List of Papers Browse by Subject Areas Author List
Abstract
Gliomas, especially high-grade gliomas, have a high mortality rate. O6-Methylguanine-DNA Methyltransferase (MGMT) status is crucial for gliomas treatment and prognosis. Traditional diagnosis relies on invasive tissue analysis, which is often infeasible for high-risk patients. While machine learning and deep learning methods using multi-sequence Magnetic Resonance Imaging (MRI) images and radiomics provides a non-invasive alternative, existing methods suffer from low accuracy and poor generalization due to challenges in extracting features from the integrated multi-sequence representation. To address this issue, we propose a Multi-modal feature extraction and Global-aware feature Graph-based deep learning network (MGG-Net), integrating convolutional neural network (CNN) and graph convolutional network (GCN) for multi-modal and multi-scale feature learning. Specifically, MGG-Net consists of multiple CNN-GCN stages, responsible for processing MRI image features and radiomic features at different scales. CNN blocks are used to extract fine-grained and sequence-specific local features from each MRI sequence. These features are then fed into a GCN, which models long-range dependencies and extracts high-level global representations. Finally, the fused multi-scale features extracted are used for classification. Experimental results demonstrate that MGG-Net outperforms previous approaches, effectively leveraging multi-scale and multi-modal information for improved MGMT status classification.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3708_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuHao_MGGNet_MICCAI2025,
author = { Liu, Haoyang and Zeng, Yuwen and Zhang, Xiaoyong and Zhou, Wentong and Nagai, Arata and Kanamori, Masayuki and Endo, Hidenori and Homma, Noriyasu},
title = { { MGG-Net: A Multi-Modal Feature Extraction and Global-Aware Feature Graph-Based Deep Learning Network for MGMT Status Classification in Glioma } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15962},
month = {September},
page = {359 -- 368}
}
Reviews
Review #1
- Please describe the contribution of the paper
Summary and Contributions: This paper introduces MGG-Net, a hybrid deep learning architecture combining convolutional neural networks (CNNs) and graph convolutional networks (GCNs) to classify MGMT methylation status in gliomas. The method processes each MRI modality with CNNs and then aggregates local features globally using dynamic axial GCNs. Radiomic features are also integrated to enhance feature representation. The model achieves strong classification results on the UCSF-PDGM dataset using multi-modal MRI.
- 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.
Major Strengths:
- Innovative Architecture: The integration of CNNs for local extraction with GCNs for global reasoning is well-justified and novel for MGMT classification.
- Radiomics Fusion: The use of radiomic features alongside learned embeddings adds value to the multimodal analysis.
- Multi-scale Learning: The model processes data across multiple scales, enhancing feature richness.
- Good Performance: Experimental results show good performance.
- 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 Weaknesses:
- No External Validation: The model is only tested on the UCSF-PDGM dataset. This limits claims of robustness and applicability across institutions with different acquisition protocols and patient demographics. Without external benchmarking, it is hard to judge its readiness for clinical deployment.
- Interpretability: Although the model integrates GCN and radiomics, it does not provide visualization or interpretability metrics to show what features are most predictive. This limits clinical acceptance and makes it difficult to perform error analysis.
- Code and Training Pipeline Gaps: While the architecture is explained, the paper lacks clarity on training duration, GPU/memory requirements, and optimizer details. This poses a challenge to reimplementation or benchmarking by other researchers.
- Clinical Impact Discussion: The model’s potential role in informing treatment decisions (e.g., MGMT as a predictor of temozolomide response) is not discussed, missing an opportunity to contextualize its real-world impact.
- Limited Dataset Size and Diversity: Although the UCSF-PDGM dataset is well-curated, using only this single-source dataset limits model generalization. The sample size is also modest for training a hybrid CNN-GCN architecture.
- Institutional Bias and Splitting Strategy: The paper does not specify whether stratification was done at the patient level. If patients are split across folds or subsets, site-specific features may be learned, reducing true generalization.
- Superior Alternatives in Literature: Other recent papers have explored MGMT prediction using transformer-based models or multimodal ensembles and have reported that there is no MRI signal for MGMT classification. Acknowledging these would help position this method realistically.
- 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure 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
Clarity of Presentation: Well-structured and logically written. Figures help clarify the architecture. Some parts (e.g., radiomic feature embedding) would benefit from added clarity. Comments on Reproducibility: Moderate reproducibility: while the architecture and dataset are well-described, lack of code and configuration details limit full replication. Constructive Comments:
- Evaluate the model on an external or public dataset for generalization.
- Include ablation studies on the contribution of radiomic vs CNN vs GCN components.
- Discuss interpretability techniques (e.g., SHAP, attention maps).
- Explore lightweight alternatives for real-time clinical integration. Comments on Experiments: Experiments are sound and support the claims. However, detailed ablation and stratified performance (e.g., by tumor type or location) would enhance credibility.
- 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?
