List of Papers Browse by Subject Areas Author List
Abstract
Accurate segmentation of aortic dissection (AD) lumens in CT angiography (CTA) is essential for quantitative morphological assessment and clinical decision-making. However, reliable 3D delineation remains challenging due to limited long-range context modeling, which compromises inter-slice coherence, and insufficient structural discrimination under low-contrast conditions. To address these limitations, we propose BiM-GeoAttn-Net, a lightweight framework that integrates linear-time depth-wise state-space modeling with geometry-aware vessel refinement. Our approach features a Bidirectional Depth Mamba (BiM) module to efficiently capture cross-slice dependencies and a Geometry-Aware Vessel Attention (GeoAttn) module that combines orientation-sensitive anisotropic filtering with lightweight attention for vessel refinement. Experiments on the current multi-source AD CTA dataset show that BiM-GeoAttn-Net achieves a Dice score of 93.35% and an HD95 of 12.36mm, outperforming representative CNN-, Transformer-, and SSM-based baselines in overlap metrics while maintaining competitive HD95. These results suggest that coupling linear-time depth modeling with geometry-aware refinement provides an effective and computationally efficient strategy for 3D AD segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6549_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/bbbwang/BiM_GeoAttn_Net.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaYua_BiMGeoAttnNet_MICCAI2026,
author = { Zhang, Yuan AND Liu, Lei AND Zhang, Jialin AND Zhang, Ya-Nan AND Wang, Ling AND Mu, Nan},
title = { { BiM-GeoAttn-Net: Linear-Time Depth Modeling with Geometry-Aware Attention for 3D Aortic Dissection CTA 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
The paper presents a mamba-based segmentation model for 3d segmentation of aortic dissections in CTA images. In contrast to other mamba-derived models, the proposed one uses orientation-sensitive anisotropic filtering in geometry aware vessel attention module. Compared to Unet, trasnfomer and mamba based models, the presented architecture is shown to produce better segmentation performance with smaller requierements for training time and memory. The ablation study shows that each component in the proposed model contributes positively to overall segmentation 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 paper presents a novel architecture for 3D aortic dissection segmentation in CTA images. The design of the model considers anatomical properties of the organ. Ablation study is provided to see part-based performance improvement
- 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 performance of the proposed method is evaluated at a single dataset.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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?
Although the method is validated on a single dataset, the design of the model respects the geometry of the anatomy, which in turn generates better segmentation performance
- Reviewer confidence
Very confident (4)
- [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.
Although the reasoning of using directional attention on vessels is well anatomically motivated, the superiority of the performance of the proposed method should be shown on more datasets. Also, performance comparison regarding directional attention realised with existing methods should be reported.
Review #2
- Please describe the contribution of the paper
The paper proposes BiM-GeoAttn-Net, a lightweight architecture for 3D aortic dissection CTA segmentation. The proposed Bidirectional Depth Mamba (BiM) module is designed to achieve long-range modeling with linear complexity in 3D segmentation, while the Geometry-Aware Vessel Attention (GeoAttn) module is used to refine vascular details. Experiments on the adopted CTA dataset show that the method achieves better performance than the baseline methods with a relatively low parameter count.
- 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 is generally well organized, and the proposed method is clearly designed around the main challenges of 3D aortic dissection CTA segmentation, including insufficient inter-slice continuity along the depth direction and ambiguous boundaries. To address these issues, the authors introduce the BiM and GeoAttn modules in a task-oriented manner. The BiM module uses 1D sequences formed along the depth axis together with bidirectional Mamba to achieve near-linear-complexity long-range modeling and improve slice-wise consistency. The GeoAttn module employs orientation-sensitive convolutions and subsequent attention operations to enhance boundary and detail information. The reported experiments on the adopted dataset suggest that the method is effective while maintaining a relatively lightweight parameter count.
