Abstract

Glioblastoma exhibits diverse, infiltrative, and patient-specific growth patterns that are only partially visible on routine MRI, making it difficult to reliably assess true tumor extent and personalize treatment planning and follow-up. We present a biophysically-conditioned generative framework that synthesizes biologically realistic 3D brain MRI volumes from estimated, spatially continuous tumor-concentration fields. Our approach combines a generative model with tumor-infiltration maps that can be propagated through time using a biophysical growth model, enabling fine-grained control over tumor shape and growth while preserving patient anatomy. This enables us to synthesize consistent tumor growth trajectories directly in the space of real patients, providing interpretable, controllable estimation of tumor infiltration and progression beyond what is explicitly observed in imaging. We evaluate the framework on longitudinal glioblastoma cases and demonstrate that it can generate temporally coherent sequences with plausible changes in tumor appearance and surrounding tissue response. These results suggest that integrating mechanistic tumor growth priors with modern generative modeling can provide a tool for generating controlled synthetic data to support downstream neuro-oncology workflows. In longitudinal extrapolation, we achieve a consistent 75\% Dice overlap with the biophysical model while maintaining a constant PSNR of 25 in the surrounding tissue. Our code is available at: https://github.com/valentin-biller/lgm.git

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4612_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/valentin-biller/lgm.git

Link to the Dataset(s)

N/A

BibTex

@InProceedings{BilVal_TumorFlow_MICCAI2026,
        author = { Biller, Valentin AND Bubeck, Niklas AND Zimmer, Lucas AND Erdur, Ayhan Can AND Nagar, Sandeep AND Meyer-Baese, Anke AND Rueckert, Daniel AND Wiestler, Benedikt AND Weidner, Jonas},
        title = { { TumorFlow: Physics-Guided Longitudinal MRI Synthesis of Glioblastoma Growth } },
        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

    The authors propose a physics-guided generative approach, for synthesizing glioblastoma MRIs, conditioned on tumor concentration fields. The authors propose a flow-based generative approach, constrained on a biophysical growth modelling of tumor concentration over time.

  • 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 idea of the paper is very interesting and, to my knowledge, novel. The biophysical integration in the synthesis of tumor brain MRI is, in general, meaningful and promises more robust and plausible results.
    • The evaluation of not only the synthetic tumor tissue, but also the anatomical tissue surrounding it, is of bis importance, since this is a typical pitfall for other generative approaches.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The major weakness seems to be evaluation and results. The quantitative evaluation in Tab. 1 compares only against one other method, even though multiple approaches are shown to be visually compared in Fig. 3.Picking only Med-DDPM seems as an unfair comparison.
    • Visually, the results do not look convincing. The normal anatomical tissue seems to be generated realistically; however, the tumor tissue looks very different to the ground-truth (Fig. 3 and Fig. 5). An in-depth discussion on this issue would be helpful.
  • 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The method is very interesting, however, the results do not seem convincing enough yet.

  • 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 addresses many of the issues mentioned in my review. Overall, the approach is very interesting.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a physics-guided generative framework for longitudinal brain tumor MRI synthesis by integrating a biophysical tumor growth model with a latent flow-matching generative model. Specifically, the method leverages spatially continuous tumor concentration fields derived from a reaction-diffusion model to condition a generative network, enabling controllable synthesis of tumor appearance and temporal progression. The approach aims to bridge the gap between biophysical tumor modeling and image-space generation, allowing the synthesis of anatomically consistent and temporally coherent tumor growth trajectories from cross-sectional MRI data. Experimental results suggest improved image fidelity and tumor consistency compared to baseline generative 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.

    1.The paper explores an interesting and potentially impactful direction by combining a biophysical tumor growth model with generative modeling. The integration of spatially continuous tumor concentration fields, rather than discrete segmentation masks, is conceptually appealing and may allow finer-grained modeling of tumor infiltration patterns.

    2.The manuscript is clearly written and generally easy to follow, which helps convey the high-level idea of the framework.

