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

Abstract

Accurate white matter (WM) lesion segmentation is essential for diagnosing and monitoring neurological disorders such as white matter hyperintensities and multiple sclerosis. However, this task remains highly challenging owing to the small and spatially scattered nature of lesions and their low contrast with surrounding tissues. Recently, the Segment Anything Model (SAM) and its medical variants, often augmented with CNNs or Mamba, have shown promising potential for medical image segmentation, yet they struggle to reliably capture subtle lesion patterns and indistinct boundaries that demand fine-grained regional discrimination. In this work, we propose a novel Hierarchical Region-Aware Multi-Granularity Mamba (HiRAM) incorporated into the SAM decoder. HiRAM explicitly decomposes the features into region-specific representations (i.e., interior, boundary, and background regions) and performs confidence-driven, non-causal state-space propagation. By incorporating region-aware modeling that prioritizes semantically reliable features and hierarchically integrates multi-granularity learning, our method enhances boundary sensitivity and fine-grained lesion delineation while preserving global contextual consistency. Comprehensive evaluations across two public challenge datasets reveal that HiRAM outperforms state-of-the-art methods, achieving unmatched segmentation accuracy and robustness. Code is available at https://github.com/deepnoid-ai/HiRAM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2898_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/deepnoid-ai/HiRAM

Link to the Dataset(s)

MWSC dataset: https://wmh.isi.uu.nl/ MSLesSeg dataset: https://doi.org/10.6084/m9.figshare.27919209

BibTex

@InProceedings{LeeDah_Hierarchical_MICCAI2026,
        author = { Lee, Dahye AND Oh, Kwanseok},
        title = { { Hierarchical Region-Aware Multi-Granularity Mamba for White Matter Lesion 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 proposes a region-aware extension of state-space modeling within a SAM-based segmentation framework. It introduces a hierarchical module that decomposes feature maps into interior, boundary, and background regions, and applies confidence-driven, non-causal Mamba-based propagation over region-specific sequences.

  • 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 combination of region decomposition, confidence-based ordering, and SSM-based propagation is not straightforward, and prioritizing high-confidence regions in state evolution is interesting.
    • The reported improvements over baselines (e. g. , nnUNet, SAM variants, and Mamba-based models) are not small for WMH/MS tasks.
    • The contribution of each component (region priors, permutation, auxiliary loss) is nicely analyzed.
  • 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.
    • While the paper emphasizes non-causal and region-aware modeling, the core pipeline still follows a standard pattern using the SAM encoder, CNN backbone, modified decoder, and SSM/attention hybrid.
    • Reordering tokens by confidence only replaces spatial bias with heuristic ordering. It is unclear whether this is theoretically justified or just empirically helpful.
    • Although multiple methods are included, it is unclear whether nnUNet is properly tuned, whether recent lesion-specific architectures are fairly represented, or whether multi-modal methods are excluded in a way that simplifies the task.
    • Using only FLAIR simplifies the problem and avoids the more realistic missing-modality challenge, which is a major issue in WM lesion segmentation literature.
    • The method introduces multiple interacting components (RPG, MUG, RA-SSM, auxiliary loss, permutation), but the paper does not clearly explain why this level of complexity is necessary and how the training and inference time look compared to SOTA.
    • The use of Gumbel-Softmax and hard assignment into regions raises concerns about training stability and sensitivity to hyperparameters, which are not discussed.
    • The method is evaluated on two datasets (WMH and MS), both involving closely related tasks (binary lesion segmentation on FLAIR MRI). While this is a reasonable starting point, it does not convincingly demonstrate that the proposed region-aware Mamba design generalizes beyond this narrow setting to multi-class segmentation tasks, different anatomical structures, or more heterogeneous, real-world clinical datasets.
    • The reported results for several baselines appear lower than what is typically observed in the original challenge benchmarks (WMH segmentation). Prior works on the same dataset have reported competitive Dice scores using relatively simple architectures.
    • There is no mention of confidence or statistical significance. Given the modest dataset sizes, the reported gains may not be robust.
  • 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.

    (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 addresses a relevant problem, and the gains are promising. However, I am not convinced by the methodological contribution and experimental validation. The core idea feels like an extension of existing SAM-CNN-Mamba hybrids. I may be mistaken, but the reported baseline numbers seem somewhat lower than what I recall from the original WMH challenge and follow-up work.

  • 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.

