acceptodds
Under review as a conference paper at ICLR 2027

Modeling Uncertainty from Missing Modalities for Brain Tumor Segmentation

Abstract

Brain tumor segmentation from multimodal MRIs is challenging when one or more modalities are missing. Many existing methods focus on multimodal representation learning or knowledge distillation but pay limited attention to how missing modalities affect predictions and the uncertainty associated with them. In this work, we study uncertainty from missing modalities under partial observation. From a probabilistic perspective, we characterize uncertainty from missing modalities by the total conditional variance of the posterior under full observation and show that this quantity equals the expected squared change between the class posteriors under full and partial observations. Building on this insight, we adopt Dirichlet modeling for predictions under missing modalities and learn the Dirichlet variance by matching it to the discrepancy between predictions under full and partial observations. This allows the estimated uncertainty to reflect the effect of missing modalities on the current prediction. Extensive experiments on BraTS2020 and BraTS2024 demonstrate that our method achieves strong performance across diverse missing-modality settings. Uncertainty analyses show closer agreement between the estimated uncertainty and the prediction change caused by missing modalities.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
missing modalities, uncertainty estimation, multimodal learning, brain tumor segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_XHAuOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses missing modalities in multimodal MRI brain tumor segmentation. The authors train a Dirichlet-based uncertainty estimator using the difference between predictions under full and partial observations, and evaluate segmentation performance on BraTS2020 and BraTS2024. However, the uncertainty evaluation and some experimental details require further clarification.

Critical strengths

The paper observes that missing modalities can substantially change predictions even when model confidence remains high. Studying their effect on the current prediction is worthwhile. The theoretical analysis also motivates the training objective: the authors distinguish the prediction change for an individual sample from the expected change given a partial observation, and explain why the former can serve as a supervision signal for uncertainty learning under squared loss.

Critical weaknesses
  1. The segmentation setup and the stability of the results need further clarification.
    • For BraTS2024, the paper reports only WT, TC, and ET results. It does not explain how the resection cavity (RC) label is handled or report segmentation results for RC.
    • The main text does not report results across multiple random seeds, variances, or confidence intervals. For modality combinations with small Dice differences, it remains unclear whether the improvements are consistent.
  2. The current evidence does not fully establish the utility of the uncertainty estimates.
    • The uncertainty evaluation primarily measures agreement with the training target. RMSE and UCE compare the estimated uncertainty with the difference between model predictions under full and partial observations. This shows that the model fits the proposed proxy target, but does not establish whether its uncertainty estimates identify actual segmentation errors or reflect clinical risk.
    • Competitive uncertainty estimation baselines are missing. The learned concentration is compared only with several fixed concentrations, making it difficult to assess its advantage over predictive entropy, model ensembles, or other uncertainty estimation methods.
Questions for authors

I would appreciate clarification on the weaknesses above, particularly the following:

  1. How is the RC label handled in the BraTS2024 experiments?
  2. Can the authors compare their approach with existing uncertainty estimation methods and assess whether its uncertainty estimates predict segmentation errors or performance degradation caused by missing modalities?

If these concerns are adequately addressed, I would be willing to consider raising my rating during the rebuttal period.

2 comments
AuthorsOpenReview ↗

Thank you for your thoughtful review. We address your concerns as follows.

Q1.How is the RC label handled in the BraTS2024 experiments?

A1. In our experiments, the RC label was mapped to the background and excluded from evaluation.

Q2.Can the authors compare their approach with existing uncertainty estimation methods and assess whether its uncertainty estimates predict segmentation errors or performance degradation caused by missing modalities?

A2. We compared our method with two uncertainty estimation approaches: predictive entropy and maximum softmax probability (MSP). We evaluated segmentation error detection and risk–coverage performance using AUROC, average precision (AP), and the area under the risk–coverage curve (AURC). Experiments covered four single-modality settings, with results averaged across the WT, TC, and ET regions. Segmentation errors were identified by comparing predicted and ground-truth region masks. The results are presented in the table below and have been added to the revised manuscript.

