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

Abstract

State Space Models (SSMs) have recently demonstrated outstanding performance in long-sequence modeling, particularly in natural language processing. However, their direct application to medical image segmentation poses several challenges. SSMs, originally designed for 1D sequences, struggle with 3D spatial structures in medical images due to discontinuities introduced by flattening. Additionally, SSMs have difficulty fitting high-variance data, which is common in medical imaging. In this paper, we analyze the intrinsic limitations of SSMs in medical image segmentation and propose a unified U-shaped encoder-decoder architecture, Meta Mamba UNet (MM-UNet), designed to leverage the advantages of SSMs while mitigating their drawbacks. MM-UNet incorporates hybrid modules that integrate SSMs within residual connections, reducing variance and improving performance. Furthermore, we introduce a novel bi-directional scan order strategy to alleviate discontinuities when processing medical images. Extensive experiments on AMOS22, ACDC, and BTCV datasets demonstrate the superiority of MM-UNet over state-of-the-art methods. MM-UNet achieves a Dice score of 91.5% and 89.0% on AMOS22 task 1 and task 2, surpassing nnUNet by 1.7% and 1.0%, 93.2% Dice on ACDC, and 87.1% Dice on BTCV. These results confirm the effectiveness of integrating SSMs in medical image segmentation through architectural design optimizations. We release our code at https://github.com/bxie9/MM-UNet.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6340_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/bxie9/MM-UNet

Link to the Dataset(s)

N/A

BibTex

@InProceedings{XieBin_MMUNet_MICCAI2026,
        author = { Xie, Bin AND Yan, Yan AND Agam, Gady},
        title = { { MM-UNet: Meta Mamba UNet for Volumetric Medical Image Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes MM-UNet, a hybrid U-shaped framework specifically designed for 3D medical image segmentation. The architecture integrates State Space Models , specifically Mamba, with CNNs. The primary contributions are: The authors introduce a “Hybrid Module” that embeds Mamba blocks within a residual connection following two sequential CNN layers. This design aims to leverage the local inductive bias of CNNs to provide a smoother, lower-variance feature distribution, which addresses the inherent difficulty of SSMs in fitting high-variance intensity data in medical images. To adapt the 1D sequence-modeling nature of Mamba to 3D volumetric data, the paper proposes a 1D Bi-directional scanning strategy. The authors perform a comprehensive ablation study on nine different scanning configurations to identify the optimal balance between spatial coherence and computational efficiency. The framework incorporates an 8-fold Test-Time Augmentation strategy and 3D Gaussian weighting during inference. These techniques are designed to suppress boundary-related errors and enhance the spatial consistency of the final segmentation maps. MM-UNet is evaluated on three diverse and major 3D medical imaging benchmarks: AMOS22, BTCV, and ACDC. The results demonstrate that MM-UNet achieves state-of-the-art performance, outperforming established CNN-based and Transformer-based models.

  • 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 effectively identifies limitation of SSMs when applied to medical imaging: the difficulty in fitting high-variance intensity data. The proposed Hybrid Module, which places Mamba blocks after dual CNN layers and within a residual structure, is a well-motivated architectural adaptation. This design successfully leverages the local inductive bias of CNNs to provide a more stable feature distribution for Mamba’s global sequence modeling. One of the standout aspects of this work is the extensive ablation study on 3D scan orders. Rather than arbitrarily choosing a scanning path, the authors rigorously evaluated nine different configurations. This thorough investigation provides valuable empirical insights for the community regarding the optimal way to serialize volumetric medical data for SSM-based architectures. The MM-UNet demonstrates consistently strong performance across three major and diverse 3D medical imaging datasets: AMOS22, BTCV, and ACDC. Achieving competitive Dice scores against established heavyweights like nnU-Net and Swin-UNETR highlights the practical potential of the proposed framework in clinical segmentation tasks. The paper provides interesting visualizations of Mamba’s attention maps. By comparing forward and backward scanning patterns, the authors offer a clear visual demonstration of how the bidirectional scanning strategy captures complementary receptive fields. This adds a layer of interpretability to the model, showing how it maintains spatial coherence despite the 1D flattening of 3D volumes. The inclusion of practical inference-time strategies, such as 3D Gaussian weighting to suppress boundary errors, shows a clear focus on clinical feasibility. This attention to detail suggests that the method is designed with real-world medical image characteristics in mind, where boundary precision and spatial consistency are paramount.

