List of Papers Browse by Subject Areas Author List
Abstract
Accurate segmentation of medical images remains a intricate challenge due to the complex geometry of target structures and their blurred boundaries. While recent approaches leverage topological and semantic information by introducing graph or hypergraph, they face two critical limitations: i) fixed patch nodes inherently ignore object boundaries, causing semantic ambiguity by mixing heterogeneous features; and ii) restricting reasoning to bottlenecks or skip connections fails to fully inject global contexts for fine-grained segmentation. To address these issues, we propose S3H-Net, a novel architecture that integrates Superpixel-based Semantic-Spatial Hypergraphs. Specifically, distinct from patch-based methods, we utilize a boundary-aware Simple Linear Iterative Clustering (SLIC) algorithm for node construction, encouraging contour-aware regional grouping and alleviating heterogeneous feature mixing. Furthermore, we design Hybrid Semantic-Spatial Hypergraph (HSSH) blocks to construct a parallel Semantic Encoder, which captures multi-scale high-order semantic guidance continuously. To explicitly bridge the semantic gap between the non-local hypergraph topology and local convolutional features, we introduce a Cross-Covariance Semantic Injection (CCSI) module. By exploiting inter-channel attention, this module effectively injects the reasoned global context into local features for comprehensive semantic awareness. Extensive experiments on five medical imaging modalities demonstrate that S3H-Net significantly outperforms state-of-the-art methods. Our code is available on GitHub.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2653_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Saury997/S3H-Net
Link to the Dataset(s)
CVC-ClinicDB dataset: https://polyp.grand-challenge.org/CVCClinicDB
ETIS-LaribPolypDB dataset: https://www.kaggle.com/datasets/nguyenvoquocduong/etis-laribpolypdb
ISIC2018 dataset: https://challenge.isic-archive.com/data/#2018
USG dataset: https://www.cancerimagingarchive.net/collection/breast-lesions-usg
DSB2018 dataset: https://www.kaggle.com/competitions/data-science-bowl-2018
PROMISE12 dataset: https://promise12.grand-challenge.org
BibTex
@InProceedings{YanZon_S3HNet_MICCAI2026,
author = { Yang, Zongjian AND Ma, Lanxiang AND Gao, Zeli AND Zhang, Yuqi AND Ma, Jiquan},
title = { { S3H-Net: Injecting Hybrid Semantic-Spatial Guidance via Superpixel Hypergraphs for Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper tackles the limitation of fixed patch-based nodes (e. g. , in hypergraph networks), which do not adequately account for object boundaries. It also addresses the issue of modeling global context only at bottlenecks or through skip connections. To this end, the authors propose an architecture that utilizes superpixel-based hypergraphs, where nodes are constructed using the boundary-aware SLIC superpixels algorithm. In addition, they introduce hybrid hypergraph blocks (HSSH) to build an encoder that continuously captures multi-scale, high-order semantic information. To bridge the semantic gap between hypergraph (global) and convolutional (local) features, the authors propose a CCSI module. The proposed architecture is evaluated across five medical imaging modalities and achieves strong performance.
- 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 effectively integrate multiple concepts and improvements to address several key challenges: (1) the complex geometry of target structures and their blurred boundaries, (2) the limitation of fixed patch-based nodes, which inherently ignore object boundaries, and (3) the restriction of global reasoning to bottlenecks or skip connections. To tackle these issues, they propose a set of complementary enhancements. Specifically, the architecture employs a dual-encoder design, where a CNN captures local features while a hypergraph encoder models global, high-order relationships. For node construction, the method leverages a boundary-aware and differentiable superpixel algorithm, enabling better alignment with object boundaries. Furthermore, the authors introduce an injection module to effectively fuse features from the two encoders. The authors also provide a link to the code.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The authors do not report the model complexity in terms of the number of parameters or FLOPs. This is particularly important given that the architecture employs two encoders, one of which is a hypergraph module that is typically computationally expensive.
- 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.
(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?
There are relatively few papers on hypergraph-based segmentation, and this work provides multiple contributions: (1) a dual-encoder architecture, including a hypergraph encoder; (2) a method that utilizes a boundary-aware node construction algorithm; and (3) a mechanism to fuse the two encoders while reducing the semantic gap.