\hline \mathrm{Metric} & \mathrm{Entropy} & \mathrm{MSP} & \mathrm{Ours} \\\\ \hline \mathrm{AUROC}\uparrow & 0.5795 & 0.6084 & \mathbf{0.6163} \\\\ \hline \mathrm{AP}\uparrow & 0.6326 & 0.6645 & \mathbf{0.6753} \\\\ \hline \mathrm{AURC}\downarrow & 0.5651 & 0.5541 & \mathbf{0.5511} \\\\ \hline \end{array}$$ Q3.The main text does not report results across multiple random seeds, variances, or confidence intervals. For modality combinations with small Dice differences, it remains unclear whether the improvements are consistent. **A3.** Due to time constraints, we could not repeat all experiments. **We conducted three additional runs with different random seeds on BraTS2024 and compared their mean and standard deviation with the original results.** We report our results averaged over three random seeds alongside the original results of the competing methods, as summarized in the table below. $$\begin{array}{l|c|c|c|c} \hline \mathrm{Method} & \mathrm{WT} & \mathrm{TC} & \mathrm{ET} & \mathrm{Mean} \\\\ \hline \mathrm{RFNet} & 78.59 & 47.82 & 47.04 & 57.82 \\\\ \hline \mathrm{mmFormer} & 77.35 & 42.84 & 41.56 & 53.92 \\\\ \hline \mathrm{IMS}^{2}\mathrm{Trans} & 75.65 & 49.69 & 47.87 & 57.73 \\\\ \hline \mathrm{M}^{2}\mathrm{FTrans} & 74.48 & 47.49 & 46.79 & 56.25 \\\\ \hline \mathrm{MSTKD} & 80.52 & 48.84 & 48.02 & 59.13 \\\\ \hline \mathrm{Ours} & \mathbf{83.42\pm0.58} & \mathbf{51.18\pm0.87} & \mathbf{51.25\pm1.01} & \mathbf{61.95\pm0.48} \\\\ \hline \end{array}$$ **Overall, our method achieves competitive performance across all three tumor regions and the highest mean Dice score, demonstrating a favorable overall performance advantage over the compared methods.**

I thank the authors for their response to my review comments and for their detailed clarification of the relevant points. Their response has helped me better understand the paper. After considering it carefully, I believe that my current overall rating still reflects my assessment of the paper, so I am inclined to maintain my original rating.

Reviewer_6MroOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates uncertainty estimation in brain tumor segmentation with missing MRI modalities. It establishes a probabilistic relationship between missing-modality uncertainty and the expected squared discrepancy between full- and partial-modality posteriors. Based on this analysis, the authors propose a Dirichlet-based uncertainty learning framework that uses prediction discrepancies as supervision, combined with knowledge distillation and Brier loss. Experiments on BraTS2020 and BraTS2024 demonstrate competitive segmentation performance across different modality combinations, while uncertainty calibration and ablation studies support the effectiveness of the proposed approach.

Critical strengths
  1. Clear motivation and theoretical foundation. The paper addresses an important but relatively underexplored problem of uncertainty caused by missing modalities. Its probabilistic analysis provides a principled connection between prediction discrepancies and uncertainty estimation.

  2. Consistent experimental improvements. The method achieves competitive segmentation performance across different missing-modality settings on BraTS2020 and BraTS2024. The uncertainty analysis and ablation studies further support the effectiveness of the proposed approach, although broader validation is still needed.

Critical weaknesses

W1. Limited methodological novelty. The framework mainly combines established techniques, including Dirichlet modeling, knowledge distillation, and Brier loss. The main novelty lies in using full-partial prediction discrepancy for uncertainty supervision, but the advantage over simpler uncertainty regression approaches remains unclear. W2. Gap between theoretical formulation and practical modeling. The theoretical uncertainty is defined as a conditional variance, but the proposed Dirichlet variance has a restricted range determined by the predicted probabilities. It is unclear whether this parameterization can adequately represent the target uncertainty in all cases. W3. Insufficient validation of uncertainty modeling. The uncertainty evaluation mainly compares learned and fixed Dirichlet concentrations using prediction discrepancy as the reference. Comparisons with established uncertainty estimation methods and evaluations of error detection or prediction reliability are needed. W4. Unclear contribution of uncertainty learning to segmentation. The paper states that uncertainty learning is detached from segmentation features, yet the corresponding loss improves segmentation performance. The gradient flow and mechanism behind these improvements require clarification. W5. Incomplete comparisons with recent methods. The experiments do not include several relevant 2026 methods, such as UniME (CVPR 2026), Virtual Nodes Guided Dynamic Graph Neural Network (CVPR 2026). Uni-Encoder Meets Multi-Encoders: Representation Before Fusion for Brain Tumor Segmentation with Missing Modalities Virtual Nodes Guided Dynamic Graph Neural Network for Brain Tumor Segmentation with Missing Modalities