  • 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.

    The authors explicitly state that an 8-fold Test-Time Augmentation via axis-flipping was employed during inference for MM-UNet. However, it is fundamentally unclear whether the baseline models (especially SwinUNETR, VMUNet, and UMamba) in Tables 1-3 were evaluated under the exact same 8-fold TTA protocol. If the baselines were evaluated without TTA, the marginal performance gains (e. g. , +0.2% over Swin-UNETR on BTCV) are heavily confounded by this inference-time trick rather than the proposed architectural novelty. The authors should provide a fair comparison table without TTA for all methods, or explicitly confirm that all reported baseline results also utilize the exact same TTA strategy. There is a logical paradox between the authors’ architectural claims and their inference strategy. To justify the 8-fold TTA, the authors argue that “This is analogous to applying multiple scan orders in Mamba”. However, this directly contradicts their own findings in the ablation study (Section 2.3), where they demonstrated that extending to multiple scanning pairs (e. g. , 3 pairs or 6 pairs) yields lower Dice scores (0.908 and 0.907, respectively) because “multiple pairs provide no additional benefit and may introduce redundancy”. Furthermore, they explicitly claimed earlier that “a single pair of opposite scan orders is sufficient to represent the input”. The authors argue that SSMs struggle with high-variance data and propose embedding Mamba within a residual block after two CNN layers because “pre-trained feature maps behind CNN layers. . . exhibit lower variance”. This mathematical justification is questionable. If the fundamental limitation of SSMs is strictly the statistical variance of the input features, standard Normalization techniques (e. g. , LayerNorm, InstanceNorm)—which mathematically standardize the features to have unit variance—would entirely resolve this issue without the need for CNNs. In the abstract and contributions, the authors state: “We introduce a bi-directional scan order strategy to mitigate discontinuities. . . “. Presenting this as a novel introduction is inaccurate and overlooks the established consensus in the recent SSM literature. Multi-directional scanning strategies are already a fundamental design choice in visual SSMs to address 1D-to-2D/3D structural disruption.

  • 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    My recommendation (Borderline/Weak Reject) is primarily driven by a discrepancy between the reported empirical success and the underlying logical consistency of the proposed methodology. While the model achieves competitive SOTA results, the heavy reliance on 8-fold Test-Time Augmentation (TTA) during inference introduces a significant confounding factor. Given the marginal 0.2% improvement over baselines in certain benchmarks, it is unclear whether the gains stem from the architectural design or the intensive inference-time ensemble. A fair, TTA-free comparison is essential to validate the core contribution. There is a fundamental contradiction between the authors’ theoretical justification and their empirical findings. The ablation study suggests that multiple scanning pairs introduce “redundancy” and degrade performance, yet the final inference strategy utilizes 8-way flipping, which the authors claim is “analogous to multiple scan orders.” This inconsistency undermines the theoretical robustness of the proposed bidirectional scanning strategy The argument that CNNs are required to “lower variance” for SSMs overlooks the fact that standard Normalization layers already mathematically force unit variance. Furthermore, the 1D Bi-Scan strategy is a logical adaptation of existing 2D Mamba paradigms rather than a foundational breakthrough.

  • 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.

