List of Papers Browse by Subject Areas Author List
Abstract
Colorectal polyp segmentation can assist doctors in screening colonoscopy images, which is crucial for the prevention of colorectal cancer. Although deep learning has significantly advanced polyp segmentation, three issues remain: (1) Most polyp segmentation methods only extract Euclidean features such as shape and texture, while neglecting non-Euclidean features, such as the geometric topology between the polyp and its surrounding tissue; (2) Non-Euclidean features vary across different regions, but most feature fusion methods overlook both the non-Euclidean topological structures and the differences between internal, edge, and background regions. (3) Low-level features are not fully exploited, and the differences between low- and high-level features are not effectively addressed. To resolve these issues, we propose Hybrid Graph Mamba (\ourmodel{}) based on Mamba and Graph Convolutional Network (GCN). Our model first uses the pyramid vision transformer to extract features at different levels. Next, we propose hybrid graph Mamba modules to process low-level features from multiple directions using quad-directional Mamba and extract non-Euclidean features with GCN. A boundary discrimination fusion module is also designed to handle high-level features, extracting semantic information for the interior, edges, and background to improve the fusion of low- and high-level features. Finally, a bidirectional Mamba decoder combines bidirectional Mamba and dilated convolutions to aggregate multi-scale features, minimizing information loss and producing the final prediction. Extensive experiments on five benchmark datasets demonstrate that \ourmodel{} significantly outperforms eight State-Of-The-Art models. Our code is publicly available at https://github.com/YueyueZhu/HGM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1195_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YueyueZhu/HGM
Link to the Dataset(s)
CVC-300 (60 test samples), CVC-ClinicDB (62 test samples), CVC-ColonDB (380 test samples), ETIS-LaribPolypDB (196 test samples), Kvasir (100 test samples): https://drive.google.com/file/d/1Y2z7FD5p5y31vkZwQQomXFRB0HutHyao/view?usp=sharing
Kvasir-SEG (900 train samples) and CVC-ClinicDB (550 train samples): https://drive.google.com/file/d/13Zij1HbKTn9PKTW9bM19_fXtfQKWdCtD/view?usp=sharing
BibTex
@InProceedings{ZhuYue_Hybrid_MICCAI2025,
author = { Zhu, Yueyue and Lv, Haolin and Chen, Geng and Zhang, Zhonghao and Jiang, Haotian and Xia, Yong},
title = { { Hybrid Graph Mamba: Unlocking Non-Euclidean Potential for Accurate Polyp Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15969},
month = {September},
page = {276 -- 285}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents the Hybrid Graph Mamba (HGM), an innovative deep learning model that combines Mamba and Graph Convolutional Networks (GCN) for effective polyp segmentation in colonoscopy images. Its key contributions include the creation of a Hybrid Graph Mamba Module (HGMM), which leverages Quad-directional Mamba and GCN to capture non-Euclidean features, a Boundary Discrimination Fusion Module (BDFM) that adeptly handles diverse image regions, and a Bidirectional Mamba Decoder (BMD) that integrates multi-scale features while preserving critical information.
- 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 successfully combines state-space models (Mamba) with graph neural networks, creating a unique architecture that addresses different aspects of the segmentation problem. This takes thoughtful consideration of the limitations of these individual architectures and leverages their complementary strengths.
-
The empirical evaluation across five diverse datasets demonstrates the model’s robust performance, particularly excelling on challenging datasets like ETIS where it achieves significant improvements.
-
The paper provides detailed ablation studies that convincingly demonstrate the contribution of each proposed component.
-
- 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.
-
While the paper repeatedly emphasizes the importance of non-Euclidean features, it fails to precisely define what these features represent in the context of polyp segmentation. The claim that GCN extracts “geometric and topological structures formed by a polyp and its surrounding tissues” lacks empirical validation. The paper doesn’t provide visualizations or analyses demonstrating that the extracted features are indeed capturing meaningful topological relationships rather than just providing an alternative feature transformation.
-
The authors make a significant simplification in the GCN’s adjacency matrix by arbitrarily setting “every 16 units along each axis” to one. This design choice seems ad hoc and lacks theoretical justification. How does this predetermined graph structure capture the actual topological relationships in polyp images, which vary significantly in size and shape? This raises questions about whether the GCN is truly modeling non-Euclidean geometries or merely serving as another feature transformation layer.
-
No computational complexity analysis is provided.
-
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The idea is interesting and novel with minor confusions and weaknesses.
- Reviewer confidence
Somewhat confident (2)
- [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 proposed the Hybrid Graph Mamba Model, which introduced non-Euclidean features into polyp segmentation for the first time and finally proved its effectiveness.
- 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.
- Novel Methodology: Noticed about the long-neglected non-Euclidean geometric features in polyp segmentation, proposed a novel hybrid architecture which combine Mamba and GCN.
- Well-Designed Modules: The BDFM distinguishes regional features, QM captures multi-directional details, and BMD fuses multi-scale information, synergistically improving segmentation accuracy.
- Clinical Relevance: Superior performance on challenging scenarios (e.g., ETIS dataset) and accurate boundary detection reduce the risk of missed diagnoses.
- 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.
