Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Data-driven longitudinal monitoring of tumor treatment response from computed tomography (CT) is challenged by temporally sparse and irregular follow-ups between patients. Naïve interpolation or unconstrained generative augmentation to fill missing scans often produces anatomically and temporally unrealistic tumor evolution, reducing realism and utility for tasks that reason about response trajectories. We propose OxFlow, a framework combining geometry conditioning via optimal transport (OT) mask interpolation and flow matching to synthesize smoothly varying tumor morphology. Our approach models tumor shape evolution as Wasserstein geodesic interpolation in mask space, yielding smooth and geometrically consistent intermediate shapes. These masks are then used to condition a flow matching model to produce anatomically coherent CT images. We evaluated the proposed framework on a longitudinal CT cohort of patients with lung cancer. Quantitative evaluations confirmed that our approach produced geometrically smooth tumor trajectories and highly consistent anatomical images. Furthermore, we observed that utilizing these synthesized sequences of temporal scans enhanced model accuracy for predicting immunotherapy response compared to standard augmentation techniques. Overall, this work highlights the value of incorporating geometric constraints into generative modeling for longitudinal medical imaging and suggests a general strategy for geometry-aware medical image synthesis.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5716_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{YanJia_Spatiotemporal_MICCAI2026,
        author = { Yang, Jiannan AND Jiang, Jue AND Ma, Tengfei AND Veeraraghavan, Harini},
        title = { { Spatiotemporal Modeling of Tumor Dynamics via Optimal Transport for Enhanced Longitudinal Analysis } },
        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 paper proposes OxFlow to solve the longitudinal monitoring of tumor treatment response in computed tomography (CT). OxFlow synthesizes smoothly varying tumor morphology by combining (1) geometry conditioning via optimal transport mask interpolation and (2) flow matching. The method models the tumor shape volution in the masks space, and predicts the intermediate tumor shapes in a smooth and topology-consistent manner. The proposed framework is evaluated on a longitudinal CT cohort of patients with lung cancer.

  • 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.

    Strengths:

    1. The task of monitoring tumor progression in the image space is relevant and challenging to the audience of MICCAI.
    2. Flow matching using x-prediction sounds reasonable for this task.
    3. The visualization in Figure 2 is nice.
    4. The experiments on durable response prediction is relevant and meaningful.
  • 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.

    Weaknesses:

    1. I am not very convinced by the experimental setup and baselines in the majority of quantitative experiments.
    2. Regarding the experiment setup: Table 1 and Figure 3 are largely assuming what the authors call “ideal progression” of tumor, which basically looks like a linear growth. Unless that is experimentally observed or validated, it’s a strong and unnecessarily correct assumption. I will be happy if the authors can provide evidence justifying this assumption (or correction on my understanding if I misunderstood). Otherwise, what would have been more valuable is if the datasets have intermediate observations between the start and the end states, which are hidden from the model and only used for evaluation for the predicted tumor progression.
    3. Regarding the baselines: Table 1 and 2 are essentially presenting results of ablation studies, but a comparison with existing methods seems to be missing. For example, ImageFlowNet [1], BrLP [2], ∆-LFM [3], IMMFM [4], among many others, are established methods in the field that worth comparing against.

    [1] Liu, Chen, et al. “Imageflownet: Forecasting multiscale image-level trajectories of disease progression with irregularly-sampled longitudinal medical images. “ ICASSP 2025. [2] Puglisi, Lemuel, et al. “Brain Latent Progression: Individual-based spatiotemporal disease progression on 3D Brain MRIs via latent diffusion. “ Medical Image Analysis 2025. [3] Chen, Hao, et al. “Learning Patient-Specific Disease Dynamics with Latent Flow Matching for Longitudinal Imaging Generation. “ ICLR 2026. [4] Islam, Mohammad Mohaiminul, et al. “Longitudinal Flow Matching for Trajectory Modeling. “ AISTATS 2026.

  • 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.

  • 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?

    As mentioned in the weaknesses, I am not very convinced by the experimental setup and baselines in the majority of quantitative experiments.

  • 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.

