Abstract

Neural cellular automata (NCA) provide a lightweight alternative to encoder–decoder segmentation networks. However, it can be difficult to decide when a prediction should be trusted. Here, we study uncertainty estimation for NCA-based medical image segmentation without modifying the underlying architecture or retraining the model. Our approach is motivated by viewing the NCA as a dynamical system where convergent attractors correspond to confident predictions. Concretely, we propose resilience, a simple measure that leverages the intrinsic iterative structure of NCAs by probing the stability of the final prediction under small perturbations of the automaton state. Predictions that return to the same solution are deemed confident, while those that change substantially are flagged as uncertain. We evaluate uncertainty by its ability to predict segmentation quality using selective prediction metrics (∆Dice@9 and AURC) and ranking metrics (AUROC and AUPRC). Across multiple medical segmentation benchmarks, resilience identifies failure cases more reliably than baselines, improving trust and safety in NCA-based models.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3943_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/marrlab/resilience

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SadAri_Measuring_MICCAI2026,
        author = { Sadafi, Ario AND Deutges, Michael AND Navab, Nassir AND Marr, Carsten},
        title = { { Measuring Prediction Uncertainty in Neural Cellular Automata } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a training-free uncertainty estimation method tailored to Neural Cellular Automata (NCA) for medical image segmentation. The central idea exploits NCA’s iterative dynamical-system structure: after running the NCA for T steps to obtain the final state S_T and prediction m, Gaussian noise is injected into the latent state, the automaton is allowed to relax for an additional T’ steps to produce a new prediction m’, and uncertainty is defined as u_res = 1 − IoU(m, m’).

  • 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.Conceptually clean and NCA-native. Framing the NCA as a dynamical system and probing the basin of attraction around the final state is a natural fit, more so than transplanting feed-forward uncertainty methods (MC dropout, deep ensembles) onto NCA. The intuition is well-explained. 2.Truly training-free and architecture-agnostic. No dropout layers, no retraining, no modification of the readout — the method can be dropped onto an existing NCA segmentation pipeline at zero cost. This is genuinely valuable for practitioners.

  • 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.Missing the most relevant baselines. The paper compares only against NCA-internal uncertainty signals, but does not include any of the standard uncertainty methods adapted to NCA: MC dropout (which can be applied to NCA by adding dropout to the local update network and retraining — this is not “architecture-changing” in any prohibitive sense), deep ensembles (training multiple NCAs with different seeds), or input-perturbation TTA with stronger augmentations. Without at least one of these as a reference, it’s unclear whether the dynamical-system framing actually buys anything over more conventional baselines. 2.The noise magnitude σ and the relaxation length T’ are the two parameters that fully determine the behavior of the method, yet neither value nor sensitivity is reported. A simple sweep would let readers calibrate the method and would address the obvious concern that the results are tuned per-dataset. 3.Resilience uses one noise sample per image. A natural extension is to compute mean and variance over multiple perturbations, which would likely be both more informative and more stable. The paper does not justify why a single sample is preferred or report results for multiple samples. 4.The NCA baseline is intentionally weak. The paper states that no dataset-specific tuning is performed, leading to mIoU as low as 0.43 on Kvasir-SEG. Uncertainty methods are easier to evaluate when models make many errors, so the strong gains may not transfer to a well-tuned NCA. At least one experiment with a competitively trained NCA would address this concern. 5.Computational cost is not reported. “Highly efficient” is claimed in the conclusion but never quantified. Resilience requires T’ additional NCA iterations beyond the standard T, which is a substantial overhead (potentially 50–100% more compute per image), and this should be made explicit and compared against the cost of TTA. 6.Distribution shift evaluation is mentioned but underreported. Section 3.1 mentions augmenting the test sets of ClinicDB and Kvasir-SEG with geometric/photometric perturbations, but Table 2 does not separate clean vs. perturbed results, so the reader cannot tell whether resilience holds up under shift specifically.

  • 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.