    The authors’ rebuttal inadequately responds to the core technical concerns and fundamentally fails to address the critical issue of insufficient technical novelty. These two major unresolved flaws make the entire rebuttal unconvincing.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a unified 3D encoder–decoder framework called MM-UNet (Meta Mamba UNet). The main contributions are:

    Clear theoretical insights: The authors point out two key limitations of state space models (SSMs/Mamba) for 3D medical images: spatial discontinuity caused by flattening, and the difficulty of modeling high-variance data. Careful hybrid design: By systematically exploring combinations of CNN and Mamba, they find that placing Mamba inside residual connections helps reduce input variance and improves performance. Improved scanning strategy: A simple and effective bi-directional scan is introduced to reduce the spatial discontinuity introduced when flattening 3D data into 1D sequences. Strong experimental results: The model achieves state-of-the-art performance on several challenging 3D medical datasets, including AMOS22, BTCV, and ACDC.

  • 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.Addresses core limitations of Mamba in 3D imaging: Unlike many works that directly apply 2D Mamba to 3D data, this paper takes a more principled approach. It analyzes, from a theoretical perspective, how SSMs struggle with high-variance data (linked to the properties of the HiPPO operator), and addresses this by placing Mamba inside residual connections to stabilize the input. This problem-driven design is both thoughtful and academically meaningful. 2.Thorough architecture and scanning ablations: The paper builds a “meta” framework that includes pure CNN, pure SSM, and hybrid models for systematic comparison. It also evaluates more than ten different 3D scanning strategies (e.g., DHW, WHD, Zigzag, 3D Window Scan), making the final choice of “1D BiScan” well-supported and convincing. 3.Strong performance on challenging 3D clinical datasets: The model achieves excellent results on large-scale and difficult 3D datasets such as AMOS22 (Task 1 & 2) and BTCV. For instance, it reaches 91.5% Dice on AMOS22 Task 1, outperforming nnUNet by 1.7%, demonstrating strong capability in handling complex 3D anatomical structures.

  • 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.

    Over-reliance on test-time augmentation (TTA): The paper uses 8× flip-based TTA during inference (as shown in Fig. 4), averaging predictions over eight passes for each 3D patch. While this can boost Dice scores, it also significantly increases inference time. Since no results on inference time or FPS are reported, it’s unclear how much of the performance gain comes from the model itself versus the extra computational cost introduced by repeated inference. The term “Meta” feels somewhat overstated: The model is named “Meta Mamba UNet,” but in practice, it mainly involves manually designed building blocks (e.g., EncMetaBlock, DecMetaBlock) and systematic ablation. While this is a thorough and well-executed study, the use of “Meta”—which often implies meta-learning or neural architecture search—may be somewhat misleading.

  • Please rate the clarity and organization of this paper

    Good

  • 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/en/REVIEWER-GUIDELINES.html

    Here’s a smoother, more natural version:

    It would be helpful if the authors could include a more detailed table in the camera-ready version, reporting the model size, FLOPs, and actual inference time. In particular, comparing performance and time cost with and without 8× TTA would make the efficiency claims much clearer and strengthen the paper from a practical deployment perspective.

    It is also recommended to release the code to improve reproducibility and support further research.

  • 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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The authors do not blindly follow the popularity of Mamba. Instead, they clearly identify its inherent limitations in 3D medical imaging, such as spatial discontinuity caused by flattening and the difficulty in modeling high-variance data. By embedding Mamba modules within residual connections and introducing a bi-directional 3D scanning strategy, the paper presents a solution that is both theoretically grounded and practically effective.

    The strong SOTA performance on challenging benchmarks like AMOS22 and BTCV further highlights its potential for real clinical applications. Although the use of 8× test-time augmentation may raise concerns about computational efficiency, and the term “Meta” might be somewhat overstated, these issues do not undermine the core contribution of this work as a solid Mamba–CNN hybrid architecture.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    After reading the rebuttal, I maintain a positive final recommendation.

    The authors addressed my main concern regarding the use of 8× test-time augmentation by clarifying that the same inference pipeline was used for their method and the re-implemented baselines. This reduces my concern that the reported performance gain is mainly due to an unfair inference-time advantage. The additional discussion on model complexity and the comparison with and without TTA also helps clarify the practical cost of the method.

