List of Papers Browse by Subject Areas Author List
Abstract
Weakly supervised segmentation enables model training from plane-level labels. Existing methods often rely on 2D encoders, neglecting the volumetric nature of medical data. We propose TranSamba, a hybrid Transformer-Mamba architecture designed to capture 3D context via cross-plane modeling. TranSamba augments a Vision Transformer backbone with Cross-Plane Mamba blocks, leveraging linear-time modeling for efficient information exchange across neighboring planes. This exchange improves in-plane self-attention and subsequent attention maps for object localization. TranSamba maintains linear time complexity and constant space complexity with respect to the input volume depth. Extensive experiments on three datasets covering diverse modalities and pathologies show that TranSamba achieves state-of-the-art performance, demonstrating the generalizable efficacy of cross-plane modeling. Code is available at: https://github.com/YihengLyu/TranSamba.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4858_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/YihengLyu/TranSamba
Link to the Dataset(s)
Brain Tumor Segmentation (BraTS) dataset: http://medicaldecathlon.com/
Kidney Tumor Segmentation (KiTS) 2023 dataset: https://kits-challenge.org/kits23/
Left Atrium Segmentation Challenge (LASC) dataset: https://www.cardiacatlas.org/atriaseg2018-challenge/
BibTex
@InProceedings{LyuYih_Hybrid_MICCAI2026,
author = { Lyu, Yiheng AND Xu, Lian AND Arrow, Coen AND Bennamoun, Mohammed AND Boussaid, Farid AND Dwivedi, Girish},
title = { { Hybrid Transformer-Mamba for Weakly Supervised Volumetric Medical Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes TranSamba, a hybrid Transformer-Mamba architecture for weakly supervised volumetric medical segmentation from plane-level labels. The key idea is to overcome the limitations of purely 2D encoders by introducing a Cross-Plane Mamba (CPM) block that enables efficient information exchange across neighboring slices, while preserving the strength of Transformer class-to-patch attention for localization. The method is designed to capture 3D context with linear-time cross-plane modeling, and the paper shows improved segmentation/localization performance on three datasets spanning different organs and modalities.
- 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.
Well-motivated problem and useful technical idea. The paper addresses an important limitation in weakly supervised volumetric segmentation: many existing methods still rely on slice-wise 2D encoders and therefore fail to fully exploit volumetric context. The introduction of cross-plane modeling is a sensible and relevant way to address this gap.
Strong experimental evaluation. The experiments cover three datasets and include ablations on cross-plane modeling, the number of planes, and the hybrid layer design, along with both quantitative and qualitative analyses. The object-size analysis is also a valuable addition beyond standard benchmark reporting.
Consistent empirical gains. TranSamba achieves the best reported performance on all three test datasets. The ablation results further suggest that these improvements are driven by cross-plane modeling rather than simply by increased model complexity.
Reproducibility. The availability of code and the detailed experimental settings further support reproducibility.
- 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 novelty is solid but somewhat incremental, since hybrid Transformer–Mamba backbones are already an active direction and the main contribution is the specific cross-plane design for weakly supervised volumetric localization rather than a fundamentally new paradigm. The empirical comparison could be more complete, as Table 3 mainly includes CAM-style and earlier baselines, while the paper’s own references suggest additional recent relevant methods that might also merit discussion or comparison. In addition, the study is limited to binary segmentation, which restricts the demonstrated generality of the approach.
- 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.
(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?
I am leaning weak accept because the paper addresses a real and important limitation in weakly supervised volumetric segmentation, proposes a technically coherent solution, and demonstrates consistent improvements across multiple datasets. The method is easy to understand, well ablated, and practically relevant for scenarios where voxel-wise annotation is expensive. However, I am not rating it higher because the novelty feels more architectural/incremental than fundamental, the SOTA comparison set appears incomplete for such a strong claim, and the demonstrated scope is still limited to binary segmentation with weaker performance on small objects. Overall, I think the paper is above threshold.
- 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.
I maintain my original weak accept. The rebuttal clarifies the novelty and scope, especially that the main contribution is cross-plane Mamba modeling for weakly supervised volumetric localization. This partially addresses my novelty concern. However, some limitations remain, including incomplete comparison with recent baselines and binary-only evaluation. Overall, the method is coherent, well ablated, and shows consistent gains.