    (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 proposed method is conceptually clean and the empirical results across five datasets are encouraging, but the paper has two structural weaknesses that I find hard to overlook for a methods contribution: the absence of any conventional uncertainty baseline (notably MC dropout, which is the standard reference in the field), and the lack of sensitivity analysis for the two hyperparameters that fully control the method’s behavior. The weak NCA baseline also makes it difficult to know whether the gains generalize to a competitively trained model. The idea is good and worth pursuing, but in its current form the experimental case is incomplete. I would consider raising my score if the rebuttal adds at least an MC-dropout-on-NCA comparison and σ/T’ sweeps.

  • 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

    The paper introduces a method to estimate uncertainty in Neural Cellular Automata (NCA) by perturbing NCA states. Stable segmentation generation with perturbations is correlated with low uncertainty, while instability or variability in the segmentation output resulting from perturbed states is associated with higher uncertainty. This was confirmed quantitatively by estimating the Dice gain using the 90% most confident samples, where the proposed method shows the highest gains overall in both Dice and failure detection compared to competing methods for estimating uncertainty in NCA. An important contribution of the paper is providing a single uncertainty score per image, which assists in evaluating the segmentation’s reliability and in flagging samples for additional human review (failure detection/deferral).

  • 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.

    First to introduce uncertainty estimation method within NCA segmentation approaches, therefore providing a method of assessing the risk associated with NCA-based segmentation with quantitative measures for detecting segmentation failures as presented in this paper.

  • 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.
    • Introduction: As NCAs are not well-known within the community, a more detailed explanation into the mechanism in which NCAs produce segmentations from input images would have been useful. This can be added in the third paragraph in the introduction. What advantages do NCAs provide over standard encoder-decoder-based architectures.
    • Table 1 summarizes the results for the NCA baseline performance on the different datasets. It would have been useful to also have a standard UNet performance to compare it to, or at least discuss why it would not be useful within the scope of the paper.
  • 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

    Questions:

    • Section 2.2: ○ Is the perturbation done only twice? Or is the explanation in section 2.2 only for demonstrative purposes? If it is only twice why? And if it were done several times, how would the u_res score be computed?
    • Section 3.4: For failure detection how was the u_res thresholded in order to determine the threshold for failure (not accepting a segmentation from the model) What criteria was used to determine this threshold.
    • What is the computational cost of introducing perturbations to the states to estimate uncertainty?

    Minor Comments:

    • Page two: “…sharp final mask can still be unreliable if it has not truly settled,…” settled is not the correct or appropriate term to use here.
    • The method “resilience” was italicized in some places and not in others (like in the abstract/introduction). The authors should be consistent by italicizing the method’s name everywhere for ease of legibility.
    • Section 2.1: “Given a 3 color input image” reword to “..RGB input image”
    • Tta (test time augmentation) should be capitalized
  • 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?

    Introduction of uncertainty for NCA-based segmentation is a valuable contribution for improving the safety of deployment of segmentation models.

  • 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



Review #3

  • Please describe the contribution of the paper

    This paper introduces “resilience,” a novel, training-free method for estimating uncertainty in Neural Cellular Automata (NCA) based medical image segmentation models. The core idea is to treat the NCA as a dynamical system and probe the stability of its final prediction. This is achieved by taking the final state of the NCA, adding a small Gaussian perturbation, and allowing the model to run for a few more iterations. The uncertainty is then quantified as one minus the Intersection-over-Union (IoU) between the original and the “relaxed” segmentation masks. The authors evaluate their method on five medical segmentation benchmarks, comparing it against other NCA-native uncertainty baselines. They show that resilience consistently improves selective prediction (∆Dice@90) and is competitive in failure detection (AUROC/AUPRC).

  • 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.

    This paper addresses the important and timely topic of uncertainty quantification. The core idea of “resilience” as a proxy for stability in a dynamical system is elegant and intuitive.

  • 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.Inadequate Contextualization and Baseline Comparison: The paper positions NCAs as a “lightweight alternative” to standard encoder-decoder architectures like U-Net but fails to provide any comparison to them. This is a critical omission. The reported Dice scores in Table 1 are quite low for several datasets (e.g., 0.60 for Kvasir-SEG, 0.74 for ClinicDB). It is highly likely that a standard U-Net would achieve significantly better segmentation performance. If the primary claim is that NCAs are a viable alternative, the authors must demonstrate this through direct comparison.