    I believe the rebuttal did not adequately address the widespread concern among reviewers on the “ideal progression” (or what the authors decided to term “geometric reference trajectory”. The validity of the core evaluation relies on the assumption of the reference trajectory, but the authors have not provided evidence that the reference trajectory is reasonably derived from real disease progression trajectories. I would recommend the authors consider showing the population-level normalized volume progression from the actual data to justify the selection of the reference trajectory. Otherwise it feels like injecting an unjustified prior into the model.



Review #2

  • Please describe the contribution of the paper

    The paper proposes a spatiotemporal modeling framework of tumor dynamics named OxFlow by combining the optimal-transport mask interpolation with flow-matching for image synthesis. The key idea of the paper is to decouple tumor geometry and appearance by firstly generating temporally consistent intermediate tumor masks via Wasserstein geodesics, and then conditioning a generative model for the synthesis of corresponding CT images. The authors have evaluated the proposed method on a longitudinal CT cohort of n=229 lung cancer patients, with experiments assessing both synthesis quality and downstream immunotherapy response prediction. Results demonstrate the improved geometric smoothness of tumor trajectories and increased AUC in response prediction comparing to baseline 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 paper introduces the usage of optimal transport-based interpolation to overcome the limitation in prior longitudinal generative models, and the constraint by Wasserstein geodesics on tumor evolution is conceptually strong

    • The integration of optimal transport and flow matching for image synthesis is novel, and the ControlNet-based conditioning strategy is well-justified and improves the interpretability

    • The paper includes ablation studies and also the experiments show the results of downstream clinical prediction, which further strengthen the paper as the improvement in AUC supports the claim that better temporal modeling could enhance the clinical utility

  • 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 study is conducted on a single-center cohort of n=229 patients, which is relatively limited for developing the method as well as evaluating the model generalizability; it would be desirable to include external validation or cross-institution testing, or showcase on other types of tumors

    • As the authors also pointed out, the proposed method operates on 2D slices with 2.5D formulation, which ignores the full 3D tumor structure and temporal consistency across slices, and should be considered as a major limitation for the tumor evolution modeling

    • The Dice scores for synthesized images as shown in Table 2 seems to be only marginally better than baselines

    • It is unclear that whether interpolation between scans of the same patient could leak temporal information into validation folds, especially under the cross-validation scheme

  • 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.

    (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?

    The major factors for evaluating this paper are the novel methodological advances of combination of optimal transport and flow matching for tumor evolution modeling. Although the current paper may have limitations, it would be beneficial to future work to build upon and further apply to larger cohort or other tumor types.

  • 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.

    Based on the rebuttal, I will remain my decision and score for the paper.



Review #3

  • Please describe the contribution of the paper

    The paper proposes a method for image level interpolation of spatio-temporal time series data. The model enables continual sampling of a smooth change trajectory between observed ground truth images.

  • 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.

    Evaluation of trajectory smoothness, and the match between a propagated mask and a Unet segmentation of the synthesized image suggest favorable performance.

    A downstream task of progression free survival also shows that the model contributes relevant information.

    The overall aim of the paper is interesting and relevant, as the image level completion of disease progression or treatment response trajectories is gaining attention.

  • 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.

    To my understanding, the paper does not present direct evidence that the synthetic images at intermediate time points are similar to actual observed data at this time. For this time series with at least three available time points would have been necessary.

  • 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.

  • 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 presents interesting and relevant results, that will spark discussion at the conference.

  • 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 paper isninteresting,! and might spark rleevant discussions in the community. After the rebuttal i stay with my initial accept assessment.



Review #4

  • Please describe the contribution of the paper

    This paper proposes OxFlow, a framework that integrates Optimal Transport (OT) with flow-matching generative modeling to capture tumor dynamics in longitudinal CT data. The key innovation is modeling tumor evolution as a Wasserstein geodesic in mask space, enabling geometrically consistent intermediate states, and using these as conditions to guide image synthesis. Additionally, the framework explicitly decouples geometry from appearance via ControlNet conditioning, introducing structured temporal constraints into generative modeling.

  • 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.

    Compared to conventional interpolation or unconstrained generative approaches, OxFlow produces smoother, temporally coherent, and anatomically plausible tumor trajectories. Importantly, its performance gains come from improved trajectory consistency rather than per-image fidelity, leading to better generalization in downstream tasks such as treatment response prediction. This demonstrates strong clinical utility and highlights the importance of incorporating geometric priors in longitudinal medical image synthesis.