    After reading the rebuttal, I maintain my decision to reject for the following reasons: 1.The authors provide mathematical justification for confidence-based token ordering via the SSM recurrence. However, the core contribution remains an extension of existing SAM-CNN-Mamba hybrids. Semantic region grouping and confidence-based ordering do not represent a fundamentally new methodological departure. The novelty is incremental relative to the complexity introduced. 2.The framework introduces multiple interacting components without convincingly demonstrating that this level of complexity is necessary. 3.The method is evaluated exclusively on two datasets involving essentially the same task: binary lesion segmentation on FLAIR MRI. For a paper presenting a general region-aware Mamba design as a methodological contribution, this is insufficient. The authors acknowledge broader generalization to multi-class tasks and other anatomical structures as future work, but this does not address the concern; it confirms it. 4.Given the modest dataset sizes, particularly the 22-scan test set for MSLegSeg, the reported gains may not be statistically robust.



Review #2

  • Please describe the contribution of the paper

    The authors propose a hierarchical region-aware multigranularity Mamba module (HiRAM) and build upon it a dual-branch parallel network for white matter lesion segmentation. Based on the outputs of the CNN branch and the SAM encoder in the first stage, HiRAM decomposes the features into foreground interior, foreground boundary, and background regions, and leverages region confidence to guide the model toward prioritizing the modeling of critical regions. The proposed method is evaluated on one white matter hyperintensity segmentation dataset and one white matter segmentation dataset, where it demonstrates superior performance compared with the competing methods included in the study.

  • 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 propose a novel and well-motivated quasi-two-stage segmentation framework. Based on the outputs of the first-stage CNN branch and SAM branch, the method derives three region categories, namely foreground interior, foreground boundary, and background, and these region categories are subsequently used to guide later feature propagation and representation learning. This design intuitively reflects an emphasis on lesion boundary delineation, and achieves more refined segmentation results than the other baselines on the two datasets presented in the paper.

  • 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.

    Major weaknesses: 1.In the RAE Block, the authors employ a cross-attention (CA) + MLP + CA design to process the features before feeding them into HiRAM. While this design appears intended to maintain continuous feature interaction, the paper does not provide sufficient motivation or empirical analysis to support this architectural choice. It would strengthen the paper if the authors could better explain why this specific design is necessary and whether it offers advantages over simpler alternatives. 2.For the compared method WMH-seg in Ref. [1], the reported TPR on the MWSC dataset reaches 0.92, which is substantially different from the result presented in Table 1.The paper does not clarify whether this discrepancy is caused by differences in experimental settings, implementation details, or other factors. If the discrepancy is due to reimplementation or evaluation protocol differences, I would encourage the authors to explicitly describe the settings used for the compared methods in the Implementation Details section to ensure a fair and transparent comparison. 3.Both MWSC and MSLegSeg are medical imaging datasets, for which MedSAM weights would intuitively seem more appropriate. However, according to Table 1, better DSC and TPR are achieved with MedSAM weights on MWSC, whereas on MSLegSeg, SAM weights lead to consistently better performance across all reported metrics. This makes it somewhat unclear whether the main conclusion of the paper should favor SAM or MedSAM initialization. Further discussion or analysis of this observation would improve the interpretability of the experimental results. Minor weaknesses: There are some presentation issues in the manuscript. For example, in the last paragraph of Page 5, “see Fig. 2(b))” appears to be an incorrect reference and should likely be “Fig. 1”. Reference: [1] He, H., Jiang, J., Peng, S., He, C., Sun, T., Fan, F., Song, H., Sun, D., Xu, Z., Wu, S., et al. A robust automated segmentation method for white matter hyperintensity of vascular-origin. NeuroImage 315, 121279 (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 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?

    The paper proposes a white matter lesion segmentation framework built upon a region-partitioning design, and demonstrates its effectiveness by outperforming the compared baselines on two relevant datasets. Nevertheless, my overall score is limited by the following concerns. First, the motivation and analysis for some design choices in the core RAE Block are not sufficiently developed. Second, some experimental results are inconsistent with those reported for the compared methods, which raises questions regarding the experimental settings. Third, although the best DSC scores on the two datasets are achieved with SAM and MedSAM weights, respectively, the paper does not provide sufficient analysis of this observation or a clear overall conclusion. Taken together, I am inclined toward a Weak Accept rather than a clear accept at the current stage.

