List of Papers Browse by Subject Areas Author List
Abstract
Accurate polyp segmentation is crucial for early detection and timely prevention of cancer. Although existing methods have achieved remarkable progress, they still struggle in real-world clinical scenarios due to strong polyp heterogeneity, fuzzy boundaries, and the high risk of missing small polyps. To address the challenges, in this paper, we propose to jointly exploit spatial structural priors and frequency-domain cues, and construct an adaptive frequency-guided parallel mamba framework for polyp segmentation. Extensive experiments on five public benchmarks demonstrate that our method can achieve excellent balance among segmentation accuracy, efficiency, robustness, and cross-domain generalization. It achieves consistently leading performance compared with state-of-the-art methods, particularly in challenging cases with fuzzy boundaries, small polyps, and cross-domain shifts, while maintaining competitive computational cost. The code is available at https://github.com/B00MCLAP/Polyp-AFMamba.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4551_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JiaAiw_Adaptive_MICCAI2026,
author = { Jiang, Aiwen AND Yu, Hongqian AND Li, Xue AND Xu, Zhiqiang AND Cui, Rui},
title = { { Adaptive Frequency-Guided Parallel Mamba Network for Polyp 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
Use both spatial structure information and frequency domain cues for better segmentation. Frequency domain cues are the key. It use frequency domain info to guide a parallel Mamba network. The results are consistently good.
- 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.
Use DCT (Discrete Cosine Transform) to extract low frequency components which contributes more to the object boundary. This is a well developed technology in digital image processing such as JPEG compression. There, DCT identifies low frequency components that are essential to human visualization. High frequency components can be discarded.
Comparison with SOTA methods show the proposal is effective.
- 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.
Fig 1 drawing quality could be improved.
- 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 provide sufficient information for 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
Is the frequency guide applicable to other types of networks?
DCT transform will also identify the boundary of non-object-of-interest. How do you deal with that?
- 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?
Use DCT transform to identify object boundaries is a good use of existing technology from DSP.
- 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.
The authors have satisfactory addressed my questions.
Review #2
- Please describe the contribution of the paper
The paper proposes Polyp-AFMamba, a framework blending spatial and frequency-domain cues for superior polyp segmentation. It introduces an Adaptive Parallel Visual 2D-Mamba layer to maintain spatial continuity efficiently and a Dynamic Frequency Decoupling module for sample-specific spectral separation. By combining high-frequency edge enhancement with low-frequency regional priors , the model achieves SOTA performance on five benchmarks , effectively tackling small polyps and fuzzy boundaries while maintaining FPS.
- 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 innovative dual-domain network design uniquely combines 4-direction parallel Mamba scanning for spatial information with dynamic DCT frequency features to capture both global structure and fine details.
- The superior boundary refinement and small polyp detection are handled through specialized high-frequency modules that solve the common problem of fuzzy edges that many standard CNNs fail to address.
- experimental results demonstrate that the model almost achieves the highest Dice and mIoU scores (comparisons with 10 state-of-the-art methods) across five public datasets and maintains exceptional robustness when generalized to unseen data. The optimal balance of accuracy and efficiency is evidenced by a high processing speed of 78.58 FPS.
- 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 paper suffers from limited novelty as core ideas like DCT-based frequency decoupling and edge-guidance modules closely overlap with recently published works such as FAGF-Net, Polyp-Mamba (2025).
Y. Li, Z. Zheng, W. Ren, Y. Nie, J. Zhang and X. Jia, “Frequency Aware and Graph Fusion Network for Polyp Segmentation,” ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Seoul, Korea, Republic of, 2024, pp. 1586-1590
Zhu, Xingguo & Wang, Wei & Zhang, Chen & Wang, Haifeng. (2024). Polyp-Mamba: A Hybrid Multi-Frequency Perception Gated Selection Network for polyp segmentation. Information Fusion. 115.102759.
-
The absence of an open-source code commitment makes it nearly impossible to replicate the complex results, especially with intricate modules like adaptive dynamic frequency masking.
-
The paper has a notable gap in that it neither discusses nor experimentally compares with direct competitors such as Polyp-Mamba and FAGF-Net, thereby weakening its claim of being the most advanced technology currently available.
-Some technical details are still unclear, especially how the dynamic frequency masks are optimized. The paper doesn’t clearly describe the loss design either, so it’s hard to tell whether these masks and region priors are supervised by auxiliary losses or just learned indirectly from the main segmentation loss
- 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 provide sufficient information for 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?
The paper presents good experimental results and a reasonable spatial-frequency combination design, but it has some significant limitations such as low novelty due to overlap with recent works, lack of direct comparison with related methods, and unclear technical descriptions, especially loss design, thus limiting reproducibility. Overall, the paper has potential but needs further clarification and additions to be more convincing.
- 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.