  • 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 assumption of a linear “ideal progression” in Fig. 3 is questionable. Tumor growth or regression in real clinical scenarios is often highly non-linear and patient-specific. Using a linear trajectory as the reference may oversimplify the underlying biological dynamics and could bias the interpretation of interpolation quality. 2.The details of the downstream classification task are insufficiently described. It is unclear whether the classifier is trained directly on synthesized CT images, how real and synthetic data are combined. More clarity is needed to ensure reproducibility and fair evaluation. 3.The evaluation in Table 1 may lack clinical realism. Tumor evolution is inherently irregular and heterogeneous, and enforcing metrics such as monotonic volume change or smooth trajectories might not fully reflect true disease progression. A more rigorous evaluation using datasets with densely annotated longitudinal ground truth would strengthen the claims. 4.The quantitative results in Table 2 suggest only modest improvements over baselines. Given the added complexity of the proposed framework, the performance gains may not be sufficiently compelling. 5.In Fig. 2, including real intermediate scans (if available) would provide stronger evidence of temporal realism. 6.The inference-time construction of the background condition is not clearly specified. It is unclear whether the background is derived from the initial scan, the final scan, or some combination. 7.The evaluation is limited to a single retrospective cohort. Validation on public datasets would improve the generalizability and reproducibility of the proposed method. 8.The novelty of the work is somewhat limited. While the integration of OT and flow matching is interesting, many individual components (e.g., diffusion/flow-based generation, mask conditioning, and OT interpolation) are well-established. The contribution appears more as an application and integration rather than a fundamentally new methodology.

  • 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.

    (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 paper addresses an important problem in longitudinal medical imaging and proposes a reasonable integration of optimal transport and flow-based generative modeling. However, the overall contribution is limited, as most components are existing techniques combined in a relatively straightforward manner. The experimental evaluation is also not fully convincing: key design choices (e.g., classification setup and background construction) are insufficiently detailed, and some evaluation protocols (e.g., linear “ideal” progression and trajectory metrics) rely on assumptions that may not reflect real tumor dynamics. Additionally, performance gains over baselines are modest, and validation is restricted to a single private dataset, raising concerns about generalizability.

  • 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.

    The use of OT to simulate intermediate tumor masks is not sufficiently justified. Tumor evolution is inherently irregular and heterogeneous, and a geometrically smooth OT trajectory may not reflect the actual biological progression of tumors. Moreover, the current evaluation metrics mainly assess trajectory smoothness or geometric consistency, but they do not validate whether the synthesized intermediate masks are accurate with respect to real intermediate tumor states. Therefore, I suggest that the authors include datasets with real intermediate scans and corresponding tumor masks to directly evaluate the accuracy of the generated intermediate masks. Given these unresolved concerns, I recommend rejection, as the paper still requires substantial revision and validation.



