Abstract

Automated guidewire navigation in percutaneous coronary intervention can reduce radiation exposure and clinician workload. However, existing methods suffer from low sample efficiency and struggle with long-horizon guidewire navigation. World models offer a promising solution, but scarce paired angiographic-3D data leaves end-to-end world model approaches unexplored. To address these challenges, we propose CAG-WM, the first synthetic-data-driven world model for autonomous coronary guidewire navigation. Specifically, we construct dynamic synthetic coronary angiography–3D data and train a world model within a SOFA-based simulation environment to learn vessel–guidewire dynamics and perform internal multi-step rollouts for long-horizon navigation. CAG-WM is trained on 100 synthetic cases with full-view angiographic renderings and corresponding 3D voxelized coronary models, and evaluated on 20 multi-center clinical cases and 10 public datasets. Experimental results show that CAG-WM achieves a success rate exceeding 80% in clinical cases, significantly outperforming competing methods. Even in unsuccessful cases, the guidewire remains close to target lesion. Despite slightly longer per-step inference time, CAG-WM achieves the shortest total navigation time (about 2 min) thanks to its higher success rate, which reduces premature termination at the maximum step limit. Code and data is available at: https://anonymous.4open.science/r/CAG-WM-FEA0.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4072_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/4072_supp.zip

Link to the Code Repository

https://github.com/CaoYue2004/CAG-WM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{CaoYue_CAGWM_MICCAI2026,
        author = { Cao, Yue AND Zhang, Zhuo AND Peng, Xiaoshan AND Lan, Guipeng AND Xiao, Shuai AND Yang, Jiachen},
        title = { { CAG-WM: A Synthetic-Data-Driven Coronary World Model for Autonomous Guidewire Navigation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16893},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The main contribution of this paper is the proposal of CAG-WM, a comprehensive world model framework for autonomous coronary guidewire navigation, which systematically integrates synthetic coronary angiography 3D data construction, a SOFA-based physics simulation environment, and a TD-MPC2-based world model to enable long horizon planning from angiographic observations to control actions. The key innovation lies in leveraging synthetic data to address the scarcity of real paired data, while explicitly modeling guidewire vessel dynamics to support multi step imagination based reasoning, ultimately leading to significantly improved navigation success rate and efficiency on real clinical data, and demonstrating the feasibility of applying world models to complex endovascular intervention tasks.

  • 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.Well motivated problem with clear clinical relevance:The paper tackles autonomous coronary guidewire navigation, a high impact and technically challenging task in PCI that involves long horizon planning under partial observability. 2.Novel integration of synthetic data, physics simulation, and world models:A key contribution is the coherent integration of synthetic angiography 3D data generation, SOFA-based physical simulation, and a TD-MPC2-based world model into a unified training pipeline. While each component exists individually, combining them to form a closed loop system for learning guidewire vessel dynamics is original and practically valuable, especially for data scarce medical robotics tasks. 3.Demonstrated effectiveness on real clinical data despite synthetic training:The method is trained entirely on synthetic data yet shows clear performance gains over RL and existing world model baselines on multi center clinical datasets. This suggests that the approach has promising generalization ability and provides an encouraging step toward real world deployment in endovascular intervention tasks.

  • 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.Limited analysis of synthetic-to-real generalization:While the method is trained on synthetic data and evaluated on real clinical cases, the paper provides limited analysis of the domain gap between synthetic and real angiography. For instance, it would strengthen the work to include additional studies on how differences in imaging characteristics or anatomical variability affect performance. Prior sim-to-real studies often include such analyses or adaptation strategies, and incorporating similar evaluations here could further support the claimed generalization ability. 2.Methodological novelty is primarily at the system level:The proposed approach builds upon existing world model frameworks, particularly TD-MPC2 , with task specific adaptations in observation design and reward formulation. While this integration is well-motivated and effective for the application, the methodological contribution is mainly at the system and application level rather than introducing a fundamentally new world model formulation. Clarifying this positioning could help better align expectations regarding novelty. 3.Clarity and fairness of baseline comparisons could be improved:The proposed method leverages structured inputs such as guidewire points, vessel centerlines, and target locations (Sec. 2.3) . It is not entirely clear whether all baselines are provided with the same input representations, or whether some operate on different modalities. Providing additional clarification or controlled comparisons with consistent observation spaces would make the performance gains more directly attributable to the proposed modeling approach.

  • 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.

  • 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

    No

  • 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?

