List of Papers Browse by Subject Areas Author List
Abstract
Accurate survival prediction for adult diffuse gliomas requires robust integration of multi-parametric MRI (mpMRI) and heterogeneous clinical variables across institutions. We propose a multimodal survival framework that combines diffusion-guided imaging representation learning, language-based clinical encoding, and interaction-aware fusion. For imaging, we learn robust representations from standard mpMRI sequences using dual-branch 3D diffusion-based self-supervision. Structured clinical variables are summarized into short patient-level text descriptions and encoded with a frozen biomedical language model, reducing the need for manual covariate harmonization across cohorts. Multimodal fusion is performed using an attention-based graph with a learnable virtual node to capture cross-modal interactions. The model is optimized end-to-end with a Cox proportional hazards loss, diffusion denoising loss, and a stop-gradient consistency regularizer. We evaluate our method on three public cohorts (UCSF-PDGM, UPENN-GBM, RHUH-GBM) under cross-cohort protocols. Our method consistently yields significant risk-group separation across internal and external tests, including the difficult small-sized RHUH cohort.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3503_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/minjicho246/DIVER-Surv
Link to the Dataset(s)
UCSF-PDGM dataset: https://www.cancerimagingarchive.net/collection/ucsf-pdgm/
UPENN-GBM dataset: https://www.cancerimagingarchive.net/collection/upenn-gbm/
RHUH-GBM dataset: https://www.cancerimagingarchive.net/collection/rhuh-gbm/
BibTex
@InProceedings{ChoMin_DIVERSurv_MICCAI2026,
author = { Cho, Minji AND Ra, Sinyoung AND Chung, Jiwon AND Park, Hyunjin},
title = { { DIVER-Surv: Diffusion and Virtual-Node Graph Fusion for Cross-Cohort Glioma Survival } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.The authors introduce a dual-branch 3D diffusion-based self-supervised objective to learn robust imaging embeddings from multi-parametric MRI (mpMRI). By regularizing the 3D U-Net encoder with a denoising diffusion probabilistic model (DDPM) style objective and a consistency loss, the framework promotes the learning of acquisition-invariant features that are less sensitive to inter-site variability, such as differences in scanner protocols.
2.The study proposes a method to handle heterogeneous clinical variables by summarizing them into natural-language patient descriptions via an LLM, which are then encoded by a frozen biomedical language model. These clinical embeddings are integrated with imaging features using an attention-based graph with a learnable virtual node, allowing the model to capture complex cross-modal interactions rather than relying on simple feature concatenation.
- 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.Interaction-Aware Multimodal Fusion: The framework utilizes an attention-based graph with a learnable virtual node to capture complex, non-linear interactions between mpMRI sequences and clinical context. This architecture surpasses standard concatenation methods by enabling sample-specific message passing, which refines global patient representations and maintains significant risk-group separation even under institutional domain shift.
2.Harmonization-Free Clinical Encoding: By converting heterogeneous clinical variables into natural-language summaries via an LLM and encoding them with a frozen biomedical language model, the authors provide a practical solution to inconsistent data naming across cohorts. This approach demonstrates strong clinical feasibility, allowing the model to integrate diverse institutional data without requiring manual covariate alignment or site-specific feature engineering.
- 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 dual-branch 3D diffusion-based objective significantly increases computational overhead, yet the authors do not provide a detailed analysis of training time or memory requirements compared to standard models. Additionally, while an LLM summarizes clinical data, there is no formal verification process to ensure the accuracy of these narrative outputs or to check for potential information loss during the text generation phase.
2.Data Constraints and Resolution Limitations: The evaluation relies on the RHUH-GBM cohort, which consists of only 40 patients, limiting the statistical power for external validation. Furthermore, resizing all MRI volumes to 120x120x64 may discard high-resolution phenotypic features essential for precise glioma prognosis and risk stratification.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
-
Since the framework utilizes a 3D U-Net within a dual-branch diffusion-based setup, the authors should provide data on training time and memory consumption.
