Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Multimodal medical image segmentation is often challenged by two factors: (i) missing modalities that disrupt cross-modal feature consistency, and (ii) residual inter-modality misalignment that degrades feature fusion. To this end, we propose MoCaf-Mamba (Modality Completion and Alignment in feature space), a Mamba-based framework that unifies feature-space modality completion and deformable feature alignment for missing-modality segmentation. Specifically, the \textbf{M}odality \textbf{C}ompletion in \textbf{F}eature \textbf{S}pace (\textbf{MCFS}) predicts missing-modality representations from available modalities and is trained with explicit reconstruction constraints to preserve segmentation-relevant semantics. The \textbf{D}eformable \textbf{F}eature \textbf{A}lignment (\textbf{DFA}) is integrated hierarchically into the encoder and predicts a bounded dense displacement field for feature-level warping, mitigating geometric discrepancies during fusion. Finally, the \textbf{S}pace \textbf{T}oken \textbf{M}ixer (\textbf{STM}) aggregates completed and aligned multi-scale features via a dual-branch design with a global consensus token, producing a robust representation for decoding under arbitrary missing patterns. Experiments on MM-WHS, BraTS2023, and Prostate158 show that MoCaf-Mamba consistently improves missing-modality segmentation and achieves state-of-the-art results across diverse missing settings. The code is available at: https://github.com/wekiqs/MoCaf-Mamba

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3091_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/wekiqs/MoCaf-Mamba

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhoYon_MoCafMamba_MICCAI2026,
        author = { Zhou, Yongsong AND Wang, Gui AND Shen, Linlin},
        title = { { MoCaf-Mamba: Modality Completion and Alignment in Feature Space for Missing-Modality 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 MoCaf-Mamba, a unified framework for multimodal medical image segmentation under missing-modality settings. The main contribution lies in jointly addressing two key challenges: modality completion and inter-modality misalignment in the feature space. Specifically, the method introduces MCFS, DFA, and STM components within a Mamba-based architecture, and the proposed approach achieves robust segmentation performance under arbitrary missing-modality conditions.
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 three-module design (MCFS, DFA and STM) is well organized, with each component playing a clear role, resulting in an effective and coherent overall framework. 2.The paper provides anonymized open-source code and conducts experiments on three datasets under various missing-modality settings, thereby supporting the reliability and reproducibility of the results.
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

1.The cascaded dependency between modality completion and feature alignment is insufficiently justified. While the paper reasonably integrates two key challenges—modality completion and inter-modality misalignment—into a unified framework, the current pipeline performs feature-space modality completion prior to feature alignment. This raises an important question: how can the method ensure that the completed features are sufficiently accurate and stable to support reliable downstream alignment? If the completion quality is suboptimal, the errors may propagate directly to the DFA module and negatively affect the subsequent alignment process. At present, this dependency is not discussed or validated in sufficient depth. 2.The justification for adopting Mamba as the backbone remains insufficient. The current ablation study mainly validates the contribution of the proposed modules. However, since the baseline is already Mamba-based, it does not clearly demonstrate the specific advantage of Mamba over more conventional backbones such as CNNs or Transformers. A direct comparison with non-Mamba backbones would strengthen this point. 3.The paper lacks feature-space visualization and analysis. The current evaluation mainly focuses on segmentation performance and output examples. Since the core contribution lies in feature-space completion and alignment, visualizing feature representations before and after completion/alignment, especially to demonstrate the effect of completion quality on subsequent alignment, would provide more direct evidence for the effectiveness of the proposed mechanism and improve the interpretability of the paper.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

N/A
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

N/A
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

Overall, the paper addresses a well-motivated and practically important issue. The idea of jointly handling modality completion and feature alignment is meaningful, the architecture is reasonably designed, and the experimental results are generally sufficient. However, several concerns remain: completion is performed before alignment, so the quality of completed features may directly affect downstream alignment, but this dependency is not thoroughly discussed; and feature-space visualization is missing to directly support the effectiveness of the proposed completion and alignment mechanisms.
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.

Although the rebuttal is helpful in clarifying several aspects of the submission, it does not sufficiently change my overall assessment, and I therefore maintain a Reject recommendation.

On the positive side, the authors clarified several implementation details that were ambiguous in the original manuscript. In particular, they explained that the framework uses fixed and ordered modality slots with availability masks to handle arbitrary missing-modality patterns, and clarified that STM performs attention only along the modality-token dimension rather than over spatial voxels. The rebuttal also clarifies that DFA predicts modality-specific displacement fields instead of using a single shared field, and provides a more concrete motivation for adopting Mamba by comparing it with CNN- and Transformer-based alternatives. These clarifications improve the readability of the method and partially address concerns regarding implementation and computational complexity.

