List of Papers Browse by Subject Areas Author List
Abstract
Multimodal MRI is essential for accurate brain tumor segmentation. However, acquiring all modalities at inference is often challenging in practice, which causes intrinsic uncertainty due to unavoidable information loss. Without modeling this uncertainty, existing methods encode incomplete evidence into deterministic representations that appear plausible but lack reliability. In this regime, we propose a probabilistic representation framework that models representations as Gaussian distributions, where their mean captures task information and their variance measures uncertainty from missing evidence. To make variance reflect information deficiency, we regularize the mean from each partial configuration toward its full-modality counterpart, while scaling the variance with the discrepancy between their aligned means. We further introduce a set-inclusive strategy that exploits the hierarchical structure of modality subsets and enforces an ordering constraint to maintain their consistent uncertainty relationships. Extensive experiments on BraTS 2018 and 2020 demonstrate that our approach offers superior performance over baselines across diverse missing-modality scenarios. Code and model checkpoint are available at https://github.com/atlas-sky/SIUM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1877_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/atlas-sky/SIUM
Link to the Dataset(s)
BraTS 2018: https://www.med.upenn.edu/sbia/brats2018/data.html
BraTS 2020: https://www.med.upenn.edu/cbica/brats2020/data.html
BibTex
@InProceedings{BaeSeu_SetInclusive_MICCAI2026,
author = { Baek, Seunghun AND Park, Jihwan AND Sim, Jaeyoon AND Lee, Hoseok AND Lee, Seungjoo AND Kim, Won Hwa},
title = { { Set-Inclusive Uncertainty Modeling for Robust Brain Tumor 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
This paper proposes a probabilistic representation framework that models data embeddings as Gaussian distributions, incorporating uncertainty to enhance multimodal brain tumor segmentation under missing-modality scenarios. Furthermore, it provides a theoretical analysis of how uncertainty influences model optimization.
- 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 data representations are modeled as Gaussian distributions, where the mean captures task-relevant information and the variance encodes uncertainty; this uncertainty is further leveraged to guide multimodal segmentation under missing-modality conditions. 2.A theoretical analysis is provided to demonstrate how uncertainty influences the optimization of representation learning. 3.A Set-Inclusive Modality Masking and Uncertainty Guidance strategy is proposed, which exploits the hierarchical structure of modality subsets and incorporates uncertainty-based ranking to regulate uncertainty levels.
- 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.Both the mean and variance are estimated using a 3D CNN. The manuscript should specify the network depth and discuss whether the number of layers affects the estimation accuracy, as well as the overall model performance. 2.The manuscript does not specify how the values of α and β are determined.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 demonstrates a certain degree of novelty, with a well-organized structure, clear presentation, and well-motivated design. It also provides theoretical justification for the proposed approach. The experimental evaluation is comprehensive and effectively demonstrates the superiority of the method.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have already provided clarifications regarding the model details and experimental settings.
Review #2
- Please describe the contribution of the paper
Brain tumor segmentation with missing MRI modalities is a clinically relevant setting. The main contribution is a probabilistic representation framework that models each modality configuration as a Gaussian embedding, where the mean captures task information and the variance reflects uncertainty caused by missing evidence. The paper further introduces a set-inclusive design that leverages the hierarchical relation among modality subsets to regularize uncertainty across different missing-modality settings.
- 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 proposed method is meaningful and novel for this setting, especially the combination of probabilistic embedding and set-inclusive uncertainty regularization. The experiments are thorough for this line of work, covering multiple missing-modality configurations and comparisons against competitive baselines.
- 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, Clinical motivation is clear, but the link from clinical insight to the method could be strengthened. 2, The method presentation is overly machine learning, it could be revised with more medical imaging perspective. 3, The connection between the proposed uncertainty modeling, optimization, and prediction is not explicit enough. 4, The related work is insufficient, especially with respect to more recent work in missing-modality segmentation [1][2] and uncertainty-aware medical image analysis [3][4]. [1] Bridging the gap in missing modalities: Leveraging knowledge distillation and style matching for brain tumor segmentation [2] Hypergraph tversky-aware domain incremental learning for brain tumor segmentation with missing modalities [3] DFuse-Net: Disentangled feature fusion with uncertainty-aware learning for reliable multi-modal brain tumor segmentation [4] A reliable framework for brain tumor segmentation via multi-modal fusion and uncertainty modeling
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 like this work and believe it makes a meaningful contribution to the missing-modality brain tumor segmentation. I think this work has the potential to extend more general incomplete data scenarios. The only concern is about presentation.
