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

Abstract

Retinal vessel segmentation and artery/vein (A/V) separation are central to fundus-based vascular biomarker analysis, yet performance often degrades under domain shift when acquisition-dependent appearance overwhelms thin-vessel evidence. We propose NS$^3$Net, a frequency-explicit segmentation framework that couples Mamba-based state-space modeling with region-adaptive spectral processing. Its core NS$^3$MoE performs frequency-plane mixture-of-experts routing to refine high-frequency Fourier magnitudes while preserving phase, enhancing vessel boundaries and thin capillary cues. A low-frequency residual path preserves global appearance to reduce artifact-driven false positives. We further introduce supervised two-view routing-consistency training, which aligns latent routing distributions between original and photometrically perturbed labeled views without unlabeled data or target-domain adaptation. Experiments on six public datasets show consistent ID/OOD gains for vessel segmentation and A/V separation, particularly in boundary-sensitive performance, while maintaining a lightweight profile of 10.03G FLOPs. Our code is available at \url{https://github.com/Thanaporn09/NS3Net.git}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6038_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Thanaporn09/NS3Net.git

Link to the Dataset(s)

N/A

BibTex

@InProceedings{VirTha_NonStationary_MICCAI2026,
        author = { Viriyasaranon, Thanaporn AND Choi, Jang-Hwan},
        title = { { Non-Stationary Spectral State-Space Networks with Region-Wise Frequency Routing for Robust Fundus Vessel Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16884},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper addresses the critical clinical issue of reduced robustness in retinal vessel segmentation and arteriovenous classification under domain shifts by proposing the NS³Net network, which integrates frequency-domain expert blending, visual Mamba, and routing consistency training. The method features a novel design and offers significant advantages in terms of lightweight architecture.

  • 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.The experimental design is comprehensive. The authors conducted validation on two major tasks—binary vascular segmentation and arterial-venous segmentation—across six public datasets, reporting both in-domain and out-of-domain generalization performance. They employed multiple evaluation metrics, including DSC, IoU, and HD95, and utilized multiple random seeds and cross-validation to ensure the reliability of the results. 2.The proposed method strikes a balance between accuracy and efficiency. The model contains only 4.68 million parameters and requires 10.03 GFLOPs of computational power, significantly less than mainstream baselines such as nnUNet and TransUNet. 3.The paper is logically structured and easy to understand.

  • 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.Fixed-spectrum designs lack adaptability. The low-pass cutoff frequency, bandwidth division, and router grid size are all manually designed and fixed. 2.The perturbations used in routing consistency training are limited. Only luminance perturbations are employed, failing to account for real-world variations such as resolution, field of view, and blurring, which limits the potential for improving robustness.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    To address the issue of domain drift in fundus vascular segmentation, the authors propose NS³Net, which combines frequency-domain expert routing with a state-space model. The method’s design demonstrates clear innovation and specificity. The paper conducts systematic experiments on six public datasets, achieving leading performance in both vascular segmentation and arteriovenous classification tasks while maintaining a lightweight architecture with 4.68 million parameters and 10.03 GFLOPs, resulting in an excellent balance between accuracy and efficiency. Although the work has limitations, such as fixed frequency-domain parameters and a single perturbation factor, the overall approach is well-defined, the experiments are thorough, and the conclusions are relatively reliable. It provides valuable insights for research on domain generalization in medical image segmentation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    The authors propose a cleverly designed vessel segmentation and artery/vein (A/V) separation model, specifically an encoder-decoder UNet-style architecture (NS^3Net) that leverages a novel twist using FFT/high/low frequency features in the proposed NS^3MoE block, which is the key novelty in the architecture. NS^3Net is comprised of carefully constructed modules, specifically the NS^3Block which contain two submodules: NS^3MoE, which performs a FFT then separates the high and low frequencies and uses high freq’s for a region-wise MoE module, and a Mamba SSM. These blocks are set up in the classic bottleneck’d UNet shape with gated residual fusion between hierarchical layers.

