List of Papers Browse by Subject Areas Author List
Abstract
Multimodal MRI is valuable for brain tumor segmentation because different modalities highlight different tumor subregions. In clinical practice, however, complete modalities are often unavailable due to shortened protocols or when contrast cannot be used. As a result, segmentation is often performed with incomplete-modality inputs, and the key challenge is to fuse available modalities. Existing approaches often introduce long-range dependency modeling for fusion, yet two challenges remain. First, 3D features are typically flattened into a 1D token sequences for global interaction, which could weaken spatial structure. Second, missing modalities are commonly replaced by zeros during fusion. Fusion then has fewer valid modality features to aggregate, which makes the fused features sparse in effective information. In this paper, we propose \textbf{HK-Fuse}, a Hilbert-interleaved Kimi fusion framework for incomplete-modality brain tumor segmentation. HK-Fuse organizes multimodal features with 3D Hilbert scanning to preserve spatial structure. It performs cross-modal fusion with Kimi Delta Attention, which supports a selective mechanism to reduce the impact of zero-filled inputs. We further introduce Key-Voxel Routing to apply cross-modal interaction selectively on informative key voxels. Extensive experiments on BraTS2023 validate the effectiveness of HK-Fuse across all 15 missing-modality scenarios. Code is available at https://github.com/weizhizhang7/HK-Fuse.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5288_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/weizhizhang7/HK-Fuse
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaWei_HKFuse_MICCAI2026,
author = { Zhang, Weizhi AND Li, Shumeng AND Zhang, Jian AND Qi, Lei AND Lin, Ling AND Yu, Qian AND Fang, Yuqi AND Shi, Yinghuan},
title = { { HK-Fuse: Hilbert-Interleaved Kimi Delta Attention for Incomplete-Modality Brain Tumor 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
This paper proposes the HK-Fuse framework for incomplete-modality MRI brain tumor segmentation. The method contains three core components: (1) A serialization scheme based on the 3D Hilbert space-filling curve, which maps multimodal voxels into a 1D sequence organized in a fixed-slot, modality-interleaved ordering; (2) The HK-Block based on Kimi Delta Attention (KDA), which leverages a channel-wise decay mechanism to suppress the information propagation of zero-filled missing-modality tokens; (3) Key-Voxel Routing (KVR), which reduces the computational cost of cross-modal fusion at high-resolution feature levels through Top-K sparse routing. Experiments are conducted on the BraTS2023 dataset, covering all 15 missing-modality combinations, achieving significant performance improvements on the enhancing tumor (ET) region.
- 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 paper is well-structured with good readability and clearly presented formulations. The clinical problem of missing modalities caused by shortened scanning protocols or contraindications to contrast agents is well-defined and practically significant. The method design is reasonable. The use of 3D Hilbert scanning for feature serialization to address the loss of locality caused by flattening 3D features is a thoughtful design. The design choice of suppressing zero-filled missing-modality inputs through the channel-wise selective decay mechanism of Bi-KDA is well-motivated. The introduction of Key-Voxel Routing at high-resolution stages for Top-K sparse routing of key voxels has certain rationality. The experiments are relatively thorough. Comprehensive validation is conducted on the BraTS2023 dataset across all 15 possible missing-modality combinations, with comparisons against multiple state-of-the-art incomplete-modality methods covering different technical paradigms. Ablation studies analyze the individual contributions of core components (HK-Block and KVR), and the efficiency trade-off between model parameters and segmentation performance is also discussed.
- 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.