    This paper tackles an important and clinically relevant problem and presents a well engineered integration of synthetic data, physics-based simulation, and world models, with empirical results showing clear improvements over existing baselines. However, the strength of the contribution is somewhat limited by the lack of thorough analysis of synthetic-to-real generalization, the primarily system level novelty, and some ambiguity in baseline fairness. Overall, the work is promising and potentially impactful.

  • 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 #2

  • Please describe the contribution of the paper

    This paper addresses autonomous coronary guidewire navigation in PCI, a long-horizon continuous-control problem. It proposes CAG-WM, a synthetic-data-driven coronary world model with three main components: (1) a synthetic data pipeline that generates paired CAG–3D coronary data through implicit 3D generation, segmentation, multi-view reconstruction, and centerline extraction; (2) a SOFA-based physics simulation for guidewire–vessel interaction; (3) a TD-MPC2-style world model that performs latent dynamics modeling and planning to output advancement, retraction, and rotation actions. The model is trained on 100 synthetic cases and evaluated on 20 multi-center clinical cases and 10 public cases, with comparisons against SAC, TD3, CEM-MPC, DreamerV3, TD-MPC, and TD-MPC2.

  • 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) Technical novelty and innovation: Applying a world model to coronary guidewire navigation is a relevant and reasonably novel direction. The integration of synthetic full-view CAG, reconstructed 3D anatomy, and SOFA-based physical simulation is a meaningful systems contribution. Figures 1 and 2 present the high-level workflow clearly. (2) Experimental rigor and validation: The paper evaluates against three classes of baselines: model-free RL, planning-based methods, and world-model-based methods. It reports three metrics on two test sources (InCAG and VMR), and includes a reward ablation study. At a minimum, the experimental surface is broader than many papers in this area. Tables 1 and 2 attempt to characterize performance, failure quality, and runtime. (3) Clarity of presentation: The manuscript is generally well organized. The problem setup, framework overview, experimental protocol, and result sections are easy to follow. Figure 1 provides a useful system overview, and Figure 3 offers qualitative visualizations of navigation trajectories and training-space coverage.

  • 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 main concern is that the algorithmic novelty appears limited. The core seems closer to “TD-MPC2 + task-specific observation engineering + reward design + synthetic simulation environment” than to a substantially new world-model formulation. The paper claims “the first synthetic-data-driven world model,” but the actual methodological advance over TD-MPC2 is not sharply defined. For a top-tier venue, the manuscript needs to separate system integration novelty from algorithmic novelty much more clearly. (2) The manuscript repeatedly claims the method is end-to-end from angiographic observations to actions, but Sec. 2.3 explicitly states that it first detects the stenosis target, tracks guidewire points, and extracts vessel centerlines before encoding them with TCNs and an MLP. Sec. 3.1 further states that all perceptual information required by the observation is pre-extracted prior to training. This is not a standard end-to-end setup; it is a staged pipeline with substantial structured intermediate supervision. The wording is therefore misleading. (3) The mathematical and physical descriptions are under-specified. Eq. (2) is not explained in enough detail for reproducibility, and the dynamic vessel-wall deformation model is especially unclear. The manuscript claims dynamic vascular properties, but it does not establish whether the displacement field reflects realistic cardiac/respiratory motion or merely synthetic perturbations. At present, the dynamic environment seems plausible as a simulation device, but not yet clinically validated.

  • 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.

  • 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

    The coronary artery at the top center of Figure 1 seems to have been artificially created? The branching of the small blood vessels doesn’t look very realistic.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The major factors behind my recommendation are the mismatch between the paper’s claims and the supporting evidence, the limited isolation of its true methodological contribution, and the lack of sufficient experimental rigor. In particular, the claimed novelty beyond existing TD-MPC2-style world models is not clearly established, and the “end-to-end” characterization is overstated given the method’s reliance on pre-extracted structured inputs.

  • 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

    The authors propose CAG-WM, a framework for autonomous coronary guidewire navigation in PCI. The framework consists of 2 parts: (i) an EG3D-based[1,2] training data generation strategy; (ii) a TD-MPC2[3] backbone with modifications on input encoder and reward design. The experiments are conducted on a private dataset (InCAG, 20 cases) and a public dataset (10 cases), with customized evaluation/benchmark procedure. Compared to the top-3 most similar works [4,5,6], the novelty and contribution of this work lies in its data generation strategy and the open codebase.

