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

Abstract

Multi-parametric brain MRI is essential for glioma assessment, yet in clinical workflows, one or more modalities (FLAIR, T1, T1c, T2) are frequently missing, substantially degrading multimodal tumor segmentation. Existing solutions span multiple distilled specialists, single “catch-all” networks that fuse available modalities but often rely on naive zero-filling, and feature reconstruction approaches that recover missing representations but may suffer from mismatches between reconstructed and observed features, limited adaptability to missing-modality patterns, and additional sensitivity in pre-/post-treatment settings. We propose ReMiX-Seg (Reconstruction + Mixture-of-eXperts for Segmentation), a unified framework that combines lightweight token-level completion for missing modalities with mask-aware mixture-of-experts refinement to adapt fusion and denoising to each observed subset. To further stabilize training, we introduce Expert-wise Consistency Alignment, encouraging expert-specific agreement between reconstructed and observed feature distributions. Crucially, ReMiX-Seg is evaluated across both BraTS2023 (Adult Glioma Pre-treatment) and BraTS2024 (Adult Glioma Post-treatment), demonstrating robustness across domain-shifted settings. ReMiX-Seg achieves state-of-the-art performance on both benchmarks and remains consistently robust across all non-empty modality combinations, supporting practical deployment under incomplete MRI. The code is available at https://github.com/highquanglity/ReMiX-Seg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2915_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/highquanglity/ReMiX-Seg

Link to the Dataset(s)

BraTS2023 Adult Glioma Pre-treatment dataset: https://www.synapse.org/brats2023 BraTS2024 Adult Glioma Post-treatment dataset: https://www.synapse.org/brats2024

BibTex

@InProceedings{NghVan_ReMiXSeg_MICCAI2026,
        author = { Nghiem, Van Quang AND Lyu, Si-Qin AND Lin, Che},
        title = { { ReMiX-Seg: Latent Modality Completion Meets Expert Routing for Robust Glioma Segmentation } },
        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

    The authors propose ReMiX-Seg, a unified network for incomplete multimodal brain tumor segmentation. It addresses missing modalities in the latent space rather than image space via a Context-Gated Imputer (CGI) that uses cross-attention to reconstruct missing tokens. Furthermore, it introduces a Serialized Mamba-Mixture-of-Experts Fusion (SeMoF) block to adaptively route and fuse features based on the availability mask. Training is stabilized using an Expert-wise Consistency Alignment (ECA) loss to minimize the Jensen-Shannon divergence between reconstructed and teacher expert distributions.

  • 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.By performing missing-modality completion entirely at the token level via the CGI, the framework completely bypasses the massive computational overhead associated with pixel-level generative synthesis (e. g. , GANs or Diffusion models). 2.The explicit integration of a load-balancing loss (L[bal]) and the ECA loss prevents the notorious “routing collapse” problem common in Mixture-of-Experts architectures, ensuring all experts are utilized and aligned with the fully-observed teacher. 3.Testing the framework on both pre-treatment (BraTS 2023) and post-treatment (BraTS 2024) cohorts is a massive strength. Demonstrating robustness against the domain shift introduced by post-operative anatomy (Resection Cavity) elevates the clinical relevance of the empirical claims. 4.The authors rigorously report performance across all 15 possible non-empty modality combinations, transparently showcasing the framework’s degradation curve under severe missingness.

