Abstract

Synthetic contrast enhancement offers fast image acquisition and eliminates the need for intravenous injection of contrast agent. This is particularly beneficial for breast imaging, where long acquisition times and high cost are significantly limiting the applicability of magnetic resonance imaging (MRI) as a widespread screening modality. Recent studies have demonstrated the feasibility of synthetic contrast generation. However, current state-of-the-art (SOTA) methods lack sufficient measures for consistent temporal evolution. Neural cellular automata (NCA) offer a robust and lightweight architecture to model evolving patterns between neighboring cells or pixels. In this work we introduce TeNCA (Temporal Neural Cellular Automata), which extends and further refines NCAs to effectively model temporally sparse, non-uniformly sampled imaging data. To achieve this, we advance the training strategy by enabling adaptive loss computation and define the iterative nature of the method to resemble a physical progression in time. This conditions the model to learn a physiologically plausible evolution of contrast enhancement. We rigorously train and test TeNCA on a diverse breast MRI dataset and demonstrate its effectiveness, surpassing the performance of existing methods in generation of images that align with ground truth post-contrast sequences.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4096_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2025/supp/4096_supp.zip

Link to the Code Repository

https://github.com/LangDaniel/TeNCA

Link to the Dataset(s)

MAMA-MIA dataset: https://www.synapse.org/Synapse:syn60868042/wiki/628716 Duke-Breast-Cancer-MRI dataset: https://www.cancerimagingarchive.net/collection/duke-breast-cancer-mri/

BibTex

@InProceedings{LanDan_Temporal_MICCAI2025,
        author = { Lang, Daniel M. and Osuala, Richard and Spieker, Veronika and Lekadir, Karim and Braren, Rickmer and Schnabel, Julia A.},
        title = { { Temporal Neural Cellular Automata: Application to modeling of contrast enhancement in breast MRI } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},
        page = {603 -- 613}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    (1) propose a novel Temporal Neural Cellular Automata (TeNCA) method for modeling contrast agent dynamics. (2) Supports dynamic time step (Δt), allowing the model to simulate the changes of contrast agent at any time interval. (3) Evaluate model on a multi-center 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.

    (1) The manuscript proposes a novel Temporal Neural Cellular Automata (TeNCA) method to model temporally sparsely sampled DCE-MRI image sequences, which is a direction that has not been explored in existing contrast synthesis methods. (2) As can be seen from the video, the model can simulate the dynamic changes of DCE well and has potential for clinical application.

  • 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 summary and comparison of existing work in the field of breast DCE-MRI synthesis is limited. Recent works on single-time-point prediction should also be summarized, such as,

    • Deep learning to simulate contrast-enhanced breast mri of invasive breast cancer.
    • Synthesis of Contrast-Enhanced Breast MRI Using T1- and Multi-b-Value DWI-based Hierarchical Fusion Network with Attention Mechanism (2) The evaluation of this task should consider more clinical applications, such as evaluating its ability to locate tumors and analyzing the corresponding time-intensity curves. Since the MAMA-MIA dataset has tumor masks and pCR labels, the authors can perform clinical evaluation of both tumor localization and TIC analysis. (3) Lack of ablation study.The role of each module of the method is not clear. (4) Sequential consistency lacks quantitative evaluation. It can be evaluated by metrics such as total variation (TV). (5) The calculation of the metrics needs further clarification. The PSNR is too high. From the visualization, it can not achieve a PSNR of 32. So does LPIPS.
  • 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.

  • 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?

    (1) The summary and comparison of existing work in the field of DCE-MRI synthesis is limited. (2) The evaluation of this task should consider more clinical applications, such as evaluating its ability to locate tumors and analyzing the corresponding time-intensity curves.

  • 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 study has clinical relevance. While some details could be clearer, the manuscript could be accepted.



Review #2

  • Please describe the contribution of the paper

    ​​(1) A novel Temporal Neural Cellular Automata (TeNCA)​​ framework that addresses the challenge of training on ​​temporally sparse, non-uniformly sampled medical imaging data​​ through adaptive loss computation and physical time-step mechanisms, enabling physiologically plausible modeling of contrast enhancement dynamics (novelty in temporal modeling); (2) Rigorous validation on a ​​multi-center, multi-protocol breast MRI dataset​​ with diverse acquisition times and subcohorts, demonstrating robustness in complex clinical scenarios.

