List of Papers Browse by Subject Areas Author List
Abstract
Existing methods for multimodal MRI segmentation with missing modalities typically assume that all MRI modalities are available during training. However, in clinical practice, some modalities may be missing due to the sequential nature of MRI acquisition, leading to performance degradation. Furthermore, retraining models to accommodate newly available modalities can be inefficient and may cause overfitting, potentially compromising previously learned knowledge. To address these challenges, we propose Replay-based Hypergraph Domain Incremental Learning (ReHyDIL) for brain tumor segmentation with missing modalities. ReHyDIL leverages Domain Incremental Learning (DIL) to enable the segmentation model to learn from newly acquired MRI modalities without forgetting previously learned information. To enhance segmentation performance across diverse patient scenarios, we introduce the Cross-Patient Hypergraph Segmentation Network (CHSNet), which utilizes hypergraphs to capture high-order associations between patients. Additionally, we incorporate the Tversky-Aware Contrastive (TAC) loss to effectively mitigate information imbalance both across and within different modalities. Extensive experiments on the BraTS2019 dataset demonstrate that ReHyDIL outperforms state-of-the-art methods, achieving an improvement of over 2% in the Dice Similarity Coefficient across various tumor regions. Our code is available at ReHyDIL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2774_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/reeive/ReHyDIL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanJun_Hypergraph_MICCAI2025,
author = { Wang, Junze and Fan, Lei and Jing, Weipeng and Di, Donglin and Song, Yang and Liu, Sidong and Cong, Cong},
title = { { Hypergraph Tversky-Aware Domain Incremental Learning for Brain Tumor Segmentation with Missing Modalities } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15970},
month = {September},
page = {288 -- 298}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes ReHyDIL for brain tumor segmentation with missing modalities with MRI modalities sequentially acquired.
Further it introduces the Tversky-Aware Contrastive (TAC) loss to balance the learning between different modalities.
In addition it also presents the Cross-Patient Hypergraph Segmentation Network (CHSNet), which captures high-order associations across patients, enhancing feature representations and improving segmentation performance.
- 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 presents a novel ReHyDIL framework to address the challenge of missing modalities in brain tumor MRI segmentation. Key strengths include:
-
A novel combination of Domain Incremental Learning (DIL), hypergraph-based segmentation, and a new Tversky-Aware Contrastive (TAC) loss.
-
The use of the Cross-Patient Hypergraph Segmentation Network (CHSNet) to capture high-order associations between patients.
-
The development of the TAC loss to handle inter-modality imbalance.
-
Strong empirical results on the BraTS2019 dataset, demonstrating improved performance compared to state-of-the-art methods.
-
- 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.
Methodology: The paper focuses on the important problem of brain tumor segmentation with missing modalities. It mentions “Existing methods…methods often assume that all MRI modalities are available during training, which is a condition often unmet in practice”. However, the paper later introduces a replay buffer and a balanced queue, which suggests that some data from previous modalities is retained and used during training anyway. The motivation is not entirely clear to me. The related work section could be strengthened to provide more context on this. Any clarification on this would be highly appreciated.
Related work: While the authors cite relevant work, a more in-depth analysis would help to better differentiate ReHyDIL’s contributions and explicitly highlight how it is different from prior work such as mmF and others which also assume missing modalities.
Replay Mechanism: The paper argues that other methods struggle to retain data effectively. Yet, ReHyDIL employs a “balanced queue” mechanism that functions similarly to prior recall or experience replay strategies. This raises a question about the novelty of this component and appears to contradict the paper’s claims regarding the advantanges of this method over prior methods in retaining past information.
Issues with Clarity and Motivation of Solutions: The paper’s clarity could be improved. The motivation behind specific design choices, particularly in the CHSNet architecture and other details isn’t always clearly articulated. This lack of clear explanation can make it challenging for readers to fully grasp the rationale behind the proposed solutions and their effectiveness. The transition from Introduction to Method is quite abrupt and directly jumps into the implementation without providing the necessary motivation.
Results: There appears to be a mismatch in the reported results when comparing Table 1 of this paper with that of mmF. Clarification on this matter is essential to ensure the accuracy and reliability of the comparative evaluation. Any clarification on this would be very helpful.
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
See weaknesses and strengths
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper presents the Replay-based Hypergraph Domain Incremental Learning (ReHyDIL) approach for brain tumor segmentation with missing modalities. This method enables the model to be trained in a sequential modality order.
- 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 authors apply a domain incremental learning (DIL) method to brain tumor segmentation with missing modalities. This approach only requires a sequential training strategy.
- The authors verify that the Tversky similarity is more suitable than the cosine similarity for mitigating intra-modality imbalance.
- 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.
- It is recommended that the authors cite some more recent related works, such as RFNet[1], TMFormer[2], and ACDIS[3]. These works also address brain tumor segmentation with missing modalities.
- Why did the authors choose the training order of “T1->T2->FLAIR->T1CE”? What would be the impact if a different order were used?
- What would be the impact if the model is trained with the Tversky loss while randomly discarding modalities, similar to other methods dealing with missing modalities? [1] RFNet: Region-aware Fusion Network for Incomplete Multi-modal Brain Tumor Segmentation, ICCV2021. [2] TMFormer: Token Merging Transformer for Brain Tumor Segmentation with Missing Modalities, AAAI2024. [3] Anatomical Consistency Distillation and Inconsistency Synthesis for Brain Tumor Segmentation with Missing Modalities, ECAI2024.
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is novel, but there are some minor drawbacks.