However, my main concern remains. The rebuttal does not fully address the central dependency between feature-space modality completion and subsequent deformable feature alignment. The authors argue that completion should be performed before alignment because aligning absent or zero-filled features is ill-posed, which is a reasonable motivation. However, this does not fully resolve the remaining concern about how errors from MCFS may affect the subsequent DFA module. Since DFA operates on features that are partly completed by MCFS, the alignment process may be sensitive to the quality of these completed latent representations. If MCFS introduces inaccurate, biased, or unstable features, it remains unclear how such errors would influence the predicted deformation fields and the following fusion stage. The rebuttal explains why completion-before-alignment is a reasonable design choice, but it does not sufficiently discuss this potential error propagation, which is central to the proposed completion–alignment pipeline.

Overall, I appreciate the authors’ efforts in the rebuttal, and I agree that the problem is important and reasonably motivated. Nevertheless, since the reliability of the cascaded completion-before-alignment design is central to the proposed framework and remains insufficiently clarified, I do not believe the paper crosses the acceptance threshold in its current form.

Review #2

Please describe the contribution of the paper

To address the missing modality problem in multimodal medical image segmentation and the feature issues of the remaining modalities under modality absence, this paper proposes a unified framework combining Modality Completion in Feature Space (MCFS), Feature Alignment (DFA), and a Fusion Module (STM), constructs the MoCaf-Mamba model based on Mamba, and conducts extensive experiments on multiple multimodal datasets, demonstrating the effectiveness of the proposed method.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

1.The motivation and problem formulation of this paper are very well presented; modality missing and residual cross-modal misalignment are indeed practical issues encountered in clinical settings. 2.The methodological framework of the paper is relatively complete, primarily addressing the proposed problem through a three-stage pipeline (MCFS, DFA, STM). 3.Feature space completion is indeed more computationally efficient compared to GANs and diffusion-based methods. 4.The experiments in this paper are fairly extensive, having been conducted on multiple datasets.
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.

Major Questions: 1.The authors do not clearly explain why Mamba is used as the backbone. 2.In MCFS, a shared token is used to represent modalities; however, this approach may be overly simplistic and could potentially ignore inter-individual differences. 3.In the ablation study, the performance gain after introducing DFA is minimal, which weakens the significance of this module. 4.In Figure 1, the content mainly consists of stacked three-dimensional shapes, and the arrows and directions are unclear, making it difficult to quickly understand the overall workflow of the paper. 5.In the related work section “Multimodal Fusion and Alignment.”, the mentioned models such as SwinUNETR and TransUNet are essentially multi-level information fusion models rather than true multimodal fusion architectures.

Weakness: 1.In the Method section, below Equation (1), the term $F_{k}$ appears but is not defined or present in Equation (1). 2.In STM, each modality employs a separate self-attention mechanism. If there are four modalities, this requires four attention modules; especially in 3D voxel settings, this may significantly increase the number of model parameters. 3.“Results on BraTs2023” appears twice, suggesting that the authors did not carefully proofread the manuscript. 4.From Table 1, the proposed method shows limited improvement over the best-performing method on WT and TC, and the single-modality performance is also inferior to the best model.
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 has provided an anonymized link to the source code, dataset, or any other dependencies.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

N/A
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

1.I strongly recommend that the authors revise Figure 1.2.I suggest that the authors thoroughly reorganize the related work section. 3.The authors should include comparisons of the number of parameters, FLOPs, and throughput, as well as more visualization figures.
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 presents a clear motivation, a complete methodological framework, and sufficiently comprehensive experiments; however, it still suffers from the aforementioned issues, such as deficiencies in writing and experimental design. Therefore, I ultimately assign this score.
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’ response has addressed my concerns; however, the authors should incorporate the corresponding revisions comprehensively in the final version.

Review #3

Please describe the contribution of the paper

Paper addresses the problem of missing modality and inter-modality misalignment issue in multi-modal medical image segmentation. They do it by proposing three novel modules in Mamba framework that is MCFS, DFA and STM. It has been shown that on range of datasets, the proposed approach MoCAf-Mamba leads to better results than several state of the art methods.
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.