- 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.
The paper has a good contribution.
Review #2
- Please describe the contribution of the paper
This paper presents a well-motivated and technically model that achieves strong empirical results across diverse medical imaging modalities. The main contributions are threefold: a boundary-aware SLIC algorithm that generates semantically pure nodes aligned with anatomical contours, a Hybrid Semantic-Spatial Hypergraph (HSSH) that captures high-order correlations through complementary semantic and spatial hyperedges, and a Cross-Covariance Semantic Injection (CCSI) module that bridges the gap between non-local topological reasoning and local spatial representations.
- 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 boundary-aware SLIC is the most original contribution of the paper. Formulating the edge probability as a barrier integral (Eq. 1–2) yields an elegant and fully differentiable design that enforces strict boundary adherence at the node construction stage. (2) The hybrid hypergraph construction, which combines FCM-based semantic hyperedges with dilated spatial hyperedges, is well-motivated and reasonable, albeit somewhat incremental relative to prior hypergraph designs. (3) Comprehensive experiments across five medical imaging modalities demonstrate that S³H-Net consistently outperforms state-of-the-art methods, with statistical significance testing further supporting the reliability of the improvements.
- 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)While the proposed framework demonstrates a certain degree of novelty as a whole, each of its major components is either an existing method or a straightforward extension thereof. In the superpixel module, although the introduction of boundary constraints via a composite-trapezoidal approximation of an edge-aware line-integral barrier shows some innovation, the overall architecture still follows the SSN framework. In the hypergraph reasoning module, the construction strategy is essentially a combination of two existing approaches (FCM-based semantic edges and dilated-window spatial edges). Similarly, the CCSI module closely resembles channel-wise co-attention, which has already been explored in existing works such as DuSFE[1], MA-Neck[2].
[1] “DuSFE: Dual-channel squeeze-fusion-excitation co-attention for cross-modality registration of cardiac SPECT and CT.” Medical image analysis 88 (2023): 102840.[2] “MA-Neck: Mutual attention-based feature enhancement for lightweight object detection.” Pattern Recognition (2025): 112254. 2) The choice of hyperparameters in the implementation section warrants further justification. In particular, the values λ_xy = 0.17 (SLIC compactness) and λ_b=0.5 (boundary sensitivity) are stated to be “set empirically” but without supporting evidence. It is unclear whether these values were determined through grid search, ablation studies, or simply adopted from prior work. A sensitivity analysis would strengthen the reproducibility of the method and help readers understand the robustness of performance to these settings. Relatedly, the number of sampled points K used in the trapezoidal approximation of the barrier integral (Eq. 2) is never specified.
3) The experiments section lacks a sensitivity analysis for the scale factor λ in the hybrid incidence matrix H=[H_sem||λ*H_spa]. Although the authors report λ=5 as an empirical choice, no evidence is provided to justify this value, and it is unclear how sensitive the model is to this balance between semantic and spatial hyperedges.
4) The paper would benefit from visualizations of the generated superpixels and the constructed hypergraphs overlayed on medical images. Such qualitative evidence is needed to substantiate the motivation presented in Fig. 1 that superpixels of anatomical boundaries are more faithfully than grid patches, and that the learned hyperedges capture meaningful high-order semantic correlations.
5) The overall presentation of the manuscript requires significant refinement. The writing style feels somewhat non-standard and repetitive, characterized by an over-reliance on “We”-led sentences. The authors are encouraged to revise the manuscript to achieve more varied sentence structures and to adopt more precise academic terminology. On the of Page 5 “Hypergraph Construction” reads: “…we employ the Fuzzy C-Means (FCM) algorithm inspired by [9]…”. References should not be used as grammatical components of the sentence. Phrasings such as “…in [1]…” or “Reference [1] states…” are not acceptable. The authors should rephrase such sentences and ensure proper reference formatting throughout the manuscript. A careful pass for English constructions, grammatical errors, and formatting inconsistencies is strongly recommended to improve overall presentation quality.
- 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.
(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?
The proposed method appears to build mainly upon a combination of existing techniques. I will reconsider my evaluation based on the authors’ response.
- Reviewer confidence
Very confident (4)
- [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 has addressed the concerns, and the paper can be accepted.