- 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 novelty of the GeoAttn module appears insufficient. The idea of using anisotropic convolutions to process features was already discussed in earlier works such as AH-Net [1], and is a relatively mature design idea. Similarly, the idea of using spatial attention for residual enhancement, followed by channel attention through global average pooling / max pooling + FC, has also been discussed in works such as CBAM [2]. This makes the module look more like a combination of existing methods. If the authors could explain more clearly what is genuinely novel in GeoAttn, it would help improve the readability of the paper. 2.The number of baselines is limited, and the most recent compared method seem to be from 2024.All selected baselines are generic medical image segmentation methods. More recent medical image segmentation methods from 2025 should be included for comparison, especially more task-relevant works such as DB-SNet [3] for vascular segmentation. 3.The analysis of HD95 is insufficient given the stated motivation of the method. Although the proposed method achieves the best Dice and IoU, it does not consistently obtain the best HD95.In the reported table, the HD95 value of 12.36 is only the third best. For vascular or aortic segmentation, HD95 is an important metric because it better reflects boundary accuracy and structural continuity than overlap-based metrics alone. This point is particularly relevant here, since the paper explicitly motivates BiM and GeoAttn as modules for improving inter-slice consistency and boundary/detail refinement. The authors should provide a clear analysis and explanation of this point. 4.The ablation study for BiM and GeoAttn could be more fine-grained. The current ablation is not sufficient to fully demonstrate the contribution of the BiM and GeoAttn design. For BiM, I think the paper should include experimental results with single-directional scanning to demonstrate the rationality and effectiveness of the proposed method. For GeoAttn, it is necessary to separately evaluate the contribution of the anisotropic multi-branch convolutions and the contribution of the attention branches based on global average pooling / global max pooling, rather than only reporting the effect of the module as a whole. References [1] Liu, S., Xu, D., Zhou, S.K., Mertelmeier, T., Wicklein, J., Jerebko, A., Grbic, S., Pauly, O., Cai, W., Comaniciu, D.: 3D anisotropic hybrid network: Transferring convolutional features from 2D images to 3D anisotropic volumes. In: Frangi, A.F., Schnabel, J.A., Davatzikos, C., Alberola-López, C., Fichtinger, G. (eds.) Medical Image Computing and Computer Assisted Intervention – MICCAI 2018.LNCS, vol. 11071, pp. 851–858.Springer, Cham (2018). [2] Woo, S., Park, J., Lee, J.-Y., Kweon, I.S.: CBAM: Convolutional block attention module. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) Computer Vision – ECCV 2018.LNCS, vol. 11211, pp. 3–19.Springer, Cham (2018). [3] Yang, M., Lyu, J., Hu, J., Bian, X., Zhang, Y., Su, S., Lou, X.: DB-SNet: A dual branch network for aortic component segmentation and lesion localization. Computerized Medical Imaging and Graphics 124, 102592 (2025).
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a lightweight framework for 3D aortic dissection CTA segmentation and reports competitive results, but I have several concerns that limit my overall score. The novelty of some modules may be insufficient and appears more like a combination of existing ideas. In addition, the baseline comparison is relatively limited and not sufficiently recent, and the ablation study could be further strengthened. More importantly, the reported results do not fully support the main motivation of the method. The paper states that BiM and GeoAttn are designed to improve inter-slice consistency and boundary/detail refinement; however, the method does not achieve the best HD95, which is an important metric for vascular or aortic segmentation because it reflects boundary accuracy and structural continuity more directly than overlap-based metrics. This inconsistency should be analyzed more clearly. For these reasons, I rated the paper as Weak Reject.
- 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 motivation and technical content of this paper are sufficiently solid. Therefore, at the rebuttal stage, I recommend accepting this paper.
Review #3
- Please describe the contribution of the paper
The article presents a variant of mamba architectures for the segmentation of aortic dissection lumens from CTA images. The main contributions are a bidirectional module that captures the context between slices and a geometry aware attention to better delineate boundaries and the target structures.
While the methodology is sufficiently detailed, the background section requires further development to better contextualize current state-of-the-art methods for the target segmentation task. Furthermore, although the experimental results demonstrate the proposed method’s efficacy, the inclusion of a formal statistical analysis would significantly strengthen the presentation and validate the findings.
- 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 address the challenges of aortic dissection lumen segmentation, specifically focusing on low-contrast regions and indistinct boundaries.
- The reported results indicate superior performance over multiple state-of-the-art methods, encompassing Transformer, U-Net, and Mamba architectures.
- The paper effectively combines state-of-the-art and very recent methodologies, a mamba based architecture with geometry based attention.
- 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.
- In the Introduction, the review of related works concerning aortic dissection (AD) lumen segmentation in CT imaging is insufficiently presented. The authors should incorporate additional literature to better contextualize the task and provide a more comprehensive bibliography.
