acceptodds
Withdrawn from ICLR 2027

ResiMamba: Budgeted Recurrent State Residency for Efficient MRI Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-range spatial modeling is essential for accelerated MRI reconstruction, yet its recurrent computation is commonly assigned too coarsely: every spatial token in a state-space scan is executed as a recurrent step. We observe that this is largely unnecessary—sparse recurrent access recovers most of the PSNR gain provided by dense recurrence: one resident token per local window recovers about half of this gain and a 6.25% effective budget over 80%, with strongly diminishing returns beyond. Building on this, we separate a stream’s eligibility to carry recurrent state from the token-level residency that decides which of its positions are executed as scan steps, and propose ResiMamba, a learned, budgeted mechanism that grants direct recurrent access to only a small fraction of tokens. A window-constrained allocator selects resident tokens under a fixed local budget; the remaining tokens are not discarded but routed through a dense evidence pathway that both condi- tions the shortened recurrent scan and contributes to the reconstruction. Across five MRI datasets, highly constrained residency retains quality close to dense re- currence: on fastMRI 4×, the 5% and 20% budgets reach 2.62× and 2.14× end- to-end speedups at only 0.10 and 0.02 dB PSNR cost. ResiMamba turns recurrent state access into a quality–latency design axis for efficient MRI reconstruction, with each budget trained as a separate operating point.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
mri recon

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_k2EPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates how much recurrent computation state-space models need for accelerated MRI reconstruction. It proposes ResiMamba, an architecture that enables controlled evaluation of different recurrence budgets.

Critical strengths

The paper provides a detailed analysis of recurrence in state-space models for MRI reconstruction, supported by extensive ablation studies.

Critical weaknesses
  • The relevance of the central question to accelerated MRI is insufficiently motivated. After reading the introduction, the paper feels primarily like an extensive ablation study. The research problem and its significance for advancing accelerated MRI remain unclear.

  • The description in the introduction (lines 74–83) is too abstract to help the reader understand the architecture. The method section is similarly dense, making it difficult to follow the key operations. The operators in lines 137–140 should be defined explicitly in the main text, and the roles introduced in lines 141–144 need a more concrete explanation.

  • A short background section explaining state-space models and their recurrent updates would improve accessibility. The accelerated MRI forward model, the reconstruction problem and its connection to the proposed architecture should be explained more clearly.

  • The fastMRI knee experiments use k-space simulated from magnitude images. This raises serious inverse-crime concerns, as this setup can simplify the reconstruction problem and yield overly optimistic results (Shimron et al., 2022). These experiments therefore provide limited evidence of performance on acquired k-space.

  • The comparisons omit strong, established reconstruction methods such as End-to-End VarNet (Sriram et al., 2020). This makes it difficult to judge the contribution relative to existing high-performing MRI reconstruction methods.

References

Shimron, E. et al. "Implicit data crimes: Machine learning bias arising from misuse of public data". Proceedings of the National Academy of Sciences, 2022.

Sriram, A. et al. "End-to-end variational networks for accelerated MRI reconstruction". Medical Image Computing and Computer Assisted Intervention, 2020.

Questions for authors
  • Why is the amount of recurrence a particularly relevant question for accelerated MRI, and how do the findings advance the field?

  • How is k-space obtained for the datasets listed in lines 286–291? Please clearly distinguish acquired and simulated k-space for the

  • How does ResiMamba compare with End-to-End VarNet under the same acquisition and evaluation settings?

Reviewer_nkuQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The work proposes ResiMamba, which aims to improve the computational efficiency of the conventional Mamba in dense prediction task of MRI reconstruction. The main contribution is the architectural design of several modules, such as the carrier construction, resident selection etc., which filters information and selects position of the feature grid for sparse recurrence computation. Experiments were conducted on several MRI datasets for MRI reconstruction, which shows competitive performance against baselines with lower latency, along with competitive generalisation performance. Ablation study has also been performed to validate the design. Experiments with varying target ratio shows the effectiveness of the ResiMamba with reduced latency.

Critical strengths

Critical strength lies in the objective of studying the question of how much of recurrence a dense reconstruction needs in a Mamba framework. The goal sounds reasonable and useful, for understanding the mechanism of the state space models.

Critical weaknesses

There are several major weaknesses about this paper.

