Abstract

Acute ischemic stroke patients are at high risk of stroke-associated pneumonia (SAP) and hemorrhagic transformation (HT), where early prediction from brain MRI can enable timely interventions. However, when modeling long-range dependencies in MRI sequences, most existing CNN- or Transformer-based models either over-emphasize local patterns or incur high computational cost. Therefore, we propose N6-TreeMamba, a Mamba-based framework that combines a 6-neighborhood tree scanning (N6-TS) strategy to efficiently capture hierarchical long-range interactions. To further align the model with clinical reasoning, we design a radiology knowledge-guided anchoring (RKA) module that leverages white matter hyperintensity (WMH) segmentation to select scanning anchors, guiding the model to attend to clinically relevant regions while maintaining global coverage. Experiments on an in-house stroke MRI dataset demonstrate that N6-TreeMamba consistently improves performance while achieving lower computational complexity.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0441_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://anonymous.4open.science/r/N6Treemamba-925

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhaChe_N6TreeMamba_MICCAI2026,
        author = { Zhao, Chenyang AND Yu, Yaoyao AND Xia, Tianyi AND Yuan, Mengxuan AND Yang, Yunjun AND Hu, Chuanfei},
        title = { { N6-TreeMamba: Radiology Knowledge-Guided 6-Neighborhood Tree Scanning for 3D MRI-Based Stroke Complication Prediction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper studies 3D multi-sequence MRI-based prediction of post-stroke complications (SAP and HT) and proposes N6-TreeMamba, a TreeMamba-based framework for volumetric stroke prognosis. Its key technical components are N6-TS, a 3D 6-neighborhood tree scanning strategy for more locality-preserving token ordering, and RKA, a WMH-guided anchoring module that biases scanning toward clinically relevant regions.

    The contribution is therefore best viewed as a task-specific 3D adaptation of tree-scanning Mamba for stroke MRI, rather than a fundamentally new backbone. The evaluation on an in-house clinical dataset strengthens the paper’s practical relevance, although this should be considered an empirical strength rather than a core methodological novelty.

  • 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 paper addresses a clinically meaningful problem. Predicting SAP and HT from brain MRI is relevant to early post-stroke management, so the task itself has potential clinical value. 2: The experimental section includes relevant Mamba-based comparisons and ablations. The paper compares against several scan-strategy baselines, including GrootVL, and provides ablations for N6-TS, RKA, and anchor composition, which helps isolate where the reported gains come from. 3: The method is reasonably adapted to 3D MRI. The proposed N6-TS replaces naive linearization with a 3D tree-based scan order, which is a sensible way to better preserve volumetric neighborhood structure during selective scanning. 4: The evaluation is conducted on an in-house clinical dataset. Although this is not a methodological novelty, it does make the study more practically grounded than a purely benchmark-driven evaluation.

  • 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 appears limited relative to prior tree-scanning Mamba work. The main technical idea is a task-specific adaptation of TreeMamba-style scanning to 3D MRI, with WMH-guided anchor selection, rather than a fundamentally new sequence modeling framework. 2: The empirical validation is too limited for the strength of the claims. The study is conducted on only one in-house dataset with 188 patients in total, including only 41 SAP cases and 25 HT cases, and evaluation is based on a single 8:2 train/test split. This is a very limited experimental setting for a clinical prediction paper, especially for HT where the number of positive cases is particularly small. There is no cross-validation, repeated runs, confidence interval, significance testing, or external validation, so it is difficult to determine whether the reported gains are robust or split-specific. 3: The experimental reporting is somewhat inconsistent. The table caption mentions F1, but the table reports only Acc and AUC. This reduces confidence in the completeness and rigor of the presentation

  • 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.

    (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 assign a 4 (Weak Accept). The paper is overall of good quality, with a coherent technical narrative, strong visual presentation, and a method that is reasonably well motivated for 3D MRI-based stroke complication prediction. In particular, the use of a 3D tree-based scan design together with WMH-guided anchoring is a sensible and clinically motivated adaptation for volumetric MRI, and the inclusion of evaluation on an in-house clinical dataset makes the study more practically grounded than a purely benchmark-driven paper.

    That said, my recommendation is only marginally above the acceptance threshold. The main concerns are that the methodological novelty is somewhat incremental relative to prior tree-scanning Mamba-style work, and the empirical validation remains limited to a single in-house cohort with one train/test split and no external validation or robustness analysis. Still, I find the paper promising, technically meaningful, and potentially impactful if the limitations in validation are better addressed.

  • Reviewer confidence

    Somewhat confident (2)

  • [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



Review #2

  • Please describe the contribution of the paper

    This paper proposes N6-TreeMamba, a radiology knowledge-guided framework for 3D MRI-based stroke complication prediction. The main contribution lies in integrating clinically meaningful priors into long-range sequence modeling, rather than relying solely on generic architectural improvements.

    Specifically, the paper introduces (1) a radiology knowledge-guided anchoring (RKA) module that leverages WMH as clinically relevant cues to guide model attention, and (2) a 6-neighborhood tree scanning strategy (N6-TS) that preserves 3D spatial locality while enabling efficient long-range dependency modeling within a Mamba-based framework.

    Overall, the work is well-motivated and grounded in clinical reasoning, addressing the challenge of capturing both global context and localized pathological signals in 3D MRI. Importantly, the contribution goes beyond incremental architectural modification by explicitly incorporating domain knowledge into the modeling process, which enhances both interpretability and practical relevance. In particular, the proposed tree-based scanning strategy provides an effective solution for adapting Mamba’s inherently 1D sequential modeling to 3D volumetric data, compensating for the loss of spatial structure during sequence linearization.