    The authors report both in-distribution (ID) and out-of-distribution (OOD) results which demonstrate top performance among reported methods for ID, and similarly performant generalization capabilities for OOD test sets, helped by the proposed “two-view routing-consistency”, which is an auxiliary loss designed to encourage expert routing consistency/alignment under augmented views.

    The key innovations provided by the authors is in the NS^3MoE submodule to not only leverage the frequency domain, but then apply region-based experts for FFT-transformed high frequency feature maps, which is novel and theoretically motivated extension to other FFT-based approaches. Also, their multi-dataset in-distribution/OOD evaluations and ablations make a convincing case for the proposed architecture over existing (mostly UNet variant) models for vessel segmentation, at just a fraction of the model size/FLOPs.

  • 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 authors compare to 7 different common segmentation architectures and across 4 OOD test sets (and the 2 in-distribution sets) to make the case that their approach has good performance and generalizes well. At just a fraction of the model size (4.68M params), the small model size is impressive on its own, especially more so it is at or above other methods. The low cardinality of their train dataset, combined with strong OOD performance is also a notable strength, showing this approach can work well in data scarce regimes.

    • In terms of architectural innovation, the main novelty is the NS^3MoE module due to its usage of the FFT (not novel), frequency partition (not novel), then region-based routing of high freq features (is novel), and this is an interesting and theoretically well motivated approach, as the authors contend that these high freq features can “sharpening boundaries and [recover] thin capillaries”, which provides sufficient rationale to justify the approach. Understanding this approach would benefit future researchers ability to work with FFT/high/low frequencies in creative new and useful ways, and could motivate many more follow up ideas along these lines. Although frequency-based approaches are not entirely novel [1], and neither is separation of high/low freq’s [2, 3], I would score innovation as moderately high, but not necessarily paradigm shifting or meta-defining.

    • With regards to scientific communication, everything was presented clearly, and this is a high-quality and well written paper all around. The reproducibility is strong given the attention to detail in explanations of methods. One small suggestion/idea for the authors could be that sections 2.4 and onward could be part of an “Experiments” section to delineate methods versus experiments. I also see in Fig 2b the experts are listed “Ex1”, “Ex2”, “Ex2”, I think the authors meant Ex3 for the last one; also in the conclusion “pperforms”. In general the presentation quality is high and these extremely minor details should not detract from otherwise impressive work, I only point this to the author to make this an even stronger paper in case of publication

    [1] Qiao, Q., Wang, W., Qu, M., et al. (2024). Medical image segmentation via single-source domain generalization with random amplitude spectrum synthesis. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2024. [2] Li, M., Yan, Z., Yan, M., et al. (2025). PFESA: FFT-based parameter-free edge and structure attention for medical image segmentation. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2025. [3] Zhou, Z., He, A., Wu, Y., Yao, R., Xie, X., & Li, T. (2024). Spatial-frequency dual domain attention network for medical image segmentation. In 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) (pp. 4076–4081).

  • 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 marginal improvement of this method over other traditional methods is relatively low and within the margin of error for datasets, but still notably good given the results span across many datasets and approaches. The ablations for the frequency-based MoE module show just very slight gains (+0.19 Dice), though it combined with Routing-Consistency (RC) show a noticeable bump in performance

    • If the authors are scoping the work to just vessel segmentation (which seems they are) instead of more general approaches, then the impact is relatively limited. If the authors are supposing that the region-aware MoE approach is good for segmentation in general (or fine-structure seg), then more focus on that would have been needed (i.e. other tasks). But since the novelty is relatively narrow (i.e. FFT twist in NS^3MoE submodule), and the focus tightly scoped, the overall impact is limited

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    The authors present all details clearly in the methods and everything was able to be found. But one clarification I’m wondering is on the region-wise component: I see in the MoE diagram (Fig 2b), it takes the feature map, applies FFT, then separates high/low freq features. I see the MoE has 4 experts (J=4), I see you “set the router grid to hg × wg = 8 × 8”, and I see the Region Universal Router architecture has avg pooling, convs, and activations which yields the top-k selections. Can the authors help me put it all together to understand what regions of the (FFT/high-freq.) feature maps are being referred to and their flow through the network? Another way that may help is to explain the dimensionality/shape, at each step, but mostly the High Freq Features as it inputs to the Region Universal Router.

