Abstract

Atrial fibrillation (AF) is a prevalent cardiac arrhythmia often treated with catheter ablation procedures, but procedural outcomes are highly variable. Evaluating and improving ablation efficacy is challenging due to the complex interaction between patient-specific tissue and procedural factors. This paper asks two questions: Can AF recurrence be predicted by simulating the effects of procedural parameters? How should we ablate to reduce AF recurrence? We propose SOFA (Simulating and Optimizing Atrial Fibrillation Ablation), a novel deep-learning framework that addresses these questions. SOFA first simulates the outcome of an ablation strategy by generating a post-ablation image depicting scar formation, conditioned on a patient’s pre-ablation LGE-MRI and the specific procedural parameters used (e.g., ablation locations, duration, temperature, power, and force). During this simulation, it predicts AF recurrence risk. Critically, SOFA then introduces an optimization scheme that refines these procedural parameters to minimize the predicted risk. Our method leverages a multi-modal, multi-view generator that processes 2.5D representations of the atrium. Quantitative evaluations show that SOFA accurately synthesizes post-ablation images and that our optimization scheme leads to a 22.18\% reduction in the model-predicted recurrence risk. To the best of our knowledge, SOFA is the first framework to integrate the simulation of procedural effects, recurrence prediction, and parameter optimization, offering a novel tool for personalizing AF ablation. The code is available at our repository: https://github.com/cys1102/SOFA

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3966_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/cys1102/SOFA

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ChuYun_SOFA_MICCAI2025,
        author = { Chung, Yunsung and Lim, Chanho and Bidaoui, Ghassan and Massad, Christian and Marrouche, Nassir and Hamm, Jihun},
        title = { { SOFA: Deep Learning Framework for Simulating and Optimizing Atrial Fibrillation Ablation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},
        page = {497 -- 506}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper aimed to address an important clinical issue, how to improve ablation outcomes for patients with atrial fibrillation.

  • 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 aimed to address an important clinical issue, how to improve ablation outcomes for patients with atrial fibrillation. The authors proposed a deep learning approach to predict AF recurrence risk.

  • 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.

    There are several major issues with the study: It proposes a deep learning approach that combines pre-ablation MRI and ablation lesion information to predict the risk of AF recurrence after ablation. However, the underlying mechanisms behind this approach are unclear. It appears to establish only an association rather than a causal explanation. I struggle to understand how combining pre-ablation MRI and ablation lesion data can meaningfully predict AF recurrence. If the goal is to detect ablation gaps, then the ablation lesion data alone (PVI) should be sufficient.

  • Please rate the clarity and organization of this paper

    Poor

  • 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.

  • 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.

    (1) Strong Reject — must be rejected due to major flaws

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

    There are several major issues with the study: It proposes a deep learning approach that combines pre-ablation MRI and ablation lesion information to predict the risk of AF recurrence after ablation. However, the underlying mechanisms behind this approach are unclear. It appears to establish only an association rather than a causal explanation. I struggle to understand how combining pre-ablation MRI and ablation lesion data can meaningfully predict AF recurrence. If the goal is to detect ablation gaps, then the ablation lesion data alone (PVI) should be sufficient.

  • 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

    This paper proposed a novel deep-learning-based framework named SOFA, which is capable of predicting post-ablation images and atrial fibrillation recurrence subject to pre-ablation images and ablation-specific parameters (i.e., duration, temperature, power, and impedance). In addition, the authors propose to optimize the procedural parameters for reduced recurrence risk by exploiting the differentiability of the pipeline. In a 5-fold cross-validation, the authors demonstrate the effectiveness of SOFA to predict significantly better post-ablation images and comparable recurrence predictions when compared to custom baselines. In addition, the authors showed that the parameter tuning can reduce the estimated recurrence risk by approximately 22%.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    (1) The paper is interesting and very relevant, because AF recurrence prediction is an important open challenge.

    (2) The proposed method is very novel since it combines post-ablation image synthetization with recurrence risk prediction. In particular, the use of multiple 2D rendered views of the underlying 3D data and the capability to directly optimize ablation parameters to reduce the recurrence risk is interesting and, to the best of my knowledge, it is fundamentally different to existing approaches.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    (1) Even though the paper was very interesting to read, the paper lacks clarity and a lot of details to fully assess the contribution. For instance, it is not clear to me what the pre- and post-ablation images comprise or what the ablation parameters comprise (the text mentions ablation duration, temperature, power, and impedance, but figure 3 suddenly mentions force and not impedance). Similarly, it is not clear to me whether the feature embeddings of phase 2 are different from the fused features of phase 1.

    (2) While the use of multiple views generated from the 3D data is interesting and novel, it is not clear to me what justifies the use of it if 3D models are available. Moreover, I believe that the generated views are not ideal for the task at hand, because it is apparent that the shading by the underlying renderer is distorting colors in high-curvature areas (= shadows). Could the authors please share their thoughts on this topic?

