacrificing the efficiency of selective state-space models. The leakage diagnostic provides a new analysis tool for hidden-state behavior in such networks. The multi-benchmark evaluation across anatomies strengthens generalizability claims.

major comments (2)
  1. [§3.2] §3.2, SOR definition: the routing logic that constructs the resident carrier and routes non-resident evidence to B/C modulation and the output outlet is described at a high level but lacks an explicit equation or pseudocode showing how the three ownership paths are computed from the input feature map; this makes it impossible to verify that the separation is parameter-efficient and does not introduce new inconsistencies across unrolled stages.
  2. [Table 2] Table 2 (main results): while average PSNR/SSIM gains are stated, the table omits per-benchmark standard deviations across runs or statistical significance tests; without these, the claim of 'consistent' improvement over Mamba baselines cannot be assessed for robustness across the five datasets with varying coil configurations and sampling patterns.
minor comments (2)
  1. [§4.3] §4.3, leakage diagnostic: the two-level outer-band energy measurement is introduced but the exact frequency cutoffs and normalization are not specified, hindering exact reproduction of the diagnostic curves shown in Figure 4.
  2. [Abstract] Abstract and §1: the phrase 'competitive computational efficiency' is used without reference to concrete metrics (FLOPs, latency, or memory) or direct comparison numbers against the Mamba baseline; a short efficiency table would clarify this.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for minor revision. We address each major comment below and will revise the manuscript to incorporate the requested clarifications and additional analyses.

read point-by-point responses

  1. Referee: [§3.2] §3.2, SOR definition: the routing logic that constructs the resident carrier and routes non-resident evidence to B/C modulation and the output outlet is described at a high level but lacks an explicit equation or pseudocode showing how the three ownership paths are computed from the input feature map; this makes it impossible to verify that the separation is parameter-efficient and does not introduce new inconsistencies across unrolled stages.

    Authors: We agree that the current description of the State-Ownership Router (SOR) in §3.2 is at a high level and would benefit from explicit equations and pseudocode. In the revised manuscript we will add a formal definition of the three ownership paths, including the computation of the resident carrier from the input feature map, the routing of non-resident evidence to affine modulation of the B/C state interfaces, and the output correction outlet. These additions will explicitly show the parameter count and confirm that no new cross-stage inconsistencies are introduced. revision: yes

  2. Referee: [Table 2] Table 2 (main results): while average PSNR/SSIM gains are stated, the table omits per-benchmark standard deviations across runs or statistical significance tests; without these, the claim of 'consistent' improvement over Mamba baselines cannot be assessed for robustness across the five datasets with varying coil configurations and sampling patterns.

    Authors: We acknowledge that reporting standard deviations and statistical significance would strengthen the robustness claims. In the revised manuscript we will augment Table 2 with per-benchmark standard deviations obtained from three independent runs with different random seeds and will add a footnote or supplementary section reporting paired statistical tests (e.g., Wilcoxon signed-rank) against the Mamba baselines. This will allow readers to assess consistency across the diverse coil and sampling configurations. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained architectural proposal

full rationale

The paper proposes SO-Mamba as a new Mamba variant for unrolled MRI reconstruction, introducing the State-Ownership Router (SOR) to assign roles to resident carrier evidence, state-interface modulation, and output correction. This is presented as a design motivated by the need to handle stage-dependent iterates in data-consistency-coupled solvers, not as a mathematical derivation from first principles or fitted parameters. Central claims rest on empirical improvements over baselines on five external public MRI benchmarks with diverse anatomies and sampling patterns; no equations, predictions, or uniqueness theorems are shown to reduce by construction to the paper's own inputs, self-citations, or ansatzes. The leakage diagnostic is likewise an introduced monitoring tool rather than a self-referential result.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim depends on the effectiveness of the newly introduced SOR and diagnostic; these are design choices rather than derivations from external benchmarks or first principles.

free parameters (1)
  • SOR learnable parameters
    The router contains trainable weights that are fitted during end-to-end training on MRI data.
axioms (1)
  • domain assumption Unrolled iterative reconstruction with explicit data-consistency steps remains a valid solver framework.
    The paper builds directly on existing unrolled MRI pipelines without re-deriving their convergence properties.
invented entities (1)
  • State-Ownership Router (SOR) no independent evidence
    purpose: Assigns reconstruction evidence to recurrent carrier, state interfaces, and output correction paths.
    New component invented to enforce the ownership rule inside each Mamba stage.

pith-pipeline@v0.9.0 · 5823 in / 1287 out tokens · 53786 ms · 2026-05-22T06:56:15.276302+00:00 · methodology

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

14 extracted references · 14 canonical work pages · 5 internal anchors

  1. [1]

