List of Papers Browse by Subject Areas Author List
Abstract
Multimodal brain tumor segmentation typically leverages multiple MRI modalities, yet incomplete modality acquisition is common in clinical practice due to protocol heterogeneity and scan failures. Although recent methods maintain segmentation accuracy under missing modality conditions, they frequently overlook prediction reliability, leading to miscalibrated confidence estimates that hinder clinical adoption. Existing calibration techniques are largely modality-agnostic or assume that prediction difficulty decreases monotonically as additional modalities become available. However, in brain tumor segmentation, prediction difficulty depends primarily on which modalities are absent rather than how many, leading to combination-specific and spatially heterogeneous calibration errors. To address this, we propose Missing Modality-Aware Local Temperature Scaling (MMA-LTS), a post-hoc voxel-wise confidence calibration method. It estimates a spatially adaptive temperature field conditioned on a modality-availability learnable token and a voxel-wise difficulty score. Experiments on BraTS 2020 and FeTS 2024 show that MMA-LTS improves calibration while preserving the segmentation accuracy of state-of-the-art models across diverse missing-modality scenarios, thereby enhancing trustworthiness toward clinical deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1236_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
BraTS 2020 dataset: https://www.med.upenn.edu/cbica/brats2020/data.html
FeTS 2024 dataset: https://www.synapse.org/Synapse:syn54079892/wiki/626481
BibTex
@InProceedings{LeeSol_Missing_MICCAI2026,
author = { Lee, Sol AND Kim, Hyunji AND Hong, Sungrae AND Han, Donghee AND Yi, Mun Yong},
title = { { Missing Modality-Aware Calibration for Trustworthy Brain Tumor Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses the largely overlooked issue of confidence calibration in missing-modality multimodal brain tumor segmentation and proposes a novel post-hoc calibration framework, MMA-LTS, while revisiting a fundamental assumption in existing methods. Specifically, the authors argue that conventional calibration approaches implicitly assume that prediction reliability increases monotonically with the number of available modalities. However, in practice, different MRI modalities contribute unevenly to distinct tumor subregions, making prediction performance and uncertainty primarily dependent on which modalities are missing rather than how many, thereby inducing spatially heterogeneous calibration errors. To address this, MMA-LTS introduces three key components: (1) a learnable modality availability token to explicitly model modality combination dependencies; (2) a voxel-wise difficulty score (VDS) based on Mahalanobis distance to capture spatial uncertainty; and (3) a temperature estimator that integrates spatial-difficulty gating with FiLM-based conditioning to produce a voxel-wise temperature field for locally adaptive scaling. Extensive experiments demonstrate that MMA-LTS consistently reduces calibration error (ECE) across multiple datasets and segmentation backbones without compromising segmentation accuracy, validating its effectiveness in addressing both modality-combination-specific and spatially heterogeneous miscalibration.
- 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 main strengths of this paper lie in its clinically relevant problem setting and its focus on an often overlooked yet important issue in missing-modality segmentation, namely the reliability of model confidence under incomplete modality conditions rather than segmentation accuracy alone. Methodologically, the novelty of MMA-LTS stems from its combination of modality-combination awareness and spatially adaptive calibration. Unlike conventional approaches that rely on modality count or global temperature assumptions, the proposed method explicitly models which modalities are missing via learnable modality tokens, capturing fine-grained modality dependencies. In addition, it introduces a voxel-wise difficulty score (VDS) to characterize spatial uncertainty and generates a voxel-wise temperature field for locally adaptive calibration, which is better suited to address modality-specific and spatially heterogeneous miscalibration. The method is implemented as a post-hoc module without requiring retraining of the segmentation backbone, making it practically appealing from an engineering perspective. Experimentally, the paper provides relatively comprehensive validation across multiple datasets, segmentation backbones, and modality combinations, along with ablation studies supporting the effectiveness of each component. Overall, the strengths of this work lie in a well-motivated problem, a targeted methodological design, and reasonably thorough empirical 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.
