Abstract

Thermal ablation is an increasingly utilized treatment modality for both secondary and primary hepatic tumors. However, it presents significant challenges in treatment planning, particularly when employing multiple applicators. Numerical methods for evaluating the effectiveness of an ablation procedure plan can assist in this task, but they are often computationally intensive or too simplistic, making them impractical for interaction or fast optimization loops in automatic planning. This paper introduces Chained Neural Cellular Automata (C-NCA), a deep learning approach that allows to quickly estimate cell death in thermal ablation procedures. The C-NCA model is trained on a dataset generated by a numerical simulation. When compared to existing methods, the C-NCA achieves comparable accuracy with substantially reduced computation time, thereby making it suitable for interactive planning, instant visualisation, fast automatic planning or even real-time surgical replanning, and potentially enhancing clinical workflows.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4370_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)

https://figshare.com/s/a4d1e9a9ededdcdeef39?file=52676279

BibTex

@InProceedings{MehJon_CNCA_MICCAI2025,
        author = { Mehtali, Jonas and Verde, Juan Manuel and Essert, Caroline},
        title = { { C-NCA : Chained Neural Cellular Automata for Fast and Accurate Thermal Ablation Estimation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper describes a deep learning based cellular automata surrogate model for thermal ablation simulation.

  • 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 is clearly written and appropriately describes the motivation and key contributions of the presented work.
    • The work considers a broad spectrum of thermal techniques and appropriately presents the key formulas for numerical simulation.
  • 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 state of the art, regarding other relevant works on thermal treatment simulation, both through numerical techniques and faster models, is not sufficiently addressed.
    • The suitability of the proposed cellular automata based method is not convincingly presented, in terms of how much it is expected to model of the original numerical methods, or whether there are theoretical limitations that can limit their applicability in certain cases.
    • Results are reported on a synthetic dataset, which is constructed with simulation of the numerical model the proposed method is trained on. More realistic datasets would be needed to validate the proposed technique.
  • 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?

    The presented work has limitations in the justification of its adequacy for the proposed simulation task. Further, the sole use of experiments on simulated data (of the same models the method is trained on) is a limitation.

  • 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.

    The rebuttal addresses only part of the issues raised by the reviewers (understandable due to limited space, but this also points at the relatively high number of issues the reviewers found). Although an interesting paper, I lean towards rejection.



Review #2

  • Please describe the contribution of the paper

    This work proposes the use of Chained Neural Cellular Automata (C-NCA) to enable fast predictions of the expected ablation zone in RF ablation procedures. C-NCA iteratively updates an initial 3D seed state, which comprises the initial tissue survivability, the ablation duration (encoded per voxel of the ablation device), and three inhomogeneous tissue parameters (perfusion rate, density, and specific heat). The proposed method is trained on 2500 synthetic simulations generated by the Finite-Difference-Method and with varying numbers of ablation probes and varying regions of interest. On a test set of 500 additional simulations, C-NCA demonstrated high accuracy and interative run-times (>250fps). In addition, the authors present an ablation study on the number of NCA update steps.

  • 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 scope of the paper is very relevant as fast estimation of the expected ablation volume is extremely important to precisely plan ablation procedures

    (2) The use of neural cellular automata in the field of RF ablation zone modeling is novel and interesting

    (3) The method demonstrates high accuracy at very high run-times, which bodes well for the integration into planning software

    (4) The authors shared the synthetic database (seed state + final tissue survivability)

  • 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) Even though the paper is very interesting to read, the paper is lacking clarity in several places, making it difficult for the reader to follow

    (2) The paper is missing several details and without access to code it is not easily possible to replicate the results

    (3) Few claims by the authors are slightly overexaggerated, e.g., the authors claim that they are pioneering the use of NCA for estimating heat propagation, but their model is not modeling the evolution of temperature fields.

    (4) The paper would greatly benefit from a deeper discussion of the method and results and for providing better context. For instance, the authors compare their run-time against the reported numbers by Meister et al (2022). The referenced work, however, cannot be directly compared since Meister et al. predict the temperature evolution and also consider the heat advection subject to the underlying blood velocity field. Could C-NCA also account for the advection? Could it account for Dirichlet or Neumann boundary conditions?

  • 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
    • General comment: Please be careful with your wording of the contributions. For instance, the statement in the abstract “The method leverages personalized patient details, such as anatomical and tissue features, to effectively mimic heat distribution and the consequent tissue death.” is not necessarily true, because you are not modeling the evolution of the heat distribution, but the growth of the ablation zone.

