Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Liver cirrhosis plays a critical role in the prognosis of chronic liver disease. Early detection and timely intervention are essential for reducing mortality rates. However, the intricate anatomical architecture and diverse pathological changes of liver tissue complicate the accurate detection and characterization of pathological liver structures in clinical settings. Existing methods underutilize spatial anatomical details in volumetric MRI data, thereby hindering their clinical effectiveness and explainability. To address this challenge, we introduce a novel Mamba-based network, SRMA-Mamba, designed to model the spatial relationships within complex anatomical structures of MRI volumes. By integrating the Spatial Anatomy-Based Mamba module (SABMamba), SRMA-Mamba performs selective Mamba scans within pathological liver tissues and combines anatomical information from the sagittal, coronal, and axial planes to construct a global spatial context representation, enabling efficient volumetric segmentation of pathological liver structures. Furthermore, we introduce the Spatial Reverse Mamba Attention module (SRMA), designed to progressively refine boundary details in the segmentation map, utilizing both the coarse segmentation map and hierarchical encoding features. Extensive experiments demonstrate that SRMA-Mamba surpasses state-of-the-art methods, delivering exceptional performance in 3D pathological liver segmentation. The source code is available at \url{https://github.com/JunZengz/SRMA-Mamba}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5682_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/JunZengz/SRMA-Mamba

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZenJun_SRMAMamba_MICCAI2026,
        author = { Zeng, Jun AND Trinh, Quoc-Huy AND Nayak, Deepak Ranjan AND Tomar, Nikhil Kumar AND Bagci, Ulas AND Jha, Debesh},
        title = { { SRMA-Mamba: Spatial Reverse Mamba Attention Network for Pathological Liver Segmentation in MRI Volumes } },
        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 addresses volumetric segmentation of pathological liver structures in MRI. The authors propose SRMA-Mamba, a Mamba-based encoder-decoder framework that combines multi-plane anatomical scanning with a reverse-attention-based refinement mechanism. The method aims to capture richer 3D spatial context by aggregating information from sagittal, coronal, and axial planes, while progressively refining segmentation predictions in the decoder. Experiments on the CirrMRI600+ T1W and T2W datasets show improved performance over several representative CNN-, Transformer-, and Mamba-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.

    1.The idea of leveraging multiple anatomical planes is intuitive for volumetric MRI and is well motivated as a way to capture richer spatial context for 3D segmentation.

    2.The proposed method demonstrates improved results over several strong baselines on both T1W and T2W subsets of CirrMRI600+, including gains in overlap- and boundary-related metrics. The reported parameter and FLOP comparisons suggest that the method may offer a relatively efficient alternative to some stronger volumetric baselines.

    3.In addition to the main benchmark comparisons, the paper includes ablation studies and model complexity analysis, which help assess the contribution of different components.

  • 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 methodological novelty is not yet sufficiently clear. The overall design is reasonable, but the manuscript could more clearly articulate what is new beyond existing multi-plane context modeling and decoder refinement strategies in medical image segmentation. Multi-view or multi-plane feature aggregation has been explored in various forms before, and reverse-attention-style refinement is also not entirely new. In its current form, the method appears closer to an incremental architectural extension within a Mamba-based framework than to a clearly distinct new formulation [18].

    2.The role of Mamba relative to prior context modeling approaches is not fully clarified. While the high-level motivation is understandable, the paper does not yet make sufficiently clear what specific modeling advantage the proposed Mamba-based design offers over other bidirectional or recurrent/context-aggregation strategies, beyond being a different implementation choice.

    3.Some technical details remain under-specified. The proposed modules, especially the anatomical selective scan and the decoder refinement process, are described at a relatively high level. For example, the exact scan directions, fusion strategy across planes, and the role of the reverse attention map could be explained more precisely. This affects both clarity and reproducibility.

  • 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.

    (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 main concern is that the methodological novelty is not yet sufficiently clear or convincing in its current presentation. The proposed framework appears to combine several familiar ideas—multi-plane context modeling, Mamba-based sequence processing, and reverse-attention-style refinement—without yet fully clarifying what the distinctive methodological contribution is beyond these existing design elements. In addition, the manuscript does not sufficiently differentiate itself from closely related prior or concurrent work on Mamba-based pathological liver segmentation (ref [18] in the proceeding).

    Overall, the work is promising and not far from the bar, but its current form is not convincing that it provides sufficient methodological distinctiveness and experimental depth for acceptance.