    3.The experimental section includes both qualitative and quantitative evaluations across multiple datasets and settings

  • 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.Unclear experimental details regarding VAE encoder. Some important implementation details are not sufficiently clarified. In particular, it is unclear whether the conditioning components (modality m, tissue segmentation c_s, and tumor concentration map c_tc) are encoded using a shared pretrained VAE encoder or separate encoders. This distinction is important, as it affects both the representation consistency and the validity of the conditioning mechanism.

    2.Concerns about fairness in downstream evaluation. The fairness of the downstream segmentation comparison is not fully convincing. In particular, when comparing against Med-DDPM, the conditioning is based on a hard segmentation mask obtained by thresholding the tumor concentration map. The choice of threshold (e.g., 0.2, 0.6) may significantly influence the resulting mask quality and therefore the Dice score. Given that Med-DDPM reports substantially lower Dice scores, it is difficult to rule out whether this performance gap is partially due to differences in conditioning quality rather than inherent model capability. A more controlled comparison would strengthen the validity of the conclusions.

    3.Limited alignment between generated tumors and ground truth. From the qualitative results (e.g., Fig. 3 and Fig. 5, left), the generated tumor regions still exhibit noticeable discrepancies compared to the ground truth. While the generated tumors may appear visually plausible, the spatial alignment and shape fidelity are not sufficiently accurate in several cases.

    4.Lack of reliable quantitative evaluation for tumor realism. The evaluation of tumor realism relies largely on visual inspection and indirect metrics, without a dedicated quantitative measure that directly assesses whether the generated tumors are clinically meaningful or biologically accurate. This limits the ability to judge the practical utility of the method in real clinical scenarios.

    5.Anatomical consistency of non-tumor regions is questionable. Although the tumor appearance is visually reasonable, the results in Fig. 3 and Fig. 5 (left) suggest that non-tumorous anatomical structures may be altered or distorted in some cases. This raises concerns about whether the method consistently preserves anatomical correctness, which is critical for clinical applicability.

  • 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

    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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    My recommendation is mainly driven by concerns about the limited strength and persuasiveness of the experimental evaluation. See the weakness from 2-5.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    1.The postoperative Lumiere Dice remains relatively low (0.42 vs. 0.22), which still raises concerns regarding the reliability and potential clinical applicability of the generated longitudinal progression.

    2.The authors only report tissue-specific Dice scores (CSF 0.65, GM 0.80, and WM 0.82) for their own model, without comparisons to other baselines. Therefore, it is difficult to assess whether these values are sufficiently strong to support the claimed anatomical realism and preservation of non-tumorous structures, given 0.65 is not a good score on CSF.

    3.I would suggest additional experiments to better demonstrate the clinical utility of the proposed framework. For example, evaluating whether the synthesized data can improve downstream tasks would provide stronger evidence of practical value. With stronger clinical validation, the paper would be more convincing for publication.



Review #3

  • Please describe the contribution of the paper

    This paper proposes TumorFlow, a physics-guided generative framework for longitudinal glioblastoma MRI synthesis.

    Key contributions:

    Combines biophysical tumor growth modeling (reaction-diffusion PDE) with generative modeling (flow matching in latent space) Conditions generation on continuous tumor concentration fields instead of segmentation masks Enables controllable, temporally consistent tumor growth simulation from single time-point MRI Demonstrates longitudinal MRI synthesis with temporal coherence and anatomical consistency

    The method integrates:

    Pretrained 3D VAE Latent flow matching model Biophysical tumor growth simulation ControlNet-style conditioning

  • 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.Strong Novelty and Conceptual Contribution Combines: Physics-based tumor growth models Modern generative modeling (flow matching) Conditioning on continuous tumor concentration fields is a significant improvement over mask-based conditioning This is a clear step forward in controllable medical image generation

    2.Clinically Meaningful Problem Addresses key limitation: MRI cannot capture full tumor infiltration Provides: Longitudinal progression simulation Potential for treatment planning and monitoring