    • Introduction, contribution #5: Could the authors please elaborate what reference is considered as the state-of-the-art? If the referenced work by Meister et al. is considered, how do you compare the accuracy of both works (the referenced paper reports DICE scores for the ablation zones and RMSE for the temperature fields)?

    • Section 2.1, NCA extension: Could the authors please provide additional insights into why common components of conventional NCA pipelines (like Sobel filtering, stochastic updates, and life masking) were eliminated?

    • Section 2.1, feature description: (a) The abstract and introduction mention that anatomical features are used, but the list of input features does not list masks denoting anatomical structures. Do I see it correctly that the anatomy is (implicitly) encoded as spatial changes in the tissue parameters (e.g, perfusion rate)? (b) How do you define the applicator geometry? Is the same probe geometry (= type of ablation probe) used in all simulations? (c) How is the ablation duration encoded when multiple ablation probes are applied sequentially (e.g., 5 probes, each applied for 5 mins, for a total of 25 mins)? (d) Why is the inhomogeneous thermal conductivity not included in the initial seed state? (e) Several studies suggest that there is a temperature dependence of the tissue properties. Can this method be extended to account for this effect? (f) Please add the limits used to normalize the feature channels.

    • Section 2.1, C-NCA: To my understanding, the neural network starts with the initial seed state and then updates all voxel states iteratively. While I understand that the network is trained to converge to the expected tissue survivability, it is not clear to me what happens to the other features of the initial state. Could the authors please provide additional information? In addition, could the authors please clarify whether only the final expected ablation zone is used for the loss computation or whether all intermediate predictions are used for training the network?

    • Section 2.2: What anatomical structures and their associated tissue parameters were used from the IT’IS database?

    • Section 2.2, Eq. 2: Could the authors please provide additional information on how this model is used? How is V_i computed?

    • Section 3.1: Unfortunately, I am having a hard time understanding how the 20 CT scans are used to obtain 3500 simulations. For instance, the sentence “The same process was used to build the 500 volumes of the test subset, but with two different anatomies, unseen during training and testing, resulting in 5 more cases.” is confusing me. Could you please provide more information on how the data is split, augmented, and how many simulations per case were run? What geometry have the added tumors? In addition, could the authors please share their thoughts on whether the trained C-NCA would also work in instances where multiple probes are simultaneously applied (instead of a sequential ablation)?

    • Shared synthetic data: I looked at several examples and for instance under training -> synthetic_0 the applicator end point seems to be located outside of the liver. Could you please share your thoughts if this could be a problem? In addition, it seems like the full ablation probe was modeled. What tissue parameter are you prescribing for the probe?

    • Paragraph “Solution Convergence”: Please clarify that NCA steps refers to the variable N_c

    • Section 3.2, last paragraph: I believe it would be better to mention the choice of the power setting in section 3.1

    • Section 3.3: The referenced paper by Meister et al. simulates 10 min of real intervention (with heating only in the initial 7 mins) within 0.06s. Please correct your statement. Most importantly, I believe the paper could greatly benefit from a deeper discussion of the results and from more context when comparing the proposed method against alternative approaches. For a fair assessment of different methods, please acknowledge differences in the biophysics model, in the prediction task, and, when comparing run-times, differences in the used hardware.

    • Table 1: In your comparison of FDM and C-NCA, how is the time step of the FDM selected?

    • Ablation study on number of NCA update steps: While I understand the choice of keeping N_c x N_s constant, do I understand it correctly that the method with 1 NCA update has only 1/3 of the trainable parameters compared to the variant with 3 NCA updates? For a fair comparison, it may be good to make sure that both networks have the same capacity to learn the complex underlying function.

  • 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 proposed method for fast emulation of ablation zone growth is very interesting, relevant, and the experiments and results are convincing. However, the manuscript is lacking clarity and important details, hence the method would not be reproducible with the presented information. In addition, the paper would benefit from a deeper discussion.

  • Reviewer confidence

    Very confident (4)

  • [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 done a great job in addressing my questions and concerns. Even though they couldn’t comment on each and every point due to the character count limit, the extremely clear and comprehensive answers further convinced me that this is a high-quality contribution. In particular, the proposed method is very novel and it tackles a real and important clinical challenge in a creative way. In addition, the evaluation presents promising results both in terms of accuracy and run-time.