  • 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.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper studies pathological liver segmentation in volumetric MRI and proposes a Mamba-based framework named SRMA-Mamba. The core design consists of: 1.Spatial Anatomy-Based Mamba (SABMamba) with an Anatomy-Based Selective Scan (ABSS) module to model 3D spatial relationships by scanning along sagittal, coronal, and axial planes; 2.Spatial Reverse Mamba Attention (SRMA) in the decoder, which uses coarse segmentation maps and hierarchical features to progressively refine boundaries. Experiments are conducted on the T1W and T2W subsets of the CirrMRI600+ dataset. The method achieves strong results, including mDSC 92.95 / HD95 17.52 / ASSD 3.05 on T1W, and mDSC 86.25 on T2W.

  • 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 task is clinically meaningful. Automatic segmentation of cirrhotic/pathological liver structures from MRI has clear clinical value, especially as a non-invasive alternative to biopsy and for supporting disease assessment. The problem is well motivated. 2.The method is structurally complete and the motivation is understandable. The idea of mimicking radiologists’ multi-planar reading habits by scanning along three orthogonal anatomical planes is intuitively reasonable for volumetric MRI. The encoder-side SABMamba and decoder-side SRMA together form a coherent coarse-to-fine framework. 3.The main experimental results are generally strong. On T1W, SRMA-Mamba improves over the strongest baseline SegMamba from 91.80 to 92.95 mDSC, while reducing HD95 from 26.97 to 17.52.On T2W, it also achieves the highest mDSC of 86.25, suggesting a degree of stability across sequences. 4.The complexity analysis is a notable strength. Table 4 shows that SRMA-Mamba has substantially lower parameter count and FLOPs than SegMamba (17.22M / 149.14 GMac vs. 64.24M / 2379.03 GMac), while also achieving better segmentation performance. This is an important practical advantage of the paper. 5.The paper includes ablation studies. The ablations provide initial evidence for the contributions of SABMamba, SRMA, and ABSS to the final performance.

  • 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 experimental scope is narrow, and the evidence for generalization is limited All experiments are conducted on a single benchmark, CirrMRI600+, although it includes both T1W and T2W subsets. Since the dataset is described as single-center, multi-vendor, multi-sequence, it is more diverse than a single-device dataset, but it is still insufficient to establish broader generalization claims. For a method whose main claim is improved spatial modeling for 3D medical segmentation, it would be helpful to provide at least one stronger form of evidence, such as:

    • evaluation on an external or cross-center dataset;
    • cross-sequence transfer experiments;
    • stratified analysis on smaller lesions or more boundary-ambiguous cases. As it stands, the results mainly support the claim that the method is effective on this benchmark, rather than demonstrating broad clinical robustness. 2.The ablation design is still not sufficiently rigorous Although the paper provides an ablation table, there are several issues. First, “w/o ABSS” is defined as using only a single anatomical plane, but the paper reports only the best result among the three single-plane variants. This likely overestimates the performance of that ablation setting and does not allow a fair assessment of the true gain from ABSS. A more rigorous design would report all three single-plane results, or at least their average. Second, the text claims that “incorporating the SABMamba module … yields a 16.64% improvement in mDSC,” but Table 3 changes multiple components simultaneously, including SRMA and ABSS. Therefore, this number cannot be cleanly attributed to SABMamba alone. Third, “w/o SABMamba” replaces the encoder with a UNetr encoder, and “w/o SRMA” replaces the decoder with a UNetr decoder. These substitutions do not isolate the target modules cleanly, because they alter the overall architectural family as well. As a result, the ablations are not fully controlled. 3.The claim of “efficiency” is not yet fully supported The paper repeatedly emphasizes “efficient volumetric segmentation,” and supports this mainly through parameter count and FLOPs. However, for 3D medical image segmentation, readers would typically also expect:
    • actual inference time;
    • GPU memory usage;
    • training time;
    • scalability across different input sizes. This is particularly important because the method includes multi-plane selective scanning and progressive reverse-attention refinement. Reporting only FLOPs is not enough to fully establish practical deployment efficiency. 4.The evidence for the claimed advantage of anatomical spatial modeling could be more direct The core selling point of the paper is that multi-plane scanning improves 3D spatial modeling, but the current evidence is mainly the final segmentation improvement. More direct analysis would strengthen this claim, for example:
    • which types of cases benefit most from three-plane scanning versus single-plane scanning;
    • whether the gains are larger for anisotropic lesions, boundary-complex cases, or smaller lesions;
    • whether different anatomical planes contribute differently to different types of pathological patterns. Such analysis would make the anatomical motivation more convincing. 5.The gains on T2W are relatively modest, and this is not sufficiently discussed On T2W, SRMA-Mamba improves over SegMamba by only about 0.89 mDSC, and it does not achieve the best ASSD (SwinUNETRv2 reaches 4.26, better than SRMA-Mamba’s 4.57). This suggests that the advantage on T2W is much less pronounced than on T1W. The paper does not really discuss this point. It would be useful for the authors to analyze:
    • why the method is not consistently best on all boundary-distance metrics for T2W;
    • whether this is related to sequence-specific contrast, noise, or lesion morphology;
    • whether SRMA or SABMamba behaves differently across modalities.
  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper proposes a reasonably complete Mamba-based framework for pathological liver segmentation in volumetric MRI. Its main strengths are:

    • the task is important and clinically meaningful;
    • the method is systematically designed;
    • the T1W results are strong;
    • compared with SegMamba, the method shows advantages in both performance and complexity. However, the current work also has several notable weaknesses:
    • the experiments remain limited to a single benchmark, with limited evidence of generalization;
    • the ablation studies are not yet fully clean and controlled;
    • the arguments for efficiency and anatomical spatial modeling could be strengthened further.
  • 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.

    The rebuttal partially improves clarity but does not substantially change my assessment, because the main concerns about generalization and controlled ablation remain unresolved.



Review #3

  • Please describe the contribution of the paper

    The paper addresses the challenging task of volumetric pathological liver (cirrhosis) segmentation in MRI scans. The authors observe that while recent State Space Models (SSMs), such as Mamba, offer linear-time complexity and global receptive fields, their standard implementation flattens 3D volumetric data into 1D sequences, inherently disrupting the complex 3D spatial topology necessary for identifying irregular cirrhotic structures. To solve this, the authors propose SRMA-Mamba. The core contributions are twofold: First, they introduce the Spatial Anatomy-Based Mamba (SABMamba) and Anatomy-Based Selective Scan (ABSS) modules. These modules maintain 3D spatial context by unfolding and scanning the MRI data across three orthogonal anatomical planes (sagittal, coronal, and axial) and subsequently merging the features to build a global spatial representation. Second, they incorporate a Spatial Reverse Mamba Attention (SRMA) module in the decoder, which utilizes the coarse segmentation map alongside hierarchical encoding features to progressively refine uncertain boundaries and fine details of the cirrhotic tissue.

  • 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.

    Effective Multi-Planar 3D Context Modeling: Adapting Mamba to 3D medical imaging is a non-trivial challenge due to the loss of spatial adjacency during sequence flattening. The proposed ABSS module resolves this by explicitly performing selective scans across the sagittal, coronal, and axial planes. This biologically inspired multi-planar approach is highly suited for irregular pathological structures.

    High Computational Efficiency: The paper demonstrates that SRMA-Mamba achieves superior segmentation results compared to existing 3D SSM baselines (like SegMamba) while requiring significantly fewer computational resources. With only 17.22M parameters and 149.14 GMacs (compared to SegMamba’s 64.24M parameters and 2379.03 GMacs), the model is lightweight and efficient for volumetric tasks.

    Comprehensive Clinical Benchmark: Evaluating the method on the large-scale CirrMRI600+ dataset across both T1-weighted and T2-weighted modalities provides robust evidence of the model’s clinical viability.

    Clear and Convincing Ablation Studies: The ablation experiments (Table 3) systematically isolate the individual and combined impacts of the SABMamba, SRMA, and multi-planar ABSS modules, mathematically justifying each architectural design choice (e.g., proving that multi-planar scanning is far superior to single-plane scanning).

  • 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.

    Unexplored Modality Performance Gap: There is a significant performance disparity between the T1W dataset (92.95% Dice) and the T2W dataset (86.25% Dice). While the authors briefly acknowledge that T2W presents “challenging cases” resulting in false positives, they fail to investigate why this gap exists or how the architecture might be adapted to better handle the specific signal characteristics and artifacts inherent to T2W MRI.

    Limited Algorithmic Novelty in Reverse Attention: While applying reverse attention within a Mamba framework is novel in this specific context, the underlying principle of reverse attention (progressively subtracting a coarse prediction map from deep features to mine edge details) is a highly established technique in medical image segmentation (e.g., PraNet, CaraNet). The theoretical novelty of the SRMA module is therefore somewhat incremental.