    3.Elegant Integration of Physics and Learning Reaction-diffusion model (Eq. 2, page 4) provides: Interpretability Biological grounding Generative model adds: Realistic imaging appearance Strong alignment between mechanistic modeling + data-driven learning

    4.Longitudinal Consistency Modeling Novel use of: Backward-forward latent transport (Eq. 3) Produces: Temporally coherent sequences Figures (page 2 & 7): Show consistent tumor evolution while preserving anatomy

    5.Strong Experimental Setup Large dataset: ~3600 subjects across multiple cohorts Includes: External datasets (UCSF, TCGA, Rembrandt) Evaluation includes: FID, KID, MS-SSIM Dice for tumor adherence Temporal consistency metrics

    6.Clear Improvement Over Baselines Outperforms: Med-DDPM baseline Diffusion and ViT ablations Table (page 5): Dice: 0.739 vs 0.344 baseline Large improvements in FID/KID

    7.High-Quality Presentation Clear figures: Fig. 1: longitudinal progression Fig. 2: architecture Fig. 5: real patient comparison Well-structured and readable

  • 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.Evaluation Relies on Proxy Metrics (Major Concern) Uses: FID, KID, MS-SSIM Dice via pretrained segmentation model These do not fully validate clinical realism or utility No: Radiologist evaluation Downstream task validation (e.g., prognosis, segmentation improvement)

    2.Weak Ground Truth for Longitudinal Evaluation True longitudinal tumor progression: Not available Evaluation uses: Biophysical model as pseudo-ground truth This creates a closed-loop validation problem

    3.Limited Comparison to Recent SOTA Compared to: Med-DDPM Missing: Recent latent diffusion / flow-based medical models Other physics-informed generative approaches

    4.Complexity and Reproducibility Concerns Pipeline includes: VAE Flow matching Biophysical model ControlNet conditioning High complexity may: Limit reproducibility Require significant domain expertise

    5.Assumptions in Biophysical Model Fixed parameters: Diffusion coefficient (D) Proliferation rate (ρ) Real tumor growth is: Highly patient-specific Influenced by treatment

    6.Limited Real Longitudinal Validation Lumiere dataset used (post-op cases) Model not trained for post-surgery scenarios Limits real-world applicability

  • 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

    This is a strong and innovative paper that pushes forward the integration of physics-based modeling and generative AI in medical imaging.

    The idea of conditioning on tumor concentration fields is particularly impactful and could enable more realistic and controllable simulation.

    However, the evaluation remains the main limitation:

    Metrics are indirect No clinical validation Longitudinal ground truth is missing

    To strengthen the paper:

    Include clinical or expert evaluation Evaluate downstream utility (e.g., segmentation, planning) Compare with recent generative models Explore patient-specific parameter estimation

  • 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 presents a novel and well-motivated approach combining physics-based tumor modeling with generative learning, addressing an important clinical problem.

    Strengths:

    Strong conceptual novelty Good experimental validation High clinical relevance

    Weaknesses:

    Evaluation relies on proxy metrics Limited real longitudinal validation

    Overall:

    Methodological contribution is strong enough for acceptance But lacks full validation for a higher score

  • 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.

    Addressed the comments to my satisfaction.