Questions for authors

refer to the weakness

2 comments
AuthorsOpenReview ↗

We sincerely appreciate your valuable comments. Our responses are provided below.

Q1. Limited methodological novelty. The framework mainly combines established techniques, including Dirichlet modeling, knowledge distillation, and Brier loss. The main novelty lies in using full-partial prediction discrepancy for uncertainty supervision, but the advantage over simpler uncertainty regression approaches remains unclear.

A1. We replaced the Dirichlet concentration head with a scalar regression head trained using the same prediction-discrepancy target.

\hline \mathrm{Method} & \mathrm{WT} & \mathrm{TC} & \mathrm{ET} & \mathrm{Mean} \\\\ \hline \mathrm{Scalar\\ regression} & 82.89 & \mathbf{50.77} & 50.27 & 61.31 \\\\ \hline \mathrm{Dirichlet} & \mathbf{84.08} & 50.46 & \mathbf{52.25} & \mathbf{62.26} \\\\ \hline \end{array}$$ Compared with the scalar regression, our method achieves higher Dice scores in WT, ET, and overall, while the scalar head performs slightly better in TC. Q2. Gap between theoretical formulation and practical modeling. The theoretical uncertainty is defined as a conditional variance, but the proposed Dirichlet variance has a restricted range determined by the predicted probabilities. It is unclear whether this parameterization can adequately represent the target uncertainty in all cases. **A2.** We evaluated variance-bound violations under four single-modality settings on 270 held-out BraTS2024 cases. **Although violations occur, their rates remain low, supporting the suitability of our Dirichlet parameterization.** $$\begin{array}{l|c|c|c|c} \hline \mathrm{Voxel\\ group} & \mathrm{T2\\ (\\%)} & \mathrm{T1\\ (\\%)} & \mathrm{T1ce\\ (\\%)} & \mathrm{FLAIR\\ (\\%)} \\\\ \hline \mathrm{All\\ brain\\ voxels} & 4.067\pm2.694 & 4.399\pm2.787 & 3.657\pm2.929 & 3.181\pm2.093 \\\\ \hline \mathrm{High\\ confidence\\ voxels} & 1.701\pm1.410 & 2.091\pm1.771 & 1.799\pm1.929 & 1.568\pm1.126 \\\\ \hline \end{array}$$ Q3. Insufficient validation of uncertainty modeling. The uncertainty evaluation mainly compares learned and fixed Dirichlet concentrations using prediction discrepancy as the reference. Comparisons with established uncertainty estimation methods and evaluations of error detection or prediction reliability are needed. **A3.** We compared our method with two conventional uncertainty estimation approaches: predictive entropy and maximum softmax probability (MSP). We evaluated segmentation error detection and risk–coverage performance using AUROC, average precision (AP), and the area under the risk–coverage curve (AURC). Experiments were conducted under four single-modality settings, with results averaged across the WT, TC, and ET regions. Segmentation errors were identified by comparing the predicted and ground-truth region masks. $$\begin{array}{l|c|c|c} \hline \mathrm{Metric} & \mathrm{Entropy} & \mathrm{MSP} & \mathrm{Ours} \\\\ \hline \mathrm{AUROC}\uparrow & 0.5795 & 0.6084 & \mathbf{0.6163} \\\\ \hline \mathrm{AP}\uparrow & 0.6326 & 0.6645 & \mathbf{0.6753} \\\\ \hline \mathrm{AURC}\downarrow & 0.5651 & 0.5541 & \mathbf{0.5511} \\\\ \hline \end{array}$$ Q4. Unclear contribution of uncertainty learning to segmentation. The paper states that uncertainty learning is detached from segmentation features, yet the corresponding loss improves segmentation performance. The gradient flow and mechanism behind these improvements require clarification. **A4.** We re-examined our implementation and corrected the description. The concentration head receives shared decoder features without gradient detachment, **while gradients are detached from both full- and partial-observation probabilities in the prediction-discrepancy target and from the partial-observation probabilities in the Dirichlet variance numerator.** We have clarified these gradient paths in the revised manuscript. Q5. Incomplete comparisons with recent methods. The experiments do not include several relevant 2026 methods, such as UniME (CVPR 2026), Virtual Nodes Guided Dynamic Graph Neural Network (CVPR 2026). Uni-Encoder Meets Multi-Encoders: Representation Before Fusion for Brain Tumor Segmentation with Missing Modalities Virtual Nodes Guided Dynamic Graph Neural Network for Brain Tumor Segmentation with Missing Modalities **A5.** We reviewed three relevant 2026 studies, including AdaMM (MIA 2026). For Virtual Nodes Guided Dynamic Graph Neural Network (CVPR 2026), we could not identify a publicly available GitHub repository, and its experimental setup differs from ours. **Given the need for a comparable evaluation and timeline, we reproduced AdaMM on BraTS2024.** The results are summarized below. $$\begin{array}{l|c|c|c|c} \hline \mathrm{Method} & \mathrm{WT} & \mathrm{TC} & \mathrm{ET} & \mathrm{Mean} \\\\ \hline \mathrm{AdaMM} & 71.9 & 41.5 & 42.8 & 52.07 \\\\ \hline \mathrm{Ours} & \mathbf{84.08} & \mathbf{50.46} & \mathbf{52.25} & \mathbf{62.26} \\\\ \hline \end{array}$$
Reviewer_6MroOpenReview ↗

