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

Abstract

Incomplete multi-modal MRI data severely degrades clinical brain tumor segmentation. Existing solutions utilizing generative synthesis or implicit disentanglement often suffer from hallucinated artifacts or weak feature discriminability. To address these limitations, we propose a novel Sample-Agnostic Retrieval Network (SAR-Net), shifting the paradigm from “fabricating’’ missing features to “retrieving’’ latent style prototypes from a global distribution. We explicitly disentangle images into modality-agnostic “content’’ and sample-agnostic “styles’’, utilizing a Dynamic Style Memory Bank optimized via Momentum Contrastive Learning to ensure high discriminability. To impute missing modalities, a Global Consensus Voting mechanism retrieves compatible style prototypes, which are then fused with structural content via adaptive modulation. Extensive evaluations on BraTS 2020, together with ablation studies on BraTS 2018, demonstrate that our framework outperforms state-of-the-art methods under diverse missing-modality settings. The code is available at https://github.com/CoCodeRepo/FS2R.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1730_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/CoCodeRepo/FS2R

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZheShe_SARNet_MICCAI2026,
        author = { Zheng, Shenhai AND Zhou, Weifu AND Tang, Xue AND Li, Laquan AND Li, Weisheng},
        title = { { SAR-Net: Sample-Agnostic Retrieval Network for Robust Brain Tumor Segmentation with Missing Modalities } },
        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 introduces the SAR-Net framework for brain tumor segmentation under missing MRI modalities. Instead of synthesizing missing modalities or relying on modality-invariant representations, the method reconstructs missing modality features by retrieving style prototypes from a memory bank. The reconstructed features are then integrated into a segmentation network with transformer-based fusion.

  • 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 use of Gram matrices to represent style and the separation between structural and statistical components are well-motivated, and the integration with contrastive learning is reasonable.
    • The component-wise and loss-wise ablations support the contribution of each module, and the incremental gains from retrieval and contrastive learning are demonstrated.
    • The experiments consider all missing-modality combinations.
    • Gains in more challenging regions (TC/ET) are more than those of WT, suggesting the method may indeed improve boundary precision.
  • 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 method largely combines existing ideas (Gram matrix-based style transfer, memory banks with MoCo-like contrastive learning, and feature modulation via AdaIN).
    • The assumption that Gram matrices capture modality-specific style independent of anatomy is borrowed from natural image literature. However, in MRI, intensity distributions are not just stylistic, and scanner/protocol variability, as well as pathology, may be entangled.
    • The selected baselines are reasonable, but recent studies in diffusion-based imputation or foundation models with multimodal learning are not addressed.
    • Given the added architectural complexity (memory bank, contrastive learning, retrieval, transformer), the improvements achieved over the baseline are relatively small.
    • The paper criticizes generative models for being computationally expensive, but does not provide runtime comparisons, memory consumption, or training stability discussion.
    • No cross-dataset evaluation is provided to show how well this generalizes to unseen distributions or whether performance depends on dataset-specific factors.
    • The paper argues that retrieval avoids hallucination, but does not provide qualitative evidence (e. g. , failure cases, visual comparisons).
    • The paper lacks implementation details to ensure reproducibility. Several aspects are missing, including training hyperparameters (learning rate schedule, batch size, number of epochs), memory bank configuration (size, update frequency, queue length per modality), momentum coefficient, and temperature in contrastive learning, details of data preprocessing and augmentation, and architectural specifics (e. g. , encoder backbone, feature dimensions, transformer configuration).
  • 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 provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    The method seems stable and yields consistent improvements in clinically relevant regions. However, implementation details are missing, and the claimed novelty or general impact is not convincing. The framework builds on established components, and the gains over prior work are relatively incremental.

  • 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.

    After considering the rebuttal, I am moving my score to Accept for the following reasons:

    • The authors provided full implementation details.
    • The authors acknowledged the limitation of the Gram matrix/style analogy in the MRI context, conceding that intensity distributions are entangled with scanner, protocol, and pathology effects, and committing to moderating the relevant language in revision.
    • The hallucination concern is adequately reframed: since SAR-Net operates at the feature level rather than synthesizing missing-modality images, image-level hallucination is not applicable, and the original claim is a definitional rather than empirical one.



Review #2

  • Please describe the contribution of the paper

    Addressing the prevalent clinical challenge of missing modalities in multi-modal MRI, this paper proposes SAR-Net, a method designed to effectively enhance inference accuracy under incomplete data conditions. Distinguishing itself from generative synthesis and representation learning approaches, this method innovatively introduces a retrieval-based paradigm. It imputes missing modality features via similarity retrieval, thereby circumventing the generation of hallucinated artifacts. Architecturally, the authors innovatively decouple structural information and explicitly disentangle latent styles from spatial structures. Furthermore, they employ momentum updates and contrastive learning to efficiently maintain a memory bank, ultimately achieving high-performance inference in missing-modality scenarios.