-
To ensure reproducibility and assess the reliability of the clinical summaries, the authors should include the exact prompt template used.
-
The MRI volumes were resized to 120x120x64.The authors should discuss whether this downsampling affects the capture of fine-grained phenotypic tumor features, as high-resolution structural insights are often vital for precise glioma prognosis.
-
- 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?
My recommendation of Weak Accept is justified by the framework’s innovative integration of LLM-based clinical encoding and diffusion-guided representation learning to address multi-institutional survival analysis challenges. The use of an LLM to serialize heterogeneous clinical variables into natural-language narratives provides a practical, harmonization-free solution for integrating inconsistent data across cohorts. Furthermore, the dual-branch 3D diffusion objective and virtual-node graph fusion effectively enhance acquisition robustness and capture complex cross-modal interactions. Despite limitations regarding the small RHUH-GBM cohort size and downsampled image resolution, the model’s consistent ability to yield significant risk-group separation across three independent datasets demonstrates strong clinical feasibility and generalizability.
- 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 #2
- Please describe the contribution of the paper
This paper proposed a multimodal framework for glioma survival analysis using mpMRI and clinical data. Specifically, the authors employ a diffusion-guided strategy to learn robust MRI representations across sites, utilize an LLM-based encoder to represent heterogeneous clinical variables, and introduce an attention-driven virtual-node graph to enable interaction-aware multimodal fusion.
- 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.This paper focuses on a clinically significant topic: the limited generalisability of survival prediction models across institutions due to variations in imaging acquisition protocols and heterogeneous clinical data. 2.This paper proposed a clinically practical framework to encode clinical data and mpMRI. Instead of using diffusion for image generation, the model leverages denoising objectives and consistency regularization to learn acquisition-invariant features, which is particularly relevant for handling inter-site variability in medical imaging. 3.The paper includes rigorous cross-cohort evaluation protocol. The model is trained on a single institution and tested on independent external datasets, providing a realistic assessment of generalization performance under domain shift. 4.The paper includes comprehensive ablation studies that systematically evaluate the contribution of each component, including imaging, clinical encoding, and fusion strategies.
- 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.Lack of methodological novelty: while the overall framework is well-designed, the individual components (including diffusion-based representation learning, LLM-based clinical encoding, and graph-based fusion) are not novel on their own. The main contribution lies in their integration rather than a fundamentally new methodological advance. 2.Result not consistently the best: the proposed method does not consistently outperform all baselines across evaluation settings. For instance, in some cross-cohort scenarios, competing methods (e.g., DeepHit) achieve higher C-index values. Although the authors argue comparable clinical stratification, the lack of consistent superiority weakens the overall claim of improved performance. 3.The transformation of structured clinical variables into text representations introduces additional complexity, yet the benefits over standard tabular encoding are not sufficiently justified. It remains unclear whether this step improves performance or introduces potential noise. Ablation test on standard tabular encoding (ML methods) should be applied and compared. 4.Limited interpretability: despite targeting a clinically relevant task, the model lacks interpretability analysis, such as feature attribution, modality importance, or case-level explanations. This limits the transparency and potential clinical adoption of the method.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes a framework for a clinically significant problem: cross-cohort generalisation in glioma survival prediction using multimodal MRI and clinical data. The work is well-motivated and addresses a realistic challenge in neuro-oncology, where data heterogeneity across institutions often limits the applicability of machine learning models.
The proposed framework is technically sound and integrates multiple components in a coherent way, including diffusion-guided MRI representation learning for improved robustness, LLM-based encoding of heterogeneous clinical variables, and an attention-based virtual-node graph for multimodal fusion. In particular, the emphasis on cross-cohort evaluation, where the model is trained on one institution and tested on independent external datasets, is a strong aspect of the work and provides more realistic evidence of generalisation performance.
