Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Medical image restoration (MedIR) demands precise modeling of anisotropic spatial dependencies, where directional anatomical patterns are frequently degraded by conventional methods. We propose Directional Adaptive Shuffle Mamba (DASMamba), a state-space model architecture that addresses this challenge through two novel components: (1) the Directional Adaptive Shuffle Module (DASM), which captures long-range dependencies via directional adaptive random shuffle and selective scanning, and (2) the Dual-path Feedforward Network (DPFN), enhancing feature representation through multi-scale learning and dynamic channel fusion. By integrating these modules into a hierarchical U-shaped architecture, DASMamba achieves state-of-the-art performance on MRI super-resolution, CT denoising, and PET synthesis tasks while maintaining linear computational complexity. Our framework’s ability to preserve diagnostically critical structural details underscores its clinical value.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0433_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/cc111mp/DASMamba-MedIR

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ChaSim_Directional_MICCAI2025,
        author = { Chan, Simon C. K. and Shi, Lulin and Huang, Bingxin and Wong, Terence T. W.},
        title = { { Directional Adaptive Shuffle-Based Visual State-Space Models for Medical Image Restoration } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15972},
        month = {September},
        page = {160 -- 170}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduce DASMamba, an SSM-based framework that can capture anisotropic structures via adaptive shuffling in the SS2D module. The experiment results show the effectiveness of the DASMamba.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper designs a novel 2D-Directional Adaptive Shuffle Scan Block to preserve anisotropic structures via adaptive shuffling while modeling global contexts through selective scanning.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. The concept of shuffling in SSM has already been proposed in prior work [1][2]. It would be beneficial to discuss how your ASM design differs from and improves upon these existing methods. Furthermore, the current architecture appears quite similar to MambaIRv2 [2], so it would be helpful to clarify your unique advantages relative to that approach.
    2. The DPFN module in your paper has a structure reminiscent of certain previously proposed designs [3][4] for image restoration. Please highlight its specific differences and improvements beyond those prior works.
    3. In medical imaging, the term “anisotropic features” often refers to 3D volumes. In your Introduction, it might be valuable to include more discussion on how the idea transfers to 2D images and why it serves as an effective motivation in 2D medical image restoration.
    4. Although your paper states that the proposed method achieves linear complexity, it appears there is also a use of self-attention (MHSA) that typically has quadratic complexity. Please elaborate on how your approach remains computationally efficient despite including self-attention.
    5. Consider moving Figure 1 to page two or three to improve readability and flow of the text.
    6. Figure 3 is inserted as an image and appears unclear when enlarged. It would be helpful to ensure a higher resolution or an alternative format that enhances readability.
    7. In Figure 2, consider reducing the image size slightly and adding visual examples from the other two tasks. This will provide readers with a more comprehensive view of the method’s performance.

    Cao K, He X, Hu T, et al. Shuffle mamba: State space models with random shuffle for multi-modal image fusion[J]. arXiv preprint arXiv:2409.01728, 2024. Guo H, Guo Y, Zha Y, et al. MambaIRv2: Attentive State Space Restoration[J]. arXiv preprint arXiv:2411.15269, 2024 Li J, Fang F, Mei K, et al. Multi-scale residual network for image super-resolution[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 517-532. Chen X, Li H, Li M, et al. Learning a sparse transformer network for effective image deraining[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2023: 5896-5905.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This study shows a degree of methodological innovation by integrating an 2D-Directional Adaptive Shuffle Scan Block to enhance global features. However, there is still room for further discussion regarding the novelty, methodological details, and the validation of results. Therefore, I rate it as “weak reject.”

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    While the authors clarified the distinction between their approach and previous methods in the rebuttal, demonstrating a certain degree of novelty, the improvements reported in the ablation study—particularly those attributed to DPFN and 2D-ASM—appear limited. Moreover, given that the model is positioned as an all-in-one solution for multiple tasks, conducting ablation experiments on only a single task does not sufficiently validate the proposed architecture.

    Additionally, in the qualitative results, the lack of comparison with the method that performs second-best in the quantitative evaluations raises concerns about fairness and completeness in the evaluation.

    Based on these considerations, I do not believe the paper is ready for acceptance and thus recommend rejection.



Review #2

  • Please describe the contribution of the paper

    This paper introduce a Mamba-based method for medical image restoration. The two key contributions are: (1) a direction adaptive shuffle module for modeling global depedencies and (2) a dual-path feedforward network for multi-scale learning and dynamic channel fusion.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    1. This paper introduce a lightweight Mamba model for medical image restoration.
    2. A novel direction adaptive shuffle module is proposed for modeling global depedency.
    3. A novel dual-path feedforward network is proposed for learn multi-scale information and dynamic channel fusion.
    4. Extensive experiments have been conducted to validate the effectiveness of the method.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1. A primary limitation of this study is found in the ablation experiments, which demonstrate only marginal improvements from the proposed component. This is likely due to the small size of the AAPM dataset, making significant improvements difficult to achieve. I recommend conducting ablation studies on the MRI super-resolution task, where a larger dataset might better showcase the component’s benefits.

    2. Several important references on general medical image restoration are missing. Aggregating various MedIR tasks to develop comprehensive algorithms is a recent trend, so it is crucial to include a complete set of citations. For example: [1] Huang J, Yang L, Wang F, et al. Enhancing global sensitivity and uncertainty quantification in medical image reconstruction with Monte Carlo arbitrary-masked mamba. Medical Image Analysis, 2025, 99: 103334. [2] Yang Z, Chen H, Qian Z, et al. Region attention transformer for medical image restoration. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024: 603–613.