- Furthermore, explicitly articulating the advantages of the proposed model over these existing methods would better highlight the manuscript’s contributions.
- Given that the results are currently derived from a single data split, performing statistical significance testing (e.g., a t-test) is necessary to substantiate the performance claims and strengthen the presentation of the results
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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
Equations 1-2 include “Rev” which is not defined. or “IN” in equation 7.Please recheck equations with their definitions. global max pooling GMP etc The authors should consider releasing their code publicly to support the transparency and reproducibility of their findings.
- 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?
Overall, this paper presents a compelling approach to the challenging task of aortic dissection lumen segmentation by integrating a Mamba-based architecture with attention mechanisms. The methodology is sound, and the initial results show promising improvements over established benchmarks like U-Nets and Transformers. However, the manuscript requires further refinement in its contextual positioning and statistical rigor. Specifically, the introduction needs a more comprehensive literature review to properly frame the contribution, and the results must be supported by statistical significance testing to ensure the robustness of the comparative findings. Addressing these concerns will significantly strengthen the paper’s impact.
- 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 integration of a Mamba-based architecture with attention mechanisms remains a sound and compelling approach that yields promising improvements over established benchmarks for aortic dissection lumen segmentation. However, some minor limitations persist from the initial submission, specifically the overall literature review in the introduction remains somewhat limited, and the results would still benefit from more comprehensive statistical significance testing to fully support the robustness of the comparative findings. Despite these remaining weak points, the core methodological contributions and the overall refinement of the paper are sufficient to justify its publication.
Author Feedback
We sincerely thank the reviewers and meta-reviewer for their careful evaluation and constructive feedback. We appreciate the positive assessment of the clinical relevance, anatomy-aware design, and computational efficiency of our framework. We further clarify several important points regarding the methodological contribution, experimental interpretation, and positioning of the paper.
Reviewer #1: Generalization and dataset setting
We agree that generalization is important for AD CTA segmentation. Although publicly available expert-annotated 3D AD CTA datasets remain limited due to the complexity of TL/FL annotation and the clinical urgency of Type-B AD, our experiments were conducted on a curated multi-source dataset composed of two independent public cohorts with strict patient-level splitting to avoid data leakage. Despite this challenging setting, BiM-GeoAttn-Net consistently outperforms representative CNN-, Transformer-, and Mamba-based baselines while maintaining competitive HD95 and lightweight computational cost.
Reviewer #2: Methodological novelty and experimental interpretation
We appreciate the reviewer’s comments regarding GeoAttn. We agree that anisotropic convolutions and attention mechanisms have been explored in prior works such as AH-Net and CBAM. Our contribution lies not in introducing new primitive operators, but in the anato-my-aware integration of geometry-guided refinement for AD CTA segmentation. GeoAttn employs plane-specific anisotropic filtering to target ambiguous vessel boundaries under anisotropic volumetric spacing. Positioned after BiM, it complements depth-wise global coherence with local geometry-aware refinement, enabling the two modules to jointly address inter-slice discontinuity and boundary ambiguity.
Compared with methods emphasizing either generic volumetric modeling or local vessel enhancement, our framework explicitly couples anatomy-aligned depth-wise modeling with geometry-aware refinement, enabling improved cross-slice continuity and robust boundary delineation simultaneously. This also differentiates our framework from recent task-related methods such as DB-SNet.
We agree that HD95 is highly relevant for vascular segmentation. Although Attention U-Net achieves the lowest HD95, its Dice and Recall are substantially lower than our method, indicating more missed vessel regions and fragmented predictions. Since HD95 measures only the 95th percentile boundary distance, incomplete predictions may occasionally obtain favorable HD95 values despite inferior volumetric continuity. In contrast, our method achieves the best Dice, IoU, Recall, and Precision while maintaining competitive HD95, suggesting a stronger balance between overlap accuracy, structural continuity, and boundary fidelity.
The current ablation already isolates the effects of BiM, GeoAttn, and their combination. The results indicate that BiM mainly improves depth-wise coherence, while GeoAttn enhances geometry-aware boundary refinement. Their combination yields the strongest overall performance, supporting the complementary roles of the two modules.
Reviewer #3: Related work and experimental support
We appreciate the reviewer’s suggestions regarding related work positioning and experimental support. The paper positions the proposed framework relative to CNN-, Transformer-, and SSM/Mamba-based AD CTA segmentation methods, highlighting the motivation of using anatomy-aware depth-wise modeling and geometry-guided refinement to address depth-wise discontinuity and ambiguous vessel boundaries.