Paper is well-written. Topic is interesting, combining images with tabular data needs lot of attention.
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 issue of modality mis-alignment needs to be better motivated, perhaps adding a diagram or illustrative figure to explain this would be beneficial. I found the introduction section lacking substance such as problem explanation, motivation, limitations, etc.

MCFS is neatly explained, however, have you tested if more than one modality is missing? It relies deeply on Mamba generator – which makes reader question the novelty of the block and how is it different from earlier works published in this domain.

Fig 1 is good (well done), but after STM blocks, it is just too hard to follow due to small fonts.

The novelty and motivation for DFA block needs to be strengthened.

The paper is missing comparison with some recent works like DRIFA and DRIFA-Net and some other models.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

N/A
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

N/A
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

Please see the weakness listed above.

Main issue is the novelty of the paper is limited. The experiments are missing comparison with other SOA methods. Also the number of datasets could be more.
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 #4

Please describe the contribution of the paper

The paper introduces MoCaf-Mamba, a new architecture specially designed for robustness to missing modalities. MoCaf-Mamba is an Encoder-Decoder architecture with skip connections enhanced with 3 blocks which constitute the methodological contribution of the paper. MCFS is a small block combining a channel mixing layer and a Mamba-based feature generator. MCFS aims at reconstructing feature maps from missing modalities. The DFA block is a small CNN block predicting deformation fields aiming at better aligning the feature maps from the different modalities, hence compensating for potential residual miss-alignment post modality-fusion. This block is added at every stage of the decoder. Finally, STM is a dense channel attention layer spatially mixing the information from the different modalities. Authors validate their method on 3 different datasets: BRATS 2023, MM-WHS and Prostate158.Results are shown competitive/better than other baselines considered. An ablation study also highlights the contribution of the different blocks.
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 introduces 3 meaningful block with clear purpose. These blocks are integrated in a standard architecture and shown useful. The problem tackled is relevant. The experimental section is particularly informative. Results on 3 datasets are presented, showcasing convincing performance enhancement. The ablation study also confirms each bloc’s contribution.
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 my opinion, the main weakness of the paper is the lack of clarity of the method section. I am unable to fully understand the method from looking at the fig. 1 and reading the method section. In particular I have the following questions:
- How are missing modality feature maps generated when more than 1 modality is missing? Indeed, both fig. 1 bottom left and equation (1) seems to indicate that \phi_{fuse} takes as input all modalities except the one missing, hence expecting an input with K-1 channels. How is such input produces when 2 modalities are missing. In addition regarding this bloc, how are channels ordered? Does \phi_{fuse} have to learn to be agnostic to channel permutations?
- Another point that, while clearly explained, lack of justification is why is the DFA block needs to be introduces at each layer? Shouldn’t the deformation correction be very related from one level to the next (i.e. just a down/up sampled version of one another)? Also, from equation (3), it seems that the same displacement field is used to resample all modality feature map. In such setup, this block would not be able to do what is claimed, i.e. compensation for modality fusion residual error. To do so, different deformation fields should be used for different modality feature maps. It is also unclear what sort of deformation field is predicted.
- Finally, it is unclear how modality tokens are obtained for the STM block. It is only said in the text: “For each voxel v, we form modality tokens”. The new term E_{spatial} is not clearly defined. From fig.1 bottom right, one can understand that it is a residual branch with a few conv layers but it is not clear from the text.
This lack of clarity in the method negatively impact the paper and should be improved in the corrected version of the paper.
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 has provided an anonymized link to the source code, dataset, or any other dependencies.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

N/A
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

While the code released is anonymous, I would recommend to stick to English comments to be “even more anonymous”.

There is also a formatting issue in the abstract, please have a look.
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The result section is convincing and the method has some clear contributions. Yet, clarity in the method section is lacking to be accepted as is. Yet, I believe these points can be easily addressed. by the authors.
Reviewer confidence

Somewhat confident (2)
[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.

Authors addressed most of my points in their rebuttal.

Author Feedback

We thank the Area Chair and Reviewers for their constructive feedback. We will revise the manuscript to improve method clarity, Fig. 1, related work, visualization, and proofreading.

1.Mamba & non-Mamba backbone comparison (R1W2&R2W1&R3W3): We use Mamba because 3D missing-modality segmentation requires long-range volumetric context with manageable cost. 3D CNNs have limited effective receptive fields, while 3D Transformers require quadratic token complexity. Mamba provides global modeling of large anatomical context to infer missing latent semantics, with linear complexity. We have additionally conducted comparisons with ResNet3D-18 (CNN) and Swin3D (Transformer), which achieved Avg Dice of 0.7155 and 0.7284, respectively, both below MoCaf-Mamba. We will add these results and Params/FLOPs/throughput comparisons.

