List of Papers Browse by Subject Areas Author List
Abstract
Myocarditis, an acute cardiac disorder progressing rapidly to life-threatening heart failure, requires precise lesion segmentation from Cine Magnetic Resonance Imaging (Cine-MRI) for timely intervention. Current segmentation accuracy is limited by two key challenges: 1) spatiotemporal discordance between myocardial motion patterns and evolving pathological features and 2) morphological complexity (irregular borders, scattered lesions). In this paper, we propose the MG-Mamba, a framework integrating deep state space models with graph-based spatiotemporal analysis. The architecture employs Mamba blocks to establish initial intra-/inter-frame dependencies in Cine-MRI sequences. For Challenge 1, we improve the detection of subtle abnormal motions through multi-step cross-frame analysis, extending beyond conventional adjacent-frame analysis. For Challenge 2, we further implement multi-scale patch division and constructs inter-patch graphs to concurrently capture global lesion distribution and local geometric patterns. Extensive evaluations on SYC-QC and SYC-SX clinical datasets demonstrate MG-Mamba’s superior segmentation accuracy over ten state-of-the-art benchmarks, significantly advancing myocarditis diagnostic precision. The code is available at https://github.com/userZ-CY/MICCAI.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1349_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/userZ-CY/MICCAI
Link to the Dataset(s)
SYC-QC: private
SYC-SX: private
BibTex
@InProceedings{YuChe_Multiscale_MICCAI2025,
author = { Yu, Chengjin and Zhang, Hao and Yan, Yuanting and Zhang, Dong and Lv, Sangyin and Pu, Cailing},
title = { { Multiscale Graph and Multi-Step Cross-Frame Mamba for Myocarditis Lesion Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15962},
month = {September},
page = {443 -- 452}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a novel MG-Mamba framework for myocarditis segmentation, which addresses the challenges of dynamic motion representation and morphological complexity in cardiac imaging.
This paper proposes the MCMMA module for dynamic motion representation and MGGA module for morphological complexity.
The method is evaluated on real-world datasets collected from two hospitals. Experiments demonstrate that the proposed approach outperforms ten 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.
-
The MCMMA module captures abnormal motion patterns at different temporal scales by designing multi-step cross-frame scanning sequences.
-
The MGGA module constructs graph structures at multiple scales, simultaneously capturing local details and global distribution information, addressing the complexity problem of lesion morphology.
-
The research is validated using real clinical datasets from two hospitals, demonstrating the applicability of the method.
-
The feature representation capability of the model is visually demonstrated through t-SNE technology.
-
- 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 the Introduction, references [8] and [9] are claimed to relate to segmentation, but these works actually focus on super-resolution and reconstruction, which may mislead readers.
-
The paper lacks theoretical or clinical evidence explaining why temporal features are relevant for myocarditis in cine CMR, and how they correlate with pathological manifestations. This weakens the rationale for using spatiotemporal modeling in this context.
-
The reproducibility of this work is challenged. Several important hyperparameters are not described, including the frame sampling interval, learning rate, learning rate decay strategy, optimizer and other settings.
-
The paper does not specify the value of the loss weighting coefficient α, nor does it provide an ablation study to quantify the relative contribution of each loss component.
-
The chosen stride values in the multi-step temporal scan are not explained, and no ablation experiments are provided to analyze their impact.
-
There is no ablation study on the scan order, and the necessity or effectiveness of each scan direction remains unverified. It is also unclear why specific scan orders (e.g., order 1, 2, 3) are chosen and whether alternative or reversed scan orders would yield different results. This reduces confidence in the claimed benefits of the temporal modeling strategy.
-
The use of three specific patch sizes (8×8, 10×10, 16×16) is not justified, and there is no ablation study to explore how patch scale affects performance.
-
The data description is unclear. For example, it’s not specified how lesion masks were obtained or from which modality. This information is important, as the mask source can affect both the task setup and model performance.
- The paper does not describe how the dataset was divided into training, validation, and test sets. This information is essential for evaluating the validity of the reported performance.