The rebuttal addresses most of my main concerns, particularly regarding the technical clarity of the frequency-guided design and the optimization details of the dynamic masking strategy. While the overall novelty remains somewhat incremental relative to recent frequency-guided Mamba-based methods, the paper presents a coherent framework with strong empirical performance, competitive efficiency, and clearer methodological justification after rebuttal.
Review #3
- Please describe the contribution of the paper
This manuscript proposes Polyp-AFMamba, a Mamba-based polyp segmentation framework designed to improve robustness to heterogeneous appearance, blurred boundaries, and small lesions. The authors introduce three core modules: 1）An adaptive parallel visual 2D-Mamba layer, which is designed to model long-range dependencies while better preserving 2D spatial continuity. 2）A dynamic frequency decoupling module, which aims to separate low-frequency and high-frequency components using sample-adaptive masking. 3）A low-frequency-guided region-prior adaptive decoding scheme, equipped with dynamic kernels for region-aware refinement to enhance regional consistency. The method is evaluated on five public polyp segmentation benchmarks and demonstrates state-of-the-art performance with competitive computational cost.
- 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 motivation is clear and well justified, as it addresses practical challenges in polyp segmentation such as heterogeneous appearance, blurred boundaries, and missed small targets. The manuscript also attempts to improve not only accuracy, but also robustness and efficiency, which is valuable for real-world deployment. 2.The proposed architecture integrates spatial continuity preservation and frequency-aware enhancement into a unified Mamba-based segmentation pipeline. The interaction among APV-2DMamba, dynamic frequency decoupling, and low-frequency-guided decoding is conceptually coherent, making the overall design more systematic than a simple single-branch modification. 3.The authors conducted extensive experiments to evaluate the proposed method, which is commendable. Although the evaluation metrics are limited to Dice and mIoU, the reported results are overall consistent.
- 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.Compared with recent polyp segmentation methods, the novelty of this work appears limited and is more of an engineering-style innovation through the combination of existing ideas. PolyMamba[1] has already clearly demonstrated the central idea of improving Mamba-based polyp segmentation by combining spatial priors with high-frequency enhancement, while Polyp-Mamba[2] also explored a Mamba-based framework with multi-frequency perception and DCT-based spectral modeling for polyp segmentation. In comparison with these methods, the present paper appears closer to a reconfiguration of existing design directions than to a clearly new methodology. 2.A core claim of the paper is that APV-2DMamba preserves 2D spatial continuity better than standard 1D flatten-then-scan visual Mamba. However, this motivation has already been well captured in 2DMamba[3], which explicitly emphasizes preserving 2D topology rather than folding images into 1D sequences. Therefore, the paper needs to explain more clearly, both conceptually and empirically, what APV-2DMamba contributes beyond existing 2D selective scanning or multi-directional spatial Mamba formulations. 3.The manuscript repeatedly claims advantages in handling small polyps, blurred boundaries, and cross-domain generalization, but these claims are mainly supported by qualitative examples rather than targeted quantitative analysis. This is particularly reflected in the fact that the evaluation metrics are limited to Dice and mIoU. 4.The ablation study only verifies the usefulness of the model, but lacks further analysis of key heuristic hyperparameters and design choices, such as the selection of top_k_ratio. 5.The method description in the manuscript lacks clarity. In particular, the MHEE module is not fully described in the main text. 6.The experimental section lacks statistical reporting, such as variance/confidence intervals and significance testing. In addition, the code has not been made publicly available. [1] Fu R, et al. “PolyMamba: Spatial-Prior Guided Mamba for Polyp Segmentation with High-Frequency Enhancement.” MICCAI. 2025.[2] Zhu X, et al. “Polyp-mamba: A hybrid multi-frequency perception gated selection network for polyp segmentation.” Information Fusion. 2025.[3] Zhang J, et al. “2dmamba: Efficient state space model for image representation with applications on giga-pixel whole slide image classification.” CVPR. 2025.
- 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 provide sufficient information for 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.
(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?
Although the paper proposes a coherent polyp segmentation framework and demonstrates some empirical gains, its overall contribution to the polyp segmentation field appears limited. The work mainly focuses on combining functional modules such as Mamba-based polyp segmentation, 2D Mamba spatial modeling, and frequency-guided dynamic decoding, while the methodological novelty is not sufficiently strong. Although the experimental validation is reasonable, it is still insufficient to clearly demonstrate superiority over recent methods, and some implementation details are lacking.
- 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
Author Feedback
We thank all reviewers for the constructive comments. Following MICCAI rules, we only clarify the manuscript without new results or external material.