  • 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 primary strength of this paper lies in its innovative extension of the Neural Cellular Automata (NCA) framework to propose TeNCA, a temporal modeling approach that effectively addresses the challenge of handling temporally sparse and non-uniformly sampled data in dynamic contrast-enhanced MRI (DCE-MRI). The core novelty resides in tightly coupling NCA’s iterative update process with physical temporal progression through adaptive loss computation and dynamic time-step conditioning. This design enables the model to generate physiologically plausible, continuous contrast enhancement sequences, significantly improving temporal consistency—a critical gap in current methods that treat iterative steps merely as a means to reach static outputs. The authors rigorously validate TeNCA on a diverse, multi-center breast MRI dataset with heterogeneous imaging protocols, demonstrating superior performance over U-Net and diffusion models in key image quality metrics while achieving remarkable parameter efficiency , a crucial advantage for clinical deployment. Particularly compelling is TeNCA’s robustness against hallucination artifacts prevalent in diffusion-based approaches, as evidenced by its lower LPIPS compared to CC-Net, ensuring reliability essential for medical applications. The work further strengthens its contribution through comprehensive temporal stability analysis, revealing TeNCA’s unique capability to maintain consistent performance across varying post-contrast phases—a critical requirement for modeling dynamic contrast kinetics. By bridging NCA’s inherent iterative nature with physical time evolution, this study establishes a novel paradigm for temporal medical image synthesis with broader implications for dynamic imaging modalities.

  • 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 primary weakness of this paper lies in insufficient validation of the generated images’ authenticity and clinical applicability. While TeNCA demonstrates strong performance on image similarity metrics (SSIM, PSNR), distribution metrics (FID, FRD) reveal persistent gaps in global feature alignment with real data, particularly in modeling complex physiological mechanisms of contrast agent kinetics. The experiments employ a single temporal resolution (Δt=8 seconds) without sensitivity analysis, potentially limiting protocol adaptability. Despite emphasizing reduced hallucination risks compared to diffusion models, the evaluation lacks pathological correlation assessments through radiologist blind reviews or histopathological alignment. Although multi-center data is included, preprocessing through linear intensity normalization and fixed-size cropping may compromise scanner-specific signal characteristics, affecting model generalizability. The method comparison remains limited by omitting state-of-the-art spatio-temporal generative models like video diffusion architectures, weakening comprehensive demonstration of advantages.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    Fixed temporal resolution (Δt=8s) without sensitivity analysis or inter-patient variability consideration restricts adaptability to diverse protocols and biological heterogeneity.Discrepancies in FID/FRD suggest incomplete biological realism. While criticizing CC-Net’s hallucinations, the paper inadequately addresses how TeNCA ensures fidelity-safety balance in medical synthesis.Uniform temporal progression assumptions overlook tumor heterogeneity and angiogenic variability, potentially failing to discriminate malignant/benign enhancement patterns crucial for diagnosis.

  • 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.

    Thank you for the authors’ response. While most of the issues were addressed, I remain unconvinced on the key points.



Review #3

  • Please describe the contribution of the paper

    The authors propose a method to predict dynamic contrast enhanced breast MRIs (DCE-MRI) at a low temporal resolution given only the pre-contrast image. The authors utilize Neural Cellular Automata (NCA) to predict the DCE-MRI at s\cdot\delta t with s NCA update steps resulting in a video showing how the contrast agent interacts with the breast leading to the contrast enhanced 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.
    • The manuscript is well written
    • TeNCA generates visually pleasing videos without hallucination outperforming all baselines
    • The authors train and evaluate their method on the union of two commonly used datasets
  • 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 pseudo-code should have some comments or at least additional text next to the variables. e.g.: for datum index j in 0…m do
    • I think line 7 in the pseudo-code is wrong. It should be either: t\gets l\delta t or t\gets t+\delta t
    • The description of the baseline (last sentence Sec. 4.3) is not clear to me
    • Fig. 1 has a lot of white space, and the NCA architecture is way too small. Also, I needed some time to understand that the images in the gray box are a batch. Maybe some additional text would be helpful.

    minor weaknesses:

    • The sentence in the introduction: “The method images changes in tissue enhancement over time” is very difficult to understand
    • The font of math symbols and text in the pseudo-code should be different
    • The indices of S_vis^0 and S_vis^N are swapped compared to earlier notation (page 4 bottom)
    • You mention \delta t = 8s. The data is up to 1024s long. Does that mean you train TeNCA for 128 steps? What is the hardware demand of training an NCA for so many steps?
  • 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.

