List of Papers Browse by Subject Areas Author List
Abstract
Medical image segmentation is critical for disease monitoring and treatment planning, with U-Net serving as a fundamental architecture in this domain. Building on its effectiveness, recent advancements have integrated mechanisms such as Transformers and Mamba into U-shaped structures, achieving promising results. However, these methods typically operate in Euclidean space, limiting their ability to capture complex geometric structures. Additionally, they often apply uniform attention strategies across scales, failing to appropriately model fine-grained spatial details in shallow layers and high-level semantic dependencies in deeper layers. To address these limitations, we propose a novel framework, Hierarchical Hyperbolic Self-Attention Network (H$^2$SA-Net), which integrates geometric representation learning with a dual-granularity attention mechanism. Our key contribution is a depth-adaptive stratified Hyperbolic Self-Attention (HSA) module. In shallow layers, pixel-wise HSA refines local spatial context, while in deeper layers, channel-wise HSA enhances global semantic correlations. To maintain geometric consistency, we employ the Einstein Midpoint for feature aggregation. Enhanced by a wavelet-based backbone and a Hyperbolic Multinomial Logistic Regression (MLR) head, the proposed framework effectively balances local texture preservation with global semantic understanding. Extensive experimental results on diverse datasets demonstrate the effectiveness of our method.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1498_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jjscc/H2SA-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JinSha_Hierarchical_MICCAI2026,
author = { Jin, Shaocheng AND Wang, Rui AND Hu, Chen AND Xu, Tianyang AND Wu, Xiao-Jun AND Zhou, Tao},
title = { { Hierarchical Hyperbolic Self-Attention Network 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
The author propose H2SA Net, a UNet variant that replaces standard self attention with hyperbolic self attention operating in the Poincare ball. . A stratified module applies pixel-wise attention in shallow layers and channel-wise attention in deeper layers. A hyperbolic multinomial logistic regression head performs final classification. Writing is good. Presentation is clear.
- 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.
hyperbolic math clean and well presented. Ablations study curvature and stratification. Evaluated three distinct tasks: polyp, thyroid nodule, brain tumor.
- 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.
link between anatomical hierarchy and hyperbolic feature space seems quite asserted . . not empirically validated. Gains over Euclid baselines look marginal and untested for significance. Also missing nnUNet. 2D only. but the most important weakness is that it is only a tiny change in network architecture to answer a very old established question. Novelty of work thus in question.
- 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?
theory seems sound but the experiment valid look thin. Marginal gains, missing key baselines, and lack proper statistics. seem an incremental rather than compelling contribution.
- 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 addressed the major empirical gaps competently. The authors added nnU Net comparison, statistical significance testing across three random seeds, efficiency analysis showing minimal overhead, and Euclidean ablations isolating the hyperbolic contribution. The brain MRI patient level split was clarified. While the core contribution remains incremental, the experimental validation is now sufficiently complete to support the claims.
Review #2
- Please describe the contribution of the paper
The paper proposes H2SA-Net, which combines a wavelet-based U-shaped backbone with hyperbolic self-attention and a hyperbolic MLR segmentation head. The main design choice is to use pixel-wise HSA in shallower layers and channel-wise HSA in deeper layers. Experiments on polyp, thyroid, and brain tumor segmentation show consistent gains over the listed baselines.
- 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 problem is relevant and well motivated. 2.The method is fairly coherent, and the depth-adaptive attention design is reasonable. 3.The experiments are broader than average for this format, with multiple datasets and ablations. 4.The reported gains are consistent across the benchmarks.