  • 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 SeMoF module serializes modality tokens and processes them with Mamba. However, Mamba’s foundational mathematical advantage [$] is strictly in handling ultra-long sequences with O(L) complexity. Serializing features across only 4 modalities (even if patchified) does not typically reach the sequence lengths where SSMs outperform standard self-attention. Please, explicitly justify the mathematical or computational necessity of using Mamba over standard Transformer attention for this specific fusion step. Does it actually improve throughput, or is it an architectural embellishment? 2.The framework is heavily compounded, integrating ConvNeXt, Transformers, Cross-Attention (CGI), Mamba, and Mixture-of-Experts (MoE) into a single pipeline. While empirical results are strong, it is difficult to isolate the fundamental driver of success. So, provide a precise parameter-count and FLOPs comparison against baseline models (like IM-Fuse or DC-Seg) to prove this compounded architecture does not incur prohibitive computational debt. 3.BraTS 2024 introduces the Resection Cavity (RC) label. Unlike tumor biology which follows general anatomical patterns, surgical cavities are uniquely human-made and highly unpredictable. Reconstructing latent features for a missing structural modality (like T1c) in a post-operative brain carries a high risk of hallucinating structural tissue that isn’t there. Please, clarify if the CGI module exhibits any ‘hallucination’ tendencies specifically regarding the RC class when highly structural modalities are missing, and discuss the clinical safety of latent imputation for post-surgical anatomy.

    [$] Gu, Albert, and Tri Dao. “Mamba: Linear-time sequence modeling with selective state spaces. “ Conference on Language Modeling (2024).

  • 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?

    The paper effectively addresses missing MRI modalities using a mathematically sound Mixture-of-Experts (MoE) approach and provides robust validation across both pre and post-treatment cohorts (BraTS 2023/2024). However, my primary reservation is architectural over-engineering. Combining ConvNeXt, Transformers, Mamba, and MoE risks unnecessary computational bloat. Specifically, deploying Mamba, a model designed for ultra-long sequences, to fuse a limited number of multimodal tokens lacks clear mathematical justification compared to standard self-attention.

  • 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



Review #2

  • Please describe the contribution of the paper

    This paper proposed a unified single-network framework ReMiX-Seg, the main contributions of the paper include:1.It proposed several modules like Shift-Gated Token Refiner, Context-Gated Imputer, Serialized MambaMixture–of–Experts Fusion and Expert-wise Consistency Alignment for robust brain tumor segmentation with incomplete modalities; 2.It was the first missing-modality segmentation framework evaluated systematically in both BraTS2023 and BraTS2024 settings; 3.It achieved state-of-the-art performance on both benchmarks and remains consistently robust across all non-empty modality combinations.

  • 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.This paper presents comprehensive implementations across different domains (pre-treatment and post-treatment glioma cohorts) and compares the proposed approach against a diverse array of state-of-the-art (SOTA) methods to convincingly demonstrate its effectiveness.

    2.The methodological descriptions and their corresponding diagrams are clear and rigorous. Furthermore, the proposed method is driven by a well-defined motivation and demonstrates a solid degree of novelty.

    3.The paper provides thorough ablation experiments regarding the proposed loss functions and the number of Cross-Attention layers within the CGI module. These studies effectively substantiate the individual contributions of each loss component to the overall method.

  • 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.Missing Hyperparameter Ablations: The paper lacks ablation studies for the various loss weighting coefficients; however, given the strict page limits, this omission is understandable.

    2.Unsubstantiated Claims of Cross-Cohort Robustness: Although the authors claim that “ReMiX-Seg achieves state-of-the-art (SOTA) performance and demonstrates consistent robustness across all non-empty modality subsets under cross-cohort distribution shift,” the actual experimental setup fails to evaluate this cross-cohort distribution shift. The paper utilizes both the BraTS2023 and BraTS2024 datasets; however, according to the experimental details, these two datasets are trained and evaluated entirely independently. This intra-dataset setup does not capture or reflect the impact of distribution shifts. To genuinely demonstrate the method’s robustness against distribution shifts, the authors should employ a cross-dataset evaluation protocol—for instance, training on the BraTS2023 dataset and evaluating on the BraTS2024 dataset.

  • 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 paper is very well-developed and innovative in terms of motivation, methodology, and experiments. My only concern is that its experiments do not support its claim that “ReMix-Seg demonstrates strong generalization under domain shift”. In particular, it emphasizes that ‘ReMiX-Seg is among the first missing-modality segmentation frameworks evaluated systematically in both BraTS2023 and BraTS2024 settings’ as one of the paper’s contributions. However, its experimental results can only prove that the method is independently applicable to the two datasets, but it does not overcome domain shifts.