  • 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 demonstrates several key strengths. 1.First, it proposes an innovative retrieval-based approach to solve the missing modality problem, designing specialized feature extraction and retrieval methods tailored to the specific challenges of missing modalities in medical segmentation tasks. 2.Furthermore, in the context of multi-modal medical image segmentation, the authors accurately identify the distinction between structural information and modality-specific styles, developing a corresponding disentanglement module. To efficiently maintain the memory bank, the study introduces contrastive learning to further enhance the separability and distinctiveness of different modality styles. 3.Finally, the paper is well-structured and logically coherent, and the experimental results robustly demonstrate the excellent performance of the proposed model.

  • 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.

    However, the manuscript still has several weaknesses that warrant further clarification. 1.Missing Loss Function Weights: The paper does not specify the weighting coefficients for the various loss functions. The authors should discuss how different weighting ratios might affect the model’s performance. 2.Inconsistent Dataset Usage: There is a discrepancy in the datasets used for evaluation: the ablation study is conducted on the BraTS 2018 dataset, while the SOTA comparison uses the BraTS 2020 dataset. Even though these datasets share some overlapping cases and exhibit domain similarity, this experimental design choice lacks a proper justification. 3.Limited Reproducibility: The paper lacks comprehensive implementation details and does not provide reproducible source code, which hinders the overall reproducibility of the method.

  • 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 provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

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

    I recommend this manuscript for direct acceptance. Addressing the prevalent clinical challenge of missing modalities in multi-modal segmentation, this paper innovatively introduces a retrieval-based paradigm. To tackle the core challenges of feature extraction and memory bank maintenance inherent in retrieval systems, the authors propose highly targeted designs: explicitly disentangling features into structural content and modality-specific styles, and employing contrastive learning to maintain the dynamic memory bank. These innovations enable the model to efficiently retrieve and impute missing features. Comprehensive experimental results robustly demonstrate that SAR-Net consistently outperforms existing state-of-the-art methods across various missing-modality scenarios, particularly under severe missingness, highlighting its significant potential for real-world clinical applications. Furthermore, the manuscript is well-structured, logically coherent, and supported by thorough experiments. It is a highly meritorious work worthy of 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.

    Thanks to the author for the additional response, which addressed some of my concerns.



Review #3

  • Please describe the contribution of the paper

    SAR-Net proposes a “Disentangle-Contrast-Retrieve” paradigm for brain tumor segmentation under missing MRI modalities. It decouples images into modality-agnostic structural content and sample-agnostic styles (via Gram matrices), maintains a momentum-updated Dynamic Style Memory Bank optimized with contrastive learning, and retrieves missing modality styles via Global Consensus Voting across available modalities. Retrieved styles modulate structural content via AdaIN-inspired adaptive feature modulation for final segmentation.

  • 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.Methodological contribution is novel. Framing missing modality imputation as style retrieval rather than generation is well-motivated. 2.Disentanglement with Gram matrix is more explicit than adversarial or implicit alignment methods. Visualizations confirm that. 3.Comprehensive evaluation reported. All 15 missing modality combinations were reported with both DSC and HD95 showing consistent gains in clinically critical TC and ET regions.

  • 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 authors claim to have tested on BraTS 2024, but there is no result shared in the paper. 2.TMFormer [24], KMD [9], and CCSD [18] are cited but not compared against despite being directly relevant and recent. (citation indexes from paper). 3.Table 1 is the BraTS 2020 results. But ablation is on BraTS 2018.The standard convention is to share results across all datasets and to show ablation results on a subset. That is not followed here, and the claim that it surpasses all BraTS 2018, 2020, and 2024 is not supported.

  • 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.

    (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 retrieval paradigm is novel and the ablations are thorough, but the missing BraTS 2024 results, absent recent baselines, and inconsistent experimental evaluation (not supporting all the claims) make this paper incomplete.

  • Reviewer confidence

    Very confident (4)

  • [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 all reviewers for your valuable comments to improve our work. To Meta Reviewer: Below we have provided clarify experimental discrepancies, BraTS 2024 results, and methodological novelty.