However, the overall methodological novelty is somewhat limited, as the framework mainly combines existing techniques rather than introducing a fundamentally new approach. In addition, the performance improvements over strong baselines are not consistently observed across all settings, and in some cases competing methods achieve comparable or better results. While the authors attempt to justify this through clinical metrics such as risk stratification, a more detailed analysis of performance variability would strengthen the claims.
Furthermore, certain design choices, such as the use of LLM-based clinical encoding, are not sufficiently justified, and the model lacks interpretability analysis, which is important for clinical adoption. The relatively small size of one external cohort also limits the strength of the validation.
Overall, I find this work to be well-executed and practically relevant, with strong experimental design and clear motivation. Despite its limited novelty and some inconsistencies in performance, the focus on cross-cohort generalisation and the comprehensive evaluation make it a valuable contribution to the field.
- Reviewer confidence
Somewhat confident (2)
- [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
1.A multimodal survival prediction method for adult diffuse gliomas that combines diffusion-based MRI representation learning, language-based clinical variable encoding, and graph fusion. 2.Attention-based graph for multimodal fusion, where modality-specific embeddings interact through message passing, and a learnable virtual node aggregates cross-modal information for risk prediction. 3.Cross-cohort evaluation on three independent glioma datasets (UCSF-PDGM, UPENN-GBM, RHUH-GBM).
- 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 paper has a rigorous setup where external cohorts are never used for hyperparameter tuning or model selection, and evaluates in both directions (UCSF -> UPENN/RHUH and UPENN -> UCSF/RHUH). Including the small RHUH cohort (n=40) as an external stress test adds to this result. 2.The paper reports C-index with 95% bootstrap confidence intervals, hazard ratios, and log-rank p-values. Kaplan-Meier curves are provided for qualitative assessment of risk stratification. This is quite comprehensive, which reinforces confidence in the evaluation 3.Table 2 isolates the contribution of each component and shows that each addition contributes incrementally to both internal and external performance. The ablation is well-structured and clearly presented.
- 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 total evaluation involves roughly 1200 patients across three cohorts. While cross-cohort evaluation is rigorous in design, the small sample sizes limit the statistical power of the reported confidence intervals and p-values. The wide CIs on RHUH cohort (eg 0.504 - 0.753) make it difficult to draw strong conclusions about generalization. The paper would benefit from adding more cohorts or adding a discussion on the statistical limitations arising from these sample sizes. 2.The imaging ablation study (Table 2, row 2) uses a diffusion-guided encoder, but the comparison baselines (DeepSurv, DeepHit) replace diffusion with a standard MSE reconstruction loss. The paper does not isolate whether these gains come specifically from the denoising diffusion objective or simply from having any self-supervised regularization on the encoder. An ablation comparing diffusion denoising against other self-supervised objectives (e.g., contrastive learning, masked autoencoding) on the same backbone would clarify this. 3.The clinical text summaries are generated using Gemini 2.0 Flash, a proprietary model whose outputs may change across API versions. This makes it hard to reproduce, as different runs or future replication attempts could produce significantly different patient summaries, which would propagate to different clinical embeddings.
- 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.
- 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.
(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 addresses a well-motivated and practically important problem. The experimental validation is quite rigorous: true cross-cohort evaluation with held-out external cohorts, appropriate metrics with confidence intervals, and ablations that demonstrate incremental value from each component. The consistent significant risk-group separation across all six evaluation settings, including the difficult 40-patient RHUH cohort, is a strong result that most baselines fail to achieve. While concerns remain around the diffusion objective’s specific contribution versus simpler self-supervised alternatives and the use of proprietary LLM harming reproducibility, the overall quality of the methodology, evaluation design, and presentation is solid.
- 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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 method is well-motivated and solid. The novelty is moderate, as it mainly integrates existing components and shows some variability in performance across settings. Concerns remain regarding limited interpretability, small external cohort size, and insufficient justification of proposed components.