reply to authors

Thank you for the additional experiments and clarifications. The rebuttal addresses several of my concerns, particularly regarding gradient flow and comparisons with recent methods. The new results also provide further evidence for the effectiveness of the uncertainty modeling. However, my main concern regarding methodological novelty remains, as the motivation for the proposed Dirichlet-based uncertainty modeling over simpler alternatives is not sufficiently clear. Overall, I appreciate the authors' efforts but will maintain my original rating

Reviewer_LE1yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper defines uncertainty from missing modalities through the expected squared difference between full- and partial-observation class posteriors. The authors adopt a Dirichlet distribution to estimate uncertainty under missing modalities and capture its effect on the current prediction. While the theoretical formulation follows from a standard conditional variance identity, its connection to clinically meaningful predictive uncertainty remains insufficiently justified.

Critical strengths

The paper provides a probabilistic analysis linking uncertainty from missing modalities to the expected prediction discrepancy between full and partial observations, which offers an interesting perspective. The experimental results demonstrate competitive segmentation performance across different missing-modality settings, supported by ablation studies and additional uncertainty analyses.

Critical weaknesses

My main concern is that fully observed multimodal data may not always be available in real-world settings, even during training. Why should prediction change be treated as an appropriate uncertainty supervision signal rather than simply a measure of prediction inconsistency? My comments are listed below.

The paper defines uncertainty from missing modalities through the expected squared difference between full- and partial-observation class posteriors. However, for me, prediction disagreement does not necessarily indicate uncertainty or unreliability, as both predictions may be incorrect. How does the reliance on full-modality observations affect the practical applicability of the proposed method?

The methodological novelty beyond combining these existing techniques is unclear to me. For example, the theoretical propositions follow from established properties of conditional expectation, variance decomposition, and squared-loss regression; Section 4.1 adopts standard Dirichlet uncertainty modeling, while the framework combines knowledge distillation and the Brier loss.

It is unclear whether the resulting uncertainty captures information loss from missing modalities or simply reflects model disagreement. The authors should clarify how these objectives interact and how gradients propagate during training.

The uncertainty estimates are trained to match prediction changes and evaluated primarily against the same type of target. This does not independently validate uncertainty quality. Comparisons with existing uncertainty estimation methods and evaluations based on segmentation error detection, calibration, or risk–coverage analysis are needed.

Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Thank you for your constructive feedback. Our responses are as follows.