  • 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

    This paper proposes HiRAM (Hierarchical Region-Aware Multi-Granularity Mamba), a new decoder module integrated into a SAM-CNN hybrid framework for white matter (WM) lesion segmentation covering both WMH and MS. The core idea is to move away from the fixed spatial scan orders used in existing SAM-Mamba methods, which are argued to be misaligned with the non-causal 2D structure of medical images. Instead, HiRAM decomposes intermediate features into three semantic regions (interior, boundary, background) via differentiable Gumbel-Softmax routing, sorts tokens within each region by prediction confidence, and processes them using a region-adaptive SSM (RA-SSM) that injects learned region priors into the output projection. A subsequent SS2D pass enforces global spatial coherence. The framework also introduces a region decomposition loss (Ldec via KL divergence) to provide explicit supervision for the region assignment. Experiments on MWSC (WMH challenge, 170 subjects) and MSLegSeg (MS, 115 scans) show consistent improvements of approximately 4–5% DSC over the best competing methods.

  • 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 ablation study is one of the stronger ones. Table 2 shows a clean, stepwise progression from baseline (69.9% DSC) through SS2D alone (+0.3%), RA-SSM (+4.7%), Ldec (+2.9%), adding πr (+1.5%), adding Pr (+1.4%), to the full model (82.1%). Each component makes a meaningful difference and the gains compound sensibly. It’s clear that RA-SSM is the engine and the other pieces genuinely help on top of it — this isn’t the kind of ablation where everything contributes 0.2% and you can’t tell what matters. 2.Strong and consistent experimental results. A +5.1% DSC improvement over the best baseline on MWSC and +3.3% on MSLegSeg are meaningful margins, not marginal ones. The AVD improvements are also substantial (13.1 vs 18.1 on MWSC), which matters clinically for lesion volume estimation. The qualitative results in Fig. 2 and the region decomposition visualization in Fig. 3(b) genuinely support the quantitative claims. 3.The semantic decoupling formulation for prior banks (Eq. 2: b_r = e ⊗ E_r) is a small but nice design choice, sharing a common embedding across regions while allowing region-specific projections keeps parameter count low and ensures regions exist in a unified semantic space.

  • 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.

    1.2D slice-based processing for volumetric MRI. The framework takes I ∈ R^{H×W×3} (2D), which means it processes FLAIR MRI as independent axial slices. WMH segmentation challenge winners almost universally exploit 3D context, the spatial continuity of lesions across slices and hemisphere-level patterns are important cues. 2.MSLegSeg is a very small test set (22 scans). The standard deviations in Table 1 for this dataset are quite large for several baselines (e.g., MedSAM ±9.6%, TP-Mamba ±5.3%). The proposed method’s notably lower std (±2.4% for OursSAM) is encouraging but could partly reflect favorable data splits rather than true robustness.

  • 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.

    (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?

    This is a solid paper with a genuinely interesting idea at its core. The confidence-driven region-aware Mamba sequencing is novel, well-motivated, and the ablation study makes a convincing case that it contributes substantially. The performance gains over a competitive set of baselines (+4–5% DSC) are among the larger margins I’ve seen in recent WM lesion segmentation papers. These are the reasons I lean toward acceptance.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    For the reasons mentioned earlier.



Author Feedback

We would like to thank the reviewers for their valuable comments.

Method Novelty (R1,3): We acknowledge that the macro-architectural composition is shared with prior SAM-Mamba hybrids. Our contribution lies in rethinking how SSM propagation can be reorganized for small, scattered, low-contrast lesions in medical images, where causal scan orders struggle to align with their irregular spatial distribution. In this regard, HiRAM departs from prior work in two concrete respects: features are grouped by semantic regions instead of spatial adjacency, and tokens are ordered by confidence, not position—a formulation, to our knowledge, underexplored in prior SAM-Mamba literature. To elaborate, unrolling SSM yields hₜ = Σₖ A̅^{t-k}B̅xₖ, where early tokens propagate via the highest powers of A̅ and shape every downstream state. By prioritizing reliable features at the start of this recurrence, π_r enables a more robust state propagation—which, in our view, is a natural fit for SSM dynamics rather than an arbitrary heuristic. This stabilizing effect also helps explain the lower variance on MSLegSeg (±2.4% vs ±5.3–9.6%), particularly on hard cases.