    That said, I still think some concerns remain. In particular, the rebuttal includes several results or analyses that were not clearly present in the original submission, such as the no-TTA comparison, complexity numbers, multi-seed BTCV results, and additional variance analysis. I understand that some of these may be considered additional metrics or clarifications, but I would still base my final judgment mainly on the submitted manuscript. The term “Meta” also still feels somewhat overstated.

    Overall, the paper presents a solid and well-motivated Mamba–CNN hybrid architecture for 3D medical image segmentation, with strong empirical performance and useful ablation studies. The rebuttal sufficiently reduces my main concern about TTA fairness, although not all weaknesses are fully resolved. Therefore, I recommend acceptance.



Review #3

  • Please describe the contribution of the paper

    The paper proposes MM-UNet, a unified U-shaped encoder-decoder framework that integrates State Space Models (SSMs / Mamba) with CNNs for 3D medical image segmentation. The authors claim four contributions: (1) A meta-architecture framework in which encoder, bottleneck, and decoder “meta-blocks” are interchangeable, allowing existing Mamba-based U-Nets (VM-UNet, SwinUMamba, U-Mamba, SegMamba) to be expressed as special cases within a single unified design space. (2) An empirical analysis of two intrinsic limitations of SSMs when applied to 2D/3D medical imagery: (a) difficulty fitting high-variance signals due to HiPPO’s polynomial-projection smoothing, and (b) discontinuities introduced when flattening multi-dimensional volumes into 1D sequences. (3) A hybrid CNN to Mamba block (MetaSSM) embedded inside residual connections in the encoder and bottleneck, motivated by the observation that feature maps inside residual branches exhibit lower variance than those outside, which is favorable for SSMs. (4) A bi-directional scan order strategy (DHW + flip(DHW)) that mitigates flattening-induced discontinuities, with an ablation showing it outperforms 2D, 3D-window, zigzag, inclined, and multi-pair scan variants. Empirically, MM-UNet reports state-of-the-art Dice on AMOS22 Task 1 (91.5%), AMOS22 Task 2 (89.0%), ACDC (93.2%), and BTCV (87.1%).

  • 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.

    Major Strengths 1.It has a genuinely useful unifying framework. Framing VM-UNet, SwinUMamba, U-Mamba, and nnUNet as specific instantiations of a common meta-architecture is a clarifying contribution. It lets the community reason about where Mamba helps (encoder/bottleneck) versus where it does not (decoder) in a controlled way rather than through cross-paper comparisons that confound architecture with training recipe. 2.It has a principled, mechanistic critique of SSMs for images. The analysis in Section 2.1 and Fig. 2(a), showing that S4 smooths high-variance points due to HiPPO polynomial projection, and that row-transition discontinuities produce nearly-inverse predictions, is the kind of diagnostic evidence that is too rare in vision-Mamba papers. Most prior work asserts that Mamba “captures long-range dependencies” and moves on; this paper actually probes when and why it fails on 2D data. 3.The variance-inside-residual observation is novel and actionable. Fig. 2(b)’s comparison of feature-map variance inside vs. outside residual connections, linked causally to SSM sensitivity to variance, provides a crisp mechanistic justification for placing MetaSSM inside the residual branch. The resulting 0.9% improvement is small but consistent with the stated hypothesis. 4.The ablations are systematic. Table 4’s scan-order ablation (B1–B9) is thorough: it tests not just “does bi-directional help” but also whether more directions help (they don’t), whether unpaired directions hurt (yes), and whether 2D/window/zigzag/inclined alternatives offer advantages (no). The negative results here are informative and strengthen the final design choice. 5.There is strong and diverse empirical evaluation. Four benchmarks (AMOS22 T1, AMOS22 T2, ACDC, BTCV) spanning CT, MRI, abdominal, and cardiac anatomy, with comparisons to nine baselines including three Mamba-based competitors and three transformer-based ones. The AMOS22 results are from the official leaderboard, which mitigates concerns about evaluation cherry-picking. Per-organ Dice is reported rather than only averages. 6.Practically useful negative result. The finding that replacing the decoder with Mamba hurts performance (89.4%) while replacing the encoder/bottleneck helps is a useful guideline for future Mamba-based segmentation architectures and explains, post hoc, why U-Mamba’s design (encoder-side Mamba) works better than VM-UNet’s (everywhere).