- 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 #3
- Please describe the contribution of the paper
The authors persent a novel framework for brain tumor segmentation that addresses the challenge of incrementally learning from sequentially arriving MRI modalities while handling missing modalities. It combines Domain Incremental Learning (DIL) using a replay buffer. The study talks about 3 main novelties: a. A Cross-Patient Hypergraph Segmentation Network designed to capture high-order relationships between patients based on tumor features. b. A Tversky-Aware Contrastive loss to mitigate inter-modality imbalance and encourage learning robust cross-modal correlations, using a balanced queue mechanism. c. Explicit use of Tversky loss and Tversky-based Focal loss to handle intra-modality imbalance (small tumor regions vs. background), specifically weighted to penalize false negatives more heavily.
- 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 tackles the issue of missing modalities in clinical practice and the need to incrementally update models without catastrophic forgetting, which standard methods often fail to address. It integrates DIL, hypergraph learning (CHSNet), and specialized contrastive/segmentation losses (TAC, Tversky-based) in a novel way for this specific problem. Incorporating hypergraphs to model similarities between patients is an interesting approach to leverage population-level information. The study explicitly addresses both inter-modality imbalance (via TAC loss and balanced queue) and intra-modality imbalance (via Tversky-based losses) with theoretically motivated components. Reports improvements (>2% average DSC) over several relevant state-of-the-art methods on the BraTS’19 dataset across multiple missing modality scenarios. Lastly, the ablation study supports the contribution of the key components (CPH, TAC).
- 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.
I feel that the overall framework is quite complex, combining multiple advanced concepts. Justifying this complexity requires very clear gains. While novel, the practical benefit and interpretation of the cross-patient hypergraph could be elaborated further. How sensitive is it to the KNN parameters and the definition of “similar tumor features”? Is the computational overhead significant? The authors have used a fixed percentage (P=10%) for the replay buffer based on a mean loss strategy. The sensitivity to this percentage and the selection strategy could be discussed. Replay buffers inherently add memory overhead.Lastly, the incremental learning follows a specific modality order (T1->T2->FLAIR->T1CE). The impact of different acquisition orders can be explored.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
-
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper tackles a relevant and challenging problem in clinical medical imaging – incremental learning for segmentation with missing modalities. It proposes a novel framework (ReHyDIL) integrating several advanced techniques (DIL, hypergraphs, Tversky-based contrastive/segmentation losses) in a coherent way. The approach is well-motivated, particularly the explicit handling of both inter- and intra-modality imbalance. The reported results demonstrate significant improvements over strong baselines on a standard benchmark. While the framework is complex, the ablation studies support the contribution of its key components. The code released by the authors aids reproducibility. Overall, the novelty, relevance, and strong results make this a solid paper suitable for acceptance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Reviewer #2 1, Replay Buffer Novelty & Motivation The balanced queue is inspired by replay mechanisms. It offers two key advantages over conventional replay approaches: (a) cross-modality balance—samples from previously seen modalities are represented equally during subsequent training, preventing bias toward the most recently observed modality; (b) hard-example prioritisation within each modality—buffer capacity is devoted to the most informative, high-loss samples rather than to randomly selected instances. 2, Comparison with Related Works Unlike earlier multimodal brain tumor segmentation approaches—which typically (i) re-train a model from scratch once missing data occur, (ii) distil knowledge from a full-modality teacher into a partial-modality student, or (iii) synthesise the absent modality via image translation. ReHyDIL tackles the problem from an incremental-learning perspective with balanced memory replay and hypergraph-based cross patient reasoning. 3, mmF Results Reporting Discrepancy The mmF results in our paper were obtained from our implementation under the BraTS2019 dataset and our experimental conditions, in order to ensure a fair comparison with ReHyDIL and other baselines. These conditions differ from those of the original mmF publication. For example, the original mmF was evaluated on another dataset (BraTS2018).
Reviewer #4 1, Supplement Recent Related Works We will add citations to RFNet [1], TMFormer [2] and ACDIS [3]. [1] RFNet: Region-aware Fusion Network for Incomplete Multi-modal Brain Tumor Segmentation, ICCV2021. [2] TMFormer: Token Merging Transformer for Brain Tumor Segmentation with Missing Modalities, AAAI2024. [3] Anatomical Consistency Distillation and Inconsistency Synthesis for Brain Tumor Segmentation with Missing Modalities, ECAI2024. 2, Modality Training Order (also Reviewer #5 comment) In the incremental learning setup, we followed the typical clinical acquisition sequence of MRI modalities to simulate a realistic scenario. This order is advantageous because each subsequent modality adds new information building on the previous ones, aligning with how radiologists progressively interpret scans. 3, Random Modality Drop Our focus is on the realistic scenario where MRI modalities arrive sequentially, leading to missing modalities; this cannot be emulated by simply dropping modalities at random during training. Random dropout assumes that all modalities are potentially observable at each iteration; in contrast, our incremental learning mimics the realistic acquisition where a new modality becomes available only after finishing the previous one.
Reviewer #5 1, CHSNet Architecture CHSNet explicitly models relationships among patients by treating each patient’s tumor feature representation as a vertex in a hypergraph. CHSNet was proposed for its ability to integrate cross-patient context, and it enhances interpretability by making the model’s use of data relationships explicit. 2, Hyperparameter Sensitivity In our implementation, the hypergraph construction is adaptive rather than fixed-K. The K values flexibly adapt to local density while preventing the graph from becoming too sparse or too dense. Similarly, for the replay buffer size, we fixed P = 10% of samples per phase as a balanced trade-off between memory usage and information retention.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A