List of Papers Browse by Subject Areas Author List
Abstract
Medical image segmentation is fundamental to computer-aided diagnosis, requiring precise extraction of anatomical structures and pathological regions. While state-space models (SSMs) have demonstrated remarkable capabilities in capturing long-range dependencies with linear computational complexity, existing methods often struggle to effectively balance global context modeling and local detail preservation in medical images. To address this limitation, this paper proposes HEDS-Net, a Hybrid Vision State-space Transformer and Enhanced Deep Supervision Network. The approach introduces a Hybrid Vision State-space Transformer (HVST) module that integrates global sequential modeling with multi-scale local enhancement through a smooth progressive fusion mechanism, an Axial Cross-Information Synthesis Bridge (AXIS-Bridge) that enhances skip connections to refine cross-scale features, and a Progressively Weighted Deep Supervision Head (DS-Head) with dynamically adjusted weights for hierarchical auxiliary supervision. Extensive experimental evaluations on four medical image segmentation datasets, i.e., ISIC17, ISIC18, Kvasir-SEG, and Synapse, demonstrate that HEDS-Net outperforms advanced baseline models on key evaluation metrics: in skin lesion and polyp segmentation tasks, it achieves precise lesion detection and background exclusion; in multi-organ segmentation, it enables stable segmentation of both small and large anatomical structures. Code is available at https://github.com/Momentisbeused/HEDS-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4167_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Momentisbeused/HEDS-Net
Link to the Dataset(s)
ISIC17: https://challenge.isic-archive.com/data/#2017
ISIC18: https://challenge.isic-archive.com/data/#2018
Kvasir-SEG: https://datasets.simula.no/kvasir-seg/
Synapse: https://www.synapse.org/#!Synapse:syn3193805
BibTex
@InProceedings{FanJia_HEDSNet_MICCAI2026,
author = { Fan, Jiaying AND Liang, Yongjie AND Cao, Junyue AND Wang, Peiyuan AND Wei, Bizhong AND Li, Yuexiang},
title = { { HEDS-Net: A Hybrid State-Space Architecture with Axial Bridge and Progressive Weighting for Medical Image 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 issue that medical image segmentation methods based on state space models (SSM) struggle to balance global context modeling and local detail preservation, and proposes HEDS-Net. This network consists of three core components: (1) Hybrid Visual State Space Transformer (HVST), which combines a Visual State Space (VSS) branch for global sequence modeling and an Atrous Dilated Channel Attention (ADCA) branch for multi-scale local enhancement in parallel within a single module, and fuses them through a weight that smoothly changes during training; (2) Axial Cross-Information Synthesis Bridge (AXIS-Bridge), which enhances the skip connections between the encoder and decoder through context-conditioned mixing and axial refinement; (3) Progressive Weighted Deep Supervision Head (DS-Head), which assigns weights that dynamically decay with training rounds to different levels of auxiliary supervision heads.
- 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.
- Accurate and timely problem identification: The paper accurately captures a key pain point in the current application of SSMs (such as Mamba) in medical image segmentation - how to effectively integrate their powerful global modeling capabilities with the local detail capture strengths of CNNs. This is a very practical and important issue.
- Comprehensive experimental verification with outstanding results: Extensive experiments were conducted on four different public datasets (covering skin lesions, polyps, and multi-organ CT segmentation). The results show that HEDS-Net consistently outperforms multiple CNN, Transformer, and SSM baseline models, including VM-UNet, in key metrics such as mIoU and Dice, especially in handling blurry boundaries and small targets.
- Thorough ablation study: The ablation experiments clearly demonstrate the effectiveness of the three modules - HVST, AXIS-Bridge, and DS-Head - individually and their synergistic effects when combined, providing strong support for the design of the proposed method.
- Convincing visualization effects: The qualitative results in Figure 5 visually demonstrate the advantage of HEDS-Net in boundary accuracy, which is consistent with the quantitative results.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The core weakness of this work lies in its limited innovation, which is an engineering integration rather than a conceptual breakthrough. The three modules of HEDS-Net are essentially direct combinations or minor improvements of existing mature technologies, lacking deep original ideas. -1.1 The HVST module: The “hybrid” architecture that combines global (SSM/Transformer) and local (CNN) branches in parallel is a very popular design paradigm in recent years. For instance, early works such as TransFuse (MICCAI 2021) and TransUNet (arXiv 2021) have already explored the parallel or serial fusion of CNN and Transformer. HEDS-Net merely replaces the global branch of Transformer with the updated SSM (VSS) and specifies the local branch as dilated convolution with channel attention. This “new wine in old bottles” approach mainly contributes incrementally in engineering implementation and tuning rather than proposing new fusion mechanisms or theories. -1.2 The AXIS-Bridge module: Fine-tuning skip connections is a common strategy in U-Net variants. The idea of using axial (horizontal/vertical) information to handle anisotropic structures (such as blood vessels, organ boundaries) is not novel either. The core of MEW-UNet (arXiv 2022, Ref [21]) is “Multi-axis representation learning”, explicitly utilizing multi-axis information. HEDS-Net’s AXIS-Bridge can be regarded as a natural extension of such works, with insufficient innovation. -1.3 The DS-Head module: Weighted deep supervision is a classic technique for stabilizing training. Although the strategy of dynamically adjusting weights is somewhat novel, its core idea (shallow supervision is more important in the early stage) is intuitive and widely recognized. This is more an effective training trick rather than an innovation in model architecture.