Review #2
- Please describe the contribution of the paper
This paper addresses weakly supervised volumetric medical segmentation from plane-level labels, where most prior methods rely on 2D encoders and do not explicitly model cross-plane context. The authors propose TranSamba, a hybrid Transformer-Mamba architecture that inserts a Cross-Plane Mamba (CPM) block before each Transformer block in a ViT encoder, aiming to enrich patch tokens through efficient cross-plane information exchange while preserving effective in-plane class-to-patch attention for localization. On three datasets spanning different modalities and targets (BraTS, KiTS, and LASC), the method reports state-of-the-art results over a range of CAM-based weakly supervised baselines, and the paper further provides ablations on cross-plane modeling, number of planes, and hybrid layer design.
- 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 proposed architecture is technically coherent and well aligned with the task. The paper explains clearly why full cross-plane self-attention would be computationally expensive and why the CPM + Transformer series arrangement is intended to preserve attention-map quality while incorporating 3D context.
2.The paper includes solid ablations that support the core design choice. The comparison between V1 (ViT only), V2 (in-plane Mamba), and V3 (cross-plane Mamba + Transformer) is useful. It helps separate the gain from merely increasing model complexity versus specifically adding cross-plane modeling. The plane-number and hybrid-layer-design ablations are also informative.
- 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.With the advent of Segment Anything Model 2 (SAM 2) and its medical scenario adaptation, high-quality segmentation can be achieved via minimal interaction (e.g., a few points or boxes). While the “plane-level labels” used in this paper are cheaper than full masks, they still require categorical labeling for each slice. Given that SAM-like models can generalize from prompts on only a few key slices, the paper fails to justify why a complex weakly supervised training pipeline is preferable to foundation model prompting or fine-tuning, especially when the total annotation effort might be comparable.
2.The methodological novelty is somewhat incremental relative to recent hybrid Mamba-Transformer literature. The paper’s contribution is meaningful, but it should be positioned more carefully relative to recent hybrid Mamba-Transformer architectures and medical Mamba segmentation work already cited by the paper, such as MambaVision [16], MAP [24], U-Mamba [26], SegMamba [42], and PathMamba [12]. The main novelty here is not the existence of a hybrid Mamba-Transformer backbone, but rather the specific use of cross-plane Mamba to support weakly supervised volumetric localization via ViT attention. That is still useful, but more incremental than the paper sometimes suggests.
3.The study is limited to binary segmentation settings. All experiments are formulated as binary segmentation tasks, even on datasets that are naturally more complex. The authors acknowledge that multi-class segmentation requires additional techniques, but this still narrows the practical scope of the contribution and makes the general claim of “weakly supervised volumetric medical segmentation” somewhat broader than the actual evidence.
- 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.
(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 gave this paper a weak reject because, although the method is technically sound and the experimental study is generally careful, I do not find the overall contribution strong enough for acceptance in its current form. The main positive aspect is that the proposed design is coherent: the paper makes a reasonable case for combining cross-plane Mamba with in-plane Transformer attention, and the ablation studies provide useful evidence that the gain comes from the intended architectural choice rather than simply increased model complexity. The empirical improvements over prior CAM-based weakly supervised baselines are also encouraging.
However, several concerns limit my enthusiasm. First, the paper does not sufficiently justify why this weakly supervised formulation is the most compelling practical direction in the current landscape, especially given the rapid progress of SAM-like foundation models and prompt-based medical segmentation, which may offer a more annotation-efficient alternative in practice. Second, the methodological novelty feels somewhat incremental: the main contribution is a task-specific adaptation of the now-common hybrid Mamba-Transformer design, rather than a fundamentally new formulation. Third, the evidence is restricted to binary segmentation tasks, which makes the scope of the claims broader than what is actually demonstrated.
Overall, I think this is a solid and well-executed paper with some interesting ideas, but the combination of limited novelty, insufficient positioning against foundation-model-based alternatives, and restricted experimental scope prevents me from recommending acceptance.
- 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 authors’ response has helped clarify some of my initial concerns. While the analysis of annotation costs concerning foundation-model-driven supervision remains unaddressed, I recognize the validity of the research motivation and the rigor of the experimental results. Given these strengths, I have decided to increase my rating from Weak Reject to Weak Accept.