    3. The PET images in Figure 3 should be rescaled according to their pixel dimensions. As presented, the PET image appears stretched, which could misrepresent the data.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Extensive experiments have validated the performance of the proposed method.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The author has addressed my concerns.



Review #3

  • Please describe the contribution of the paper

    The paper proposes a novel method that achieves superior performance in medical image restoration tasks such as MRI super-resolution, CT denoising, and PET synthesis. This approach incorporates two innovative modules: 1. the directional adaptive shuffle module, which performs directional random shuffle and selective scanning to capture long-range dependencies; 2. multi-scale learning combined with dynamic channel fusion to enhance feature representation.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    This method significantly improves CT denoising, PET synthesis, and MRI super-resolution, enhancing details while avoiding artifacts, making it widely applicable in image restoration.

    The DPFN network demonstrates considerable innovation, where the authors ingeniously leverage this module to enhance feature representation.

    The algorithm validation is comprehensive, with comparisons against SOTA models.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    A more fine-grained analysis of the method’s performance across different datasets would be beneficial—for example, identifying which types of CT denoising problems see the greatest improvement.

    Introducing anatomical prior weighting into the loss function may help focus optimization on clinically relevant structures while diminishing interference from unimportant areas.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The DPFN network demonstrates considerable innovation, where the authors ingeniously leverage this module to enhance feature representation.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The DPFN network demonstrates considerable innovation.



Author Feedback

Thank you to all reviewers for your insightful comments and constructive feedback, which we have carefully considered. We address the concerns in two groups: (i) Methodology and (ii) Other considerations. #Methodology Q1(R3) Novelty relative to prior methods. A1: DASMamba (ours) introduces distinct mechanisms for feature reorganization and adaptive scanning tailored for MedIR and therefore differs from ShuffleMamba [1], MambaIRv2 [2], VMamba[5]: Feature reorganization: While [1] shuffles 1D flattened tokens (permuting from (HW)!) and [2] uses semantic 1D reordering (ASE), (Ours) employs a 2D spatial grid shuffle. We independently permute rows (H) and columns (W) of the 2D feature map (H! W! formations) before scanning. This 2D-aware permutation, unlike 1D shuffles, is to present varied spatial arrangements to our subsequent four-directional scanning, breaking fixed orientations to better model anisotropic MedIR structures (e.g., vessels). The nature of how topology is altered and its implications for directional scanning are thus fundamentally different. Scanning mechanism: [1] Standard Mamba block on its 1D shuffled sequence. [2] Single-directional scan with ASE. Our novelty lies in the pipeline: The four-directional scan operates on the 2D spatially shuffled grid (from A1), unlike methods scanning original maps (e.g.,[5]). Group-wise prompts, derived from this shuffled representation’s avg/max pooling, modulate scan parameters in the C-matrix, making the scan adaptive to the current permuted spatial context. Scanned outputs are adaptively fused per group, then unshuffled. This 2D-ASM output is then element-wise multiplied (gated) by a parallel, unshuffled branch (Fig. 1b). The gating (scanned shuffled path modulated by an unshuffled one) preserves spatial consistency & captures MedIR long-range context. This integrated design from 2D shuffling through adaptive scanning to gating distinguishes us from [1,2] & enhances anisotropic detail. Q2(R3) DPFN comparison. A2: DPFN also uses multi-scale paths, but its fusion and efficiency differ from [3,4]. Interleaved channel fusion places 3x3 & 5x5 features adjacently for fine-grained interaction, unlike block-wise concatenation. DPFN then uniquely refines this via an additive sum of processed paths for enhanced feature blending, followed by a multiplicative gate with a parallel stream for dynamic feature selection. This, with group convolutions and efficient projections, achieves richer, adaptive multi-scale representations with minimal overhead. Table 2 confirms DPFN’s superior metrics (e.g., PSNR) over MSFN/DGFF variants. Q3(R3) Anisotropic features in 2D MedIR. A3: Even in 2 D slices, many medical structures — vessels, fiber bundles, organ edges — exhibit strong orientation-specific textures. Treating them isotopically risks blurring these cues. Ours addresses this by (i) row/column shuffles that expose multiple directions and (ii) four directional scans, so filters adapt to each dominant axis rather than averaging them away, preserving critical diagnostic patterns. We will clarify this 3 D → 2 D transfer in the final manuscript. Q4(R3) Claim of linear complexity. A4: The 2D ASM uses a fixed 8 × 8 window multi head self-attention (MHSA), keeping attention at O(HW). Because DPFN and the SSM blocks are also linear, the overall model complexity remains linear. We will state this explicitly in the paper. #Others Q5(R1). We agree with R1 that fine-grained dataset analysis and anatomical loss priors are promising and valuable future work. Q6(R2). For R2, we will include the requested ablations in future extensions, expand the citations (including those suggested), and rescale Fig. 3(c). Q7(R3). We will adopt R3’s suggestions on Fig placement, resolution, and content to improve readability and clarity. Q8(MR). Regarding the meta reviewer’s comment on Fig. 2: We apologize for any inappropriate anatomical depiction in Fig. 2 and will replace it with lung slices to address this concern.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    Figure 2 may depict inappropriate anatomical regions of a male patient. Please consider replacing it with slices from the lung or abdomen.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Although the feedback has been mixed, the approach offers a fresh perspective and is quite innovative. I recommend accepting it for further consideration



back to top