- 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 paper combines several ingredients that are already present in the surrounding literature, including wavelet-based segmentation, hyperbolic representation learning, hyperbolic attention, and hyperbolic MLR. The clearest new element is the depth-dependent use of pixel-wise versus channel-wise hyperbolic attention, but the manuscript does not yet make a strong enough case that this is a sufficiently substantial methodological advance for acceptance. 2.The paper does not fully isolate the source of the gains. The ablation table is useful, but it mainly compares the proposed design against Euclidean attention or a Euclidean classification head within the same framework. This still leaves open how much of the improvement comes from the hyperbolic formulation itself versus the broader combination of wavelet backbone, attention placement, and overall architecture. This matters because the paper’s central claim is about the value of geometry-aware modeling. 3.The experimental evidence is promising but still a bit incomplete for a method paper. The manuscript does not report variance across runs, confidence intervals, or statistical testing. Some improvements over the strongest related baseline are not very large, so it is hard to judge robustness. In addition, the brain MRI experiment is described using image counts, but the split is not stated clearly enough at the patient level, which is important for ruling out possible leakage in slice-based evaluation.
- 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?
NA
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The paper would be stronger with a sharper explanation of what is genuinely new relative to prior hyperbolic segmentation and hyperbolic attention work. I would also encourage the authors to clarify the brain MRI split at the patient level, report multi-run variance or another measure of stability, and include a small efficiency summary for H2SA-Net itself. These additions would make the empirical case more convincing.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I lean weak reject at this stage. The paper addresses a relevant problem, and the empirical section is reasonably broad. However, the main technical distinction from prior work is still not sharp enough, and the current experiments do not fully isolate the claimed benefit of the hyperbolic design. In addition, the lack of stronger robustness reporting and clearer split details makes the evidence less convincing than I would like for a methodology paper. Overall, I think this is a solid submission, but it is currently below the acceptance threshold for me.
- 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 sufficiently addresses my main concern about attributing the gains to the hyperbolic design, especially through the Euclidean-counterpart ablation and multi-seed results. I still encourage clearer reporting of the implementation and statistical details in the final version, but I no longer see these as blocking issues and support acceptance.
Review #3
- Please describe the contribution of the paper
This paper proposes H^2 SA-Net, a novel medical image segmentation framework that integrates wavelet-based encoding with hyperbolic geometry to better capture hierarchical anatomical structures. They ratified the hyperbolic self-attention module, which applies pixel-wise attention to shallow features and channel-wise attention to deep features.
- 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.
Hyperbolic Space has been widly apply to many vision language task and Computer Vision area. Howevre, it seems does not apply to segemtation and this is a good application. The author also showed a promising result in a wide range of 2D dataset.
- 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] It seems the experiemtns are mainly conducted on 2D datasets and what is the perfromacne in 3D dataset? [2] If applying hyperbolic space, it may increase the trainning time and also increase the number of parameters. How is the efficiency of the method. [3] To my knowledge, hierarchical non-Euclidean structure usually apply to classfication task but may not be true to segmentation task. Can the author provide more explain or profice some visulisation?
- 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.
(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?
Please check my weaknesses 1-3. Particularly point 3, it would be more convincing to provide some explanation so that the reader can better understand the motivations.
- 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 author addressed the majority of my questions. It showed that this method does not significantly increase computational cost while maintaining good performance.
Author Feedback
We thank all reviewers for their constructive comments. Below we address the major concerns.
(@R1,R2,R4) Novelty and geometry contribution: Our contribution is not a direct replacement of Euclidean operators with existing hyperbolic modules, but a depth-adaptive geometry-aware segmentation strategy: pixel-wise HSA refines shallow spatial features, while channel-wise HSA recalibrates deep semantic features. We will revise the motivation to avoid over-interpreting anatomical hierarchy. Although many anatomical patterns are hierarchical or nested, our method does not rely on a strict tree-structured label assumption. Instead, hyperbolic geometry provides a geometric prior for organizing dense features across network depths, from boundary/texture cues to region-level semantic dependencies. To isolate the hyperbolic contribution, we conduct Euclidean-counterpart ablations under the same backbone and attention placement. Replacing the hyperbolic components with Euclidean counterparts reduces Dice/IoU from 0.844/0.759 to 0.822/0.735 on Brain Tumor, and from 0.832/0.759 to 0.820/0.736 on TN3K. Table 4 of the main paper consistent gains when HSA is incorporated into TransUNet and U-Mamba, indicting that the improvement mainly comes from geometry-aware modeling rather than only from the wavelet backbone or architecture changes.