    If my understanding correct, the goal of the region-based approach is to “selectively refine highfrequency components” which leads to (hopefully) “sharpening boundaries and [recover] thin capillaries”? If so, this is an interesting double-refinement of features: FFT is an initial refinement step, then this region-wise MoE approach is a second step to focus the highfrequency features, which also makes intuitive sense since the high freqency feature maps contain useful new signal but not uniformly everywhere, and therefore different experts can operate on those different features. The authors can consider expanding on this, beyond what they already mention in the intro (“The experts primarily transform high-frequency components to sharpen vessel boundaries and recover capillaries”), to draw the connection between high freq features being processed in this manner and its impact on sharpening vessel boundaries/capillaries, though this may be more obvious to other readers.

  • 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 paper is clearly articulated, has implications for future research (i.e. on frequency domains and working with those features), and also implications for clinical translation or more specifically point-of-care/edge translation as this small model is capable of strong performance.

    I do have some reservations about the relatively small performance increase compared to the vanilla baseline in the ablation (Table 3), and marginal gains over other common model architectures such as the ones compared against (albeit they are mostly larger), but are also often above margin of error. The OOD results for DRIVE/CHASEDB1/HRF (commonly reported) is near or above other bespoke approaches (including relatively close to in-distribution approaches, those trained on more data, etc. [4]) beyond those listed/tested by the authors [5, 6, 7], and especially so for the model size while still being mostly task-agnostic (i.e. no hard constraint on applicability to only vessel seg), and therefore this could be a relatively impactful contribution. While LES-AV (OOD) had a relatively low Dice (51.72) compared to other researchers [8], this could also just be a function of the train dataset which is incomparable, so shouldn’t be compared directly to other research in this regard, and the authors comparisons against other architectures (Table 1, Table 2) show this model is effective in equally trained settings. I only note the external comparisons to provide a feel of where this roughly stands against other bespoke approaches. In general, and as it can be directly compared, the proposed architecture appears to provide a notable performance benefit over existing approaches, especially for the model size and seemingly task-agnostic architecture.

    I would have liked to see a bit more separation in performance either from the ablations (primarily) or the compared methods (secondarily), which would have given me more confidence that the proposed region-based MoE approach is significant beyond the bespoke NS^3Net architecture’s design, which would have pushed me to an “Accept”. I don’t think this is paradigm-shifting which would have pushed me to “Strong Accept”, but instead I think it does have a place to bolster the research base with strongly motivated theoretical ideas that combine concept (FFT+region-wise MoE) in creative ways that could provide research and translational benefits.

    The authors demonstrably put significant effort into the paper, and should be commended for their scientific communication which allows researchers like myself to understand the material and what is being presented, and why it is useful. The research contributions, while narrowly scoped to a novel frequency-based region-wise MoE, along with a bespoke surrounding network and training paradigm, do show significant gains over similar approachs at a fraction of the model size, both on in-distribution and OOD datasets. I lean accept, but would like the authors to respond to the above to help solidify my decision

    [4] Fadugba, J., Köhler, P., Koch, L., et al. (2024). Benchmarking retinal blood vessel segmentation models for cross-dataset and cross-disease generalization. arXiv.

    [5] Ahmed, I., Absar, S., Sami, A. A., et al. (2026). TFFM: Topology-aware feature fusion module via latent graph reasoning for retinal vessel segmentation. arXiv.

    [6] Hu, D., Li, H., Liu, H., et al. (2024). Domain generalization for retinal vessel segmentation via Hessian-based vector field. Medical Image Analysis, 95, 103164. [7] Dang, T. D. Q., Nguyen, H. H., Tiulpin, A., et al. (2025). Image-level regression for uncertainty-aware retinal image segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV).

    [8] Galdran, A., Anjos, A., Dolz, J., et al. (2022). State-of-the-art retinal vessel segmentation with minimalistic models. Scientific Reports, 12, Article 6174.