Justification for Recommendation: The paper introduces a strong architectural innovation with meaningful performance gains. However, there are several articles stating that there is no MRI signal for MGMT classification. Thus, additional testing is required to validate the method. It is weakened slightly by the absence of cross-site validation and code release, but merits acceptance for its methodological contributions.
- 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.
Addressed my comments.
Review #2
- Please describe the contribution of the paper
O6-Methylguanine-DNA Methyltransferase (MGMT) status plays a critical role in the diagnosis, treatment planning, and prognosis of glioma. However, the current gold standard for determining MGMT status relies on invasive procedures. This study aims to achieve non-invasive and accurate MGMT status prediction using multi-sequence MRI. The authors point out that existing methods lack robustness in multi-sequence feature extraction and thus propose MGC-Net, a new network that leverages the complementary strengths of Convolutional Neural Networks (CNNs) and Graph Convolutional Networks (GCNs) for effective multi-sequence feature learning. The proposed method is evaluated using cross-validation on a publicly available 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.
The paper is well-structured and presents a clear methodological pipeline. The authors adopt a non-shared network architecture for each MRI sequence, which is a reasonable design choice given the distinct characteristics and complementary information provided by different sequences. To aggregate features across sequences, the authors employ a Graph Convolutional Network (GCN) to model the inter-sequence relationships. The authors employed multiple performance metrics to comprehensively evaluate and compare the proposed model. Additionally, visualization experiments are conducted to highlight the advantages of using GCN.
- 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 would be helpful if the authors could elaborate on the essential differences between the proposed Multi-scale Feature Extraction (MFE) module and a standard ResNet architecture. A clear comparison would aid in understanding the unique contributions of MFE and how it offers advantages over conventional CNN backbones. The manuscript mentions that radiomics features are extracted after each stage of the MFE module. Could the authors clarify how this is implemented? Specifically, was the mask downsampled to match the feature map resolution at each stage? If so, how was this handled when the number of channels in the feature maps does not align with the mask? Additionally, the rationale for selecting the top four most relevant radiomic features should be explained. A straightforward alternative would be to treat the features from each MRI sequence as independent nodes in a graph, allowing direct modeling of inter-sequence relationships. Could the authors explain the motivation for choosing DA-GCN over this more intuitive approach? Clarifying the design choice behind the graph construction and attention mechanism would strengthen the methodological justification. In Table 1, the reported performance of vViT is based on T1c and T2 sequences only, whereas the proposed model utilizes additional sequences. This discrepancy raises concerns about the fairness of the comparison. The authors should clarify whether vViT was evaluated using the same input as the proposed model or justify the rationale for using different inputs. The best-performing results in Table 1 appear to benefit from the inclusion of radiomics features, which are not incorporated in baseline methods such as ResNet and vViT. Given that Table 2 suggests radiomics features play an important role in improving model performance, have the authors considered incorporating them into the baseline models for a more balanced comparison? The observed performance improvement when the GAFG module is appended to ResNet and vViT demonstrates its effectiveness. The authors are encouraged to elaborate on how this module enhances feature representation and contributes to the overall model performance. The performance reported in the first and fourth rows of Table 1 appears to be very similar. The authors should provide a statistical test (e.g., paired t-test or Wilcoxon signed-rank test) to determine whether the observed difference is statistically significant or not. An ablation experiment demonstrating the individual contribution of each MRI sequence would be beneficial. This could further justify the replacement of T2 with DWI and SWI and validate the necessity of including specific sequences in the final model.
- 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure 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?
A point of concern lies in the necessity of redesigning the feature extractor (MFE) for each MRI sequence, which is stated as one of the main contributions of the paper. However, as shown in Table 1, the performance difference between the first and fourth rows does not clearly demonstrate a significant advantage of MFE over standard architectures. This raises questions about the practical benefits introduced by the proposed module. Furthermore, when radiomics features are excluded from the model, the accuracy (ACC) and area under the curve (AUC) drop by 0.04 and 0.06, respectively. This suggests that a considerable portion of the model’s performance may rely on hand-crafted radiomic features, which in turn casts doubt on the capability of MFE alone as an effective feature extractor. A deeper analysis or additional ablation studies are needed to validate the actual contribution of MFE to the overall performance.
- 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.
From both the author’s rebuttal and the current version of the manuscript, it remains difficult for me to clearly assess the advantages of the proposed model. As noted by the authors themselves, the performance difference between the first and fourth rows in Table 1 is not statistically significant, which raises questions about the actual benefit of the core architectural modifications. In contrast, a more noticeable improvement is observed between the second and fifth rows, but this alone does not sufficiently demonstrate the overall effectiveness of the model. Moreover, due to the absence of fair and controlled comparisons, it is challenging to evaluate the individual contributions of key components of the proposed framework—including the use of different MRI sequences, incorporation of radiomics features, and the roles of the MFE and GAFG modules. A more systematic and well-documented ablation study would be necessary to support the claimed contributions and provide a clearer understanding of the model’s strengths.