  • 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.

    Major Weaknesses 1.The “high-variance” claim is under-quantified and possibly conflated. The central mechanistic storyi.e SSMs struggle with high variance, residual connections reduce variance, therefore put SSMs in residuals, rests on one qualitative figure (Fig. 2b) reporting three variance numbers (0.308 / 0.285 / 0.482) from a single pre-trained model. There is no statistical test, no variation across layers/models/datasets, and no controlled experiment (e.g., injecting variance into a synthetic input and measuring SSM error) to establish the causal link. A 0.285 vs. 0.308 difference is particularly thin to motivate an architectural principle. The paper would be substantially stronger with a formal analysis of SSM error as a function of input variance, or at minimum a correlation plot across many layers. 2.Inference-time 8-flip TTA is not acknowledged as a confound. Fig. 4 describes averaging predictions over 8 flip variants with 3D Gaussian weighting at inference. This is classical test-time augmentation and typically adds 0.3–1.0% Dice on its own. It is unclear whether the baselines (nnUNet, UMamba, SwinUMamba, VMUNet) in Tables 1–3 were evaluated with matched TTA. If not, part of the reported improvement is attributable to inference procedure rather than architecture. This needs to be stated explicitly and ablated. 3.there is no compute/parameter/latency comparison. The paper argues MetaSSM is efficient (“substantially fewer parameters and lower computational cost” than transformer attention) but provides no parameter counts, FLOPs, memory use, or training/inference time for MM-UNet versus baselines. For a method whose central motivation is that SSMs offer linear-time long-sequence modeling, the absence of any efficiency table is a significant omission. Without it, the reader cannot evaluate whether the gains over nnUNet (+1.7% on AMOS T1) come at 1.1× or 3× the cost. 4.Limited novelty at the component level. Each individual ingredient has close antecedents:

    a)Bi-directional scanning is central to Vision Mamba (Zhu et al., ref [26]) and VMamba (ref [17]); the paper’s “1D BiScan” is essentially the same idea applied to 3D flattening. b)CNNtoMamba hybrid blocks are exactly what U-Mamba (ref [18]) proposes. c)Placing sequence modules inside residual branches is standard practice in ResNet/Transformer literature.