Review #3
- Please describe the contribution of the paper
This paper proposes TranSamba, a hybrid Transformer–Mamba architecture for weakly supervised volumetric medical image segmentation using plane-level labels. The key idea is to introduce cross-plane modeling via Mamba blocks (CPM) to enhance token representations across neighboring slices, while retaining Transformer-based in-plane self-attention. The method generates segmentation predictions by aggregating class-to-patch (C2P) attention maps. Experiments on BraTS, KiTS, and LASC datasets show improved performance over several CAM-based 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.
This paper presents a technically sound approach for weakly supervised volumetric medical image segmentation. The idea of introducing cross-plane modeling to capture 3D context is reasonable and relevant, especially given the limitations of slice-wise processing in existing methods. The proposed hybrid Transformer–Mamba architecture is clearly designed, and the use of Mamba enables efficient cross-plane interaction with favorable computational complexity.
- 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 paper has several important weaknesses. The method essentially follows a CAM-based paradigm, where attention maps are used for localization, and can be viewed as an improved CAM generation approach. The motivation is also somewhat generic, as the limitation of 2D encoders for volumetric data is already well known and not deeply analyzed in the context of weak supervision. Moreover, several baseline results are extremely low on certain datasets; even though the proposed method shows large relative improvements, the absolute performance remains at a low level, which limits clinical relevance and raises concerns about the validity of the evaluation
- 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?
I rate this paper as weak reject (3/6). In summary, while the paper presents a technically sound and well-engineered improvement with reasonable empirical gains, the limited novelty and concerns about experimental results reduce its overall impact.
- 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 thank the AC and reviewers and address following points:
P1: Novelty against hybrid/medical Mamba (AC-a/R1/R2/R3) Hybrid Mamba: we do not claim novelty against hybrid Mamba for hybrid design in general. Rather than stacking blocks, we decompose 3D modelling into in-plane and cross-plane components and assign them different mechanisms, motivated by their computing complexities. Transformer (quadratic) correlates tokens within individual slices (in-plane) and Mamba (linear) correlates tokens across neighboring slices (cross-plane, prohibitively expensive for using Transformer). Tab2b shows the gain is not from merely inserting Mamba but from cross-plane modelling. MambaVision (general 2D images)/MAP (pre-training strategy) uses both mechanisms to process the same set of tokens; neither addresses 3D localization nor medical tasks. Medical Mamba: U-Mamba/SegMamba needs voxel labels, the exact bottleneck we tackle. PathMamba’s implementation (2D pathology; not hybrid Transformer-Mamba) remains unavailable. Summary: we claim novelty in 1.New architecture, 2.New way to process 3D data with Transformer and Mamba, 3.3D medical Weakly Supervised Semantic Segmentation (WSSS).
P2: Positioning against SAMs (AC-b/R2) WSSS and SAMs have 1.Different supervision regimes; incomparable annotation budgets, 2.Different inference style, 3.Complementary. SAMs need voxel labels for training and prompts for inference. WSSS needs slice labels (often available in clinics, e.g., radiology reports state the top and bottom bounding slices of a target; any slice within/beyond the two can be inferred as positive/negative) and has automated inference. WSSS and SAM complement each other: WSSS automatically produces localizations to serve as prompts for SAM, removing SAM’s need for user input while improving WSSS absolute performance (P3). We will explore TranSamba + SAM in future work.
P3: Low absolute performance (AC-c/R3) Our contribution improves WSSS performance, not maximizes absolute metrics, which inherently trail fully supervised models. We achieved SOTA WSSS performance. Absolute performance depends on datasets’ intrinsic difficulty, e.g., object size (voxel count), dataset scale (scan count). KiTS contains many tiny objects; controlling for object size, it matches BraTS (Fig2b/Tab4). LASC suffers from fewer training scans (n=43; but suffices for evaluation, exceeding MSD Heart n=20). For clinical relevance, SAM can augment TranSamba (P2); more scans can be added for training; both extensions are orthogonal to our contribution. We will open-source our code and trained weights.
P4: Beyond binary segmentation (AC-c/R1/R2) Multi-class medical WSSS is troublesome [9]; e.g., multiple tumor regions always co-occur on the same slice, making them indistinguishable via slice labels alone. Existing methods model region correlations via target-specific losses [6,10], which do not generalize across datasets. We proposed cross-plane modelling as a generalizable solution for CT/MRI and scoped the paper to binary WSSS to isolate and evaluate our contributions. TranSamba naturally scales to multi-class settings by allocating one class token per category [44]; the cross-plane modelling math formulation between patch tokens is unchanged; target-specific losses can still augment it for specific tasks. We designated multi-class extensions for future work.