Review #3
- Please describe the contribution of the paper
The paper proposes a boundary-aware Simple Linear Iterative Clustering algorithm for node construction，design Hybrid blocks to construct a parallel Semantic Encoder, which captures multiscale high-order semantic guidance continuously. introduce a Cross-Covariance Semantic Injection(CCSI) module，which injects the reasoned global context into local features for comprehensive semantic awareness. The module design is supported by certain theories. The model is evaluated on multiple benchmark datasets with several quantitative metrics. Experiments demonstrate that the proposed modules improve segmentation performance compared with the baseline network.Extensive experiments are conducted on across five distinct medical imaging modalities, demonstrating that the proposed method achieves competitive or superior performance compared with several former approaches. Ablation studies are provided to analyze the contributions of several key components , showing that each module contributes positively to the overall segmentation.
- 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 manuscript is clearly written and well structured, which makes the overall methodology and experimental setup straightforward to understand. There are no too many obvious grammatical or typographical errors. 2.The paper addresses a relevant problem in medical image segmentation. This is a challenging task with practical significance and a reasonable motivation, namely to handle the complex geometric shapes of the target structure and the ambiguous boundaries of these shapes. 3.The proposed network is logically designed. The experimental section is reasonably complete. The paper reports results on multiple public datasets and includes both ablation studies and a cross-dataset generalization experiment, which strengthens the empirical evaluation. 4.The method shows competitive performance against recent baselines, with consistent gains on the reported benchmarks. This suggests that the proposed design is effective in practice, even if the margins are not always large. 5.The contribution is presented in a fairly clear way. The roles of the main components are distinguishable, and the paper makes a concrete engineering contribution to biomedical 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.
1.The overall methodological novelty appears somewhat limited. The motivations behind some modules are not discussed in much depth. 2.The main weakness is limited novelty. The proposed method is a combination of several familiar ideas. The integration is reasonable, but the methodological contribution feels incremental rather than fundamentally new. The work does not introduce fundamentally new modeling principles or theoretical insights. 3.Although the reported results are competitive, the margins over the strongest baselines are not always large. In some cases, the improvements are quite modest, raising questions about whether the improvements justify the additional architectural complexity, which makes it harder to judge how meaningful the gains are in practice. 4.The paper emphasizes segmentation accuracy but does not provide much information about model complexity, inference speed, or computational cost. Since the proposed framework introduces additional modules, it would be useful to report efficiency-related metrics such as the number of parameters, FLOPs, or inference time.
- 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
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 addresses a relevant problem in medical segmentation and presents a technically reasonable model. The overall framework is coherent, and the results are generally competitive. However, the contribution is fairly incremental. The method combines several familiar ideas, and the paper does not fully show that the overall design represents a clear methodological advance rather than a careful integration of existing components. The experimental section is adequate but not fully convincing. Although the paper includes ablations and cross-dataset evaluation, the improvements over strong baselines are sometimes small, and there is no reporting of variance, repeated runs, or statistical significance. This makes it difficult to judge the robustness of the claimed gains.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The revised version still seems that the method combines several familiar ideas, and the paper does not fully show that the overall design represents a clear methodological advance.
Author Feedback
We sincerely appreciate the constructive suggestions from reviewers.
1.Novelty (meta-reviewer, R2, R3) We agree that patch representations are effective, and we did not intend to claim that patches are generally problematic. Our concern is specific to graph/hypergraph reasoning: fixed patches crossing lesion/background boundaries may become mixed atomic vertices and affect high-order aggregation. We also agree that superpixels are not semantic labels, and will replace overstatements such as “semantic purity”, “strict alignment”, and “eliminate semantic ambiguity” with more precise claims: boundary-aware superpixels encourage locally homogeneous, boundary-adaptive primitives and reduce, rather than eliminate, boundary-crossing feature mixing. S3H-Net is not a simple aggregation of independent modules, but a task-driven formulation of superpixel-based hypergraph reasoning: boundary-aware differentiable SLIC provides adaptive vertices, HSSH models semantic-spatial high-order relations, and CCSI injects the reasoned topology back into pixel-level CNN features. The submitted ablation already includes a patch-node vs. superpixel-node comparison: without superpixels, HSSH uses fixed 8×8 grid patches as basic units. Replacing these patch nodes with superpixel nodes improves ETIS/USG DSC from 93.6/85.9 to 94.8/87.0 before CCSI is introduced, and the full model further improves to 95.4/87.5.DuSFE and MA-Neck are related in channel-wise recalibration, but CCSI targets a different problem: hypergraph-to-pixel semantic injection for medical segmentation.