    The novelty is therefore in the combination and systematic justification (the meta-architecture framing and the variance-based rationale), not in any single new component. This is a legitimate contribution, but the paper overclaims in places (e.g., “first architecture capable of representing existing Mamba-based models”) when it is more accurate to say it provides a useful analysis framework. 5.The authors dont report statistical significance . Gains over the strongest baseline (UMamba) are +1.0% on AMOS T1, +0.8% on AMOS T2, +0.71% on ACDC, +1.8% on BTCV. On BTCV (only 6 validation scans) in particular, the gap could fall within run-to-run variance. No standard deviations, no paired significance tests, no multi-seed results are reported. For a paper claiming SOTA by margins of <1% on several benchmarks, this is a serious gap. 6.Fig. 2(a)’s S4 experiment is an argument about S4, not Mamba. The intrinsic-limitation analysis uses S4 (ref [7]), whereas the deployed MetaSSM uses Mamba (selective SSM, ref [5]). Mamba’s input-dependent Δ, B, C parameters were explicitly introduced to overcome S4’s inability to handle content-dependent variation. The paper does not rerun the variance/discontinuity analysis on Mamba itself, so it is unclear whether the motivating pathology fully applies to the model actually used. This is a non-trivial logical gap in the paper’s central argument. 7.There are clarity issues in Section 2.1 and Fig. 5.The scan-order figure is dense and under-labeled; the mapping between (2a)–(2p) panels and the B1–B9 ablation rows requires substantial cross-referencing. The claim in Fig. 3 that “each attention map effectively captures image patterns across the temporal dimension” is asserted rather than quantified — the figure shows that forward/backward scans have complementary coverage at depth extremes, which is expected by construction of bi-directional scanning, not additional evidence of quality. 8.There are reference and notation issues. Reference [11] appears to point to the 2019 nnUNet arXiv preprint, but the authoritative Nature Methods paper (Isensee et al., 2021) is not cited. Reference [15] (3D UX-Net) is cited in the nnUNet row of Table 1 in a way that is confusing. “SSMs” and “Mamba” are used interchangeably throughout despite being distinct (Mamba is a specific selective SSM). The term “MetaScan” appears in Fig. 5 and Table 4 but is never formally defined in the text. 9.There is ambiguidty in BTCV protocol. The paper states BTCV uses “24 train / 6 val” but reports results against baselines including UNETR and Swin-UNETR that used different splits in their original papers. Without clarification that all methods in Table 2 were re-run under identical splits, the comparison is difficult to interpret. 10.No qualitative failure analysis. Fig. 6 shows cases where MM-UNet outperforms competitors but no cases where it fails. For a clinical translation argument, understanding failure modes (small organs, boundary errors, rare pathology) would strengthen the paper.

  • 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/en/REVIEWER-GUIDELINES.html

    I encourage the authors to consider the following for a revision: (a) Run the S4 analysis in Fig. 2(a) with Mamba itself. If the selective mechanism substantially attenuates the high-variance smoothing problem, that nuances your motivation; if it does not, reporting that is scientifically interesting. (b) Report parameter counts, FLOPs, peak GPU memory, and wall-clock inference time per volume for MM-UNet and the top three baselines. This is standard for architecture papers and currently absent. (c) Clarify whether the 8-flip TTA + Gaussian weighting is applied to baselines, and provide an MM-UNet-without-TTA row. Ideally include multi-seed runs with standard deviations on at least one benchmark. (d) The meta-architecture framing in Fig. 1(b) is one of the paper’s strongest contributions. Consider formalizing it as a small notation (e.g., listing each method as a tuple of (Enc, Bottleneck, Dec) block types) earlier in the paper — it would help readers immediately. (e) Consider a failure-case figure on small or low-contrast organs (gallbladder, adrenal glands) to complement the SOTA qualitative figure. (f) Define “MetaScan” explicitly. State whether the bi-directional scan is applied independently per stage or with shared parameters across directions. (g) Reference [11] should point to Isensee et al., Nature Methods 2021 for nnUNet.

  • 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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    This paper sits in a crowded space (Mamba-for-medical-segmentation) that has already produced VM-UNet, U-Mamba, SwinUMamba, and SegMamba within the last ~18 months. What lifts it above incremental is the diagnostic orientation: rather than proposing yet another Mamba-UNet variant, the authors ask where and why Mamba helps, and they offer two concrete, testable answers (variance sensitivity » place in residual branch; flattening discontinuities » pair opposite scan orders). The meta-architecture framing genuinely clarifies the design space and lets prior methods be compared on equal footing. The empirical results are strong; SOTA on four benchmarks including AMOS22’s official leaderboard, and the scan-order ablation is notably thorough. These are not negligible achievements. In my opinion, what holds the paper back from a clear accept are: (i) the central mechanistic claim (variance sensitivity) rests on thin quantitative evidence and is demonstrated on S4 rather than on Mamba itself; (ii) the absence of efficiency metrics in a paper whose premise is SSM efficiency; (iii) the likely confound of 8-flip TTA that is not explicitly ablated; (iv) sub-1% margins over U-Mamba on several benchmarks with no significance testing. Each of these is fixable in rebuttal; none is fatal, but collectively they make the evidence for the mechanistic story softer than the paper’s confident framing suggests. On balance, the conceptual clarification (meta-architecture + where-does-Mamba-help) plus the consistent SOTA results across four diverse benchmarks earn a weak accept. A convincing rebuttal addressing TTA ablation and parameter/FLOP reporting would move this to a clear accept; failure to address those points would move it to a 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.