  • 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.

    A strong enough case has been made by the authors in the rebuttal to justify an accept

    While the performance gains are marginal, small thin-structured Dice improvements are indeed significant due to the nature of Dice scores, and that is a convincing argument. And, since it can do it in fewer params/flops, it is a significant contribution. I agree with the authors on both counts

    The authors have also clarified the methodology, and claim they update the Eqs., and thus the reservations about misunderstanding of the methods would be resolved if those are included in the final publication. This is a technically challenging method that requires precise explanations, but it appears the authors heeded the reviewers feedback and can clearly explain the approach.

    Although the scope of the paper is relatively narrow in that it only addresses vessel segmentation, the theoretical justification and results are strong enough to justify its applicability and further research directions for thin-structure segmentation in general

    Overall, I would categorize this paper as have several significant strengths, and no real or obvious weaknesses beyond what is expected from a conference paper. While it may not be a monumental shift in segmentation approaches, it does provide solid research insights using modern and domain-relevant techniques, and does so with apt scientific communication in a condensed paper

    I would echo R3’s comments on paper clean-up, and remind the authors of the sentence in the Conclusion:

    “Its core NS3MoE pperforms region-wise routing in the frequency plane”

    where “performs” has 2 p’s (pperforms)

    as a small typo that would need to be corrected in final publication. I also note the figure issues but those have been addressed in Rebuttal. But, these are small details to me and believe the other relative strengths of the paper far outweight these minor typos and those caught by R3

    Great work authors, it was a pleasure to read and review your work, and wish all the best going forward



Review #3

  • Please describe the contribution of the paper

    This paper proposes a Non-Stationary Spectral State-Space Networks (NS^3Net) for retinal vessel segmentation. The method utilizes a U-Net architecture integrated with Transformer blocks. These blocks employ NS^3MoE, which takes into account high-frequency components to enhance vessel segmentation. In addition, a consistency loss function between the original and augmented images is introduced during training. Experiments on public datasets demonstrate the effectiveness of NS^3Net compared to existing methods.

  • 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 adoption of a Vision Mamba-based architecture (VSSM) as the baseline for the retinal vessel segmentation task is a technically modern and highly commendable direction.

    • The approach of focusing on high-frequency components to capture fine vessel details, as well as the idea of leveraging consistency to mitigate the scarcity of annotated medical data, are highly relevant and practically valuable.

    • Building upon the robust VSSM baseline, the proposed method achieves highly competitive accuracy. The demonstration of outperforming existing methods on public datasets is a clear and objective strength of this work.

  • 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.In the Abstract, the authors state that they propose NS^3Net, with NS^3 MoE introduced as its core module. However, the Introduction states that NS^3 MoE is the proposed method. This inconsistency can be confusing for readers. Please clarify the text to ensure a consistent explanation throughout the paper regarding whether the primary contribution is NS^3Net or NS^3 MoE.

    2.The Introduction highlights the scarcity of annotated fundus data. To address this issue, the proposed method inputs both the original and augmented images into NS^3 Net and computes a loss based on their region-wise routing probabilities. Enforcing consistent outputs regardless of the presence or intensity of data augmentation is essentially a consistency regularization strategy, which is widely utilized in semi-supervised learning frameworks such as UniMatch. Therefore, it is necessary to explicitly compare the proposed approach with the dual-stream perturbation and strong-weak augmentation of UniMatch. Additionally, the authors should provide an experimental evaluation to demonstrate the empirical advantages of this approach over standard data augmentation techniques.

    3.In Eq. (1), the left and right sides are connected by an equals sign, which is mathematically incorrect in this context. If we abstract the logic as $A_i = A_{i-1} + V$ and $A_i = A_i + N$, interpreting them as strict mathematical equalities implies that $V$ is the difference between $A_i$ and $A_{i-1}$, and $N$ equals zero. To accurately represent an algorithmic update and avoid reader confusion, an assignment operator (e.g., an arrow $\leftarrow$) should be used instead of an equals sign.