  • 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 paper is certainly of interest for the community as its results are more visually appealing while maintaining a high temporal resolution.

  • 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 addressed my main concerns and promised the slight improvements of the manuscript in the rebuttal.




Author Feedback

We appreciate the reviewers’ insightful feedback and positive evaluation. Our proposed architecture is recognized as “novel” (R2, R3) and “innovative” (R3), offering a valuable contribution that is “certainly of interest for the community” (R1).

[R2, R3] Consideration of clinical applicability: We agree with the reviewers that an in-depth analysis of clinical applicability, i.e. time-intensity curve analysis, would further strengthen our work. As the first important step in that direction, we have focused on the algorithmic design and development of our model and demonstrating its effectiveness and performance in comparison to SOTA methods using quantitative image metrics on a diverse multi-institutional dataset. Clinical applicability testing with radiological visual inspection, while outside the scope of this presented study, is subject of our future planned work.

[R2] Reference to existing work: Ref. [A] employs a fully convolutional U-Net architecture, which is not substantially different from the multi-time-point U-Net [25] we compare to. Ref. [B] also utilizes a U-Net structure but introduces an attention mechanism, designed to fuse information from different input sequences (T1 and multiple DWI). However, our approach is trained on a single input sequence (T1). Therefore, [B] would basically also collapse into a standard U-Net model, if applied to our problem setting. Hence, in our case, both [A,B] do algorithmically not substantially differ from [25]. Despite this, discussing [B] in our related work section will provide additional context and we will do so accordingly.

[R2] Evaluation metrics: We leveraged established libraries (MONAI and torchmetrics) for computation of evaluation metrics. To ensure reproducibility, we will make the calculations available in the code repository.

[R1] Description of baseline: The ‘baseline’ refers to a model that simply copies its input without altering the images, i.e. no contrast enhancement is applied. This serves as a lower bound for evaluation, and aligns with the approach employed in [19]. We will clarify this in the paper.

[R1] Pseudo code: We thank the reviewer for spotting the typo in the pseudo code and will change it accordingly.

[R1] Figure 1: We agree with the reviewer that Figure 1 can be further improved and will revise it accordingly, incorporating the reviewer’s suggestions.

[R3] Temporal progression and resolution: Scan intervals between consecutive post-contrast DCE-MRI sequences typically range from minutes, with a mean interval of 113 seconds in the dataset used here. Thus, we consider the 8-second temporal output resolution to be small enough (8s « minutes) to capture all possible imaging protocols. Moreover, our method is not restricted to learning a uniform temporal progression; instead, it can dynamically update the images, allowing patient-specific information to be encoded in the hidden channels of the model. We have verified our approach’s ability to model such diverse update rules, resulting in superior performance on the heterogeneous dataset used in this study. Further details on this aspect will be provided in the final paper, subject to space constraints.

[R3] Contrast agent kinetics: Our method has demonstrated superior capabilities in capturing contrast agent kinetics, as evidenced by its enhanced temporal performance across image metrics, as shown in Fig. 3.

We thank the reviewers for their valuable feedback and other minor suggestions to improve the paper, which we will incorporate into our revised manuscript to the best of our abilities and space permitting. Future work will involve in-depth analysis of clinical applicability and additional ablation studies.

[A] Chung et al. “Deep learning to simulate contrast-enhanced breast MRI of invasive breast cancer.” Radiology (2022) [B] Zhang et al. “Synthesis of contrast-enhanced breast MRI using T1-and multi-b-value DWI-based hierarchical fusion network with attention mechanism.” MICCAI (2023)




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    The reviewers have provided varied assessments, leading to a broad range of scores for this submission. As a result, the authors are encouraged to submit a rebuttal to clarify and support their work. Key issues to address include the incomplete review of related literature and the limited experimental validation, particularly regarding clinical applicability beyond standard computer vision metrics. Given the constrained rebuttal space, the authors should focus on the most pressing concerns highlighted across the reviews to provide a clear and impactful response.

  • 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 authors have addressed the reviewers’ concerns. There is some lingering concern about the dynamic nature of the model, which I believe is well-addressed in the rebuttal.



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’

    While the paper presents an interesting approach and improvements are recognized in the rebuttal, there are still key concerns from the reviews regarding biological realism, fidelity-safety balance, and temporal variability. Hope the comments are helpful for future work and submissions.



back to top