    Accept

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

    This is a solid contribution to Mamba-for-medical-segmentation. My pre-rebuttal recommendation was a Weak Accept (4), with the explicit caveat that “a convincing rebuttal addressing TTA ablation and parameter/FLOP reporting would move this to a clear accept; failure to address those points would move it to a weak reject.” The rebuttal addresses both of these load-bearing concerns and several others with concrete numerical evidence. I therefore move to a confident Accept.The paper now has: (i) consistent SOTA across four diverse benchmarks under matched evaluation; (ii) low cross-seed variance; (iii) matched computational cost to the strongest baseline; (iv) a mechanistic story validated on the actual model deployed; and (v) a useful meta-architectural framing that clarifies the design space. The conceptual novelty is moderate rather than exceptional, but the work is rigorous, well-evidenced, and a useful addition to the community’s understanding of how to integrate SSMs into volumetric segmentation pipelines.



Author Feedback

We sincerely thank all reviewers for their constructive feedback and are encouraged by the three positive scores (3 Weak Accepts). We address the core concerns as follows:

R1,R2,R3: 8-flip TTA, complexity, and statistical significance. 8-flip TTA + Gaussian weighting is the default nnU-Net inference setting, and both our method and all re-implemented baselines are evaluated under the same pipeline. We also report results without TTA and complexity on AMOS22 shown below. Without TTA, the drop of our method is larger than nnUNet but smaller than UMamba. The added Mamba blocks introduce minimal overhead but consistently improve accuracy.

w/ TTA| w/o TTA | #param | GFLOPs | Method 0.884 | 0.882 | 62.19M | 319.97 | SwinUNETR 0.893 | 0.892 | 102.36M | 304.17 | nnUNet 0.905 | 0.902 | 117.22M | 363.59 | UMamba 0.915 | 0.913 | 118.12M | 368.15 | Ours

The only exception is the BTCV table, where non–Mamba-based methods are taken from the original papers due to the dataset’s widespread use, but we run three seeds to reduce variance. Our method achieves the best mean performance with the lowest variance. Except for BTCV, all experiments use 5-fold cross-validation with ensemble inference, which reduces run-to-run variance.

seed 0| seed 1| seed 2| mean±std | Method 85.06 | 85.12 | 85.36 | 85.18±0.16 | nnUNet 85.31 | 85.41 | 85.33 | 85.35±0.05 | UMamba 86.95 | 86.76 | 86.97 | 86.89±0.12 | SwinUNETR 87.05 | 87.08 | 87.02 | 87.05±0.03 | Ours

R1,R2,R3: We will release the code and model weights upon acceptance.

R1: a logical paradox between architecture and inference strategy. More scan pairs in the SSM increase parameters and optimization difficulty, which can hurt performance. In contrast, 8-flip TTA is applied only at inference: the same input is flipped along different axes, passed through a single model, and predictions are averaged. Thus, a single SSM scan pair already captures directional information via input flips, while additional scan pairs add complexity without clear gain; TTA instead provides multi-view ensembling without extra parameters, explaining its stronger effect.

R1: Normalization for high variance. Different organs and tissues often exhibit substantial intensity variation, which may contribute to segmentation. While normalization stabilizes feature distributions, it can also reduce contrast information in some cases. Placing the SSM after CNN layers allows it to operate on more structured representations, helping to further stabilize feature variance while preserving spatial and morphological cues.

R3: The high-variance claim is under-quantified. We inferred 60 3D images and computed the average variance of encoder block outputs (3.2M samples). In nnUNet without residual connections, the mean variance after two consecutive CNN layers is 0.209.In nnUNetRes, the variance inside the residual branch is 0.042, while outside is 0.097.The lowest variance occurs within the residual branch, consistent with our paper’s claim. We will include these updated results in the revision.