  • 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.Clinically grounded design. A major strength of this work is the explicit integration of radiological knowledge into the model. Using WMH as anchors in the RKA module is clinically meaningful and improves the interpretability of the method compared with purely data-driven approaches.

    2.Good balance between performance and efficiency The method demonstrates strong performance while maintaining relatively low computational complexity compared to Transformer-based baselines. This is particularly important for practical deployment in 3D medical imaging scenarios. The comparison Table 1 design/format looks excellent.

    3.Clear and modular framework The overall pipeline (RKA → N6-TS → TreeMamba) is well-organized and easy to follow. The modular design also suggests potential for extension to other tasks or modalities.

  • 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.Unclear Coherence with the existing other architecture Although the paper includes a brief discussion in the Introduction (second paragraph) and experimental comparisons in Table 1, the motivation for adopting a Mamba-based framework over other mainstream approaches (e.g., CNNs or Transformers) is not sufficiently articulated. In particular, it would improve readability to include one or two sentences in the Introduction clarifying why the proposed framework is better suited for Mamba, and whether the design could be extended to other architectures such as Transformers (Suggest: To make it more coherence, the author can focus on extending the second paragraph in the Introduction, from reviewing cnn-Transformer-> why using tree strategy in mamba).

    2.Generalization concerns. The evaluation is based on a relatively small in-house dataset with limited positive cases. Additional validation or discussion on generalizability would be beneficial.

  • 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 and well-motivated work that combines methodological design with clinical insight. The contribution is not merely incremental, but reflects a thoughtful attempt to bridge medical prior knowledge and modern sequence modeling.

  • Reviewer confidence

    Somewhat confident (2)

  • [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



Review #3

  • Please describe the contribution of the paper

    The paper proposes N6-TreeMamba, a Mamba-based framework for 3D MRI-based prediction of SAP and HT. Its main contribution is a 6-neighborhood tree scanning strategy that efficiently capture hierarchical long-range interactions, together with a radiology knowledge-guided anchoring module that uses WMH segmentation to prioritize clinically relevant regions.

  • 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 paper addresses a clinically relevant prediction task, namely early prediction of SAP and HT from brain MRI, which has potential practical importance. 2.The method is well motivated from both the modeling and clinical perspectives. 3.The paper includes comparisons with both CNN/Transformer-based and Mamba-based methods, along with ablation studies and complexity analysis, which makes the empirical evaluation reasonably comprehensive.

  • 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 validation is relatively limited. The experiments are conducted on a single in-house dataset with 188 patients, without external validation.

  • 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 studies a clinically meaningful task and proposes a well-motivated Mamba-based framework for 3D MRI-based stroke complication prediction. The method is technically solid, and the experiments are fairly comprehensive, including baseline comparisons, ablation studies, and complexity analysis. The results indicate a good trade-off between predictive performance and efficiency. My main reservation is the limited validation on a single in-house dataset without external testing. Overall, I find the paper sufficiently interesting and technically supported to justify a weak accept.

  • 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 sincerely thank all reviewers for their constructive feedback. We address the main concerns below. [R1-W2, R2-W2, R3-W1] Experimental validation, dataset scale, and external validation. We fully agree that broader validation strengthens clinical claims. To our knowledge, there is currently no publicly available 3D MRI-based dataset for stroke-associated pneumonia (SAP) or hemorrhagic transformation (HT) prediction, which constrains external benchmarking at this stage. We are actively collecting multi-center private data from collaborating hospitals, and we have committed to a journal extension in which N6-TreeMamba will be validated across multiple centers with larger cohorts and richer subgroup analyses. [R1-W1] Methodological novelty relative to TreeMamba/GrootVL. We respectfully clarify that N6-TreeMamba is not a straightforward adaptation. GrootVL constructs trees over 2D tokens with attention-based roots and is not applicable to volumetric data without modification. Our contributions are: (i) the N6-TS strategy, which builds an input-dependent minimum spanning tree on a 6-neighborhood 3D graph to preserve volumetric locality. This is a non-trivial extension since naive 3D tree construction over dense tokens is expensive and structure-agnostic; (ii) the RKA module, which replaces attention-based root selection with radiology-prior-driven anchors derived from WMH connected components, farthest-point sampling, and a dilated ring for background; and (iii) a clinical formulation linking WMH priors to SAP/HT prediction. Table 1 shows N6-TreeMamba outperforms GrootVL by +2.50% AUC (SAP task) and +1.98% AUC (HT task) with fewer parameters, indicating the design changes are substantive rather than cosmetic. We will sharpen this framing in the revised Introduction. [R1-W3] F1 inconsistency in Table 1 caption. We apologize for this oversight. The caption incorrectly mentions F1; the intended metrics are Acc and AUC, as reported in the table. We will correct the caption in the camera-ready. [R2-W1] Motivation for Mamba over CNN/Transformer and extensibility. We thank the reviewer for the helpful suggestion. We will revise the second paragraph of the Introduction to make the rationale more coherent: CNNs struggle with long-range dependencies in 3D volumes; Transformers capture global context but scale quadratically with token length, which is prohibitive for 3D MRI; Mamba offers linear-time selective scanning, making it well-suited for long-context volumetric modeling. However, Mamba’s effectiveness hinges on the 1D scan order, which motivates our tree-based scanning over naive linearization. Regarding extensibility, the RKA module is architecture-agnostic, where WMH anchors could in principle guide Transformer attention masks or CNN region-of-interest pooling. We will add a brief discussion of this in the Conclusion as a future direction. We thank the reviewers again for their thoughtful comments, which will substantially improve our work.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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.

    All reviewers agree that the paper provides a clinically meaningful, technically sound contribution to 3D MRI-based stroke complication prediction. The weakness is mostly related to empirical validation and minor methodological clarification.



back to top