To Reviewer #1
R1Q1: Frequency guide, non-object edges, Fig. 1, reproducibility. A: Thank you for the supportive assessment. Frequency guidance is not tied to Mamba; DCT-based LF/HF cues can also guide CNN/Transformer decoders, but Mamba especially benefits because visual SSMs can weaken high-frequency details and flatten 2D topology. We agree DCT may respond to non-object edges. Raw DCT edges are not segmentation outputs; they are filtered by decoder semantics and region priors to suppress background responses. We will redraw Fig. 1 and release code after acceptance while preserving anonymity.
To Reviewer #2
R2Q1: Novelty vs. FAGF-Net/Polyp-Mamba. A: We appreciate that R2 recognized the dual-domain design and strong results. Our Polyp-AFMamba is not a simple DCT+Mamba recombination. It jointly designs: (1) APV-2DMamba with four TL/BR/BL/TR parallel 2D scanning subspaces; (2) dynamic frequency decoupling that learns sample-adaptive frequency importance instead of fixed partitions; and (3) region-prior dynamic decoding where LF cues guide foreground/boundary/background-aware kernels. Unlike FAGF-Net/Polyp-Mamba-like designs, our split is learned per sample and routed differently: HF cues refine boundaries via MHEE, while LF cues guide LGRAM.
R2Q2: Direct comparison and SOTA claim. A: Broader comparison with recent frequency/Mamba models would strengthen the paper. Since new experiments are not allowed in rebuttal, we point to submitted evidence. Table 1 includes the closest MICCAI Mamba baseline PolyMamba: Polyp-AFMamba improves Dice/mIoU on CVC-ClinicDB, Kvasir, CVC-ColonDB, and EndoScene, with only a small ETIS gap. Table 2 reports 78.58 FPS and competitive complexity.
R2Q3: Dynamic mask/loss details. A: The frequency masks are optimized end-to-end: DCT is fixed; GAP + 1x1 convolutions + softmax score frequency channels; top_k_ratio=0.3 selects LF channels per sample, and the rest form HF cues. The reported 5% stochastic mask perturbation reduces spectral over-reliance. For region priors, L_total = L_pred + L_SARD, where L_pred is BCE + IoU and L_SARD supervises foreground/background/boundary masks with MSE; the weight is 1.We will clarify this omitted detail if accepted.
To Reviewer #3
R3Q1: Relation to PolyMamba/2DMamba. A: We appreciate that R3 recognized our coherent unified pipeline. We agree that 2DMamba motivates intrinsic 2D scanning, and our paper cites it. Our contribution beyond 2DMamba is the segmentation-oriented parallel subspace formulation and its integration with frequency-guided region decoding. Unlike PolyMamba, which mainly compensates HF information for Mamba, our method treats frequency as an explicit prior throughout the pipeline and jointly addresses topology preservation, frequency bias, and region inconsistency.
R3Q2: Small polyps/blurred boundaries/cross-domain robustness. A: Table 1 covers seen and unseen datasets, supporting cross-dataset robustness. Figs. 2-3 provide qualitative evidence on challenging cases: fuzzy boundaries, specular highlights, false positives, and small targets. Dice/mIoU are standard in prior polyp work; boundary/statistical analyses are valuable but cannot be newly introduced here.
R3Q3: top_k_ratio, MHEE, and reproducibility. A: top_k_ratio=0.3 selects LF channels per sample, rather than a fixed cutoff, to balance LGRAM and MHEE. Too few LF channels give weak structural guidance for LGRAM; too many mix high-frequency noise into regional cues and restrict HF details for MHEE. In MHEE, HF features generate an edge guidance map, Sobel responses are added with a learnable scale, atrous convolutions with dilation {1,2,4,8} extract boundary context, and gated residual fusion suppresses background responses. We will clarify these and release code after acceptance while preserving anonymity.
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.
Please address the concerns about reproducibility issues, missing comparisons with closely related Mamba-based polyp segmentation, and limited novelty.
- 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.
two accept and one reject.
i read through and would accept subject to minor
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 post-rebuttal assessment is generally positive, with two reviewers supporting acceptance and the remaining concerns mainly centered on novelty and presentation rather than a fundamental flaw in the method. After considering the manuscript, reviews, and rebuttal, I recommend acceptance.
The paper presents a coherent spatial-frequency Mamba framework for polyp segmentation. Its main strengths are the integration of adaptive 2D Mamba scanning with frequency-guided boundary and regional refinement, consistent performance gains across five public benchmarks, and competitive efficiency compared with recent CNN/Transformer/Mamba baselines. The rebuttal also clarified several important implementation details, including dynamic frequency mask optimization, region-prior supervision, and the relationship to PolyMamba, Polyp-Mamba, and 2DMamba.
For the camera-ready version, the authors should add the missing loss/mask optimization details, tone down overly broad SOTA/generalization claims, and discuss the lack of targeted quantitative analysis for small polyps and boundary quality.
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.
reviewers are satisfied with the rebuttal period, I have concurred with them as the clarifications and responses are satisfactorily clarifying the paper’s position among others and contributions.