    4.There is a visual inconsistency between NS^3Net architecture diagram in Fig. 1 (b) and NS^3Block in Fig. 2 (b). In Fig. 2 (b), the block is clearly depicted with a single input and a single output. However, in Fig. 1 (b), two separate arrows are shown emerging directly from the encoder block. To avoid confusion regarding the exact routing of features for the skip connection and the downsampling path, the authors should revise the diagrams to ensure consistency across all figures.

    5.There appears to be a structural inconsistency regarding the gated redisual fusion module in the architecture. The detailed schematic of this module clearly depicts a 2-input, 1-output structure. In NS^3Net, this 2-input routing is shown at the deepest layers. However, at the shallower layers, the skip connections are drawn as standard 1-input, 1-output connections. If the authors intend to apply the gated redisual fusion module to all skip connections, it seems the input arrows from the lower layers are missing in NS^3Net. Please correct the figure to ensure accurate reproducibility.

    6.In Sect. 2.2, the authors use the term “a channel-wise 2D real FFT (rFFT)” and refer to “complex rFFT coefficients.” While utilizing rFFT, which exploits Hermitian symmetry for real-valued inputs to reduce computational and memory costs, is a standard and practical implementation choice, it should not be confused with the mathematical formulation of the method. From a theoretical viewpint, the operation is simply a standard 2D Discrete Fourier Transform (DFT. I suggest the authors use general mathematical terminology, i.e., DFT, in the theoretical explanation of the proposed method, and move the specific mention of the “rFFT” usage to the Implementation Details section (Sect. 2.5).

    7.In Eq. (6), $\tilde{B}{i,j}$ can be simplified for better readability and rigor. Currently, it is presented as $\tilde{B}{i,j} = f_j(B_{i,j})$ and $f_j(B) = B + GELU(…)$. First, the notation in the second expression is inconsistent, as the subscripts $i$ and $j$ are omitted from $B$. Second, introducing an intermediate function $f_j$ makes the notation unnecessarily complicated. I recommend combining them into a single, concise equation, such as $\tilde{B}{i,j} = B{i,j} + GELU(…)$, to improve clarity.

    8.Regarding the notation for the MoE routing weights in Eq. (7), denoting the weight for the $j$-th expert at layer $i$ as $P_{i,j}$ can be misleading, as it visually resembles the notation for elements of a 2D spatial map or a matrix. Furthermore, since the convolution operation for the $j$-th expert is denoted using a superscript as $Conv^{(j)}$ in Eq. (6), it would be much more consistent to align the weight notation accordingly. I recommend changing $P_{i,j}$ to $P_i^{(j)}$. This modification separates the layer index (subscript) from the expert index (superscript) and improves the mathematical consistency and readability of the paper.

    9.There is a notational inconsistency regarding the high-frequency map between Eq. (7) and Eq. (8). In Eq. (7), the refined high-frequency map is denoted as $A_{i,out}^{hf}$. However, in Eq. (8), the reconstructed map is formulated as $A_i^{rec} = A_i^{lf} + \hat{A}i^{hf}$. Because the notation for the high-frequency component differs ($A{i,out}^{hf}$ vs. $\hat{A}_i^{hf}$), it is unclear whether these represent the exact same tensor or two distinct variables. Assuming they refer to the same refined HF map, please unify the notation throughout the equations to prevent reader confusion.

    10.There is an inconsistency between Fig. 1 (b) and Eq. (10) regarding the outputs of the proposed network $f_\theta$. Fig. 1 (b) illustrates the network taking an input $x$ and outputting only the segmentation result $\hat{y}$ (with dimensions $H \times W \times N_{cls}$). However, Eq. (10) formulates the forward pass as $(\hat{y}, P) = f_\theta(x)$, introducing an additional output $P$. While $P$ is briefly referred to as “region-wise routing probabilities,” it lacks a formal definition. Given that $P$ plays a critical role later in computing the consistency loss between the original and augmented images, an explicit explanation of $P$ is essential. If $P$ simply represents a standard softmax probability distribution, this should be clearly stated. If the specific terminology “region-wise routing probabilities” implies a specialized operation, its exact mathematical formulation must be provided. Please update the figure to reflect both outputs and add a detailed definition of $P$ in the text.