    Vague Mathematical Formulation of Feature Merging: The exact mechanics of how the 1D sequences from the three distinct anatomical planes ($f_h$, $f_w$, $f_d$) are recombined into a coherent 3D volume are insufficiently detailed. Equation 4 simply defines this as $Y = \text{CrossMerge(S6(CrossScan(f)))}$. A more rigorous mathematical explanation of how spatial alignment is maintained during this CrossMerge operation is necessary for reproducibility.

    Additionally, text size in fig1 and fig 2 are too small.

  • 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 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 authors present a highly effective and computationally efficient solution to the problem of 3D pathological liver segmentation. The primary strength of the paper lies in the design of the Spatial Anatomy-Based Mamba (SABMamba) module, which successfully overcomes a major limitation of standard SSMs in volumetric imaging by enforcing multi-planar spatial awareness. The massive reduction in computational overhead compared to baseline 3D Mamba models, coupled with state-of-the-art performance on a challenging clinical dataset, makes this a valuable contribution to the field. The paper is held back from a higher score by the incremental nature of the reverse attention mechanism and the lack of deep analysis regarding the model’s performance drop on T2-weighted scans. Additionally, clarifying the mathematical mechanics of the multi-planar feature merging would improve the manuscript.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have clarified my concerns to some extent. I am in favor of borderline acceptance.



Author Feedback

We thank all reviewers for their constructive feedback. Key points are addressed below.

Novelty & Methodology (CrossMerge Mechanism) – R1, R3, Meta SRMA-Mamba is the first to combine directional Mamba scanning with reverse attention for 3D segmentation. To clarify, SABMamba/ABSS is not simple multi-view aggregation but spatial topology reconstruction. Addressing R3 on CrossMerge: it is a coordinate-consistent spatial re-indexing process. Features f_h, f_w, f_d from orthogonal axes are mapped to original 3D coordinates (i, j, k) via inverse permutation before fusion. This ensures every voxel aggregates hidden states from sagittal, coronal, and axial contexts simultaneously, resolving 1D-SSM spatial ambiguity. This mimics radiologist Multi-Planar Reconstruction, providing an anatomically-consistent global receptive field distinct from UNetR or SegMamba.

Evaluation & Hardware Accessibility – R1, R2, R3, Meta For rigorous evaluation, we used standard 1D-linearized Mamba as a baseline (Exp #1), achieving 78.06% Dice but a clinically unacceptable HD95 of 64.01mm. Plane-wise ablation proves single-plane scans like Sagittal (92.50% Dice) fail to capture full 3D context versus our approach. SRMA-Mamba achieves 92.95% Dice , outperforming 6 SOTA baselines. While Swin-based models typically require 32GB+ VRAM, SRMA-Mamba optimizes resources, consuming only 19.5GB (20018MiB) compared to SegMamba’s 24.1GB (24734MiB). Our model exhibits superior efficiency: stable convergence at Epoch 46 with a per-epoch time of 4m 15s, whereas SegMamba requires 6m 25s and converges at Epoch 49.Being 1.5x faster and memory-efficient, our model prioritizes sustainable resource ratios vital for clinical deployment.

T1W-T2W Performance Analysis – R2, R3 T1W/T2W performance disparity (92.95% vs 86.25% Dice) is rooted in MRI physics. T1W provides high anatomical contrast. T2W is water-sensitive; perihepatic ascites and edema often obscure liver boundaries, lowering SNR. Regarding T2W ASSD, SwinUNETRv2 (4.26) is slightly lower than SRMA-Mamba (4.57) due to brute-force computational overhead. Massive VRAM allows Swin-models marginal local surface smoothness gains. However, SRMA-Mamba maintains superior global integrity (mDice) and HD95 (17.52mm vs 33.27mm for SwinUNETRv2) by capturing long-range dependencies through tri-plane ABSS, ensuring feasibility on 24GB GPUs.

Dataset & Generalization – R2, R3, Meta CirrMRI600+ is the first cirrhosis-specific, multi-vendor dataset. Hardware variability causes institutional bias; our consistent SOTA across scanners proves clinical robustness. “Hard Case” analysis (extreme nodularity, nodules <2cm) confirms SRMA-Mamba reduces boundary errors where Transformers over-smooth. This 628-volume expert-annotated dataset and our efficient 3D framework constitute a major contribution.