Insufficient explanation of the core mechanism. One of the paper’s core claims is that Bi-KDA can effectively suppress the impact of zero-filled missing-modality tokens, but this is only described qualitatively without supporting argumentation or experiments. When input tokens are zero-filled (kt=0,vt=0), the state update indeed degenerates to St=Diag(αt)St−1, i. e. , only decay is triggered without writing new information. However, this property is not unique to KDA — even if the authors’ goal is to replace the Transformer fusion module in existing multi-encoder–fusion–shared-decoder models such as mmFormer with KDA, it is still necessary to demonstrate, either theoretically or experimentally, the unique advantage of KDA over other RNN/SSMs that also possess forgetting gate mechanisms (e. g. , Mamba, RetNet) when processing zero-filled input tokens. It is recommended to include comparisons between KDA and these alternatives in the ablation study. Insufficient depth in result analysis. The analysis in Section 3 mainly remains at the level of simple restatement of numbers (e. g. , “HK-Fuse achieves 77.3% ET Dice”), lacking in-depth discussion: On which specific missing-modality combinations does the method benefit the most? Why? How is ET segmentation performance maintained when T1ce is missing (T1ce being the key modality for ET segmentation — how does the method compensate from other modalities)? Discussion of the differential contributions of different modalities to the fusion is also absent. Synergistic effects among components require further justification. The main contribution of this paper lies in combining 3D Hilbert scanning, KDA, and Top-K routing in a design tailored for the incomplete-modality scenario. There is a certain design synergy logic among the components (e. g. , Hilbert scanning provides spatial locality for KDA; the channel-wise decay of KDA naturally handles zero-filled tokens). However, the authors are recommended to further strengthen the uniqueness of their contribution in the following aspects: (1) More explicitly articulate why this specific combination is superior to other possible combination schemes (e. g. , Mamba + Hilbert, KDA + non-Hilbert scanning, etc. ); (2) Verify the synergistic effects among components through ablation experiments (i. e. , whether the gains from individual components exhibit interaction effects rather than simple additive stacking). The fixed-slot, modality-interleaved organization design lacks sufficient justification. Organizing multimodal tokens in a fixed-slot, modality-interleaved manner to facilitate cross-modal interaction is an important design choice in the method, but the paper does not sufficiently discuss the rationale behind this strategy. The authors are recommended to analyze the following questions: What are the advantages of modality-interleaved ordering over simple modality concatenation? Is there a synergistic effect with the locality preservation of Hilbert scanning? When certain modalities are missing, the interleaved ordering distributes zero-filled tokens throughout the sequence — is this more favorable for the gradual decay mechanism of KDA compared to having them concentrated?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper achieves competitive performance on the incomplete-modality segmentation task on BraTS2023, with notably significant improvements over baseline methods on the enhancing tumor (ET) region. The overall design motivation is clear and the experimental validation is relatively thorough. However, the contribution is more reflective of a reasonable combination and engineering adaptation of existing technical modules (3D Hilbert scanning, KDA, Top-K routing) for the incomplete-modality scenario, with room for improvement in both theoretical analysis and empirical verification of the synergistic effects among components.
- 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 authors addressed my concerns, I have no other questions.
Review #2
- Please describe the contribution of the paper
This paper studies the missing modality problem in brain tumor segmentation, where the main challenges are how to preserve meaningful 3D spatial structure during long-range fusion and how to reduce the negative effect of zero-filled missing-modality inputs during cross-modal interaction. This paper propose HK-Fuse, which combines Hilbert-interleaved token ordering, bi-directional Kimi Delta Attention, and a Key-Voxel Routing module for selective fusion on skip features. The evaluation on BraTS2023 across all 15 missing-modality patterns is a solid experimental setting. The improvement on ET is promising.
- 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 paper addresses a practically important problem, since missing MRI modalities are common in real-world clinical settings.
- The idea of introducing Hilbert-interleaved token ordering to keep 3D spatial structure seems to be reasonable and interesting.
- The method shows notable gains on ET, which is often the most difficult subregion in brain tumor segmentation.
- 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.
- Prior work that more directly discusses the drawbacks of zero-filled missing-modality placeholders and related token handling strategies should be discussed and positioned carefully, e.g., TMFormer.
- It is still unclear why KDA is the right mechanism here, suitable for suppressing the effect of zero-filled tokens, as opposed to a simpler SSM variant, a masked fusion module, or another gated design.
- In KVR, the scoring head seems to identify generally important voxels, but the paper presents it as targeting regions with limited effective information. These are not necessarily the same thing, and the current experiments do not verify that distinction.