    11.In Eq. (13), $L_{seg}$ is utilized as the loss for the segmentation mask, described merely as “a combined Dice and cross-entropy loss.” However, the exact mathematical formulation of this combination is crucial for reproducibility. Please specify whether these two loss components are simply summed, averaged, or combined using specific weighting coefficients (e.g., $L_{seg} = \lambda L_{Dice} + (1-\lambda) L_{CE}$). I recommend providing the explicit equation for $L_{seg}$ and detailing any hyperparameters used.

    12.Regarding the ablation study, the authors only evaluate the presence or absence of the MoE integration within the Transformer Block and the consistency loss between the original and augmented images. However, to demonstrate the effectiveness of the proposed method, it is crucial to compare it against baseline strategies, rather than performing a simple component ablation. As mentioned in my previous comment, the authors should compare their consistency-based approach with standard data augmentation techniques and established semi-supervised learning methods.

    13.The performance improvements from the proposed MoE and consistency loss are marginal, as the VSSM baseline already achieves SOTA-level results. Given this strong baseline, the contribution appears incremental. To justify these additions, the authors should analyze computational trade-offs or evaluate the method on challenging scenarios where the baseline struggles.

  • 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.

  • 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

    14.The current structure of the paper consists of Introduction, Methods, and Conclusion. However, the standard and highly recommended structure for MICCAI submissions is Introduction, Material and Methods, Experiments, and Conclusion. The experimental setups and results appear to be embedded within the Methods section. To improve readability and align with the community standards, I recommend separating the experimental content from the methodology and organizing the paper into the aforementioned four distinct sections.

    15.In the description of NS^3 Block, a module named “VSSM” is introduced without any prior explanation. Assuming this refers to the “Vision Mamba-based state-space module,” it must be defined and explained at its first occurrence to ensure clarity.

    16.In the Abstract, the abbreviated name “NS^3Net” is used. Please ensure that the full name of the proposed method is spelled out upon its first appearance in the text.

    17.In Eq. (12), the function stopgrad() is used without an explicit mathematical definition or explanation. While the text mentions “stopping gradients through $P$”, the authors should clearly state that the stopgrad() notation in the equation corresponds exactly to this operation to avoid any ambiguity.

    18.The font sizes in the figures and tables are too small, making them difficult to read. Please enlarge the text in all visual elements to improve legibility.

    19 There are widespread formatting issues and outdated citations throughout the bibliography. The reference list requires meticulous proofreading and updating. Specifically, please address the following types of errors across all references:

    • Capitalization: Many proper nouns and acronyms are incorrectly lowercased. Please ensure correct capitalization (e.g., “Swin-UNet”, “TransUNet”, “AI”, “X-ray”, “2D & 3D”). Note: If you are using LaTeX/BibTeX, you may need to enclose these terms in curly braces (e.g., {TransUNet}) to preserve capitalization.
    • Missing Information: Some published conference papers lack necessary details like page numbers (e.g., [2] should include pp. 205-218).
    • Outdated arXiv Preprints: Several papers cited as arXiv preprints have already been published in peer-reviewed venues and should be updated. For example, [6] (Gu et al.) was published at ICLR 2022, and [7] (Hao et al.) was published in IEEE TMI. Please thoroughly check and correct the entire reference list before submitting the camera-ready version.
  • 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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    I recommend a Weak Reject for this paper. Although the authors achieve strong results by adopting a modern VSSM baseline, the manuscript suffers from several critical flaws. First, the methodological explanation lacks mathematical rigor, relying heavily on code-based, implementation-specific terminology instead of proper mathematical formulations. Second, there are significant inconsistencies between the architectural diagrams and the textual descriptions, raising reproducibility concerns. Third, the empirical validation is insufficient; the authors fail to compare their proposed modules against alternative standard approaches. Finally, the actual contribution of the proposed novelties appears incremental, as the performance improvements are marginal and heavily reliant on the highly competitive baseline. To reach the acceptance bar, the authors must formalize their method and conduct more comprehensive evaluations.