2.Complexity (R1, R3) Under the submitted setting (256×256 input, NVIDIA 2080Ti), S3H-Net has 54.09M parameters and 10.33G FLOPs, with 1106.0 MB peak GPU memory and 91.17 ms/item inference time. The differentiable SLIC module itself contributes only 0.02G FLOPs and 2.91 ms/item, indicating that superpixel construction is not the main computational bottleneck.
3.Hyperparameter clarification (R2) We acknowledge that λxy, λb, and λ were insufficiently justified. λxy and λb control compactness and boundary sensitivity in differentiable SLIC and were selected from a reasonable grid based on validation behavior and qualitative superpixel stability. The scale λ in H=[Hsem||λHspa] balances semantic hyperedges and spatial priors: too small λ weakens local continuity, while too large λ makes the hypergraph overly local. We used λ=5 as a moderate balancing value. K in Eq. (2) is a numerical integration setting rather than a model hyperparameter; we set K=8, corresponding to 9 sampled locations including endpoints.
4.Practical significance and robustness (R3) Although the DSC gain is modest on saturated datasets such as DSB2018, the improvement is not merely numerical: S3H-Net consistently improves boundary-sensitive HD95 across modalities and achieves best or competitive overall performance. The submitted ablations show progressive gains from each component, and the DiffSLIC overhead is small (0.02G FLOPs, 2.91 ms/item), suggesting a practical accuracy/boundary benefit with limited extra cost.
5.Qualitative evidence for superpixels (R2) We clarify that the superpixel visualization in Fig. 1 Right bottom is an actual DiffSLIC output with 64 superpixels, not a manual illustration, and thus qualitatively supports boundary-adaptive node construction. We agree that hyperedge overlays would improve interpretability, but they would introduce new visual evidence beyond the submitted manuscript; therefore, we focus on clarifying submitted results.
6.Statistical reporting (R3) Table 1 already reports t-test p-values, and we will clarify the test protocol, paired unit, and sample size.
7.Writing (R2) We will also revise repetitive “We”-led constructions, improve terminology precision, and correct citation style; e.g., “we employ FCM inspired by [9]” will be revised to “Following prior hypergraph construction using fuzzy clustering [9], semantic hyperedges are formed by FCM.”
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 main weakness is limited novelty. The method is largely a combination of existing components: SLIC/superpixel node construction, hypergraph reasoning, semantic/spatial hyperedge construction, and channel-wise attention-like feature fusion. This concern is also raised by Reviewer #2 and Reviewer #3. The paper makes superpixels the central design choice, but this is not sufficiently convincing. Superpixels are an old and extensively studied representation. In 2026, using SLIC-style superpixels as a core DL/AI pipeline component requires much stronger justification than currently provided. There is generally nothing wrong with patch-based representations in modern vision systems. Patches are simple, stable, learnable, and widely successful. The paper’s argument that patches inherently cause semantic contamination is overstated.
Superpixels may introduce additional instability. They rely on a grouping stage before the main learning system, and this stage can introduce errors, noise, and artifacts. If superpixels incorrectly merge lesion/background regions or fragment meaningful structures, the downstream hypergraph reasoning is built on corrupted nodes.
The claims of “semantic purity,” “strict boundary adherence,” and elimination of semantic ambiguity are too strong. Superpixels are low-level perceptual regions, not semantic units. A Sobel-based edge/barrier term cannot guarantee anatomical boundary alignment, especially in noisy, low-contrast, or artifact-heavy medical images.
The experiments do not adequately isolate the effect of superpixels. A controlled patch-node vs. superpixel-node comparison is necessary, with the hypergraph module and model capacity kept comparable.
The paper compares only one superpixel construction choice. Since many superpixel algorithms exist and have different properties, the authors should compare SLIC with other superpixel methods or learned/differentiable superpixel approaches. Without this, it is unclear whether the gain comes from the proposed architecture, the chosen SLIC variant, or parameter tuning.