- How sensitive is the method to the choice of K = 200 in KVR?
- Since the paper emphasizes spatial consistency and boundary quality, boundary-aware metrics such as HD95 would be helpful.
- The performance gain is significant on ET, but marginal on TC and WT. I understand that this may be because performance on TC and WT is already close to saturation, but I still have to say that this makes it difficult to assess the effectiveness of HK-Fuse. From this perspective, BraTS 2023 may not be the most appropriate testbed.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Major Factor: Although some design of HK-Fuse is interesting and it shows promising gains on ET, the positioning against prior work, the justification for key design choices, and the experimental evidence are all somewhat incomplete. In addition, it seems to be hard to assess the overall effectiveness of HK-Fuse using BraTS 2023.For these reasons, I lean toward weak reject before rebuttal
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The analysis and experiments of the module design is not enough and convincing
Review #3
- Please describe the contribution of the paper
This paper introduces HK-Fuse, which is a novel framework proposed for the task of incomplete-modality brain tumor segmentation. It addresses the challenges of preserving 3D spatial structure and handling zero-filled missing-modality inputs, which are common in existing methods. By introducing a Hilbert-interleaved Kimi Delta Attention mechanism and Key-Voxel Routing, HK-Fuse achieves state-of-the-art segmentation performance across all 15 missing-modality combinations on the BraTS2023 dataset.
- 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.
HK-Fuse demonstrates significant innovation and practicality in addressing the critical clinical problem of incomplete-modality brain tumor segmentation. The manuscript’s strengths primarily lie in its deep understanding of existing challenges and the clever solutions proposed. Firstly, to counter the weakened spatial structure caused by flattening 3D features, HK-Fuse introduces 3D Hilbert scanning and a fixed-slot, modality-interleaved organization. This approach is crucial for preserving local spatial context during long-range interactions, which is vital for precise 3D medical image segmentation. Secondly, addressing information sparsity due to zero-filled missing modalities, the paper innovatively uses Kimi Delta Attention (KDA) combined with bi-directional scanning. Its channel-wise selective mechanism effectively suppresses the influence of zero-filled inputs, ensuring stable and efficient fusion under incomplete modality conditions. The Delta Rule-inspired update mechanism of KDA is particularly well-suited for handling such uncertainties.
Furthermore, the introduction of Key-Voxel Routing is another highlight. By selectively focusing on more informative voxels on high-resolution skip connections, it optimizes computational resources and further enhances the model’s accuracy and robustness in identifying critical tumor structures, such as enhancing tumor regions. This refined interaction strategy offers a distinct advantage when dealing with small or ill-defined tumor subregions. The paper provides strong evidence supporting the effectiveness and generalizability of its method through extensive experiments on the BraTS2023 dataset across all 15 missing-modality combinations. Qualitative results also clearly demonstrate HK-Fuse’s superiority in reducing segmentation artifacts and improving spatial consistency. The overall design of the model, combining the efficient feature extraction of U-Net with the powerful long-range dependency modeling of sequential models, forms an efficient and high-performing hybrid architecture.
- 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.
Despite its strengths, the paper has several areas that could be enhanced for greater scientific rigor and practical insight:
1): The reported performance metrics are single Dice Score values, lacking accompanying standard deviations or confidence intervals (e.g., 95% CI). Given the inherent randomness in deep learning model training (e.g., initialization, data loading, and especially the random sampling of missing modality masks during training as mentioned in the paper), reporting only single values makes it challenging to ascertain the statistical significance of small performance differences. It is difficult to distinguish true improvements from random variations, undermining the robustness claims. In scientific research, especially when comparing models with seemingly small performance gains (e.g., 0.2-0.4 percentage points Dice), reporting Mean±STD or 95% CI from multiple independent runs is crucial. This provides insight into the stability and reliability of the performance, allowing for proper statistical hypothesis testing to confirm if one model truly outperforms another beyond random chance.