P5: More baselines (AC-d/R1) Baselines we considered to be task-agnostic are selected for comparison in the paper, including CAM-based methods, and general/medical WSSS. In future work, we will extend to multi-class WSSS for comparison with more baselines using target-specific losses for specific datasets (P4).
P6: 2D encoders for 3D WSSS (R3) For WSSS, 2D encoders must be used to make predictions aligning with slice labels. Because of this reliance, prior work largely neglected the 3D nature of CT/MRI; our work offers one of the first solutions for bridging 2D encoder with 3D modelling in WSSS.
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 4/3/3, with the senior reviewer recommending weak reject on novelty and absolute-performance grounds, and the second weak reject (R2) driven primarily by foundation-model positioning. The paper is borderline-leaning-negative but the substantive concerns are largely addressable through better positioning, additional comparisons, and clarified scope rather than redesign. The paper is invited for rebuttal; authors should prioritize (a) sharper positioning of the cross-plane Mamba contribution against the cited hybrid Transformer–Mamba and medical Mamba works, (b) at least a discussion (and ideally an experiment) comparing against SAM 2-style prompt-based segmentation under matched annotation budgets, (c) an explanation for the low absolute performance of baselines on certain datasets and any preliminary evidence that the approach extends beyond binary segmentation, and (d) more recent CAM-based or attention-based weakly supervised baselines to strengthen the SOTA claim.
- 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.
Reviewers agreed that TranSamba addresses an important problem in weakly supervised volumetric medical segmentation and presents a technically sound architecture that combines cross-plane Mamba modeling with Transformer-based localization. Strengths included consistent performance gains across three datasets, thorough ablations, and strong reproducibility. The main concerns centered on the incremental nature of the novelty, as hybrid Transformer–Mamba architectures are already an active research direction, limited comparisons with recent methods and foundation-model-based alternatives, the restriction to binary segmentation tasks, and questions about the practical advantage of weak supervision relative to prompt-based foundation models. These concerns led to one Weak Accept and two Weak Reject scores before rebuttal. After rebuttal, reviewers acknowledged that the authors clarified the paper’s primary novelty—cross-plane Mamba modeling for weakly supervised volumetric localization—and better positioned the contribution relative to prior work, leading both initial Weak Reject reviewers to raise their ratings to Accept/Weak Accept. While limitations regarding baseline coverage, binary-only evaluation, and foundation-model comparisons remain, reviewers ultimately viewed the method as well-motivated, carefully evaluated, leading to an overall marginal acceptance recommendation.
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.
This submission is a borderline case. Initial reviews were mixed, whereas authors responses seem to have satisfied two of the reviewers. After reading reviewers concerns and authors rebuttal, I find that several of the raised points have been addressed. Nevertheless, other important concerns remain only partially resolved, with the submission still presenting several limitations. In particular, I personally found the technical novelty rather limited. Authors response partially clarifies their novelty claims, but the justification remains only moderately convincing. Furthermore, while I agree that MICCAI policy refrains from introducing new experimental results, the empirical validation in the current paper is not convincing either, with old baselines used for comparison (the newest baseline is from 2024).
Given these comments, and based on the current submitted version and rebuttal, I recommend the rejection of this work.
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.
The paper addresses an important problem in weakly supervised volumetric medical segmentation, where plane-level labels are more practical than voxel-wise annotations but existing 2D methods often fail to use volumetric context effectively. The proposed TranSamba is technically coherent: it uses cross-plane Mamba modeling to exchange information across neighboring slices while retaining Transformer attention for in-plane localization. The experimental results are consistent across three datasets, and the ablation studies reasonably support the main design choice, especially that the gain comes from cross-plane modeling rather than simply adding model complexity. Some limitations remain. The novelty is more architectural than fundamental, the comparison with recent foundation-model or SAM-style approaches is still not fully developed, and the current evaluation is limited to binary segmentation. However, the rebuttal clarifies the intended scope of the work and gives a reasonable positioning of TranSamba as a weakly supervised method rather than a prompt-based segmentation approach. Given the practical relevance of the problem, the clear technical design, the reproducible implementation, and the consistent empirical improvements, I believe the paper is above the acceptance bar. I would therefore support acceptance, while encouraging the authors to further discuss the binary-scope limitation and strengthen comparisons with recent weakly supervised and foundation-model-based alternatives in the final version.