Reviewer #2 points out that important hyperparameters are not justified, including SLIC compactness, boundary sensitivity, the balance factor between semantic and spatial hyperedges, and the sampled points used in the barrier integral. Sensitivity analysis is needed.
Reviewer #1 and Reviewer #3 both raise the missing complexity analysis. This is a serious issue because the method adds a dual-encoder design, hypergraph computation, superpixel construction, and injection modules. Parameters, FLOPs, memory usage, inference time, and superpixel construction overhead should be reported.
Reviewer #3 also notes that the gains over strong baselines are not always large. Given the extra complexity, the authors need to show that the improvement is practically meaningful, not just numerically higher.
The statistical evidence is incomplete. The paper should report repeated runs or cross-validation results with mean ± standard deviation, and clearly explain the statistical test, paired unit, sample size, and correction for multiple comparisons. The related work is incomplete. The authors should discuss prior work on SLIC, differentiable superpixels/SSN, superpixel-based CNN/GCN/Transformer models, Vision GNN, Vision HGNN, AHGNN, G-CASCADE, and existing co-attention/channel-fusion modules such as those mentioned by Reviewer #2. The presentation should remove overclaiming. Terms such as “semantic purity,” “strict alignment,” and “eliminate semantic ambiguity” are not supported by the current evidence. The manuscript should use more precise and less absolute wording.
Overall, the paper has major issues in novelty, methodological justification, and experimental completeness. The rebuttal should specifically address whether superpixels are truly necessary, whether they are more reliable than patches, whether the added complexity is justified, and whether the reported gains are robust under stronger ablations and fair complexity-aware comparisons.
- 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 a coherent and empirically strong medical image segmentation framework that integrates boundary-aware superpixel node construction, semantic-spatial hypergraph reasoning, and semantic injection. While the methodological novelty is moderate and several claims are overstated, the method is relevant, technically reasonable, and shows consistent improvements across multiple datasets and baselines. The rebuttal addresses key concerns regarding patch-vs-superpixel comparison, computational overhead, hyperparameter clarification, and overclaiming. Remaining issues should be handled in camera-ready revision, including toned-down claims, clearer statistical reporting, complexity analysis, limitations, and better explanation of the superpixel design.
The authors should address the following points in the camera-ready version:
1.Remove or soften “semantic purity,” “strict alignment,” and “eliminate semantic ambiguity.” 2.Add a complexity table: params, FLOPs, memory, inference time, and SLIC overhead. 3.Clearly label the patch-node vs. superpixel-node ablation in Table 2.4.Specify missing hyperparameters, especially barrier integral sampling K. 5.Add a limitation paragraph about superpixel failure cases and limited novelty. 6.Improve related work on SSN/differentiable superpixels, superpixel-based deep models, Vision GNN/HGNN, AHGNN/G-CASCADE, and channel/co-attention fusion.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The rebuttal clarifies several concerns, including the patch-node versus superpixel-node comparison and computational overhead, but the core novelty remains limited because the method largely combines superpixels, hypergraph reasoning, and channel-style feature injection. Given the extra complexity and remaining concerns about superpixel reliability, hyperparameter sensitivity, and modest gains over strong baselines, the current evidence is not sufficient to establish a clear methodological advantage.
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.
This paper received mixed reviews, mainly due to concerns about novelty, superpixel justification, and added complexity. After considering the rebuttal and updated reviews, I recommend acceptance.
The paper presents a coherent superpixel-hypergraph segmentation framework, and its main strength lies in using boundary-aware superpixel nodes to support high-order semantic-spatial reasoning. The experiments cover five medical imaging modalities, and the ablation results, including the patch-node versus superpixel-node comparison, provide reasonable evidence that the proposed design is effective.
The rebuttal addresses several important concerns by clarifying the role of superpixels, reporting complexity and inference cost, explaining key hyperparameters, and softening several over-strong claims. For the camera-ready version, the authors should incorporate these clarifications, avoid absolute wording such as “semantic purity” or “strict alignment,” report the complexity metrics, clarify the statistical testing protocol, and improve the related work and writing quality.