Q1.The paper defines uncertainty from missing modalities through the expected squared difference between full- and partial-observation class posteriors. However, for me, prediction disagreement does not necessarily indicate uncertainty or unreliability, as both predictions may be incorrect. How does the reliance on full-modality observations affect the practical applicability of the proposed method?

A1. When the training data consist entirely of incomplete-modality cases without access to full-modality references, the required prediction-discrepancy supervision cannot be directly constructed using our current method. Applying our method in such scenarios would require additional mechanisms, such as cross-modal reconstruction, modality completion, or an external full-modality reference model, to obtain the necessary full-modality reference information. These issues are beyond the scope of the current study and will be discussed as future research directions in the revised manuscript.

Q2.The methodological novelty beyond combining these existing techniques is unclear to me. For example, the theoretical propositions follow from established properties of conditional expectation, variance decomposition, and squared-loss regression; Section 4.1 adopts standard Dirichlet uncertainty modeling, while the framework combines knowledge distillation and the Brier loss.

A2. Our main methodological contribution is a dedicated uncertainty-learning mechanism for prediction-posterior changes induced by missing modalities, comprising two aspects.

(1) We construct a supervision target using paired predictions for the same case under full and partial observations. Based on conditional variance decomposition and the regression property of squared loss, we theoretically establish that the squared difference between the true posterior distributions provides a sample-level supervision signal for missing-modality uncertainty. In practice, a proxy target is constructed from the model-predicted probabilities.

(2) We formulate the Dirichlet total variance using the partial-observation predictive probabilities and a concentration parameter, and train it with the PO loss to fit the paired prediction discrepancy. This directly links the concentration parameter to posterior changes induced by missing modalities, enabling the estimation of the corresponding uncertainty.

Q3.It is unclear whether the resulting uncertainty captures information loss from missing modalities or simply reflects model disagreement. The authors should clarify how these objectives interact and how gradients propagate during training.

A3. Our response consists of two parts.

(1) We added a Brier loss restoration experiment based on ground-truth segmentation labels. As shown in Figure. 4 of the revised manuscript, higher-uncertainty regions generally exhibit larger Brier loss reductions after modality restoration, indicating that the estimated uncertainty reflects the impact of missing modalities on actual prediction performance rather than merely model disagreement.

(2) The network is jointly optimized using segmentation, Brier, knowledge distillation (KD), and partial-observation uncertainty (PO) losses. The segmentation and Brier losses provide label supervision for the full- and partial-observation branches, respectively. The KD loss uses stop-gradient full-observation predictions as teacher targets to update the partial-observation student branch, while the PO loss propagates through the concentration head to the shared feature extraction layers of the partial-observation network.

Q4.The uncertainty estimates are trained to match prediction changes and evaluated primarily against the same type of target. This does not independently validate uncertainty quality. Comparisons with existing uncertainty estimation methods and evaluations based on segmentation error detection, calibration, or risk–coverage analysis are needed.

A4. We conducted additional experiments comparing our method with two conventional uncertainty approaches: predictive entropy and maximum softmax probability (MSP). We evaluated segmentation error detection and risk–coverage performance using AUROC, average precision (AP), and the area under the risk–coverage curve (AURC). The evaluation covered four single-modality settings, with results averaged across the WT, TC, and ET regions. Segmentation errors were identified by comparing the predicted region masks with their corresponding ground-truth masks. As shown below, our method achieves the best performance across all three metrics.

\hline \mathrm{Metric} & \mathrm{Entropy} & \mathrm{MSP} & \mathrm{Ours} \\\\ \hline \mathrm{AUROC}\uparrow & 0.5795 & 0.6084 & \mathbf{0.6163} \\\\ \hline \mathrm{AP}\uparrow & 0.6326 & 0.6645 & \mathbf{0.6753} \\\\ \hline \mathrm{AURC}\downarrow & 0.5651 & 0.5541 & \mathbf{0.5511} \\\\ \hline \end{array}$$

Thanks for your response to my comments. My main question remains whether modeling prediction uncertainty with a predefined missing-modality setting and using it as an uncertainty supervision signal to improve performance is appropriate. I am concerned about the work's generalizability. Because in real-world settings (other datasets), you may only have a few available modalities, and how you can train a proper uncertainty supervision signal without knowing the full number of input modalities is not clear to me. After consideration, I would keep the current overall rating.