- 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.
The author addressed my concern, it is a good work.
Review #3
- Please describe the contribution of the paper
The paper presents a method for brain tumor segmentation in MRI using an uncertainty modeling approach in the latent space.
- 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.
Comparison on public BRATS dataset.
- 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 mathematical formulation is loose. No distinction being made for scalars, vectors, tensors, etc..
-
relationship between sigma_ki and sigma_sub or sigma_sup isn’t clear.
-
In Eq 8, for one of the proposed loss terms: the GradStop operator is a heuristic. This is a hack.
-
In Eq 9, , for one of the proposed loss terms: the loss function isn’t differentiable. Furthermore, enforcing this to be satisfied (e.g., for every component tensor) isn’t motivated well. Also, the specific term isn’t motivated well.
-
Variance magnitude hardly changes (only in third place of decimal) (in Fig 4) across missing modalities! variability around the curves missing; these trends don’t seem statistically significant.
-
in fig 4, why should higher uncertainty correlate with lower performance? this isn’t motivated well in theory or otherwise.
- The baselines are old and recent baselines are ignored, e.g., from miccai 2025.
-
Some older relevant baselines also doing uncertainty modeling are ignored, e.g., .Dorent R, Joutard S, Modat M, Ourselin S, Vercauteren T. Hetero-modal variational encoder-decoder for joint modality completion and segmentation. Springer; 2019:74–82
-
Standard deviations are missing in Table 1.Some baselines come very close to the proposed method. The improvements aren’t statistical significant.
-
Detailed empirical analysis on brats 2018 missing.
- with alpha = 0.01 « beta = 5, one of the loss functions (L_UA) doesn’t seem to make any difference. also, ablation study seems to show that L_UO hardly makes any significant improvement.
-
- 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.
(1) Strong Reject — must be rejected due to major flaws
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
-
serious weaknesses in empirical analysis, as described above.
-
theoretical formulation not very sound, as described above.
-
- 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.
Thanks for the feedback that clarifies some issues but not the major ones. I still have concerns about (a) loss-function differentiability and theoretical rigor, (b) no comparison with existing variational formulations, (c) the statistical significance (lack of std. dev.), and (d) the utility of the component analyzed through the ablation study where the improvement is tiny (as stated in the rebuttal).
Author Feedback
We sincerely thank R1 and R2 for their constructive feedback, and for recognizing the novelty, theoretical motivation, and experimental validation. However, R3 listed only “Comparison on public BraTS dataset” as a strength while giving a high-confidence strong-reject recommendation. Since several R3 claims conflict with the manuscript, we use part of this rebuttal to correct these factual issues, and appreciate R1/R2’s understanding.