    More importantly, for an uncertainty paper, there is no comparison to established uncertainty methods on U-Net architectures (e.g., MC-Dropout, Test-Time Augmentation, or Ensembling). The central question is not just whether resilience is the best uncertainty method for NCAs, but whether NCA with resilience is a competitive combination for uncertainty-aware segmentation in general. Without this broader context, the practical relevance of the proposed method is unclear. If a U-Net is more accurate and its uncertainty estimates are more reliable, the motivation for using the proposed NCA-based approach dissolves. The authors must provide a compelling justification (e.g., computational efficiency, specific robustness properties) for their method in the context of the wider literature.

    2.Insufficient and Unsound Evaluation of Uncertainty: The evaluation of uncertainty is superficial and misses key aspects that are standard in the field. My primary concern is the complete lack of analysis regarding the spatial correspondence between uncertainty and segmentation error. A good uncertainty method should produce high uncertainty in regions where the model is incorrect (e.g., at object boundaries or in mis-segmented areas). The paper only reports image-level uncertainty metrics, which obscure this crucial spatial information. The authors should have included analyses that are common for segmentation uncertainty, such as:

    • Visualizing the pixel-wise uncertainty maps and comparing them to error maps.
    • Reporting metrics that measure spatial overlap, such as the Generalized Energy Distance or pixel-wise correlation between uncertainty and error.
    • Analyzing uncertainty calibration (e.g., using Expected Calibration Error). Furthermore, Figure 2 is a missed opportunity. It shows images with high and low uncertainty but does not show the corresponding resilience maps (i.e., a visualization of the difference between the original mask m and the relaxed mask m’). This would have provided at least some qualitative insight into the spatial properties of the proposed uncertainty measure. As it stands, the evaluation does not convincingly demonstrate that the proposed u_res is a spatially meaningful or well-calibrated measure of uncertainty.

    Minor Weaknesses:

    • Lack of Ablation Studies: The method relies on several key hyperparameters, yet their influence is not studied. What is the role of the noise magnitude sigma? How sensitive is the method to the number of initial iterations T and relaxation iterations T’? A thorough ablation study is necessary to understand the method’s robustness and provide guidance for practitioners.

    • Practicality and Generalizability: The paper does not discuss how one would set a threshold on u_res to flag “bad” predictions in a real-world application. Is this threshold dataset-dependent? Moreover, the proposed method is defined for binary segmentation via IoU. It is unclear how it would be extended to multi-class segmentation problems, which are common in medical imaging. The authors should discuss these limitations.

    • Missed Baseline Opportunity: An ensemble of NCAs seems like a natural and powerful baseline for uncertainty estimation. Given that ensembling is a gold standard, its absence is conspicuous. Comparing resilience to an NCA ensemble would provide a much stronger benchmark for its performance.

    Clarity and Presentation:

    The citation formatting on page 2 ([5,10,11,12]) is unconventional; citations should typically appear at the end of a clause or sentence. One can see an over-use of em-dashes, which can make sentences convoluted and feel “LLM-esque.”

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    While the paper presents a clever idea for NCA-based uncertainty, it fails to meet the standards for publication at MICCAI due to two fundamental flaws. First, it lacks a comparison to the most dominant and relevant baselines (i.e., U-Net based methods), making it impossible to judge the practical value of the work. Second, its evaluation of uncertainty is incomplete and lacks the standard spatial analysis required to validate an uncertainty method for segmentation. Combined with the missing ablation studies and issues with referencing, the paper feels premature. The authors should be encouraged to perform the necessary experiments and analyses to properly situate their work and validate their claims before resubmission.

  • 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 paper introduces a training-free uncertainty estimation method for NCAs for medical image segmentation. The reviewers see clear methodological value in this contribution, in particular its architecture-agnostic nature and the fact that the methodological choices are well argued. Despite some concerns on the lack of comparisons against relevant baselines, the work is a solid contribution to the field.



back to top