The experimental protocol adopts strict patient-level splitting and multi-source cohort construction, reducing potential data leakage and supporting the reliability of the reported improvements across quantitative metrics and ablation settings.
Finally, we thank the reviewer for identifying unclear notations. Definitions for “Rev”, “IN”, “GAP”, and “GMP” have been clarified to improve readability and consistency.
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.
This paper presents a method for 3D aortic dissection segmentation in CTA, with modules designed to address slice-wise continuity and boundary ambiguity. Reviewers appreciated that the architecture is tailored to the anatomical characteristics of the task and found the overall method generally well organized. The reported segmentation performance is promising, and the inclusion of an ablation study was viewed positively.
At the same time, the reviewers raised several concerns that should be addressed in the rebuttal. The main issues relate to the strength of the novelty claims and the scope of the evaluation. In particular, reviewers noted that parts of the GeoAttn design appear closely related to existing anisotropic convolution and attention modules, and asked for clearer positioning of the contribution relative to prior work in aortic dissection and vascular segmentation. The experimental validation is also limited to a single dataset and apparently a single split, which weakens the evidence for generalization and calls for more cautious claims. Reviewers further suggested stronger statistical support, such as significance testing, and more fine-grained ablations to better isolate the contributions of the BiM and GeoAttn subcomponents.
Another point raised was that the current empirical results do not fully support all aspects of the motivation, particularly regarding HD95, where the method does not consistently achieve the strongest performance despite emphasizing boundary refinement and structural continuity. Clarifying this discrepancy would strengthen the paper.
Overall, the paper addresses a relevant problem and shows promising results, but the current reviews indicate that questions remain regarding novelty, evaluation depth, and experimental support. I encourage the authors to use the rebuttal to address these issues clearly.
- 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 proposes a 3D aortic dissection segmentation method in CTA with modules targeting slice-wise continuity and boundary ambiguity. The anatomical motivation is clear, the Mamba-attention architecture is well tailored to the vessel structure, and the reported performance is promising.
After rebuttal, two reviewers (R2 and R3) explicitly recommend acceptance, finding the methodological contributions sound and the improvements over benchmarks convincing. Reviewer #1 raises valid concerns about single-dataset evaluation and missing comparisons with existing directional attention methods, but these are limitations that can be acknowledged rather than disqualifying flaws. The overall technical content and ablation evidence are sufficient to justify publication.
I recommend accept. For the camera-ready version, the authors must: (1) add statistical significance testing for the main comparative results; (2) expand the related work discussion in the introduction to better position the directional attention design relative to existing anisotropic convolution and attention modules; (3) explicitly discuss the single-dataset limitation and the need for future multi-center validation; and (4) address the HD95 discrepancy by clarifying when boundary refinement helps and when it does not.
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.
After weighing the post-rebuttal reviews, I recommend acceptance. Two reviewers supported the paper after rebuttal, and I agree that the remaining concerns are valid but not strong enough to outweigh the paper’s focused contribution.
The paper addresses a clinically relevant 3D CTA segmentation problem and proposes a task-oriented design: BiM models depth-wise continuity across slices, while GeoAttn introduces geometry-aware local refinement for vessel boundaries. The experiments show clear improvements in Dice, IoU, Recall, and Precision over representative CNN-, Transformer-, and Mamba-based baselines, with competitive training cost. The ablation study also supports the complementary roles of BiM and GeoAttn. The rebuttal reasonably clarified that the main contribution lies in anatomy-aware integration rather than entirely new primitive operators.
Some limitations should still be acknowledged. The evaluation remains relatively small and based on one patient-level split, GeoAttn reuses ideas related to anisotropic filtering and attention mechanisms, and the method does not achieve the best HD95 despite its boundary-refinement motivation. For the camera-ready version, the authors should temper claims on boundary superiority and generalization, incorporate the rebuttal clarifications on the multi-source dataset and HD95 interpretation, better position GeoAttn relative to prior anisotropic/attention modules, and fix the unclear notation issues noted by the reviewers.
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 presents a compelling approach to the challenging task of aortic dissection lumen segmentation by integrating a Mamba-based architecture with attention mechanisms. The motivation and technical content are sufficiently reasonable and have some appeal, so it is recommended to accept it.