  • 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



Review #3

  • Please describe the contribution of the paper

    The paper proposes a brain glioma segmentation model that can adapt to various missing-modality scenarios. The model consists of multiple encoders, gating units, a modality fusion module, and a decoder. Compared to existing multi-expert and multi-encoder approaches, the method adopts a teacher–student framework, where a teacher network trained on full modalities guides a student network under missing-modality settings. The student is encouraged to approximate the teacher’s feature representations, thereby improving segmentation performance. To achieve this, the authors design gating units, a Mamba-based feature fusion module, and a joint loss. Extensive experiments on the BraTS 2023 and BraTS 2024 datasets demonstrate the effectiveness of the proposed method.

  • 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.
    • Effectiveness: The design motivation and necessity of each module are clearly described in the method section.
    • Comprehensive experiments: The method shows strong performance across two datasets, supported by comparative experiments, ablation studies, and qualitative visualizations.
    • Reproducibility: Anonymous code is provided, making the method reproducible.
  • 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.
    • Potential over-design: 1.The method encourages the student network to mimic the teacher encoder’s features at multiple downsampling stages, introducing two losses. The L_{\text{recon}} loss measures similarity after the context-gated imputer module, and the L_{\text{eca}} loss is applied after the SeMoF module. If the context-gated imputer already learns sufficiently good features, the additional gain from L_{\text{eca}} may be limited and could introduce a risk of overfitting, which appears to be partially reflected in the results in Table 3. 2.The formulation of losses is not fully clear. While similarity (e.g., L1 distance) is used in L_{\text{recon}}, distribution-based divergence (e.g., JS divergence) is used elsewhere. However, the paper does not clearly explain how features are transformed into probability distributions (e.g., via sigmoid or softmax), nor the necessity of this transformation. Empirical justification (e.g., demonstrating that distribution-based metrics better capture feature differences) is lacking.
    • Insufficient clarity in method illustration: 1.The overall network structure is not sufficiently clear, particularly in the overview, which makes it difficult to fully understand the composition and data flow of the model. 2.Some elements in the figures are not explained. For example, in Fig. 1 (top-left), multiple modality feature extractors are connected by a “+” symbol, but it is unclear whether this denotes element-wise addition or channel-wise concatenation.
  • 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?

    The proposed method incorporates a multi-expert design with a teacher–student framework, and is supported by thorough experimental validation and reproducibility. The following points may further improve the work: 1.On the MoE formulation:
Although the method is described as a Mixture-of-Experts (MoE) framework, it involves only a small number of experts (e.g., four), and does not fully explore key aspects such as routing mechanisms or expert specialization. The proposed routing strategy mainly involves a voting mechanism. It would be helpful to analyze whether different experts indeed learn modality-specific features. Alternatively, it may be worth discussing whether the method is more appropriately characterized as a multi-branch or dynamic adapter framework. 2.Clarity of overall design:
While each module is individually well explained, the interactions between modules and the overall feature flow are not sufficiently clear. A more holistic description of the architecture would improve readability. 3.Writing issues:
There are minor writing inconsistencies. For example:

    • On page 5, the phrase “w_e its mixture weight,” appears incomplete.
    • In the SeMoF section, the expression “z = \text{Mamba}(\text{Ser}(z_i))” includes an undefined term “Ser”, which may refer to “Serialized”, but is not explicitly clarified.
  • 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 the reviewers for the constructive comments. Following MICCAI policy, we do not add new experimental results, including loss-weight sweeps requested by R3; below we clarify submitted results and model profiling.