    [1] Chan et al., EG3D, arXiv 2021.Multi-view consistent 3D GAN: proposed for natural images e.g. faces. [2] Cao et al., ATC-CAG, AAAI 2026.Translational correction for PCI camera calibration: via an EG3D[1]-based synthetic coronary angiography generation pipeline. [3] Hansen et al., TD-MPC2, 2024.Model-based (MPC) RL algorithm: proposed for general robotic continuous control tasks. [4] Robertshaw et al., MICCAI 2025.TD-MPC2 world model for mechanical thrombectomy: applies the same backbone [3] to guidewire navigation from aortic arch to carotid/cerebral vessels in SOFA. [5] Scarponi et al., IROS 2024.SAC-based coronary guidewire navigation with dynamic vessel wall: explicitly models cardiac-motion-induced vessel deformation in SOFA. [6] Schegg et al., RoboSoft 2022.Planning-based coronary guidewire navigation: MPC-based path planning for coronary bifurcation traversal in SOFA.

  • 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.
    • Practical data generation strategy: [2] produces what I consider the most realistic synthetic coronary angiography (XA/DSA) images to date, built upon EG3D[1], though [2] only applies this capability to translational correction for PCI camera calibration. The present work directly inherits this generation pipeline from [2] and repurposes it for supporting simulation environments for guidewire navigation training. While the generation module itself is not new, I view this as a valuable contribution: the gap between synthetic training data and transferable navigation policies has been repeatedly identified as a bottleneck in prior literature [7,8], yet no prior work has offered a convincing solution. That said, I note with some puzzlement that although the pipeline produces highly realistic 2D angiographic images, the method does not directly consume these images. Instead, it extracts intermediate representations (point coordinates, centerlines) and discards the visual realism. This appears to leave the most distinctive advantage of the generation pipeline underutilized, and I would appreciate the authors’ clarification on this design choice or correct my misunderstanding (I have to say the method and experiment part is too confusing, and some parts seem to be blurred. :-) ).
    • Open codebase: I carefully went through the codebase, and confirmed that the codebase contains the full SOFA-based CAG-EVE environment plus the TD-MPC2 training/evaluation scripts, but the perception/data-generation pieces described in the paper (EG3D angiography synthesis, guidewire/centerline detectors, and the 20-case clinical geometry set) are missing, which prevents end-to-end reproduction of the reported “from angiography” results.
    • Graphical content: Figures are high quality and well designed. Videos are informative and attractive.
  • 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.On task setup, experimental description, and missing information. Even as a reviewer familiar with this domain, I was unable to determine the following key information from the paper alone, prior to inspecting the released codebase. 1.Source of 3D models for clinical cases. How are the 3D vessel models for the 20 multi-center clinical cases obtained? Via the same multi-view angiographic reconstruction pipeline in Sec 2.1? Or from CTA? If from angiographic reconstruction, how many views are used? What reconstruction method is applied? What are the spatial resolution and reconstruction error? 2.Multi-view temporal synchronization. If multi-view angiographic reconstruction is used, clinical views are typically acquired in different cardiac cycles (weak async), meaning vessel morphology differs across views due to cardiac motion. Is ECG-gating used to select images from the same cardiac phase? How is this handled? [8] (AutoCAR) specifically addresses this dynamic sparse-view reconstruction problem. 3.Coordinate dimensionality. Are the input coordinates g_i, c_j in R^2 or R^3? This is never explicitly stated. SOFA is a 3D simulator; rotation actions degenerate under 2D projection; nearest-neighbor matching between guidewire and centerline requires the same dimensionality, all suggesting 3D. But this directly bears on the claim of “from angiographic observations” (2D angiography only provides 2D information). 4.Dynamic modeling and simulation. [5] (Scarponi, IROS 2024) explicitly models cardiac-motion-induced vessel wall deformation as a core contribution. What level of dynamic modeling does CAG-WM provide in comparison? If cardiac motion is not modeled, what exactly is the “real” data experiment? Is it direct XA observation, or a static rendering?