    (3) One big strength of the proposed method is the capability to directly optimize the ablation parameters for reduced recurrence risk. The presented experiment and results, however, have several flaws and I strongly believe that the manuscript would benefit from a deeper discussion of these results. (A) From my point of view, the presented training data does not have sufficient variations for the network to capture the large solution space. Consequently, when the optimization is run, I believe that the network very quickly sees input data that it has never been trained on, thus leading to erroneous predictions and hence also non-realistic ablation parameters. (B) If I understood the method correctly, the optimization is constrained to optimize only the ablation parameters and not the ablation locations. This seems to me a major limitation since I would assume that the ablation locations should have the highest impact onto the recurrence risk. (C) The optimization requires an initial ablation configuration to be optimized, which is avaiable in the retrospective data but which is not available prior to the intervention. It is therefore not clear to me how the presented approach could be used in the clinical environment and how the initial ablation configuration should be chosen.

    (4) The mathematical notations have several flaws and I would kindly ask the authors to fix them (please see my additional comments).

  • 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

    (1) Introduction, contribution #4: What is meant by 2.5D representations?

    (2) Section 3.1: (a) What are the 3 channels of I_pred? (b) How large are H_b & W_b in relation to H and W? (c) Where does B suddently come from in dimensions of B_pre^flat? (d) For better differentiation between scalars and tensors, please use a bold font for the tensors. (e) Please also always introduce all variables, e.e.g, what are the matrices W with subscripts in Eq. 1? (f) In Eq. 3, what does the subscript i relate to?

    (3) Section 3.2: (a) It is not clear to me whether z^v is different from B_fused from phase 1. If there are differences, e.g., because the embedding is from a layer of the decoder of phase 1, please mention this in text. (b) Why are the multi-view embeddings aggregated by averaging? Wouldn’t it be more expressive if all embeddings were utilized, e.g., by concatenating them? (c) In Eq. 4, the average operation is missing and V should be 6 as previously mentioned in the text.

    (4) Section 3.3: (a) How is I_abl different from I_feat of phase 1? (b) How is the optimization of the parameters setup (e.g., learning rate and number of iterations)?

    (5) Table 1 & 2: (a) What do the standard deviations relate to? Are these relating to the variation over all folds or over a specific fold? (

    (6) Section 4.2: a) In the first sentence you mention that there are no baselines, yet in the next sentence you speak about baselines. Please clarify these two sentences. (b) Please also provide all relevant details of the comparators for reproducibility including architecture details and training strategies. For a fair comparison please also consider comparing the methods by their capacity (= number of parameters).

    (7) Figure 2 & 3: It is very difficult to comprehend both figures: (a) Pre- and post-ablation images: What does the color-coding mean? (b) Ablation parameters: It is almost impossible to see spatial differences in any of the images. I strongly believe that colored renderings with appropriate colorbars would greatly improve the readability. (c) In Figure 3, it seems that the majority of parameters are significantly reduced, which contradicts with the statement in text that the optimization promotes longer durations and increased temperatures. Please double-check that the renderings were correctly generated and/or revise the statements in text. (d) Would it be possible to see at least one more example from the optimization expeirment?

    () Unfortunately, I was not able to have a look at the uploaded code since the platform said that the files were not found (only the README file worked).

  • 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?

    Even though the proposed method is very interesting and novel, the manuscript in its current state does not allow a full assessment of the method since too many details are missing. In addition, the paper lacks a thorough discussion of the results, which I hope to see in the rebuttal.

  • 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 addressed all of my questions and concerns very well. I am absolutely certain that the authors will polish up the manuscript in terms of the initially missing clarity. Since the authors also propose a novel method with convincing results, I strongly believe that this manuscript would be of great interest to the community and hence vote for acceptance.



Review #3

  • Please describe the contribution of the paper

    This paper introduces SOFA (Simulating and Optimizing Atrial Fibrillation Ablation), a deep-learning framework that simulates the effects of ablation parameters on cardiac tissue by generating post-ablation images that depict changes in shape and scar formation, and predicts AF recurrence. The proposed method relies on a multi-modal, multi-view generator to combine pre-ablation images with procedural parameters (e.g., ablation locations, duration, temperature, power, and impedance) to simulate the effects of ablation and provide early prognostic insights.

  • 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 presents an interesting approach to the problem of Atrial Fibrillation Ablation, and the optimization of the precedure parameters. The motivations, state-of-the-art, and contributions are clearly presented. The methodology seems reproductible (see my additional comments), the results are impressive and well illustrated.

  • 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.

    It is not clear what information is available in the pre-ablation images (aside of a healthy / fibrotic tissue map), but it seems that it may be missing tissue characterization that can impact the outcome of ablation. A discussion about this should be added to clear things up.

    Regarding the validation, as actually stated in the manuscript, the reduction in the predicted risk is based on a model and will require an actual clinical validation to assess the method.

  • 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 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

    A few things could be improved for added clarity. One question I had when reading the paper is how are the ablation parameters obtained? They are represented as as feature map of dimension 4×H×W so it seems logical that they come from some imaging modality, yet this is not described in the paper (except for the fact that it comes from the DECAAF-II dataset). It would help to know if this type of data is regurlarly collected during ablation procedures or not. Knowing also the accuracy of the measurements would help further assess the results I believe.