- Lack of Quantitative Analysis on Non-Euclidean Features: While emphasizing the importance of non-Euclidean features, the paper does not provide visualizations or quantitative metrics to concretely demonstrate their contribution.
- Computational Efficiency: The incorporation of non-Euclidean feature analysis will undoubtedly affect the computational efficiency, resulting in a low inference speed and even impacting the final deployment. However, this paper does not provide a discussion on the trade-off between the sacrificed computational efficiency and the achieved improvement in performance.
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes an innovative hybrid architecture that introduces non-Euclidean features into polyp segmentation and effectively fuses multi-scale information through modular designs. The experiments validate its superiority on five datasets, particularly in challenging scenarios, demonstrating clinical relevance. However, the lack of quantitative analysis on non-Euclidean features and Insufficient discussion on computational efficiency impact the final score. Hope that more experiments and discussions will be added in the future.
- 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 #3
- Please describe the contribution of the paper
A novel deep learning model, called Hybrid Graph Mamba (HGM), which achieves accurate polyp segmentation with a mixture of Mamba and Graph Convolutional Network (GCN). Specifically, we use a pyramid vision transformer to extract features at different levels. For low-level features, we design the Hybrid Graph Mamba Module (HGMM), which focuses on detailed features using Quad-directional Mamba (QM) and extracts nonEuclidean features using GCN. For high-level features, we use the Cascaded Fusion Module (CFM) to extract semantic and positional information of polyps.
- 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.
presentation of the methodology and results very good.
- 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.
1195 - Hybrid Graph Mamba: Unlocking Non-Euclidean Potential for Accurate Polyp Segmentation Comments to the Author(s), Dear Author(s),
- It is suggested to include the authors’ contribution of the proposed work.
- In Section 3.1 (Dataset), a table summarizing dataset details, including the data split, should be provided. The authors should acknowledge the critical role of dataset size and distribution in the proposed task.
- In Section 3.1 (Implementation), the selection criteria for hyperparameters should be clearly justified, and supporting results should be provided.
- Cross-validation is required. Additionally, accuracy and loss plots should be provided to enhance the reliability of the results.
- At the end of Section 3, the limitations of the proposed approach should be discussed. If possible, a comparative analysis with existing literature or state-of-the-art methods should be included.
- 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.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
80/100
- Reviewer confidence
Very confident (4)
- [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 sincerely thank all reviewers for their thoughtful and constructive feedback.
R1 Authors’ contributions: As the MICCAI follows a double-blind review procedure, we did not disclosure the authors’ contributions in our submitted paper. We agree with the reviewer’s comment and will supplement the authors’ contributions in the final paper.
Dataset details: Our training configuration follows that of PraNet. Specifically, we sample 550 subjects from ClinicDB (612 in total) and 900 subjects from Kvasir (1,000 in total) to form the training set (1,450 subjects overall). The remaining data from ClinicDB and Kvasir, together with the CVC 300, ColonDB, and ETIS datasets, comprise the test set (798 subjects). A table summarizing dataset details will be included in the final paper.
Selection criteria for hyperparameters: Our hyperparameter choices follow those of PraNet [5] and Polyp‑PVT [3]: the input image resolution is set to (352,352), the batch size to 16, and the learning rate to 1e-4.
How to enhance the reliability of the results: We will, following the rebuttal, supplement our final paper with cross validation results as well as accuracy and loss statistic charts.
Limitations: We will include a discussion of our method’s limitations in the final paper.
R3 Definition of non-Euclidean features: The non-Euclidean characteristics of polyps reflect their underlying topological and geometric structure. As a polyp grows, it induces geometric deformations in the surrounding tissue, such as stretching and twisting-whose magnitude, direction, and spatial extent define non-Euclidean geometric features that can distinguish pathological from normal tissue. Given the colon’s inherently complex, curved anatomy, a polyp’s position cannot be adequately described by simple Euclidean distances or Cartesian coordinates. Instead, its spatial relationship to colonic folds, flexures, and the lumen trajectory constitutes essential non-Euclidean information that helps localize the lesion accurately and mitigates segmentation errors caused by Euclidean localization biases. To capture these properties, we treat each pixel as an individual node in a graph convolutional network, enabling the model to learn non-Euclidean features automatically. The visualization analysis of these learning features will be presented in the final paper.
Reason to choose the GCN’s adjacency matrix: We adopt a spacing of 16 pixels because if the interval is too large, too few pixels will participate in the GCN, weakening the learned representations. If it is too small, the computational burden becomes excessive. This module captures non-Euclidean characteristics by modeling relationships between pixels, and sparsifying the adjacency matrix in this way (rather than using an all ones matrix) further reduces the overall computational cost.
Computational complexity: Computational complexity analysis will be provided in the final paper.
R4 Quantitative Analysis on Non-Euclidean Features: The quantitative analysis of non-Euclidean features will be provided in the final paper.
The Computational Efficiency of the incorporation of non-Euclidean features: The computational cost of extracting non-Euclidean features is governed by the pixel sampling interval. Through empirical experiments, sampling one every 16 pixels, an optimal compromise: it limits computation yet ensures that information from a sufficiently broad neighborhood can be aggregated and propagated without loss of critical features. A detailed discussion of the tradeoff between computational efficiency and performance gains will be included in the final paper.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A