2.5D vs 3D – R3 ABSS is not 2.5D. Traditional 2.5D processes limited adjacent slices, losing context. ABSS scans the entire 3D volume along three orthogonal axes (Sagittal, Coronal, Axial) to capture global dependencies. Ablation confirms only tri-plane integration (92.95% Dice, 17.52 HD95) achieves the volumetric closure necessary for complex pathological morphologies.



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 3/2/4, with one reviewer recommending rejection independent of rebuttal on grounds of single-benchmark evaluation, ablation cleanliness, and incomplete efficiency analysis. The paper is borderline-leaning-negative but two reviewers remain open to rebuttal-driven reconsideration. The paper is invited for rebuttal; authors should prioritize (a) cleaner ablations that report all three single-plane variants for the ABSS comparison and isolate SABMamba and SRMA without changing the architectural family, (b) actual inference time, GPU memory, and training-time measurements alongside the existing FLOP/parameter numbers, (c) at least one cross-sequence, cross-center, or stratified-by-lesion-difficulty experiment to support the generalization claim, (d) a substantive analysis of the T1W–T2W performance gap and why the method does not lead on T2W boundary metrics, and (e) sharper articulation of what the multi-plane Mamba design contributes beyond established multi-view aggregation and reverse-attention refinement, including the precise mechanics of CrossMerge.

  • 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 the clinical task is meaningful, the multi-planar approach to bridging Mamba’s 1D scanning with 3D volumetric topology is intuitive and well-motivated, the framework is structurally complete, the T1W results are strong, and the computational efficiency advantage over SegMamba (17.22M vs 64.24M params; 149 vs 2379 GMacs) is a notable practical strength. They raised several concerns: (1) methodological novelty is incremental — multi-plane aggregation and reverse-attention-style refinement (PraNet, CaraNet) both have established precedents; (2) evaluation is limited to a single benchmark, with no external/cross-center validation; (3) the ablation is not fully clean — w/o ABSS reports only the best single-plane result, substitutions to UNetR encoder/decoder change architecture family beyond the target modules, and the 16.64% SABMamba gain conflates multiple components; (4) efficiency claims rest on FLOPs/parameters rather than wall-clock inference time, GPU memory, or training time; (5) T1W (92.95) vs T2W (86.25) gap is not investigated and ASSD on T2W trails SwinUNETRv2; (6) the CrossMerge mathematical mechanics for spatial alignment across orthogonal planes are under-specified; and (7) more direct evidence linking multi-plane scanning to specific pathological subtypes (small lesions, irregular boundaries) is missing. After rebuttal, the authors clarified CrossMerge as a coordinate-consistent spatial re-indexing process, reported hardware-level comparisons (19.5GB vs 24.1GB VRAM, 1.5× faster than SegMamba), explained the T1W/T2W gap via MRI physics (T2W water-sensitivity obscures liver boundaries), justified that their HD95 advantage over SwinUNETRv2 (17.52 vs 33.27) reflects superior global integrity, distinguished ABSS from 2.5D approaches as full-volume tri-axial scanning, and mentioned a hard-case analysis. Two reviewers raise their recommendations to Accept based on these clarifications. The most critical reviewer maintains Reject, citing that the core concerns about single-benchmark generalization and fully controlled ablation remain unresolved. Overall, the AC thinks that the merits outweigh the drawbacks and recommend accept.



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.

    Several concerns highlighted by R2 were not sufficiently addressed, while R3 was not fully convinced by the clarifications and described their recommendation as a borderline Accept. Although R1 changed their score from Weak Reject to Accept, they did not clearly justify this change, which makes me more inclined to agree with the concerns raised by R2 and R3.



Meta-review #3

  • 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 paper proposes a clinically relevant 3D liver MRI segmentation framework combining tri-plane Mamba scanning with reverse-attention refinement. The submitted results are promising, especially on T1W, and the reported FLOP/parameter comparison suggests practical efficiency. However, the work remains below the MICCAI acceptance bar. The methodological contribution is largely an architectural combination of known elements: multi-plane context aggregation, Mamba-based scanning, and reverse attention. The submitted manuscript does not sufficiently distinguish the method from recent Mamba/reverse-attention segmentation work. Key concerns also remain regarding single-benchmark validation, modest and inconsistent T2W gains, under-specified CrossMerge details, and ablations that do not cleanly isolate modules. The rebuttal clarifies some implementation intent, but relies heavily on new results/measurements not present in the submission, which cannot be considered. Given these unresolved issues, I recommend rejection despite two post-rebuttal accept opinions.



back to top