[R1] We appreciate the suggestion on model capacity. For a fair comparison, we followed the RFNet backbone used by the baselines. fθ and fϕ are single 3D convolutional layers, which will be specified. α and β were selected by hyperparameter search over [0.001, 10], and we observed stable performance around the chosen values. [R2] We will revise Sec.1 and Sec.2 to make the clinical-to-method link explicit. T1/T1c/T2/FLAIR provide complementary cues for WT/TC/ET, so missing scans remove region-specific evidence and induce uncertainty. This motivates Gaussian partial embeddings, the full-modality anchor, and subset–superset uncertainty ordering. We will clarify that σ perturbs incomplete evidence to regularize μ during training, while μ is used for deterministic prediction. For related work, we will add [1–4] and clarify the differences. Specifically, [1] addresses missing-modality gaps via KD/style transfer, [2] targets sequential modality acquisition, and [3,4] focus on fusion/predictive uncertainty, whereas ours directly regularizes representation-level uncertainty for arbitrary partial-modality sets. [R3Q1] Formulation Sec. 2.2 specifies μ,σ∈R^{C×H×W×D}, so the formulation is tensor-valued. Eq. (2) states ∂ri/∂μi=I, which implies vectorization of the tensor. The notation k∈{sub,sup} is defined in Sec. 2.2.[R3Q2] Loss Validity In Eq. (8), GradStop, i.e., gradient detachment, is a standard ML mechanism used in distillation and self-supervised learning. Eq. (9) is piecewise differentiable, analogous to ReLU/hinge-style objectives. Since uncertainty is spatial- and channel-wise, element-wise ordering is suitable for enforcing the subset–superset relation. [R3Q3] Loss Scale α=0.01 and β=5 do not imply that L_UA is negligible, because L_UA and L_UO have different numerical scales and gradient behaviors. As in objectives combining task loss and regularization, coefficients should be interpreted with each loss magnitude and role. [R3Q4] Baselines Recent baselines were not ignored. Our comparison includes DC-Seg, a MICCAI 2025 method. Dorent et al. (2019) is not included as a baseline because it follows modality completion, whereas our method is imputation-free. Moreover, on BraTS2018, Dorent et al. reports lower WT DSC (80.1) than RFNet (85.49), while ours achieves 87.10.[R3Q5] Table 1 Results Due to space constraints, we omitted standard deviations, as in recent MICCAI papers such as mmFormer and DC-Seg. [R3Q6] Empirical Analysis on BraTS2018 Table 2 shows the same trend as BraTS 2020.We focus on BraTS 2020 due to space constraints. [R3Q7] Motivation/Validity of Fig. 4 Since removing modalities reduces available information, configurations with less evidence are expected to have higher uncertainty and lower segmentation performance. Fig. 4 empirically confirms this modality-induced trend. The small absolute changes mainly come from whole-volume averaging, where near-zero background uncertainty dominates. As shown in Fig. 3, tumor regions show much larger variance, exceeding 0.2 in some areas. Given μ≈1.3, the corresponding σ>0.4 introduces non-negligible perturbations to the task representation. Thus, the learned uncertainty remains meaningful despite small global-average changes. We will clarify this averaging effect and add variability bars if space allows. [R3Q8] Ablation Study ET segmentation is the most challenging BraTS subregion, and improving ET DSC from 63.74 to 64.61 with L_UO is a meaningful gain. [R3Q9] Reproducibility Sec. 3.2 states that the code and model checkpoints will be released.
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.
Please try to address the pointed weaknesses during the rebuttal.
- 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 reviewers appreciate the motivation and novelty of the proposed method, and the initial concerns have been mostly addressed during the rebuttal. I recommend accepting 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.
After considering the rebuttal and the updated reviews, I recommend acceptance. Two reviewers support accepting the paper and find the rebuttal adequate. Although one reviewer remains concerned, the raised issues do not appear to constitute fundamental methodological flaws that outweigh the paper’s contribution. Overall, the positive assessments and the value of the work support acceptance.
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.
I have mixed opinions on this paper, but I decided to go with acceptance.
On one side, what the authors called uncertainty is typically what is currently modeled by the covariance in (hierarchical) multimodal VAE. Therefore, the criticisms of R3 are valid, since there is already a principled formalism for this type of approach that the authors seem to ignore. This is especially the case in the rebuttal, with the wrong statement: “Dorent et al. (2019) is not included as a baseline because it follows modality completion, whereas our method is imputation-free.” This approach both synthesizes missing images (modality completion) and segments observed images simultaneously; segmentation is not performed after modality completion. Covariances are there to capture this uncertainty. The authors proposed regularizing the latent spaces $zi$ to align with the full modality setting, which could have been formulated as a KL divergence. Therefore, I agree that the maths are poor and that the authors seem to ignore some of the literature.
On the other hand, I disagree with R3, as the evaluation does include a recent baseline, and the proposed method appears to achieve state-of-the-art performance. This demonstrates the effectiveness of the proposed approach.
Overall, I think that the authors tend to overclaim some mathematical concepts without clear definitions, but the results are convincing. It is a borderline case, but I lean towards acceptance because the proposed approach remains conceptually interesting and obtains good performance.