R3: Limited novelty. Our design is motivated by empirical observations and proposes problem-driven methods to address the limitations of SSMs. We explored many variants and identified the most effective design, where accuracy is the primary priority in medical image segmentation.

R3: S4 and S6 for variance analysis. We further evaluated both S4 (HiPPO) and S6 (Mamba) for reconstruction on BTCV patches (96×96×96) grouped by variance. Patches were divided into low-variance and high-variance groups based on the average variance. Both S4 and S6 achieve substantially better fitting performance on low-variance data than on high-variance data, while S6 consistently outperforms S4.These results further support our hypothesis that SSM-based models fit high-variance medical data more poorly than low-variance data.

Method | Low Var (<170k) | High Var (>170k) S4 | 4.55 | 8.81 S6 | 0.90 | 1.83

R3: We will add some failures in the revision.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper got 3 x weak accept, however, the reviewers raised a several concerns that should be addressed.

  • 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.

    The authors fully addressed main concerns provided by two reviewers. I recommend to include it in camera-ready version.



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.

    While one reviewer expressed enthusiasm about limited novelty, I agree with other two reviewers that the work is rigorous, well-evidenced, and a useful addition to the community’s understanding of how to integrate SSMs into volumetric segmentation pipelines. Thus, I prefer to accept this paper.



Meta-review #3

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

    Reject

  • Please justify your recommendation.

    The rebuttal includes additional experiments that were not permitted under the rebuttal guidelines. I believe this creates an unfair review process for other submissions that adhered to the rules and limited their responses accordingly. Independently of this issue, I do not believe the work is technically mature enough for acceptance at MICCAI.

    From a technical standpoint, several major concerns remain unresolved: 1.Important recent baselines such as MICCAI 2024 SegMamba and NeurIPS 2024 VMamba are missing from the evaluation. 2.Variants of Mamba-UNet architectures have already been introduced in prior work, and the novelty of the proposed method appears limited. The reviewers supporting acceptance seem to have overlooked this point, especially given that the reported performance metrics are extremely close to existing approaches and no statistical significance analysis was provided. 3.The paper does not adequately analyze failure cases or anatomy-specific performance variations. For example, on the BTCV dataset, the proposed method achieves substantially lower Dice performance for gallbladder segmentation compared to the baseline (0.66 vs. 0.79), yet this limitation is neither investigated nor discussed. 4.The reported Dice improvements are very small, and no statistical significance testing was performed in the original submission. The rebuttal also does not provide a convincing explanation regarding the practical importance of these gains. Furthermore, the standard deviations introduced in the rebuttal—which themselves constitute new experimental results—are minimal and suggest the differences may not be meaningful. 5.The parameter analysis presented in the rebuttal also fails to demonstrate a substantial advantage over UMamba. The parameter counts are nearly identical, and the segmentation metrics are similarly close. This again raises the question of what the actual methodological contribution is beyond another incremental combination of Mamba and U-Net style components. In addition, the complexity analysis was only performed on the AMOS22 dataset, limiting the generalizability of the claims. 6.Regarding test-time augmentation (TTA), the rebuttal states: “Without TTA, the drop of our method is larger than nnUNet but smaller than UMamba. The added Mamba blocks introduce minimal overhead but consistently improve accuracy.” However, the actual Dice differences between TTA and non-TTA settings are extremely small (approximately 0.002–0.003 across methods). Therefore, describing these gains as “consistently improving accuracy” appears overstated and potentially misleading. 7.The rebuttal further explains that 8-flip TTA is applied only during inference by averaging predictions from flipped inputs. However, it remains unclear why this process has almost no measurable impact on performance. In 3D medical image segmentation, inference from multiple viewpoints typically affects prediction quality, particularly when such augmentations are not fully represented during training. The negligible changes reported here raise questions regarding the implementation details and the interpretation of the TTA analysis.



back to top