2): While the paper addresses a critical clinical need, it lacks a detailed analysis of the computational efficiency, including concrete metrics such as inference time, training time, and GPU memory consumption. Although Figure 3 compares parameters versus Dice Score, it does not provide the comprehensive performance-cost trade-off typically expected for clinical deployment, especially for 3D medical imaging tasks which are often computationally demanding.
3): The paper fixes the Key-Voxel Routing’s Top-K value to 200 without presenting a sensitivity analysis for this critical hyperparameter. The choice of K could significantly impact performance across different tumor characteristics or datasets. An investigation into how varying K affects segmentation accuracy, efficiency, and robustness would provide deeper insight into the model’s generalizability and optimal configuration.
4) The role and specific contribution of the auxiliary decoders and multi-stage supervision strategy to the overall performance are not thoroughly analyzed through ablation studies. Without clear evidence of their impact, it is difficult to determine if these are essential components driving performance gains or merely common training practices, potentially adding complexity without proportional benefits.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper writing, organization, and experiment results are major factors of my score.
- 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.
I think the authors’ response partly addressed my questions. I prefer to give a boderlined acceptance.
Author Feedback
We appreciate the reviewers’ thoughtful feedback. We are encouraged by the comments that HK-Fuse is “well-structured” (R1), “interesting” (R2), and shows “practicality” (R3). We will release the code upon acceptance.
1.KDA & Hilbert (R1-Q1/Q3, R2-Q2): To clarify, decay without writing new information for zero-filled slots is not the unique advantage of KDA. We chose KDA for its channel-wise decay and gated writing. This is useful in our fixed-slot setting because zero-filled slots have no modality evidence, so they should write little new information, while valid-slot context should be retained or decayed selectively. With the same HK-Fuse architecture, replacing KDA with Mamba/Transformer gave 82.7/82.6 Avg. Dice, below KDA (84.3). KDA with non-Hilbert ordering dropped to 81.6.This suggests an interaction effect between KDA and the proposed sequence ordering.
2.Fixed-slot & Interleaved (R1-Q4): Simple modality concatenation follows a modality-wise order, whereas our interleaving follows a voxel-wise order, allowing features at the same spatial location to interact more directly. This voxel-wise ordering is synergistic with Hilbert scanning because Hilbert preserves 3D spatial locality in the sequence, while interleaving keeps different modalities of each voxel close. Under missing-modality settings, fixed-slot interleaving distributes zero-filled slots rather than forming long zero-only segments. This avoids repeated KDA decay over consecutive missing tokens and lets valid modality tokens refresh the state more frequently, making it more favorable for KDA’s gradual state decay.
3.Positioning & TMFormer (R2-Q1): TMFormer removes missing-modality tokens to avoid zero-map bias and redundant computation. HK-Fuse keeps fixed slots because zero-filled slots have no diagnostic content, but their positions encode modality availability and keep modality identity stable across missing settings. If missing slots are removed and visible tokens are compacted, the same sequence position may correspond to different modalities. We will cite TMFormer and clarify this distinction.
4.Missing-pattern analysis (R1-Q2): HK-Fuse improves ET in all 15 settings, with +1.7–4.1pp gains over the SOTA baseline (IM-Fuse) and the largest gains in T1ce-available incomplete cases, e.g., T1+T1ce, FLAIR+T1ce, and FLAIR+T1+T1ce. This matches T1ce’s role as the dominant ET cue. When T1ce is missing, ET is harder, but HK-Fuse still gains +2.5pp on average by using FLAIR/T2 for tumor extent/edema context and T1 for anatomical/core cues. TC also gains +1.0–1.8pp in five T1ce-missing settings, showing the benefit is not restricted to ET.
5.KVR & Top-K sensitivity (R2-Q3/Q4, R3-Q3): We agree KVR should be described more precisely as selecting informative key voxels for sparse high-resolution cross-modal interaction, rather than explicitly identifying limited-effective-information regions. For K={100,200,300,400}, Avg. Dice is 84.16/84.34/83.63/84.32.We use K=200 because it gives the best Avg. Dice among the evaluated values, with no consistent gain from larger K.
6.HD95 (R2-Q5): HD95 better reflects boundary quality and spatial consistency. HK-Fuse yielded lower Avg. HD95 than the SOTA baseline (8.13 vs. 8.61).