    • THESE INFORMATION MATTERS BECAUSE they prevent reader from assessing several key claims made in the paper, including “the first,” “real world/data,” and the purported differences from [3,4,5]. More specially, some related works I list below(I am not asking author to compare these methods bacuase they may not release thier code or data, but the thing is you should at least EITHER clarify the difference compared with them, in terms of method, task setup, eval way; OR report what is conducted in the exp and method in a clear and more descriptive language). In fact there are many and I can’t list them all, e.g. CathSim. ``` [4] Robertshaw, MICCAI 2025
    • Anatomy: Aortic arch → carotid/cerebral (MT)
    • Core algorithm: TD-MPC2
    • Simulator: SOFA
    • Vessel model: 10 real patient CTA
    • Observation input: Guidewire tip coords (3D)
    • Dynamic vessel: No
    • Test environment: SOFA simulation
    • Open-source code: No
    • Comments: Same backbone as CAG-WM; 65% SR on aortic arch [5] Scarponi, IROS 2024
    • Anatomy: Coronary
    • Core algorithm: SAC
    • Simulator: SOFA
    • Vessel model: Real CTA
    • Observation input: Guidewire tip coords (3D)
    • Dynamic vessel: Yes (cardiac motion)
    • Test environment: SOFA simulation
    • Open-source code: No
    • Comments: Directly comparable anatomy; explicit cardiac dynamics

    [6] Schegg, RoboSoft 2022

    • Anatomy: Coronary
    • Core algorithm: Planning-based (MPC)
    • Simulator: SOFA
    • Vessel model: Real CTA
    • Observation input: Guidewire + vessel geometry
    • Dynamic vessel: No
    • Test environment: SOFA simulation
    • Open-source code: No
    • Comments: Directly comparable anatomy; planning approach

    CAG-WM (this paper)

    • Anatomy: Coronary
    • Core algorithm: TD-MPC2 + TCN encoder
    • Simulator: SOFA
    • Vessel model: 100 synthetic (EG3D) + 20 clinical (source ?) + 10 VMR
    • Observation input: Guidewire coords + centerline coords + target (2D or 3D ?)
    • Dynamic vessel: “Diffeomorphic displacement field” (source and physical meaning ?)
    • Test environment: SOFA simulation (described as “real clinical cases”)
    • Open-source code: Yes
    • Comments: SR 84.7% — but is the difference due to anatomy, encoder,reward, or data volume? No discussion provided. ```

    2.On overclaiming. Below I identify several instances where the paper’s claims appear to exceed what the experiments actually demonstrate. Due to the missing details above, some of these assessments are based on my inference rather than certainty.

    • “End-to-end from angiographic observations” vs. what appears to be an image to point extraction to RL agent pipeline.
    • Evidence from the paper: The Abstract states the method “learns…from angiographic observations in an end-to-end manner.” The Introduction criticizes [4] for “lack[ing] an end-to-end mapping from angiographic images to navigation actions,” implying CAG-WM achieves this. Section 2.3 writes: “given a synthesized angiography image I_t, we detect the stenosis target p*, track the guidewire to obtain N points {g_i} [12], and extract vessel centerline C [26].” Yet the same paper states: “All perceptual information required by observation is pre-extracted prior to training.”
    • Analysis by training/testing phase: During training, all information is directly available from the SOFA simulator, guidewire node positions {g_i} are internal simulator state (no detection network [12] needed); centerlines {c_j} are pre-extracted from the 3D model in Sec 2.1; target p* is manually specified on the centerline. The paper acknowledges this. During testing, clinical vessel geometries are imported into SOFA (see below), so testing also occurs within the simulator. Guidewire coordinates, centerlines, and targets can therefore be directly read from the simulator at test time as well.
    • On the citation of [12] and [26]: The references to a guidewire detection CNN [12] and a centerline extraction method [26] in Sec 2.3 create the impression that these features are extracted from angiographic images. However, within a simulation environment, this information is directly accessible from the simulator and these vision algorithms would not need to be (and likely are not) executed. These citations appear to gesture at a future deployment pipeline rather than describe the current system’s actual data flow. I ask the authors to clarify explicitly: are [12] and [26] actually run during training or testing? Is the image encoder jointly trained? Does the angiographic image enter the network at all?
    • Perception error propagation: If the method were truly deployed clinically on live angiographic images, it would require a complete perception front-end (guidewire detection + centerline extraction + target localization), and errors in each module would propagate to navigation performance. This gap is not discussed. I suggest the authors discuss this at least qualitatively, even without new experiments.
    • Dependency: The missing coordinate dimensionality from W1 exacerbates this issue, if inputs are confirmed to be 3D, the contradiction with “from 2D angiographic observations” becomes even sharper.
    • “Real clinical cases.”
    • Independent evidence: Fig 3.A displays navigation results for clinical cases, showing both a CAG modality row and a SOFA modality row for each case. This directly confirms that testing is performed within the SOFA simulator, the guidewire is virtual, physical interactions are simulated, and observations are obtained from the simulator. Regardless of how the 3D vessel model is obtained, the operating environment itself is not “real.”
    • A more accurate description would be: “evaluated on simulated environments constructed from 20 clinical coronary geometries,” rather than “evaluated on 20 multi-center clinical cases.”