Reviewer_24xCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether the available MRI sequences can predict how much a brain tumor segmentation model’s class probabilities would change if the missing sequences were acquired. Under the assumption (Y\perp M\mid X), the authors define missing-modality uncertainty as the total conditional variance of the full-observation class posterior given the partial observation. They show that this quantity equals the conditional expected squared distance between the full- and partial-observation posteriors. In practice, the method substitutes network predictions for the true posteriors and trains a lightweight concentration head so that the total variance of a Dirichlet distribution matches the resulting prediction discrepancy. At inference, both segmentation and uncertainty estimation use only the available modalities. The paper evaluates segmentation on BraTS2020 and BraTS2024 and assesses uncertainty estimates using RMSE, UCE, and calibration curves.

Critical strengths

1.The paper identifies a meaningful uncertainty target. A model can be confident given the available sequences while still being sensitive to the missing ones. Section 3 distinguishes the conditional variance of the true posterior, the posterior change for one realization of the missing modalities, and the discrepancy between network predictions. This distinction is relevant to both missing-input learning and medical image segmentation.

2.The theoretical motivation is clear. Propositions 1–3 follow under the stated assumption and give the regression objective a precise population-level interpretation. Appendix A.4 makes explicit how the accuracy of the proxy target depends on posterior estimation errors in both branches.

3.The segmentation evaluation covers a broad range of missingness patterns. Table 1 reports all 15 nonempty modality combinations on both datasets. Mean whole-tumor Dice is 88.18 versus 86.98 for the strongest listed baseline on BraTS2020 and 84.08 versus 80.52 on BraTS2024. The appendix also reports HD95 and computational cost.

Critical weaknesses
  1. The uncertainty target lacks independent validation. Table 2 and Figure 4 evaluate agreement with the same model-derived prediction discrepancy used for training. This supports prediction of branch disagreement, while the link to the proposed posterior uncertainty remains unverified. On held-out cases, compare predicted uncertainty with the label-based Brier loss reduction after restoring the missing modalities.
  2. The gradient paths are unclear. Section 4.1 describes detaching decoder features, but Equation (14) also depends on the branch probabilities. If those paths are detached, it is unclear how the uncertainty loss improves mean Dice from 60.86 to 62.26 in Table 3. If gradients flow through the target, the loss can change the discrepancy it is trained to predict. Please specify which paths are detached and verify the gains across paired seeds.
  3. The value of the Dirichlet parameterization is unclear. Table 2 shows gains over fixed concentrations. A capacity-matched scalar regression head trained on the same features and target would test whether the Dirichlet constraint adds value. Please also report how often the target exceeds the variance bound in Equation (11).
Questions for authors
  1. In Equation (14), which of the following are detached: (q_M) in the variance numerator, (q_F) in the target, and (q_M) in the target? Do the Dice gains in Table 3 persist across paired random seeds?
  2. Which dataset, modality combinations, and voxel regions are used to compute RMSE and UCE in Table 2? How are the UCE bins defined and weighted? How large are the gains under case-level aggregation, within tumor regions, and under severe missingness?
  3. How does the Dirichlet concentration head compare with a capacity-matched direct regression head under the same supervision? How often does the variance bound become restrictive among voxels where a highly confident prediction changes class after the missing modalities are restored?
1 comment
AuthorsOpenReview ↗

Thank you for your positive rating and valuable feedback. We address your concerns as follows.

Q1. The uncertainty target lacks independent validation. Table 2 and Figure 4 evaluate agreement with the same model-derived prediction discrepancy used for training. This supports prediction of branch disagreement, while the link to the proposed posterior uncertainty remains unverified. On held-out cases, compare predicted uncertainty with the label-based Brier loss reduction after restoring the missing modalities.

A1. We added a label-based Brier loss restoration experiment on 270 held-out BraTS2024 cases. For each of the four single-modality settings, we calculate the Brier losses under partial and full observations and define their difference as the Brier loss reduction. Voxels are ranked by estimated uncertainty and divided into percentile bins. The average Brier loss reduction in each bin is presented in Figure. 4 of the revised manuscript. Higher-uncertainty voxels exhibit larger Brier loss reductions after modality restoration.