- The comparison with the most relevant works is not in-depth enough: Although the paper has compared with SSM baselines such as VM-UNet, it has not conducted a direct comparison with works like TransFuse and MEW-UNet that are more similar in architectural ideas. This makes it difficult for readers to determine the true advantages of HEDS-Net over these specific hybrid architectures. 2.The analysis of computational efficiency is missing: The paper emphasizes the linear complexity advantage of SSM, but the introduction of additional CNN branches, complex skip connections, and multiple supervision heads will inevitably increase the computational cost. There is no comparison with baseline models (especially VM-UNet) in terms of the number of parameters, FLOPs, training/inference time, etc. , which is crucial for evaluating its value in practical applications. 3.The technical depth needs to be enhanced: The paper describes more about “how to do it” but lacks in-depth exploration of the “why it works” mechanism. For example, why this specific progressive fusion strategy was chosen? How is it better than simple weighted averaging or gating mechanisms? These questions lack in-depth analysis at the theoretical or experimental level. 4.The cited literature or compared methods are slightly outdated: The cited literature is mostly from before 2022.
- 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.
(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?
1.The work is well accomplished, but it is more like a meticulously designed engineering integration solution, with limitations in terms of innovation. 2.The references or comparative literature cited are too outdated, and there may be an issue of insufficient comparison with related works.
- 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
Review #2
- Please describe the contribution of the paper
The manuscript proposes HEDS-Net, a U-shaped medical image segmentation model that combines state-space global modeling, local detail enhancement, refined skip connections, and progressively weighted deep supervision. Its main contribution is improving the balance between long-range semantic context and fine boundary preservation across medical image segmentation tasks.
- 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 reported results show that HEDS-Net outperforms representative CNN, Transformer, and SSM baselines on the metrics across datasets.
- Useful training strategy without extra inference cost. The progressively weighted deep-supervision head is a practical contribution.
- Interesting hybrid global-local formulation in HVST. The most methodologically interesting part is the HVST block, which combines a VSS branch for long-range dependency modeling with an ADCA branch for local multi-scale detail enhancement.
- The manuscript targets a real weakness of existing segmentation models which is balancing long-range semantic context with fine local boundary details. That is especially important in medical images with ambiguous borders and small 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.
-
The provided code has several critical issues. Most importantly, the HVST block appears incomplete, which would cause the module to fail at runtime.
-
The deep-supervision implementation also does not match the manuscript. The code uses only Dice loss and ignores the claimed BCE+Dice and CE+Dice formulation.
-
Several novelty claims are overstated. The manuscript presents HVST, AXIS-Bridge, and DS-Head as three distinct innovations. But the actual descriptions show that HVST is a VSS branch plus an atrous channel-attention local branch with scheduled fusion, AXIS-Bridge is gated skip refinement, and DS-Head is deep supervision with decaying weights.
-
The ablation study is helpful but not sufficiently targeted. There is no direct comparison of progressive fusion vs fixed fusion, no comparison of AXIS-Bridge vs simpler skip attention, and no comparison of progressively weighted supervision vs ordinary fixed-weight deep supervision.
-
- Please rate the clarity and organization of this paper
Poor
- 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?
- The methodological novelty is limited.
- The empirical gains are not strong enough.
- My confidence dropped further because the provided code does not cleanly match the manuscript.
- 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
The Hybrid Vision State-space Transformer and Enhanced Deep Supervision Network are proposed.
- 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 proposed may be potentially useful for medical image segmentation
- 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 experimental results seems insufficient as far as i know
- 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
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?
based on the authors’ descriptions, i judged the work
- 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 sincerely thank the Area Chair and all reviewers for their careful evaluation and constructive suggestions. We have carefully considered each concern and will revise the final version accordingly.