First, the paper is somewhat hard to follow. The terminologies used, such as carrier, access interface, outlet, are not very clear despite they are defined, and seem not quite necessary. It does not really show what exactly the meaning they give. The paper should be written in a simpler way, as the content itself is not that complex.

Second, despite the architectural design is well described, it is not clear why it leads to that specific design. Justification of those design is missing, especially those connections of components look quite complicated, and multiple modules interact. Justification for these particular way of design is needed.

Third, although the paper shows the comparison of the reconstruction and latency results, it does not really seem to answer the question they raise in the beginning: what the scan carries, and which positions execute it. Despite with the target ratio experiments given, it is expected to have more in-depth analysis regarding the two questions. It only somewhat answers how much, but not what and which, which is actually more interesting to know. Some interpretation of the results or visualisation could be helpful.

Questions for authors
  • Justification for the particular architectural design of each proposed module.
  • For those terminologies, what do they exactly refer to in plain words and in Fig. 2?
  • Eq. 3, operations are not properly defined, and what is the fixed operator referring to?
  • Results show lower latency in 'ms', how does this affect the real-world deployment, as the current approaches are in about 0.1s, which seems already quite practical?
  • Table 2, why some PSNR for 4x are not reported in ResiMamba?
  • ACDC is not raw k-space data, how is it used for MRI reconstruction?
  • The work is mainly researched in MRI reconstruction, how will the conclusion/finding generalise to other applications? Will the finding still hold in other dense prediction tasks?
  • In HiFi-Mamba work, they actually reported better results than what is shown in this manuscript. Why not compare with the better performance of that? If comparing with the HiFi-Mamba (P1) in the paper, ResiMamba actually does not really outperform that, and this seems weaken the contribution.
  • Table 4, are the results statistically significant? Comparison against generative model-based approaches are also suggested.
Reviewer_iXBxOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed a new network architecture for MRI recon. It achieves a marginal improvement.

Critical strengths
  1. Throughout experiment
  2. complete ablation study
  3. generalize well across datasets
Critical weaknesses
  1. It is a very complicated architecture.
  2. The improvement is marginal
  3. No significance test
Questions for authors

N/A

Reviewer_4mSxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how much recurrent computation is needed for MRI reconstruction. The proposed ResiMamba processes selected features through a sparse recurrent pathway while preserving the remaining information through a dense evidence pathway. Experiments on five MRI datasets show that sparse recurrence reduces inference time while maintaining reconstruction quality close to dense recurrence.

Critical strengths
  1. The motivation is clear. The paper addresses whether every spatial position needs recurrent processing, while preserving unselected information rather than discarding it.
  2. The experiments cover five datasets and include measured inference time, comparisons with depth reduction, and ablations examining the main design choices.
Critical weaknesses
  1. The title and abstract on OpenReview differ from those in the submitted PDF.
  2. The main fastMRI experiments use k-space synthesized from magnitude images, and the ACDC radial masks are applied on a Cartesian grid. Although M4Raw includes acquired multi-coil data, the full recurrence-budget study is not repeated there. This limits the evidence under realistic acquisition settings.
  3. The efficiency comparison could be stronger. The depth-reduced models remain slower than the sparse models, and inference times for simpler selection strategies are not reported. Relevant recent methods, such as MambaRecon, are also missing from the comparisons.
  4. The reconstruction gains over the strongest baselines are modest in several settings, and statistical significance is not assessed. Although case-level standard deviations are reported, paired statistical tests or confidence intervals for the improvements are missing, making it difficult to judge whether the gains are consistent across cases.
Questions for authors
  1. Can you provide the evidence-only, sparse, and dense recurrence comparison on acquired multi-coil data, such as M4Raw? Does the same quality-efficiency trade-off hold?
  2. Which validation sets were used for checkpoint and budget selection? Were these separate from the evaluation sets?
  3. Were models trained separately on each dataset? If so, do these results show consistency across datasets rather than generalisation to unseen domains? Have you evaluated transfer to an unseen anatomy or acquisition setting?
  4. Are there failure cases where sparse recurrence loses fine structures or pathological details? How does it compare with dense recurrence under motion artefacts, increased noise, or corrupted k-space?
  5. Did you consider encoding the original 2D coordinates or relative spatial offsets of selected tokens, beyond the raster-gap encoding in Appendix D.3?
  6. How do inference time, total-network FLOPs, and peak GPU memory compare with the baselines under the same experimental settings? Do these measurements include selection, packing, scattering, and data consistency?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.