  • 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.

    I would like to thank the authors for their detailed and comprehensive rebuttal. Conceptually and empirically, you have adequately addressed my concerns. The explanation regarding the HD95 improvements justifies the seemingly marginal DSC gains, and the clarification that Routing Consistency (RC) operates as a latent routing regularizer rather than a standard semi-supervised prediction-level consistency is well-taken. I recognize that the core idea (NS^3MoE) is novel and the empirical results are sufficiently strong. However, my final decision must be based on the presentation quality, mathematical rigor, and reproducibility of the initial submission. While you have promised to fix all these issues, a manuscript submitted to a top-tier conference like MICCAI should be structurally and mathematically sound upon initial submission. My overall assessment is borderline now, but since the post-rebuttal system requires a binary choice, I must reluctantly select “Reject”.



Author Feedback

We thank the meta-reviewer and reviewers for the constructive comments. [Meta, R2, R3] NS3MoE tensor flow, frequency-region definition, and expert processing: Given E_i in R^{B x C_i x H_i x W_i}, NS3MoE uses a channel-wise 2D DFT to obtain amplitude A_i and phase Phi_i; for real-valued features, these are stored in non-redundant rFFT form in R^{B x C_i x H_i x W_i’}, where W_i’=floor(W_i/2)+1.A smooth radial mask M_lp decomposes A_i into A_i^lf = A_i * M_lp and A_i^hf = A_i * (1-M_lp), both in R^{B x C_i x H_i x W_i’}. The router pools A_i^hf to an 8x8 frequency-plane region grid, where each region is a coarse neighborhood of high-frequency Fourier bins, not a spatial patch. It predicts J=4 expert logits per cell, followed by top-k masking and softmax over experts, yielding P_i in R^{BxJx8x8}. For expert processing, raised-cosine radial band-pass masks are applied to A_i^hf to obtain 4 band-isolated responses, each refined by its corresponding expert. The upsampled region-wise probabilities P_i fuse the expert-refined band responses. The fused high-frequency response is recombined with A_i^lf and Phi_i before inverse DFT reconstruction. Thus, routing is region-wise on the frequency grid, whereas expert specialization is radial-band-wise. This enables region-adaptive refinement of diverse vessel-edge and capillary evidence. [Meta, R2, R3] Marginal DSC gain and accuracy-efficiency trade-off: We agree that the absolute DSC gain over VSSM is modest. However, NS3MoE’s benefit is reflected in improved OOD robustness and boundary-sensitive behavior. By design, NS3MoE targets high-frequency vessel-edge/capillary cues while preserving low-frequency appearance and phase; this aligns with the stronger gains in HD95.In the ablation, replacing the MLP FFN with NS3MoE reduces HD95 on CHASEDB1 from 35.07 to 33.34 and on DRIVE from 38.54 to 36.81; adding RC further reduces HD95 to 30.29 and 33.51, respectively. For retinal vessels, small DSC gains can still reflect meaningful boundary/fine-structure improvements because thin vessels occupy few pixels. These gains are achieved without increasing complexity: VSSM uses 4.76M parameters and 11.12G FLOPs, whereas NS3Net uses 4.68M parameters and 10.03G FLOPs. Thus, NS3Net improves the robustness-efficiency trade-off under domain shift. [R1] Fixed spectral design and perturbation strategy: The low-pass cutoff, band division, and router grid were fixed to keep NS3MoE lightweight/reproducible and avoid dataset-specific tuning in the low-data setting. The same settings were used across ID/OOD datasets, while the router/experts remain learnable. RC focuses on photometric shifts because they are common in fundus imaging and preserve high-frequency vessel cues. We now discuss blur, FOV, resolution/geometric shifts, adaptive frequency partitioning, and router resolution as future work. [R2, R3] RC clarification and comparison with UniMatch-style consistency: RC differs from UniMatch-style consistency in setting, target, and mechanism. RC uses only labeled images, without unlabeled data, pseudo-labels, weak/strong matching, or target-domain adaptation. Unlike prediction-level consistency, RC regularizes the latent expert-routing distribution. Specifically, P={P_{s,b}} denotes routing tensors from all NS3MoE blocks, where s indexes stage and b indexes block, and L_RC averages KL(P_{s,b}^{aug} || stopgrad(P_{s,b}^{ori})) over them. Thus, RC is a supervised routing-level regularizer, not a semi-supervised method. In the ablation, all variants use the same base augmentation; the improvement from no-RC to RC isolates the benefit attributable to RC beyond standard data augmentation. [R3] Formalization: To improve rigor/reproducibility, we revised Eqs. (1), (6)-(8), and (10)-(13); used a DFT formulation; specified rFFT/irFFT as implementation details; defined P, L_seg, and stopgrad; and unified high-frequency/routing notation. We also corrected terminology, Fig. 1/2, structure, fonts, typos, and references.