Author Feedback

We thank the reviewers and metareviewer for recognizing the novelty and clinical relevance of combining biophysical tumor growth with conditional generative modeling. We focus on shared concerns: tumor/anatomical fidelity, fairness in comparison and evaluation, and implementation details. Please note that all numbers presented in the following sections are reanalyses of existing data, not new experiments. Validation and Tumor Realism. We agree that predicting the MRI phenotype from concentration fields is severely ill-posed, as GBM appearance varies across biology, acquisition, and treatment. Our goal is not voxel-identical reconstruction, but controllable generation that is spatially adherent to a tumor-growth field while preserving anatomy. To address concerns about circular validation, we compute Dice against expert annotations from the longitudinal Lumiere dataset. Despite strong domain shift (post-operative cavities were unseen during training), TumorFlow achieved Dice 0.42 vs. 0.22 for Med-DDPM. In the preoperative setting, TumorFlow achieved 0.78 vs. 0.36.These results demonstrate that, while the simple Fisher-Kolmogorov model introduces some spatial alignment error (Fig. 5), our generative framework remains highly responsive to the provided conditioning and confirms that post-operative progression remains a limitation. We now state that clinical deployment would require resection-aware concentration maps and post-operative training scans. Nonetheless, our results clearly highlight the potential for coupling biophysical models and conditional generation. Anatomical Consistency. We added non-tumor-region analyses beyond the existing PSNR results. Tissue-specific Dice outside the tumor was CSF 0.65, GM 0.80, and WM 0.82, demonstrating preservation of healthy anatomy from tissue-segmentation conditioning. We will report these results in Table 1 to detail the anatomical consistency. Also, we re-analyzed distant healthy tissue separately from tumor-adjacent tissue: PSNR was 18.66 vs. 17.74, respectively, consistent with expected local changes near simulated tumor growth and infiltration, and broad preservation elsewhere. Comparison Fairness and Scope. Med-DDPM thresholds [0.2,0.6) and ≥0.6 were fixed from biophysical concentration maps, not tuned to disadvantage the baseline; all methods are evaluated against the same target masks using the exact same thresholds, ensuring fairness. While our 192×192×144-to-240×240×155 resampling may affect segmentation metrics, we want to point out that TumorFlow still exceeds the Med-DDPM Dice reported in the original paper (0.62±0.29). We further clarify that Figure 3 presents qualitative comparisons across the four models reported in Table 1, and will revise the potentially confusing category label ‘Ablation’ (for ViT and DDPM variants) to more accurately characterize these as architectural comparisons. The limited number of SOTA baselines reflects the lack of public pretrained weights for 3D spatially conditional medical generators; the DDPM and ViT variants were therefore implemented as principled baselines. Implementation Details. Modality m is not encoded; tissue segmentation c_s and tumor concentration c_tc share the pretrained VAE encoder for representation consistency. We will clarify this and provide more detailed implementation specifications in the camera-ready version to compensate for the complexity of our framework. Biophysical Assumptions and Positioning. Fixed diffusion/proliferation parameters enable controlled synthesis and isolate the generative model’s contribution. We now make this limitation explicit and add patient-specific parameter estimation, treatment effects, and resection-aware concentrations as future work. We also reframed the first contribution to align with the demonstrated scope: training on cross-sectional preoperative data and synthesizing plausible longitudinal trajectories, rather than claiming fully validated patient-specific predictions.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper proposes TumorFlow, a physics-guided longitudinal MRI synthesis approach for modeling glioblastoma growth. Reviewers R1 and R2 consider the method interesting and novel but raise concerns about the experimental validation: the results are not sufficiently convincing, and the evaluation lacks the strength and rigor needed to support the claims. They recommend more comprehensive validation and a more persuasive experimental design. Reviewer R3 is more positive, highlighting the value of integrating biophysical modeling with generative approaches, while also noting that the validation remains incomplete. Overall, the combination of physics-based modeling and generative techniques is both novel and clinically relevant. The work has potential, provided that the experimental concerns can be adequately justified.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    After rebuttal, two reviewers now recommend acceptance, emphasizing the strong methodological novelty and improved clarifications on evaluation and implementation, while one reviewer maintains a reject mainly due to still limited clinical and downstream-task validation. However, the work is sufficiently strong: the authors have reasonably addressed fairness and anatomical preservation concerns, clearly stated the remaining limitations, and appropriately reframed the scope to controllable physics-guided generation rather than fully validated patient-specific prediction. The authors should further stress in the camera-ready that longitudinal and clinical validation remain important future work.



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.

    While there were concerns about some limitations in the evaluation, reviews were generally positive, highlighting the strengths. Recommendation would be to accept given the clarity and potential impact of the study.



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.

    The concerns regarding the experiments have not been completely addressed.



back to top