1.SeMoF/Mamba and cost (R2,R4). Mamba is not used to fuse only four modality IDs. Each MRI is encoded by a CNN branch for local 3D features, and a modality-specific Transformer captures intra-modality context at the deepest level. SeMoF performs inter-modality fusion: it serializes spatial modality tokens with a learnable fused stream, e.g., [F_t,T1_t,T1c_t,T2_t,Fuse_t] over locations t, and Mamba mixes this ordered 1D spatial-modality sequence before extracting the fused stream. Thus, Mamba is a linear mixer for serialized spatial-modality fusion, not an embellishment or replacement for intra-modality attention. Under the same 128^3 profiling, ReMiX-Seg has 71.95M total/60.45M active params and 297.84 GFLOPs, close to IM-Fuse (60.57M active, 270.85 GFLOPs) and much lower FLOPs than M3FeCon/DC-Seg (764.76/833.82).

2.MoE formulation (R4). Our MoE is compact for 3D segmentation. The 4 experts match the four-modality problem scale and form a lightweight expert bank, but no expert is hard-coded to one modality. The router uses token mean/variance, availability mask, and stream flag to produce top-2 input-dependent weights. Hence experts are mask-conditioned dynamic refiners for different missing-subset and reconstruction/fusion regimes, not fixed modality branches or voting heads. Continuous top-2 mixtures cover more than four missing-pattern behaviors while keeping computation practical. Load balancing discourages routing collapse without deploying multiple networks.

3.Domain-shift claim (R3). This is not a strict source-to-target transfer experiment because BraTS2023 and BraTS2024 are trained/evaluated independently. Our claim is systematic validation in two clinically distinct, domain-shifted benchmarks: pre-treatment BraTS2023 and post-treatment BraTS2024, the latter including RC and therapy-related anatomical changes. We will phrase this as “consistent robustness across pre-/post-treatment domain-shifted benchmarks,” not as solving cross-dataset domain generalization.

4.RC safety and latent imputation (R2). CGI completes task-level latent tokens rather than synthesizing missing images. SeMoF also receives the availability mask, so reconstructed tokens are processed as imputed evidence rather than observed anatomy, limiting the risk of treating hallucinated image content as real structure. From submitted BraTS2024 RC Table 2, ReMiX-Seg averages 64.7 Dice over the seven T1c-missing subsets, compared with 63.8 for DC-Seg and 64.2 for IM-Fuse, and achieves the best overall RC average of 68.2.These results support robust latent imputation for RC; safety validation beyond Dice is still needed before clinical deployment.

5.Losses and figure notation (R3,R4). R3 asks about loss-weight ablations; such sweeps are useful, but new experimental results are disallowed in rebuttal. The weights are already specified in the paper: lambda_rec=1.0, lambda_eca=0.4, lambda_bal=0.1.More importantly, Table 3 isolates each loss function. L_rec and ECA are complementary rather than redundant: L_rec uses Smooth-L1 before fusion to match missing tokens to complete-reference tokens, while ECA acts after SeMoF and aligns expert-response distributions between reconstructed and complete-reference streams. Softmax converts expert outputs to normalized distributions before JS divergence. Table 3 supports this design: adding ECA to reconstruction improves performance and the full objective performs best. In Fig. 1, the “+” near modality encoders denotes stacking modality features for masking, not addition. Encoders are independent; interaction starts after Stack/Mask and then CGI/SeMoF. We will make symbols explicit: Stack/Concat/Ser/Gate/residual-add, with Ser(.) denoting interleaving spatial modality tokens with fused tokens.



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.

    The reviewers are broadly positive, acknowledging a clear methodological contribution, strong empirical performance, and a comprehensive evaluation across both BraTS 2023 and BraTS 2024.The raised concerns are mostly about positioning and clarity rather than core validity: in particular, reviewers questioned whether the Mamba/MoE design is somewhat over-engineered, whether the domain-shift claims are overstated without true cross-dataset evaluation, and whether some architectural and loss-design choices need clearer justification. Overall, however, the consensus is that the technical contribution is novel and well executed, the experiments are convincing, and the weaknesses appear addressable in revision, supporting an early acceptance decision.



back to top