In its current form, the paper does not yet meet the publication bar. The main concern is that the proposed contribution is largely confined to a specific calibration module rather than representing a broader architectural or methodological advance. MMA-LTS is essentially built as a post-hoc extension on top of an existing segmentation backbone, and the claimed novelty mainly comes from combining modality-aware tokens, FiLM conditioning, and voxel-wise temperature scaling within this narrow module. As a result, the work appears more like a task-specific engineering adaptation of existing calibration techniques than a genuinely new framework that advances missing-modality segmentation at the network level. The method does not introduce a new representation learning strategy, a new missing-modality learning paradigm, or a new backbone design; instead, it refines confidence correction after prediction. This makes the overall contribution somewhat limited for a top-tier venue.
A related concern is that the practical applicability of the method seems highly restricted to the specific setting studied in the paper, namely brain tumor MRI datasets such as BraTS 2020 and FeTS 2024.The proposed design is strongly tied to the characteristics of multimodal brain MRI, where different sequences have relatively well-understood and complementary roles for different tumor subregions. It is therefore unclear whether the same formulation would remain effective for other anatomical sites, other imaging modalities, or more heterogeneous multimodal clinical settings where the relationships among modalities are less structured. Since all experiments are conducted only on heavily standardized brain MRI benchmarks, the paper does not provide sufficient evidence that the method is broadly useful beyond this narrow application domain.
More generally, the paper overstates the scope of its contribution. While the problem of calibration under missing modalities is important, the current solution is specialized to a particular post-hoc module and validated only in a very specific experimental context. Without stronger evidence of architectural generality, cross-task transferability, or applicability beyond BraTS-style brain MRI benchmarks, the impact of the work remains limited. For these reasons, I do not think the paper, in its current form, provides a sufficiently broad or substantial advance to justify acceptance.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I assign this paper a Reject because, in its current form, it does not provide a sufficiently strong contribution for acceptance. The main issue is that the technical novelty appears limited: MMA-LTS is essentially a task-specific post-hoc calibration module built on top of an existing segmentation pipeline, rather than a broader architectural or methodological advance for missing-modality learning. Its main components—modality tokens, FiLM-based conditioning, and a Mahalanobis-based voxel difficulty map—are individually standard, and the overall contribution reads more as an engineering adaptation of existing calibration techniques than a genuinely new framework. In addition, the method seems narrowly tied to the specific setting of BraTS/FeTS-style multimodal brain MRI, where the roles of different modalities are relatively well understood, and the paper does not provide convincing evidence that the approach would generalize to other anatomies, modalities, or less curated clinical scenarios.
A further concern is that an important part of the method is insufficiently specified. In the VDS formulation, the paper defines voxel feature-label pairs using y_{n,v} = Y_n(v) and then computes the class-conditional Mahalanobis distance based on that label, even though the method is intended for calibration on D_test, where voxel-wise ground-truth labels are unavailable. The manuscript does not clearly explain whether VDS is computed using true labels, predicted labels, or pseudo-labels at inference time. This ambiguity directly affects reproducibility and raises a potential validity issue. Moreover, although the paper repeatedly emphasizes “trustworthiness” and “clinical deployment,” the empirical evidence is still largely limited to ECE improvements on benchmark datasets, without broader reliability analysis, detailed failure-case discussion, or stronger clinical validation. Overall, while the problem itself is important, the current submission is weakened by limited novelty, narrow applicability, an unclear implementation detail in VDS, and conclusions that are stronger than what the presented evidence can fully support. For these reasons, I believe the paper should be rejected in its current form.
- 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
Review #2
- Please describe the contribution of the paper
1) Proposing Missing Modality-Aware Local Temperature Scaling (MMA-LTS). This is the first post-hoc confidence calibration method designed specifically for missing-modality scenarios in multimodal brain tumor segmentation. It corrects calibration errors that are combination-specific and spatially heterogeneous across different modality combinations. 2) Introducing modality availability learnable tokens and voxel-wise difficulty scores. By encoding modality availability via binary masks and estimating voxel-wise difficulty based on Mahalanobis distance, the method enables fine-grained, spatially adaptive temperature field prediction for various missing-modality combinations. 3) Extensive experimental validation. Comprehensive experiments on two public datasets, BraTS 2020 and FeTS 2024, demonstrate the significant advantages of MMA-LTS over multiple state-of-the-art segmentation models.