BraTS dataset inconsistency (R1,R2,R3): (1) using BraTS 2020 for SOTA comparison and BraTS 2018 for ablation is aimed to show more datasets results within limited space. (2)The mention of BraTS 2024 in the abstract was an editing error, for fine-grained six-label BraTS 2024, we have additional evidence: in a large-scale six-region evaluation, SAR-Net improves over M2FTrans, with DSC gains of +0.5/+0.9/+0.4/+1.0/+0.5/+0.8 for WT/TC/ET/NETC/SNFH/RC and mean HD95 reduced from 7.78 to 6.58 mm. Following rebuttal policy, we do not provide more results here.. Reproducibility and code (R1,R2): Key settings are Adam, lr=2e-4, weight decay=1e-4, poly LR decay, batch size 1, 1000 epochs, seed 1024, and 128^3 crop with rotation/intensity/flip augmentation. Memory bank uses 250 entries per modality-layer, 10 warm-up epochs, momentum 0.995, and 4/128 positive/negative samples. Gram-layer temperatures are 0.5/0.45/0.4/0.35/0.3, with weights 0.3/0.25/0.2/0.15/0.1.Total loss is unit-weighted Lcon+Lrec+Lreg+Lds+Lseg. Code will available after accept. Novelty and Gram/style statistics (R1): Our approach is not a mere heuristic but a genuine innovation grounded in a deep understanding of the missing modality segmentation problem. SAR-Net uses mature components, but not as a simple stack: Gram prototypes summarize modality statistics, contrastive learning organizes the bank, retrieval selects sample-agnostic prototypes for missing modalities, and feature modulation injects them into segmentation features. We agree MRI intensity distributions are entangled with scanner/protocol variations and pathology, and are not “style” in the natural-image sense. Here, “style” is a proxy for modality-specific statistical cues, including modality appearance, protocol-related intensity variations, and distributional information. We do not claim perfect anatomy-independent disentanglement; because complete disentanglement remains difficult, SAR-Net uses Gram prototypes to aggregate these statistics as feature-level compensation cues, rather than reconstructing images or treating prototypes as anatomical representations. We will weaken “style” wording to avoid equating MRI intensity differences with natural-image style, and discuss limitations. Complexity, efficiency, and hallucination (R1): Although average gains may appear moderate, the results highlight robustness in extreme missing-modality cases and difficult tumor regions, the focus of clinical missing-modality scenarios. SAR-Net is slower than lightweight CNN/fusion baselines due to memory bank, retrieval, and Transformer modules, but avoids iterative image-level synthesis and is algorithmically more efficient than diffusion-style imputation. At inference, it uses one feature extraction pass and compact Gram-memory lookup, without generating missing MRI images. Thus, this is a cost-benefit/algorithmic-efficiency argument, not a runtime benchmark. We will clarify this tradeoff. Here, “mitigating unreliable feature hallucination” refers to avoiding image-level hallucination from synthesized missing-modality images; SAR-Net performs feature-level retrieval/compensation, so there is no image-level hallucination issue. Uncompared related works (R3): We thank R3 for noting TMFormer, KMD, and CCSD. Our baselines focus on reproducible representative methods from different paradigms under comparable missing-modality settings. KMD is a plug-and-play Koopman module, TMFormer refines Transformer via token merging, and CCSD appeared very recently and was not included. SAR-Net focuses on retrieval-based training/architecture using sample-agnostic style prototypes. Fair comparison requires unified implementation/protocol, which cannot be added in rebuttal. We thank the reviewers again and will incorporate these clarifications.



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.

    While Reviewer 2 strongly recommends acceptance, Reviewers 1 and 3 raise significant concerns regarding the inconsistency of experimental datasets, missing results for BraTS 2024, and lack of comparison with recent relevant baselines. The authors should clarify these experimental discrepancies, provide the missing BraTS 2024 results, and justify their methodological novelty against existing style-transfer and retrieval methods to secure acceptance.

  • 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.

    All reviewers suggest to accept this submission.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper introduces an innovative retrieval-based paradigm (SAR-Net) to address missing modalities in multi-modal brain tumor segmentation, a contribution recognized by the panel as clinically valuable and methodologically well-motivated. During the rebuttal phase, the authors successfully clarified key technical implementation settings, refined the conceptual framing of feature-level style retrieval, and resolved major concerns regarding reproducibility, resulting in a strong positive consensus. While minor reservations remain regarding evaluation dataset consistency across the ablation studies, the overall methodological merits and robust performance gains warrant acceptance, provided that all promised clarifications are meticulously integrated into the camera-ready version.



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 model’s approach is novel: framing missing modality imputation as style retrieval.

    Moreover, authors successfully address reviewers’ comments, I therefore recommend acceptance.



back to top