7.BraTS2023 (R2-Q6): BraTS2023 is a commonly employed benchmark with substantially more labeled training cases than BraTS2018/2020 (1251 vs. 285/369). Although WT/TC are relatively saturated, HK-Fuse still gains about 1.0–1.8pp in five T1ce-missing TC settings.
8.Mean±STD (R3-Q1): Across repeated trainings, HK-Fuse achieved WT/TC/ET Dice of 90.42±0.07 / 85.89±0.31 / 77.34±0.18. 9.Cost & Auxiliary supervision (R3-Q2/Q4): HK-Fuse training takes 112h/72GB on one A100 (bs=4), compared with 120h for the SOTA baseline under the same setting, inference is 1.5s/case for both. The auxiliary decoder/supervision is commonly used in incomplete-modality training. Removing them gives 82.1 Avg. Dice for HK-Fuse and 81.2 for the SOTA baseline.
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.
The paper addresses an important MICCAI problem, and the proposed HK-Fuse design is promising, with particularly strong gains on ET and evaluation across all 15 missing-modality settings. At the same time, the current version does not yet justify several central design choices with sufficient clarity, especially why KDA is particularly suitable for suppressing zero-filled missing-modality tokens and how the method is positioned relative to closely related prior work (e.g., TMFormer). For rebuttal, the authors should focus on strengthening the paper through clearer technical explanation and evidence already available in the submission: provide deeper analysis of which missing-modality patterns benefit most and why, clarify the rationale for the fixed-slot interleaved ordering, discuss computational and reproducibility details more explicitly, and tone claims appropriately where empirical support is currently limited. Overall, this is borderline but worth entering a rebuttal, as a strong response that improves positioning, analysis, and justification could materially affect the final decision.
- 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.
While some questions remain regarding the depth of component analysis and design justification, the work tackles a clinically relevant incomplete-modality segmentation problem, presents a coherent framework, and demonstrates empirical performance on BraTS2023.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper addresses a clinically relevant problem in incomplete-modality brain tumor segmentation, and the proposed HK-Fuse framework shows clear practical motivation. Its main strength lies in combining Hilbert-based 3D serialization, fixed-slot modality interleaving, KDA-based sequence modeling, and Key-Voxel Routing to handle missing modalities and preserve spatial structure. The evaluation is also reasonably thorough, covering all 15 missing-modality settings on BraTS2023, with particularly notable improvements on the challenging ET region. The initial reviews raised valid concerns about the justification of KDA, the fixed-slot interleaved design, the positioning against TMFormer, and the lack of sensitivity or efficiency analysis. The rebuttal addresses these points to a meaningful extent. In particular, the authors provide additional evidence showing that replacing KDA with Mamba or Transformer reduces performance, that non-Hilbert ordering leads to a clear drop, and that the selected Top-K value is supported by sensitivity results. They also clarify the distinction from TMFormer, provide missing-pattern analysis, report HD95 and mean±std results, and give computational cost comparisons. While some concerns remain regarding the depth of module analysis and broader validation beyond BraTS2023, the response substantially strengthens the technical justification and resolves most of the major ambiguities. Overall, the method is well motivated, experimentally solid, and relevant to a practical clinical setting. Given the comprehensive missing-modality evaluation and the additional clarifications provided in the rebuttal, I believe the paper meets the acceptance bar, and I would 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.
Following the rebuttal, two out of three reviewers recommended acceptance.
The notable improvements on the enhancing tumor region across all 15 missing-modality combinations on BraTS2023 represent a meaningful empirical contribution.
R2’s remaining concerns about the depth of module justification and experimental analysis are legitimate but do not constitute fatal flaws, as the core design choices are reasonable, and the rebuttal addressed the main points raised.
The authors are expected to strengthen the camera-ready version by adding statistical measures such as standard deviations, providing a more detailed computational efficiency analysis, including a sensitivity analysis for the Top-K hyperparameter, and deepening the discussion of per-modality and per-combination results.