- 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 points out that existing methods for missing-modality segmentation often ignore the issue of confidence calibration. More importantly, the authors observe a clinically meaningful phenomenon in brain tumor segmentation: the difficulty of prediction depends more on which specific modality is missing than on the total number of missing modalities. 2)To address this, the paper introduces a spatially adaptive temperature field, predicted using modality availability tokens and voxel-wise difficulty scores. The design is logically sound and well tailored to the anatomical characteristics of brain tumor segmentation. 3)The experimental validation is thorough, covering comparison experiments, ablation studies and analyses of hyperparameter sensitivity.
- 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) It is recommended to include failure case analyses to enhance the clinical interpretability of the model. 2) Although the method encodes different modality combinations via modality availability tokens, effective calibration still requires exposure to specific combinations in the validation set during training. The generalization capability for extreme modality combinations not encountered during training has not been sufficiently validated. It is suggested to do more experiments to validate the superiority of this 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 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.
(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 novelty, experimental completeness, and writing quality all meet the acceptance standards.
- 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 authros created a plug-in post-hoc calibration method that corrects modality-combination-specific and spatially heterogeneous miscalibration in brain tumor segmentation under missing modalities.
- 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 study proposes a novel method to correct modality combinations and address miscalibration in brain tumor segmentation under missing-modality settings. The authors use robust datasets and evaluate multiple approaches.
- 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.
Lack of external validation and formal statistical comparison of performance metrics.
- 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 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.
(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?
A very well-written paper with strong methodology and results, demonstrating immediate clinical impact.
- 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 sincerely thank all reviewers for their constructive feedback. Following the meta-reviewer’s suggestions, we clarify below the key concerns raised by the reviewers.
(Reviewer #2) VDS Implementation: As stated in Section 2.1, GT labels are used only during offline estimation of class-conditional statistics (µₗ, Σ) from D_train, after which these statistics are fixed. At inference time, predicted labels serve as pseudo-labels to compute the RMD without requiring GT labels. This is feasible because the pretrained segmentation model F has frozen weights, keeping the feature space invariant, such that the distribution statistics estimated from training data remain valid reference points at inference. We will provide a more explicit description of the inference procedure.
(Reviewer #2) Novelty and Generalizability: The core contribution lies in the problem formulation itself: the observation that calibration error in brain tumor segmentation varies by which modalities are absent rather than how many has not been addressed in prior work. Although individual components are standard, their combination is essential for solving this problem, as demonstrated by our ablation study. We recognize that extension to anatomical sites beyond brain tumors remains an important direction, and will consider this in future work.
(Reviewer #2, #3): Failure Cases and Limitations: Explicit failure case analysis would strengthen the paper, and we will add a discussion focusing on residual overconfidence in boundary regions. We acknowledge the absence of additional reliability metrics and validation in real clinical settings as limitations of the current work.
(Reviewer #3) Generalization to Unseen Modality Combinations: The modality availability token encodes each modality’s status independently and combines them linearly (Eq. (3)), allowing unseen combinations to be naturally represented at inference. We nonetheless recognize that calibration performance may degrade when certain combinations are entirely absent from D_val. Due to rebuttal constraints, we could not include new experimental results, but we will explore enforcing full combination coverage in D_val or generalization validation for unseen modality combinations as potential mitigation strategies.
(Reviewer #5) External Validation and Statistical Comparison: Consistent improvements across two independent benchmarks (BraTS 2020 and FeTS 2024) provide evidence for the applicability of our method. We agree that additional external validation and formal statistical comparison of performance metrics would further strengthen the work, and will look to address these in future work.
We will make every effort to incorporate the reviewers’ suggestions into the final version.
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.
This paper received mixed scores 2, 5, 5.Concerns include the method’s generalizability and presentation clarity. AC reads the paper and agrees with the two reviewers with a score 5 that this paper is clear, well-written, and sufficiently compared with SOTAs, and well above the bar for acceptance. AC encourages the authors to further clarify concerns raised by reviewers, and consider adding discussions on method limitations and failure cases.