Author Feedback

We thank the AC and reviewers for recognizing the relevance of longitudinal tumor synthesis and geometry-aware tumor trajectory modeling. We address key concerns. AC: Clinical relevance of AUC 0.772.As clinical context, the model AUC exceeds the clinical tumor volume-change based AUC of 0.70.The goal of this work was to develop OxFlow to create a patient-specific interpolated image sequence of tumor geometry changes between fixed-time images, rather than to assess clinical utility. Downstream task shows that using the interpolated sequences in training benefits outcome prediction. AC: FM stochasticity. For given endpoint masks and alpha for a patient, OT provides a patient-specific deterministic geometric path of tumor evolution with appearance-level variations resulting from FM stochasticity. Generating multi-seed synthetic datasets and robustness to mask input variations are future work. R2/R4/R5: Trajectory realism, “ideal progression,” and real scans. We agree tumor evolution is irregular and heterogeneous, and our use of this phrase to describe smooth geometric interpolation without oscillatory effects (Tab.1/Fig.3) was confusing. We will rename “ideal progression” as “geometric reference trajectory.” The OT trajectory provides a geometrically consistent conditioning path between tumor masks from two scan times for intermediate CT synthesis. It is not intended to capture the full biological dynamics of tumors. Direct comparison to real intermediate scans requires aligning the interpolation step with scan timing; otherwise temporal mismatch and synthesis error are conflated. With such alignment, the OT-derived path could serve as a patient-specific geometric reference for analyzing response deviations. R2: Baselines/related work. We agree broader comparisons are valuable and will add the cited works. Unlike our approach, these methods largely follow a whole-image generation paradigm, where lesion/tumor evolution is implicit in image-level/latent dynamics. By modeling shape evolution in mask space before CT synthesis, OxFlow separates geometry from appearance and enables patient-specific interpolation from endpoint masks without learning implicit dynamics from large longitudinal datasets. R3/R5: Modest Dice gain in Tab.2.Dice is measured only when nnU-Net yields valid tumor segmentations, so it excludes failed cases. Tab.2 shows OxFlow has the lowest failure rate among variants. R3/R5: Downstream setup/CV leakage. The classifier is trained directly on CT slices after performing patient-level data splits, eliminating chance of CV leakage. For each training patient, real tumor slices and OxFlow-synthesized intermediate slices are appended into the same patient bag with the original durable-response label. Validation/test bags contain only real CT slices from held-out patients. The fold split is performed before adding synthetic data; synthetic slices are loaded only for training patient IDs. Interpolation pairs are formed only within the same patient and slice position, so no held-out patient contributes synthetic samples to classifier training. R5: Inference background. At inference, x_bg is constructed from the smaller-tumor endpoint scan using the same masked-background operation as training, minimizing residual tumor appearance and avoiding leakage from the larger observed tumor. R5: Novelty. Our contribution is the geometry-first decomposition: explicitly using OT to model tumor-shape evolution in mask space and conditioning CT synthesis on this trajectory, rather than leaving tumor evolution implicit in image-level generation. R3/R5: Generalization and 2.5D. This work demonstrates feasibility on a real longitudinal CT cohort. External cohorts and other tumor types are future work. We adopted 2.5D given cohort size and training stability; while adjacent-slice context and OT support cross-slice consistency in principle, a full 3D OxFlow extension is ongoing. R2/R4: Reproducibility. We will make the codebase public.



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 presented work proposes an interpolation scheme of longitudinal tumor progression data, based on optimal transport interpolation and a generative model based on flow matching. It finds that on a downstream binary classification task regarding progresion free survival after 6 months, there is benefit in using the interpolated lung tumors for this task, compared with other augmentation and synthetization techniques.

    Reviewer reports are discordant. Whereas there is some positive mentioning about the difficulty and relevance of the tackled problem, there are many concerns regarding the methodological innovation, details of the downstream classification task, modest improvements using the method over baselines, and a need for densely annotated longitudinal ground truth data. I would add to these weaknesses that it is not clear if the absolute performance of the downstream classification task is actually clinically relevant (AUC 77% still seems to be a modest predictor of progression free survival, a prediction which is highly relevant to patients). Moreover, it is not clear if in the flow matching component, several repetitions of the interpolated tumors would lead to differently created interpolations, and if that would have an effect on the downstream classification task.

    I follow the overall assessment of reviewers, which is inbetween acceptance and rejectance, and due to the in principle relevant and interesting problem area, which might spark discussion at the conference, I invite the authors to prepare a rebuttal for the work, which should structure the mentioned weaknesses according to their importance and clarify/address them.

  • 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.

    Reviewers are still very much on the borderline for this paper. Given the rebuttal, I think a number of issues have been clarified, and none that would preclude acceptance remain. Due to the interesting longitudinal application scenario, which might spark some discussion at the conference, I tend to accept this paper.



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.

    I think that the idea is interesting. Although some weaknesses still remain, I thought that authors have provided a good discussion on this early-stage but interesting research, and have provided some convincing responses to specific critiques. I strongly encourage authors to include better clarifications and revisions as suggested by reviewers to strengthen the final version. I would argue for accept.



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.

    After considering the rebuttal and updated reviews, I recommend rejection. Although the paper presents an interesting idea, the core progression assumption remains insufficiently justified, and the experimental validation is relatively limited. In particular, the current evidence does not adequately demonstrate that the proposed geometric trajectory reflects real tumor progression. Therefore, I do not think the paper meets the acceptance threshold.



back to top