    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, and Feng Liu. Hifi-mamba: Dual-stream w-laplacian enhanced mamba for high-fidelity mri reconstruction.arXiv preprint arXiv:2508.09179,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L. Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  4. [5]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    URLhttps://arxiv.org/abs/2312.00752. 9 APREPRINT- MAY22, 2026 Pengfei Guo, Yiqun Mei, Jinyuan Zhou, Shanshan Jiang, and Vishal M Patel. Reconformer: Accelerated mri reconstruction using recurrent transformer.IEEE transactions on medical imaging, 43(1):582–593,

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  5. [6]

    Jiahao Huang, Yingying Fang, Yinzhe Wu, Huanjun Wu, Zhifan Gao, Yang Li, Javier Del Ser, Jun Xia, and Guang Yang

    doi: 10.1002/mrm.26977. Jiahao Huang, Yingying Fang, Yinzhe Wu, Huanjun Wu, Zhifan Gao, Yang Li, Javier Del Ser, Jun Xia, and Guang Yang. Swin transformer for fast mri.Neurocomputing, 493:281–304,

    work page doi:10.1002/mrm.26977

  6. [7]

    Li, Berlin Chen, Caitlin Wang, Aviv Bick, J

    Aakash Lahoti, Kevin Y Li, Berlin Chen, Caitlin Wang, Aviv Bick, J Zico Kolter, Tri Dao, and Albert Gu. Mamba-3: Improved sequence modeling using state space principles.arXiv preprint arXiv:2603.15569,

    work page arXiv

  7. [8]

    URLhttps://ieeexplore.ieee.org/document/8425639

    doi: 10.1109/TMI.2018.2863670. URLhttps://ieeexplore.ieee.org/document/8425639. Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmenta- tion. InInternational Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer,

    work page doi:10.1109/tmi.2018.2863670 2018

  8. [9]

    URL https://www.sciencedirect.com/science/article/abs/ pii/S0895611123000241

    doi: 10.1016/j.compmedimag.2023.102179. URL https://www.sciencedirect.com/science/article/abs/ pii/S0895611123000241. Shanshan Wang, Zhenghang Su, Leslie Ying, Xi Peng, Shun Zhu, Feng Liang, Dagan Feng, and Dong Liang. Accelerating magnetic resonance imaging via deep learning. In2016 IEEE 13th international symposium on biomedical imaging (ISBI), pages 51...

    work page doi:10.1016/j.compmedimag.2023.102179 2023

  9. [10]

    Mamba-unet: Unet- like pure visual mamba for medical image segmentation,

    Ziyang Wang, Jian-Qing Zheng, Yichi Zhang, Ge Cui, and Lei Li. Mamba-unet: Unet-like pure visual mamba for medical image segmentation.arXiv preprint arXiv:2402.05079,

    work page arXiv

  10. [11]

    Simultaneous truth and performance level estimation (staple): an algorithm for the validation of image segmentation.IEEE Transactions on Medical Imaging, 23(7):903–921,

    10 APREPRINT- MAY22, 2026 Simon K Warfield, Kelly H Zou, and William M Wells. Simultaneous truth and performance level estimation (staple): an algorithm for the validation of image segmentation.IEEE Transactions on Medical Imaging, 23(7):903–921,

    work page 2026

  11. [12]

    Jure Zbontar, Florian Knoll, Anuroop Sriram, Matthew J Muckley, Michael Bruno, Aaron Defazio, M Parente, C Lawrence Zitnick, Daniel K Sodickson, Naz Yakubova, et al

    doi: 10.1186/s42490-019-0006-z. Jure Zbontar, Florian Knoll, Anuroop Sriram, Matthew J Muckley, Michael Bruno, Aaron Defazio, M Parente, C Lawrence Zitnick, Daniel K Sodickson, Naz Yakubova, et al. fastmri: An open dataset and benchmarks for accelerated mri.arXiv preprint arXiv:1811.08839,

    work page doi:10.1186/s42490-019-0006-z

  12. [13]

    fastMRI: An Open Dataset and Benchmarks for Accelerated MRI

    URLhttps://arxiv.org/abs/1811.08839. Jian Zhang and Bernard Ghanem. Ista-net: Interpretable optimization-inspired deep network for image compressive sensing. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1828–1837,

    work page internal anchor Pith review arXiv

  13. [14]

    For Prostate158, we adopt the same procedure, except that the sampling mask is replaced with a random mask

    A Detailed Experimental Settings A.1 Data Preprocessing The preprocessing for fastMRI and CC359 follows the data preprocessing pipeline of HiFi-Mamba. For Prostate158, we adopt the same procedure, except that the sampling mask is replaced with a random mask. For ACDC, we employ a golden-angle radial sampling mask and first perform a center crop to an imag...

    work page 2026

  14. [15]

    The CNN-based baselines include UNet and ISTA-Net

    A.6 Baselines We compare SO-Mamba with CNN-, Transformer-, and Mamba-based reconstruction baselines. The CNN-based baselines include UNet and ISTA-Net. Transformer-based baselines include TransUNet, ReconFormer, and FpsFormer. Mamba-based baselines include Mamba-UNet, LMO, and HiFi-Mamba. For reproduced baselines, we use the same data preprocessing, sampl...

    work page 2026