Design Choice (R1,2,3): As noted in §2.1, the RAE block follows the validated original SAM decoder, with a minor adjustment for CNN features: channel-wise token split into CNN/SAM branches. For Gumbel-Softmax, M̄_r is used only for index selection (Eq. 3); gradients flow exclusively via the differentiable M_r, avoiding the high-variance straight-through path entirely. Combined with L_dec’s layer-wise supervision on M_r (Eq. 7), this further prevents the region collapse that unsupervised routing can suffer from. Indeed, training proceeded stably with τ=1 (no annealing), suggesting low sensitivity. WM lesions range from large regions favoring inter-slice continuity to small punctate foci confined to 1–2 slices. This heterogeneity can limit the benefit of 3D context modeling, and MWSC’s 3mm resampling [16] further reduces inter-slice continuity. We thus adopt a 2D model, leaving a 3D design as a future direction.

Baseline Validity (R1,2): The official WMH challenge survey [16] reports a DSC range of 63–80 among its top-10 models, reflecting substantial variation even within the same benchmark. Our baselines fall within this reference span at 69–77 (excluding SAM and MedSAM, which were evaluated zero-shot), which we believe supports the validity of our comparison. All baselines used their official code (i.e., no modification), and lesion-specific models (WMH-seg, Deep-MS) retained their original multi-modal inputs. The TPR gap of WMH-seg [8] is likely attributable to differences in experimental setup: [8] used 100+ private cases for training and evaluated on the full MWSC set, whereas we follow the official 60/110 split. We will clarify these details in §3.1. SAM vs. MedSAM (R2): The dataset-wise gap reflects differences in pretraining coverage. MedSAM (train: ~1.5M medical images including WMH [21]) benefits from direct domain alignment on MWSC, while MSLegSeg lies outside its training distribution; in this regime, SAM (train: 11M natural images) provides more transferable representations. Overall, we suggest SAM preweights as a robust default, with MedSAM preferred when the target task falls within its pretraining distribution.

Scope & Clarifications (R1): FLAIR was chosen as the clinical standard for WM lesion assessment and the common modality across both datasets, enabling controlled comparison. Extensions to multi-class tasks (e.g., brain tumors) and other anatomies (e.g., lung nodules) are orthogonal directions that HiRAM could naturally accommodate, planned for future work. While Tab. 2 and Fig. 3(a) partly address our design choices, we agree that the rationale in §2 is concise and that wall-clock time would better demonstrate efficiency. We will reflect both in the paper. Thanks to clear margins in Tab. 1, we initially reported std; we will also include the Wilcoxon signed-rank tests.



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 appears proposing a hierarchical framework for white matter lesion segmentation appears quite solid. There are however multiple points that would need clarification in the rebuttal and notably the justification for the chosen architectural design, the observed differences between the results presented in the reference and in the table, some diverging results between MedSAM and SAM as well as the lack of proper statistical comparison to substantiate the claims. A rebuttal will hopefully clarify the points of interrogation raised by the reviewers

  • 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 rebuttal addresses most of the issues raised previously. While there is still some argument about the level of novelty, the performance gain remains substantial enough to warrant further discussion at the conference.



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.

    I would like to start by acknowledging that the initial idea and motivation is intriguing: can the token sequence ordering for MAMBA be altered to take into account confidence instead of spatial locations? I echo the reviewers comments in saying that this is a big strength.

    However, if I take into account how the information is presented, some of the major criticisms raised by reviewer #1 and the lack of satisfactory answers to those concerns I would also like to underscore that the complexity of the whole approach (one only needs to look at the scheme in Figure 1) is poorly justified. While the authors did address why the results do not seem to match the original works, they still did not provide any statistical analysis to back up whether the results could be due to chance (random seed choice) which would help with the arbitrary complexity of the proposal. Finally, I also agree with reviewer #1 in the lack of different enough datasets. Once again, this is a fairly complex network that is being used for a “fairly” simplified problem (binary segmentation on a single image where lesions are always hyperintense). The lack of a multi-class dataset downplays the conclusions.



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 method is somewhat incremental and its validation is limited to two closely related FLAIR lesion segmentation tasks, but the submitted manuscript provides solid empirical gains, meaningful ablation evidence, and a reasonably well-motivated region-aware Mamba design. The remaining concerns on broader generalization and statistical robustness should be noted, but they do not outweigh the paper’s technical merit.



back to top