1.Novelty and Positioning (AC, R1, R2) We acknowledge the concern that global-local modeling, skip refinement, and deep supervision have been previously explored. However, HEDS-Net is not presented as three isolated conceptual breakthroughs, nor is it merely a loose engineering assembly. Its contribution lies in a task-oriented SSM segmentation design: HVST integrates VSS global modeling with ADCA local enhancement through a scheduled local-to-global fusion; AXIS-Bridge employs decoder-conditioned gating prior to horizontal and vertical descriptor refinement; and DS-Head utilizes epoch-decayed supervision and is removed during inference. Collectively, these components address a common practical challenge in SSM-based segmentation—preserving fine anatomical details while maintaining global semantic coherence. We will clarify this scope and avoid overstating isolated novelty.
2.Comparisons, Analysis, and Related Work (AC, R1, R2, R3) We agree that related hybrid and Mamba/SSM methods should be discussed more clearly and acknowledge that some citations are outdated. Direct comparisons are already included: Table 1 reports TransFuse, TransUNet, VM-UNet, and HC-Mamba where applicable, and Table 2 includes TransUNet, Swin U-Net, TransDeepLab, UCTransNet, MEW-UNet, and VM-UNet on Synapse. Therefore, the main issue is not the absence of relevant baselines but rather that the discussion does not clearly organize them by method family. In the revision, we will relate TransFuse and TransUNet to CNN-Transformer hybrids, MEW-UNet to multi-axis modeling, and VM-UNet and HC-Mamba to SSM/Mamba segmentation, while incorporating recent medical Mamba/SSM literature without overclaiming superiority over unevaluated methods.
3.Efficiency and Additional Ablations (AC, R1, R2) We appreciate the requests for detailed parameters, FLOPs, latency, memory usage, and targeted variants, including progressive versus fixed fusion, AXIS-Bridge versus simpler skip attention, and progressive versus fixed deep supervision. We agree that computational efficiency is crucial for practical deployment and will include a discussion of this aspect in the final version. Rather than claiming a lightweight architecture, we provide the following statistics to transparently present the accuracy-complexity trade-off (input size: 256×256, single RTX 3090): HEDS-Net contains 44.61M parameters and 6.06G FLOPs; FP32 inference takes 19.85 ms per image (batch size = 1, averaged over 100 runs) with a peak memory usage of 329.7 MB. Regarding progressive fusion, we examined fixed-alpha behavior during method development to compare static fusion weights with scheduled local-to-global fusion. The observed trend supported the motivation for progressive fusion; however, this analysis was omitted due to page limitations and because Table 3 prioritized module-level effects of HVST, AXIS-Bridge, and DS-Head. We will refine the discussion of the existing ablation study and consider these more fine-grained analyses in future follow-up work where appropriate.
4.Reproducibility and Code Consistency (R2) We take this concern seriously. The publicly available code should correspond exactly to the manuscript: the HVST block must be complete and executable, and the loss functions should follow BCE+Dice for binary segmentation and CE+Dice for multi-class segmentation. We will review and correct the implementation to ensure consistency among the code, method description, and conclusions.
We believe these revisions substantially strengthen the paper and respectfully ask the Area Chair and reviewers for Camera-ready.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 addresses an important problem in medical image segmentation: balancing long-range context modeling with fine boundary preservation. The proposed HEDS-Net is technically coherent, and the reported results across four datasets are promising. In particular, the HVST block is a reasonable hybridization of global state-space modeling and local enhancement, and the training-side deep supervision strategy is practically useful.
The main strengths of the paper are: 1, The paper addresses the important challenge of balancing long-range context modeling with fine boundary preservation in medical image segmentation. 2, HEDS-Net integrates global SSM modeling, local multi-scale enhancement, refined skip fusion, and progressively weighted deep supervision into a unified U-shaped framework. 3, The method reports consistent improvements across four datasets and representative CNN, Transformer, and SSM baselines. 4, The ablation and qualitative results indicate that the proposed components contribute to the final performance.
The main remaining concerns are: The novelty is partly incremental, as the method builds on known ideas such as global-local hybrid modeling, skip refinement, and deep supervision. Comparisons with closely related hybrid architectures and recent Mamba/SSM-based segmentation methods could be more complete. Efficiency analysis would be stronger with parameters, FLOPs, latency, and memory statistics.
Overall, these concerns mainly concern positioning, completeness of analysis, and presentation/reproducibility clarification. They do not undermine the main technical idea or the empirical conclusions. Given the positive reviewer consensus, the consistent results across datasets, and the practical relevance of the proposed framework, I recommend Early Accept. The authors should address the above issues in the camera-ready version. These revisions would improve the presentation and reproducibility, but they are not severe enough to delay the paper to rebuttal given its overall strength and positive reviewer consensus.