- The paper mainly compares with natural image and generic medical image segmentation methods. However, several studies focus specifically on lesion segmentation from cine-CMR, which are more directly relevant to the proposed task, such as: a. Zhang, N., Yang, G., Gao, Z., Xu, C., Zhang, Y., Shi, R., … & Firmin, D. (2019). Deep learning for diagnosis of chronic myocardial infarction on nonenhanced cardiac cine MRI. Radiology, 291(3), 606-617. b. Xu, C., Xu, L., Ohorodnyk, P., Roth, M., Chen, B., & Li, S. (2020). Contrast agent-free synthesis and segmentation of ischemic heart disease images using progressive sequential causal GANs. Medical image analysis, 62, 101668. c. Xu, C., Xu, L., Brahm, G., Zhang, H., & Li, S. (2018). MuTGAN: simultaneous segmentation and quantification of myocardial infarction without contrast agents via joint adversarial learning. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11 (pp. 525-534). Springer International Publishing.
Including these domain-specific methods in the comparative analysis would provide a better evaluation of the proposed approach.
-
- 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.
- 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.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Despite having some innovation in method design and being evaluated on real clinical data, this work has several key issues affecting its integrity and reproducibility. In particular, the core modules lack support from ablation experiments, the theoretical foundation for temporal modeling is weak, multiple implementation details are missing, and comparisons with more relevant work are absent, all of which weaken the paper’s persuasiveness.
- 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 have addressed most of my concerns, and I am generally pleased with their responses. It is also appreciated that they have committed to releasing their code and models.
Review #2
- Please describe the contribution of the paper
The authors present a Mamba-based segmentation architecture for detecting myocarditis lesions. The proposed model, MG-Mamba, combines two modules: (1) one that addresses spatiotemporal inconsistencies, and (2) another that handles morphological complexity using multi-scale graphs. The authors compare the performance of their method against both natural and medical imaging algorithms, outperforming them on most metrics. Additionally, they provide qualitative comparisons, quantitative ablation studies of their components, and a detailed process for tuning key hyperparameters.
- 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 novel application of two distinct modules, designed to mitigate spatiotemporal inconsistencies and address morphological complexity, demonstrates superior performance compared to previous methods on the target task. Evaluation: The qualitative comparisons and quantitative analyses offer valuable insights into the performance gains. Additionally, the stability analysis and t-SNE visualizations further enhance understanding of the sources of improvement.
- 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.
Limited reproducibility: Some hyperparameters are not disclosed, and the lack of publicly available code, weights, and data significantly limits reproducibility.
Design decisions: Although addressing spatial inconsistency is a key strength of the method, the authors downsampled the sequence, which further alters the original frame stack. Details on how this downsampling was performed and its implications are not adequately discussed.
Clinical implications: While Dice and Hausdorff Distance metrics are reported, the potential impact on downstream clinical decision-making is not addressed. The article would benefit from an analysis of how performance differences between this method and others might affect clinical outcomes or diagnostic decisions.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- 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?
While the method is scientifically sound and the results outperform the state-of-the-art, the reviewer is concern on the reproducibility of this work, which greatly reduces its impact. The article would greatly benefit from more transparency on that matter. Additionally, further backing up some missing design decisions would also improve the article.
- 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.
While some of the identified weaknesses could still be clarified further, the rebuttal and amendments provided sufficiently address the main concerns. I recommend the article for acceptance.
Review #3
- Please describe the contribution of the paper
- This paper presents a clinical method for segmentation of myocarditis lesions using Cine - MRI images. This method has significant clinical potential for myocarditis lesion segmentation without contrast agents and shorter scan times.
- This paper proposes a MCMMA module to address the spatiotemporal inconsistencies in myocardial motion through multi-step cross-frame state space modeling, and a MGGA module to address the morphology complexity by constructing multi-scale graphs that aggregate global lesion distribution and local geometric information.
- Extensive experiments conducted on two clinical datasets (10,400 Cine - MRI images in total) demonstrate that the MG - Mamba is effective and superior to ten SOTA 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.
- Innovation: It proposes MCMMA and MGGA modules. MCMMA adds multi-step cross-frame state space modeling to the Mamba model, and MGGA introduces multi-scale concepts to graph neural networks. These improvements, though based on existing methods, offer new solutions for myocarditis lesion segmentation. 2.Experimentation: With 10,400 Cine-MRI images from two clinical datasets, the experiments, despite not covering all common models, are sufficient to validate the method’s effectiveness.
- Ablation Study: The ablation study is well-done. It assesses component effectiveness, and conducts feature visualization and stability analyses, providing in-depth understanding of the model for further improvement.