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 reviewers recognized the clinical importance of addressing domain shift in retinal vessel segmentation and commended your proposed NS³Net. The integration of frequency-domain expert routing with a Vision Mamba (VSSM) baseline was well-received, particularly given the model’s lightweight profile and strong in-distribution and out-of-distribution performance. While the reviews are generally positive, there is a consensus that the manuscript requires tighter mathematical formalization, structural consistency, and stronger analytical justification for its design choices.

    Please ensure your rebuttal addresses the following critical issues: 1) As highlighted by Reviewer 2, please explicitly explain the dimensionality and shape of the high-frequency feature maps as they flow through the Router, and clarify practically how these regions are defined and processed by the experts. 2) Reviewers noted that the performance improvements over the baseline VSSM are relatively marginal. Please provide a justification. Highlight specific computational trade-offs, theoretical advantages, or challenging edge cases (e.g., capturing thin capillaries) where your architecture definitively outperforms the baseline.

    Please respond to the other points raised if space allows in the rebuttal.

  • 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 reviewers generally recognized the clinical relevance of improving retinal vessel and artery/vein segmentation under domain shift, as well as the strong accuracy-efficiency trade-off of the proposed lightweight model. After considering the reviews and rebuttal, I recommend acceptance.

    The main strengths are the frequency-routed NS3MoE design, its integration with a state-space backbone, and the routing-consistency regularization for stabilizing expert selection under photometric variation. The evaluation is also convincing, covering six public datasets, both binary vessel segmentation and A/V separation, and both in-distribution and out-of-distribution settings. Although the DSC gains over the VSSM baseline are modest, the reviewers noted that the improvements in boundary-sensitive metrics such as HD95 and the consistent OOD gains provide a more meaningful indication of the method’s benefit.

    Some concerns remain around presentation and reproducibility rather than the core technical idea. The camera-ready version should incorporate the rebuttal clarifications on NS3MoE tensor flow, frequency-region routing, routing probabilities, the role of RC as a latent routing regularizer, and the DFT/rFFT formulation. The authors should also fix the architectural diagram inconsistencies, clarify the exact segmentation loss, improve figure/table readability, and moderate claims where the gains are incremental.



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.

    The paper presents NS³Net, a lightweight UNet-style architecture for retinal vessel and A/V segmentation built around the NS³MoE block, which combines FFT-based frequency partitioning with region-wise mixture-of-experts routing over high-frequency components, complemented by a Mamba SSM and a two-view routing-consistency loss. After rebuttal, R1 remains weakly positive, R2 moves to accept, and R3 maintains reject. Although the field has seen many proposed architectures for retinal vessel segmentation and the absolute DSC gains over baselines such as nnUNet remain modest, the proposed method offers a favorable accuracy-efficiency trade-off — 4.68M parameters and 10.03 GFLOPs, alongside consistent OOD performance. Notably, R3 states that the remaining reservations are primarily of presentational nature and would be resolved by the promised revisions. On this basis, I recommend acceptance, with the expectation that the authors honor the methodological and presentational commitments made in the rebuttal.



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.

    After considering the rebuttal and reviewers’ comments, the paper can be recommended for acceptance. The rebuttal provides convincing clarifications on the performance, efficiency, and method.



back to top