Q2. The gradient paths are unclear. Section 4.1 describes detaching decoder features, but Equation (14) also depends on the branch probabilities. If those paths are detached, it is unclear how the uncertainty loss improves mean Dice from 60.86 to 62.26 in Table 3. If gradients flow through the target, the loss can change the discrepancy it is trained to predict. Please specify which paths are detached and verify the gains across paired seeds.

A2. We re-examined our implementation and the corresponding descriptions in the manuscript. The concentration head receives shared decoder features without detachment. Both branch probabilities in the prediction-discrepancy target and the partial-observation probabilities in the Dirichlet variance numerator are detached. We have clarified these paths in Section 4.1 and Eq. (14).

Q3. How does the Dirichlet concentration head compare with a capacity-matched direct regression head under the same supervision? How often does the variance bound become restrictive among voxels where a highly confident prediction changes class after the missing modalities are restored?

A3. Our response consists of two parts.

(1) We replaced the Dirichlet concentration head with a scalar regression head trained using the same prediction-discrepancy target.

\hline \mathrm{Method} & \mathrm{WT} & \mathrm{TC} & \mathrm{ET} & \mathrm{Mean} \\\\ \hline \mathrm{Scalar\\ regression} & 82.89 & \mathbf{50.77} & 50.27 & 61.31 \\\\ \hline \mathrm{Dirichlet} & \mathbf{84.08} & 50.46 & \mathbf{52.25} & \mathbf{62.26} \\\\ \hline \end{array}$$ Although the scalar head performs slightly better in TC, the Dirichlet concentration head achieves more pronounced gains in WT and the most challenging ET region, highlighting the value of Dirichlet modeling. **(2)** **We evaluated variance-bound violations on 270 held-out BraTS2024 cases under four single-modality settings. High-confidence voxels with class reversals inevitably violate the variance bound, as can be verified mathematically.** $$\begin{array}{l|c|c|c|c} \hline \mathrm{Voxel\\ group} & \mathrm{T2\\ (\\%)} & \mathrm{T1\\ (\\%)} & \mathrm{T1ce\\ (\\%)} & \mathrm{FLAIR\\ (\\%)} \\\\ \hline \mathrm{All\\ brain} & 4.067\pm2.694 & 4.399\pm2.787 & 3.657\pm2.929 & 3.181\pm2.093 \\\\ \hline \mathrm{High\\ confidence} & 1.701\pm1.410 & 2.091\pm1.771 & 1.799\pm1.929 & 1.568\pm1.126 \\\\ \hline \end{array}$$ Q4. Which dataset, modality combinations, and voxel regions are used to compute RMSE and UCE in Table 2? How are the UCE bins defined and weighted? How large are the gains under case-level aggregation, within tumor regions, and under severe missingness? **A4.** Our responses are as follows. **(1)** **Table 2 reports RMSE and UCE evaluated on 270 BraTS2024 test cases** over the whole-brain region, averaged across all incomplete-modality combinations. **(2)** **UCE uses 15 equal-width bins based on estimated uncertainty.** Within each bin, the mean estimated uncertainty is compared with the mean prediction change, weighted by the proportion of valid voxels in the entire evaluation set. **(3)** We provide additional evaluations from two perspectives: **different tumor regions and four single-modality settings**. Using the fixed-concentration setting with S=1 as the baseline, the improvements are summarized in the following tables. $$\begin{array}{l|c|c|c|c} \hline \mathrm{Metric} & \mathrm{T1} & \mathrm{T2} & \mathrm{T1ce} & \mathrm{FLAIR} \\\\ \hline \mathrm{RMSE} & +4.82\\% & +4.61\\% & +1.74\\% & +4.82\\% \\\\ \hline \mathrm{UCE} & +29.97\\% & +32.81\\% & +16.42\\% & +24.99\\% \\\\ \hline \end{array}$$ $$\begin{array}{l|c|c|c} \hline \mathrm{Metric} & \mathrm{WT} & \mathrm{TC} & \mathrm{ET} \\\\ \hline \mathrm{RMSE} & +1.97\\% & +3.18\\% & +2.86\\% \\\\ \hline \mathrm{UCE} & +18.09\\% & +16.54\\% & +15.49\\% \\\\ \hline \end{array}$$

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.