2.Arbitrary missing modalities & MCFS formulation (R2W2&R3W2/R3W3&R4W1): MoCaf-Mamba can handle any number of missing modalities. Tables 1–3 include all missing subsets; e.g., three modalities are missing for single-modality BraTS inputs. In implementation, all K modality slots are fixed and ordered; unavailable slots are zero-filled, and modality dropout challenge the model with arbitrary patterns. Thus, Φ_fuse receives a fixed-dimensional ordered input and doesn’t need to be permutation-invariant. We will rewrite Eq. (1) with an availability mask and correct the notation. The Style Token is a modality-level prior, while patient-specific information comes from fused spatial features, inter-individual differences are preserved. Unlike placeholder or pixel-level reconstruction methods, MCFS performs supervised feature-space completion conditioned on available modality context.

3.Completion-before-alignment and DFA clarification(R1W1&R2W3&R3W4/W5&R4W2): Completion is placed before alignment because aligning absent/zero features is ill-posed. MCFS first reconstructs segmentation-relevant latent anchors under feature reconstruction loss, after which DFA refines residual geometric discrepancies. We will add t-SNE/feature-map visualizations before/after MCFS and DFA to better show that completion stabilizes representations and DFA reduces mismatch. We will also clarify Eq. (3): DFA does not use one shared field for all modalities. The offset predictor outputs modality-specific 3D fields Δ_i, and each feature F_i is warped by its own bounded field λ · tanh(Δ_i). DFA is applied hierarchically because residual mismatch may occur at different semantic scales, requiring scale-specific correction. Although DFA gives a moderate gain (+0.5 Avg Dice), it improves the thin alignment-sensitive PZ from 65.0 to 65.8.

STM Clarity and Complexity (R2Min2&R4W3): E_spatial denotes the output of the local 3D convolutional branch. STM forms K modality tokens plus one hybrid token at each voxel and applies self-attention only along this modality-token dimension, not over spatial voxels. Thus, the attention length is K + 1 (≤5), making the overhead small. We will clarify token construction and report Params/FLOPs/throughput.

5.Performance and Comparisons (R2Min4&R3W5): Under extreme cases, i.e. only one modality is available, MoCaf-Mamba remains competitive; e.g., T1c-only TC on BraTS2023 is 89.5 vs. 89.6 (mmFormer). More importantly, our goal is robustness under arbitrary missing patterns, where MoCaf-Mamba achieves the best average Dice across BraTS2023, Prostate158, and MM-WHS. We will compare with DRIFA-Net when available for 3D missing modality, and otherwise discuss it in related work/limitations

Presentation fixes and visualization (R1W3&R2W4/W5/Min1/Min3&R3W1/W4&R4): We will redesign Fig. 1 with clearer arrows/fonts and an illustration of residual inter-modality misalignment; reorganize “Multimodal Fusion and Alignment”; add feature-space visualizations; and fix the undefined F_k, duplicate “Results on BraTS2023”, abstract formatting, and non-English code comments.

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 and practical problem in multimodal medical image segmentation under missing modalities, and the reviewers generally find the proposed framework meaningful with promising results across multiple datasets. At the same time, the submission is currently borderline due to concerns about methodological clarity and justification, particularly regarding the use of Mamba as backbone, the role and contribution of the DFA module, and the explanation of key design choices in the method and Figure 1.I recommend moving the paper to the rebuttal phase, where the authors should focus on clearly addressing these points and clarifying the current evidence without the need for additional experiments.
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.

Although concerns remain regarding clarity and validation, the rebuttal addressed most reviewers’ concerns about implementation and motivation, and the overall balance of reviews supports acceptance, with the required revisions to better justify error propagation and improve methodological clarity.

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 rebuttal addresses the main concerns with additional backbone comparisons, clearer missing-modality formulation, modality-specific DFA clarification, and planned complexity/visualization analyses. Although some components remain incremental, the method appears well motivated, competitive across arbitrary missing-modality settings, and sufficiently validated for 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.

After considering the rebuttal and updated reviews, I recommend acceptance. The overall reviewer feedback is sufficiently positive, although some concerns remain. The rebuttal clarifies several important issues, and the authors should incorporate these clarifications and further discuss the remaining concerns in the final version.

back to top

MoCaf-Mamba: Modality Completion and Alignment in Feature Space for Missing-Modality Segmentation

Author(s):