    Also, the link to the code only allows to access the README file, but no code can be accessed. This should be fixed if the paper is accepted.

  • 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?

    I like the idea that, contrary to existing methods that select from a set of predefined strategies, the proposed solution tries to optimize the main parameters of the procedure. Also, I believe that this method could be used in other applications besides Atrial Fibrillation Ablation (at least the parameter optimization part).

  • 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

We thank all reviewers and the meta-reviewer for their insightful feedback. We appreciate the recognition of our work’s novelty and relevance (R3, R4, Meta). The constructive criticism will greatly improve our manuscript. Below, we address key concerns and outline revisions.

Common Concerns

C1. SOFA’s Core Task & Novelty (R1 & Meta) Our research questions are “Can AF recurrence be predicted by simulation?” & “How to ablate to reduce AF recurrence?”. These differ from prior works (e.g., predicting recurrence from post-ablation images or selecting predefined strategies). Scar formation (a key recurrence predictor) results from ablation parameters & fibrosis interplay. Thus, we use cross-attention between pre-ablation images (fibrosis) and ablation parameter to predict post-ablation scar and AF recurrence. In addition, we optimize ablation parameters to reduce predicted risk.

C2: Input & Methodological Clarity (R3 & R4) Inputs: Pre-ablation LGE-MRI shows fibrosis (green in figs); post-ablation shows scar (red). Ablation parameters (locations, duration, temp, power, force) as numerical values from procedures. We utilize locations (point cloud) to map other parameters in different contrasts (low: dark, high: bright) on atrial surfaces. We’ll detail 2D image rendering process in the revision.

Embeddings & 2.5D: Phase 1 (per-view image generation) & Phase 2 (aggregated patient recurrence prediction) use embeddings with identical fused information from pre-ablation/ablation parameter cross-attention. Our 2.5D multi-view representation (6 rendered views) was chosen for robust performance with our small dataset, as 3D models may need more data. Renderings can cause shading in high-curvature areas. Our multi-view approach mitigates this, as edge artifacts in one view are often clear in others, minimizing shadow impacts.

C3: Rationale for Optimization (R3 & Meta) We focus on ‘how to ablate’ (optimizing procedural parameters such as duration, temp, power, force) instead of ‘where’. This is driven by DECAAF-II (PVI vs. PVI+fibrosis): 1) Its findings suggested refining where to ablate (fibrosis-guided) did not significantly improve outcomes over PVI, underscoring the need to optimize how lesions are made. 2) The DECAAF-II dataset has limited location variability beyond its two main strategies, making parameter optimization the more feasible option.

C4: Data, Validation & Clinical Availability (R3, R4 & Meta) Comprehensive datasets (pre-/post-imaging & procedural parameters) are scarce. Thus, we used 5-fold CV with our 235-patient dataset and a 2.5D approach for robust evaluation. 5-fold CV tests on unseen data per fold (20% test), demonstrating generalization crucial for potential clinical use. For new patients, physicians provide pre-MRI and an initial plan; SOFA suggests optimized parameters for this unseen case. Clinical validation of optimization efficacy is beyond this initial study. We focus on SOFA identifying plausible parameter adjustments for model-predicted risk reduction.

C5: Baselines, Presentation, Writing & Reproducibility (All) SOFA is the first work for this integrated simulation-prediction-optimization task; hence no direct prior work exists for full comparison. We are committed to improving presentation and reproducibility. All specific points raised by reviewers (especially R3) regarding mathematical notations, equation clarity, figures (e.g., consistency of Fig. 3 with text), and other textual clarifications will be thoroughly addressed in the revised manuscript. The anonymous code link will be verified, and necessary reproducibility details provided.

Revision Plan:

  • Clarify core task, inputs & methods (C1,C2).
  • Refine optimization strategy, clinical context (C3) & data/validation (C4).
  • Comprehensive textual, notational, figure revisions, baseline clarification, and enhanced reproducibility information (C5).




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”.

    This paper proposes a DL-based approach to predict the recurrence of atrial fibrillation after catheter ablation and to optimise ablation settings to decrease recurrence risk.

    As highlighted by the reviewers, it is an interesting, very original and persuasively written approach to an important medical problem. The paper suffers from a shortage of data to train on, as mentioned by the reviewers. It seems to only provide validation outcomes, instead of outcomes on an unseen test set. It also does not compare recurrence predictions to published state-of-the-art methods that include information about left atrial size, morphology and fibrotic burden.

    I struggle to understand how this method can be used in the clinic. Atrial fibrosis information from LGE is highly controversial and used only a small number of centers. The DECAAF-II clinical trial suggests that ablations based exclusively on fibrosis maps do not bring increased success rates. This can make the ablation technique the method is based on fundamentally unappealing for clinical use.

    Finally, I think the paper could be improved if the authors considered additional relevant information as an input to SOFA, such as the demographic data they use for the benchmark recurrence risk calculation or electrocardiographic information.

  • 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



Meta-review #3

  • 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