(@R1) nnU-Net comparison: We additionally evaluate nnU-Net using the same data split and preprocessing protocol. On Brain Tumor, our method achieves 0.844/0.759 Dice/IoU, compared with 0.8267/0.740 obtained by nnU-Net.
(@R1,R2) Statistical significance and MRI split: We repeat the experiments with three random seeds (42, 1234, 3407) and now report mean±std. On Brain Tumor, our method achieves 0.8443±0.0042/ 0.7592±0.0054 Dice/IoU, while the Euclidean counterpart obtains 0.8224±0.0057/0.7351±0.0066.The improvement is statistically significant (p<0.05). The MRI split is patient-level: all slices from the same patient are assigned exclusively to either training or testing, avoiding leakage in slice-based evaluation.
(@R1,R4) 3D extension: We agree that 3D validation is important. The current submission focuses on 2D benchmarks, and we will clarify this as a limitation. Methodologically, HSA is not restricted to 2D and can be extended to volumetric segmentation by replacing pixel-wise HSA with voxel-/window-wise HSA while retaining channel-wise recalibration. We will add this discussion as future work.
(@R2,R4) Efficiency: The plain Euclidean baseline requires 61.34 GFLOPs and 46.846 ms inference latency. The Euclidean variant of our design requires 67.44 GFLOPs and 47.747 ms, while the full hyperbolic model requires 67.94 GFLOPs and 51.219 ms. Thus, compared with the Euclidean variant, hyperbolic modeling introduces only 0.50 GFLOPs and 3.472 ms extra latency, while bringing consistent gains.
(@R4) Hyperbolic hierarchy for segmentation: We agree that hierarchical non-Euclidean modeling is more common in classification. However, prior segmentation works, such as Hyperbolic Image Segmentation (CVPR’22) and Hyperbolic CLIP-based Open-vocabulary Semantic Segmentation (CVPR’25) suggest that hyperbolic representations can also benefit dense prediction by modeling pixel-level decision boundaries and hierarchical semantic alignment. In our work, anatomical hierarchy is a motivation rather than a strict label-tree assumption. We use hyperbolic geometry to model the coarse-to-fine feature organization: shallow layers capture textures, edges, and boundaries, whereas deeper layers encode region-level semantic dependencies. Accordingly, stratified HSA uses pixel-wise attention for local spatial refinement and channel-wise attention for semantic recalibration.
(@R4) Reproducibility: We will provide more details, including preprocessing, data augmentation, optimizer settings, training protocols, evaluation settings, and the patient-level split for Brain Tumor. The code and trained models will be released upon acceptance.
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.
Scores are 4/3/4.The two weak accepts (R1 faculty, R4) credit a clean and well-motivated application of hyperbolic geometry to a new task, while the weak reject (R2) is grounded in the difficulty of attributing the reported gains to the hyperbolic design specifically rather than to the surrounding architecture. The paper is borderline, and the substantive concerns are addressable in rebuttal. The paper is invited for rebuttal; authors should prioritize (a) multi-run variance or significance testing on the headline numbers, (b) an ablation that explicitly isolates the hyperbolic contribution from the wavelet backbone and attention placement, (c) clarification of the patient-level brain MRI split, and (d) an efficiency summary along with discussion of the 2D-to-3D extension and the missing nnU-Net comparison.
- 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.
Pre-rebuttal, reviewers found the method technically sound with consistent gains across multiple medical segmentation tasks, but raised concerns about incremental novelty, weak baseline coverage, and insufficient statistical and ablation rigor. After rebuttal, the authors addressed key gaps by adding stronger baselines (e.g., nnU-Net), multi-seed significance testing, Euclidean ablations isolating the hyperbolic contribution, and efficiency analysis. Reviewers concluded that while the contribution remains incremental, the empirical validation is now solid and sufficient, moving from mixed weak accept / weak reject reviews pre-rebuttal to a consensus accept post-rebuttal due to strengthened empirical validation.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper is recommended for acceptance based on its solid post-rebuttal clarifications and consistent empirical performance across multi-task medical image segmentation benchmarks.
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.
three accept after rebuttal. i agree.