Review #3

  • Please describe the contribution of the paper

    The authors present an approach to predict cell death during thermal ablation procedures using neural cellular automata (NCA). This machine learning based approach could provide an alternative to numerical simulation, which is computationally expensive. The proposed algorithm uses a grid representing the current state of the tissue, which is updated in a series of NCA update steps. The main contribution of the paper is the application of NCA to predict cell death during thermal ablation.

  • 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 presented approach is very innovative and addresses a clinically relevant problem.

    While the approach is only evaluated in simulations, the results (both speed and accuracy) are convincing, and the analysis seems appropriate.

  • 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 authors refer to chained NCAs as a major contribution which is a computationally more efficient version of traditional NCAs. However, from the description it is not entirely clear how C-NCA differ from traditional NCAs. It seems mostly like a simplified version of traditional NCAs.

  • 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.

  • 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.

    (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?

    Innovative approach addressing clinically relevant problem.

  • Reviewer confidence

    Very confident (4)

  • [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.

    Interesting and novel approach with convincing evaluation (even if only in simulation).




Author Feedback

We thank the reviewers (R1,2,3) for their constructive feedback and appreciation of our work’s novelty and relevance in CAI. Due to space limits, we focus on replying below to key comments grouped into main categories.

Temperature vs. ablation: Although not directly producing a heat map (R2), our method accurately estimates cell death induced by thermal damage by implicitly modeling temperature field effects and evolution, going beyond heat distribution, with low RMSE and high speed.

Novelty and difference C-NCA / NCA: C-NCA is a novel version of NCA, simpler but elegant, impactful and computationally more efficient while improving its expressiveness (R3). We removed stochastic updates due to a lack of evidence they improved performance (10.1109/ACCESS.2024.3382541). Sobel filtering and life masking were eliminated, as these tasks can be learned through the chained architecture, with, for example, one NCA step per filter (R2,3). The network is trained by backpropagating through all update steps, constrained only by the final ablation zone estimation. Features in the seed state are free to evolve. We observed, for example, that perfusion layers with high blood flow grow outwards, capturing heat sink effects (R2).

Data: The dataset includes 25 tumors in 7 anatomies from 3D-IRCADb-01: 12 native and 13 added near vessels (scaled/rotated). For each case, simulations were run with 1–3 non-overlapping, randomly oriented applicators of identical geometry. We used 20 cases in 5 anatomies for training/testing (150 simulations per case, total 3000) and 5 cases in 2 unseen anatomies for validation (100 simulations per case, total 500) (R2). Our simulation uses the segmented organs from the database (liver, arteries and veins, tumors), as masks with different tissue properties (density, heat capacity, thermal and electrical conductivity) in the initial state (R2). Though based on synthetic data, our evaluation uses a robust pipeline with state-of-the-art FDM and a validated cell death model, providing a solid foundation for clinical translation (R1).

Parameters: Although Ns=1 indeed has fewer trainable parameters than Ns=3, it maintains constant NcxNs​, enabling fair comparison under equal computational budgets. Our goal was to highlight improved performance under the same runtime constraints, leaving broader architectural exploration for future work (R2). For the computation of the electric potential Vi, we refer readers to (10.1016/j.apm.2016.11.032) (R2). The chosen timestep for FDM was 0.025s, which satisfies the stability criterion detailed in [7] (R2).

Sequential/simultaneous: While RFA/MWA do not typically involve simultaneous activations, modifying the applicator channel in the voxel state could enable both synchronous and asynchronous applications, which is especially relevant for cryoablation, where concurrent probe activations produce complex combined ablation patterns (R2).

Advection: Though our model assumes a static setting, its chained design can approximate propagation and collision steps like in Lattice-Gas Cellular Automata, which are known to simulate fluid dynamics. This enables potential extension to dynamic scenarios, including temperature-dependent properties. Boundary effects are implicitly learned via inert border voxels approximating first- and second-type conditions (R2).

Shared data: Edge tumors may cause some endpoints to be drawn randomly outside the liver. These low-vascular cases yield low loss and do not hinder training, allowing the model to focus on clinically relevant, complex regions (R2).

State-of-the-art: The literature in the field of accurate thermal ablation estimation is relatively limited (R1). While the comparison with Meister et al (2022) [11] may seem imbalanced due to methodological differences, it offers readers a practical sense of our method’s efficiency (R2). Given the lack of directly comparable work, we believe this is a meaningful reference point.




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”.

    N/A

  • 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’

    N/A



back to top