- 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.This paper does not provide any comparison related to efficiency, such as the amount of computation (flops), the number of parameters, and the computational speed. Without these metrics, it is difficult to comprehensively evaluate the practicality of the proposed method in real-world applications, where computational resources and time efficiency are often critical factors. 2.The information in the “implementation details” is insufficient. The number of epochs is not given, and neither is the α in the combination of the learning rate and the loss function. This lack of detailed parameter specification makes it challenging for other researchers to reproduce the experiments accurately, undermining the reproducibility and credibility of the research. 3.In this paper, many methods for natural image and medical image segmentation are used for comparison. However, it seems that some very mainstream medical image segmentation methods, such as nnUNet [1] and Swin UNETR [2], are not compared. 4.There is a lack of introduction to the related works. For example, there are many works on using Mamba for medical image segmentation and many works on using Graph Neural Networks (GNNs) for visual tasks, such as ViG [3] (which also uses K - Nearest Neighbors (KNN) to build graphs and then processes them with Graph Convolutional Networks (GCNs)). This limited review fails to situate the current study within the broader context of relevant research, making it hard for readers to grasp the novelty and contributions of the proposed approach. 5.The model’s performance might be restricted by the choice of the graph neural network architecture. The paper only uses a single type of GNN, missing out on the potential advantages offered by other variants. Graph Isomorphism Network (GIN)[4] can better capture graph-level information, Graph Attention Network (GAT) [5] utilizes attention mechanisms for feature selection, and GraphSage [6] is more efficient for large-scale graphs. Exploring these architectures could potentially improve the performance of medical image segmentation and offer new perspectives on the task.
[1] Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier - Hein, K. H. (2021). nnU - Net: a self - configuring method for deep learning - based biomedical image segmentation. Nature methods, 18(2), 203 - 211. [2] Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H. R., & Xu, D. (2021, September). Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In International MICCAI brainlesion workshop (pp. 272 - 284). Cham: Springer International Publishing. [3] Han, K., Wang, Y., Guo, J., Tang, Y., & Wu, E. (2022). Vision gnn: An image is worth graph of nodes. Advances in neural information processing systems, 35, 8291 - 8303. [4] Xu, K., Hu, W., Leskovec, J., & Jegelka, S. (2018). How powerful are graph neural networks?. arXiv preprint arXiv:1810.00826. [5] Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., & Bengio, Y. (2017). Graph attention networks. arXiv preprint arXiv:1710.10903. [6] Hamilton, W., Ying, Z., & Leskovec, J. (2017). Inductive representation learning on large graphs. Advances in neural information processing systems, 30.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- 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?
Positive aspects justifying a weak accept: 1.Innovation: The paper presents novel modules, namely MCMMA and MGGA. MCMMA enhances the Mamba model with multi-step cross-frame state space modeling, and MGGA brings multi-scale concepts to graph neural networks. These innovative ideas, although building on existing methods, offer new approaches for myocarditis lesion segmentation, which shows potential value in the research area. 2.Experimentation: The use of a relatively large dataset of 10,400 Cine-MRI images from two clinical datasets provides a solid basis for the experiments. Even though not all common models are covered, the experiments are sufficient to demonstrate the effectiveness of the proposed method to some extent. 3.Ablation Study: The well-executed ablation study assesses the effectiveness of components, and includes feature visualization and stability analyses. This in-depth analysis helps in understanding the model better and paves the way for further improvement, which is a positive aspect of the research.
Negative aspects that prevent a stronger acceptance and contribute to the weak accept decision:
1.Lack of efficiency comparison: The paper fails to provide crucial efficiency-related metrics such as the amount of computation (flops), the number of parameters, and computational speed. Without these metrics, it is difficult to evaluate the practicality of the method in real-world scenarios where computational resources and time efficiency are important, thus limiting the comprehensiveness of the study. 2.Insufficient implementation details: Key information like the number of epochs and the value of α in the combination of the learning rate and the loss function is missing. This lack of detailed parameter specification hinders other researchers from accurately reproducing the experiments, reducing the reproducibility and credibility of the research. 3.Incomplete comparison: While many methods for natural and medical image segmentation are compared, some very mainstream medical image segmentation methods like nnUNet and Swin UNETR are omitted. This omission may lead to an incomplete understanding of how the proposed method performs relative to the current SOTA methods in the medical image segmentation field. 4.Limited related work introduction: There is a lack of comprehensive introduction to relevant works. For example, works on using Mamba for medical image segmentation and using Graph Neural Networks for visual tasks are not well-reviewed. This makes it difficult for readers to understand the novelty and contributions of the proposed approach within the broader context of existing research. 5.Model architecture limitation: The model’s performance may be restricted as it only uses a single type of graph neural network. Other variants such as Graph Isomorphism Network (GIN), Graph Attention Network (GAT), and GraphSage could potentially offer advantages for medical image segmentation, and exploring them might improve the performance and provide new perspectives, but this opportunity is missed in the current paper.