Review #3
- Please describe the contribution of the paper
The paper proposes to combine CNN and GCN to integrate multi-parameteric MRI features for MGMT status classification in glioma. The CNN is introduced to extract local features from individual sequences and the GCN is applied to exploit multi-parametric information fusion. In addition, radiomics features are added to further enhance the classification 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 integration of CNN and GCN for multi-modal image feature extraction and fusion appears to be effective.
- The processing of the radiomics features is interesting.
- 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 interesting that for CNN-based method, introducing DWI and SWI leads to performance degradation, while performance enhancement is observed for the proposed method. The authors should discuss more on this point.
- It seems that the authors only fuse those features that are shared among the different sequences. Does it mean that only those features are useful? Is it reasonable?
- The recall of the proposed method is lower than those CNN-based methods. Why?
- 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.
(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 results are promising. The method is interesting.
- 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 authors have adequately responded to my concerns. Given the intriguing methodological approach and promising outcomes, I recommend accepting the paper.
Author Feedback
We appreciate the reviewers’ valuable comments and address the concerns as follows. R1: 1)5) UCSF-PDGM is currently the only publicly available glioma dataset that includes both SWI and DWI sequences, which are essential to the novelty of our method. 2)4) In Section 3.3, we provide a GCN visualization highlighting the necrotic tumor core as a key predictive region. As surgical testing is often infeasible for Grade IV glioma patients, our method offers a non-invasive alternative to predict MGMT status, aiding treatment decisions (e.g., temozolomide) without biopsy. 3) Training configurations including optimizer and key hyperparameters are provided in Section 3.1. More detailed settings and full pipeline will be released with the public code. 6) We apologize for the omission. Patients were correctly split at the patient level to prevent data leakage and ensure fair evaluation. 7) Studies suggesting MRI cannot predict MGMT status did Furthermore, refer not include DWI and SWI sequences, which are key components of our approach. ences 11 to 13 cited in our paper support the feasibility of MRI-based MGMT prediction. R2: 1) The performance difference stems from two key design aspects. First, unlike ResNet, our MFE module uses a MobileNet-inspired structure—channel expansion, depthwise separable convolution, and compression—which is well-suited for low-SNR, high-contrast but detail-poor images like DWI and SWI. Second, the GAFG module dynamically builds voxel-level graphs, capturing local anomalies and long-range dependencies, enabling effective integration of spatial and non-Euclidean information across multimodal MRI and radiomics. 2) We did not use only shared features; instead, we concatenated features from all four modalities and applied convolutional fusion, thus preserving and utilizing both modality-specific and shared information. 3) Our method’s slightly lower recall reflects an inherent recall-specificity trade-off. Unlike methods with lower precision, our model maintains balanced performance, reducing false positives. As both positive and negative MGMT statuses critically inform clinical decisions, this balanced approach is preferred over solely maximizing recall. R3: 1) This point has been addressed in our response to R2 1). Please kindly refer to that reply. 2) Radiomic features were pre-extracted from high-resolution tumor ROIs, not from downsampled masks or feature maps, and integrated at each stage via learned projections with spatial broadcasting. Four features per modality and region were selected to reduce noise and redundancy, based on prior studies. 3) Modeling each modality as one node is too coarse, while voxel-level nodes create overly large graphs that exceed memory limits and miss complex cross-modality relationships. DA-GCN addresses this by dynamically and implicitly constructing the graph, allowing scalable, adaptive modeling without fixed graph structures. 4) The vViT baseline followed the design in Reference 13, with the only change being its adaptation to 3D inputs. To ensure fair comparison, we retained the original input setting. 5) It was our oversight not to clarify in the manuscript that all results in Table 1, including those of ResNet and vViT, already incorporate radiomic features. 6) We clarify that the GAFG module was not appended to ResNet and vViT in this work. However, we appreciate the valuable suggestion and will include these comparative experiments in future studies. 7) We conducted Wilcoxon signed-rank tests as suggested. No significant difference was found on traditional sequences (Table 1, rows 1 and 4; p > 0.05), while our method showed clear advantages on DWI and SWI (rows 2 and 5). Removing radiomic features also caused no significant drop (p > 0.05), confirming the independent effectiveness of the MFE module. 8) We agree that an ablation study on individual MRI sequences would provide valuable insight. Due to official guideline, we will include this analysis in future work.
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”.
N/A
- 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.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The paper introduces a hybrid network that combines CNN to extract features individually from multiple MR sequences, extract radiomics and then combine all the features within a graph based network to predict MGMT methylation status in patients with brain gliomas. The paper is reasonably well written, although relevant details regarding the radiomic features used, how the features were downselected if any, and how they were extracted etc are left out and should be included in the manuscript for completeness. The results only show incremental benefit and don’t seem to suggest importance of including radiomics, which should be stated clearly. Overall, the methodology though complex is new and was the key factor for positive reviews by the two reviewers. The authors must state the limitations of a overly complex approach and limitations in comparisons, and lack of utility of radiomics in the final version.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
Despite positive feedback by one of the reviewers, the concerns raised by R3 regarding lack of fair and controlled comparisons diminish the impact of the paper.