Overall, due to the combination of the valuable aspects like innovation and solid experimentation along with the significant limitations in various areas, a “weak accept” decision seems appropriate.
- 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.
The rebuttal has addressed the questions I posed and the disadvantages I mentioned. The paper demonstrates high quality, and I contend that it meets the standards of MICCAI.
Author Feedback
We sincerely thank meta-reviewer and reviewers for insightful comments. Special thanks to M1, R1, R2, and R3 for their encouragement and recognition of our novelty. Reviewers appreciate our experiments “sufficient to validate the method’s effectiveness”[R1], “ablation study is well-done”[R2], and “demonstrate the applicability and feature representation capability”[R3]. Main concerns include Reproducibility, Clinical evidence/implications, Design decision of the method, Comparisons and Related work. -Reproducibility[M1,R1,R2,R3]: 1)“available code”; “implementation and hyperparameters details”; 2)“masks acquisition and data dividing”. A: 1)We’ve uploaded the code to GitHub, which will be released after the paper acceptance. We’ve added more implementation details to enhance Reproducibility, such as “epoch is 200, lr is 1e-3, α is 0.6”. 2)All masks were labeled by two experts and verified by a senior expert. For SYC-QC, 20% was randomly selected as test set, with the remaining 80% split into train/val sets via 5-fold cross-validation. SYC-SX served as an independent test set. -Clinical evidence[R3]/implications[R1]. A: 1)We’ve added discussion on clinical evidence: Refs[1-3] show inflammatory cell infiltration can lead to myocardial cell damage, edema, or fibrosis, thereby causing diffuse or focal motion abnormalities. These motion abnormalities can be imaged through cine-MRI. 2)Clinical studies show that lesion distribution and size are crucial for decision-making: extensive lesions need active anti-inflammatory treatment; focal small ones require supportive care; and ring-lesion suggests progressive fibrosis. Due to the page limit, we only report Dice, Precision, Recall and HD, we will pursue this deeper analysis in our future journal version due to its clinical importance. -Design decision of the method [R1,R2,R3]: selection of 1)“downsampled the sequence” and “temporal stride”; 2)“coefficient α”; 3)“scan order”; 4)“patch sizes”; 5)“reliance on a single GNN type”. A: Our ablation studies emphasized demonstrating the effectiveness and role of the two proposed modules, rather than focusing on the specific parameter selection in the method design. Thanks to the reviewers for highlighting this, as it’s key for model reproducibility. Here are our supplement explanations: 1)We adopted temporal strides (∆t=2,3,5). ∆t=2 is for transient abnormalities, and larger ∆t=3,5 shows cumulative effects. But too large ∆t may lose motion information; 2)Our model is most robust when α is between 0.4-0.6. Other values lead to performance drop; 3)Scan order follows Vivim[22]’s ablation study, and pre-results confirm this combination; 4)Patch size was evaluated via empirical validation on our dataset. Though deeper exploration could help, our current setting already shows strong capability. We’ll include parametric sensitivity analysis in the future journal version as suggested. 5)Our GCN-based design focuses on capturing global lesion feature through multi-scale feature aggregation, where GCN serves as a basic feature aggregation approach and works well for this task. We thank the suggestion regarding GIN/GAT/GraphSage’s advantages, which are important to our future journal version. -Comparisons: 1)“lacks comparison with mainstream/relevant methods” [R2,R3]; 2)“efficiency comparison”[R2]. A: 1)In current version, we mainly compared with the latest methods(published in 2023, 2024) to highlight our performance, sorry for the lack of mentioned comparison. 2)We’ve added the details of our method about efficiency: “15.2GFLOPs, 19.89M parameters, and 34.6FPS”. We will add the comparisons in our journal version as suggested by R2,R3. -Related work: 1)“missing related works overview”[R2]; 2)“references”[R3]. A: 1)We have reviewed more related works, such as “ViG uses KNN to build graphs and then processes them with GCN” vs “Our method further explore multi-step and multi-scale graph aggregation.” 2)We have revised the wrong references.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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”.
The authors propose the MG-Mamba framework for myocarditis lesion segmentation in Cine-MRI images. Though there existed some novelty in terms of methodology, the reviewers are concern about its reproducibility.
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
N/A
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A