    <space is limited, I can not write them all>

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

  • 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
    - **"Real clinical cases."** - Independent evidence: Fig 3.A displays navigation results for clinical cases, showing **both a CAG modality row and a SOFA modality row** for each case. This directly confirms that testing is performed within the SOFA simulator, the guidewire is virtual, physical interactions are simulated, and observations are obtained from the simulator. Regardless of how the 3D vessel model is obtained, the operating environment itself is not "real." - A more accurate description would be: "evaluated on simulated environments constructed from 20 clinical coronary geometries," rather than "evaluated on 20 multi-center clinical cases." - Since the source of the 3D models is not disclosed, I cannot even assess how faithful these "clinical geometries" are to the actual patient anatomy, if they are coarsely reconstructed from a small number of angiographic views, the distance from "real" is even greater. **I ask the authors to clearly define the scope of "real" and disclose the 3D model source.** - **"Dynamic."** - Since the source and physical meaning of the displacement field are not disclosed (see W1.4), I cannot determine whether "dynamic" refers to physiologically meaningful cardiac-motion-induced vessel deformation (as explicitly modeled in [5]), or merely to the general contact mechanics inherent in any SOFA simulation (which would not constitute an additional contribution). If the latter, the use of "Dynamic" in the Sec 2.2 title and "incorporating dynamic vascular properties" in the Abstract would be overstatements. Furthermore, the equation presented in this section appears to be standard SOFA constraint mechanics; the symbols defined therein are not referenced elsewhere in the paper. **I ask the authors to clarify the source of the displacement field: is it derived from a cardiac motion model? What are its frequency and amplitude?** - **Baseline selection and argumentation logic.** - Context: In the autonomous guidewire navigation literature, existing benchmarks and open-source platforms (CathSim, stEVE [10]) predominantly focus on aortic arch to supra-aortic navigation, where datasets and standardized evaluation protocols are relatively mature. Coronary-specific benchmarks are comparatively scarce. If the paper's primary technical contribution is world-model-based guidewire navigation, one would expect the method to first be validated on existing benchmarks with established baselines, or alternatively, to port existing domain-specific methods to the coronary setting and analyze what makes coronary navigation specifically harder and how the proposed modifications address those challenges. - Current state: Table 1 compares against six generic RL/planning methods (SAC, TD3, CEM-MPC, DreamerV3, TD-MPC, TD-MPC2). None of the domain-specific methods that the paper itself cites, [4] (same TD-MPC2 backbone), [5] (coronary-specific with dynamic modeling), [6] (coronary planning), appear in the experimental comparison or are even qualitatively discussed in terms of performance differences. I do not require the authors to add experiments with these methods (simulator setups and anatomies may differ), **but I do expect a discussion section that qualitatively analyzes the relationship to and differences from [4,5,6], and that articulates what specific challenges coronary anatomy poses over the aortic arch.** - **Other minor overclaiming.** 1.Equation (2): Sec 2.2 devotes a full paragraph to the constrained dynamics formulation (Timoshenko beam + contact constraints + Gauss-Seidel solver). The presentation reads as if this is a contribution of the paper, but it is standard SOFA simulation machinery. Moreover, the symbols defined in Eq. (2) are not referenced again in subsequent sections, making this passage serve no functional role in the paper. 2."Coronary-Specific Observation Modeling": The Sec 2.3 subsection title suggests a modeling paradigm tailored to coronary anatomy. In practice, the content describes two TCN encoders applied to coordinate sequences followed by MLP fusion, a generic sequence encoding scheme equally applicable to any tubular vascular anatomy with no coronary specificity. 3.Suggestion: I recommend the authors explicitly acknowledge the scope of experimental validation in a dedicated limitations section, and frame clinical aspirations as future work rather than realized contributions. The Conclusion already mentions "Future work will expand validation cohorts and advance evaluation on physical vascular phantoms and in vivo experiments", **I suggest extending this honest framing to the rest of the paper.** 3.**Minor issues.** 1.Self-reference error in Sec 2.2: "The 3D coronary models from Sec. 2.2 are imported into SOFA", but Sec 2.2 is the SOFA simulation section itself. This should reference Sec 2.1 (Synthetic Data Construction). 2.Likely typo in Table 2: The entry "83.3+-0.082" for CAG-WM on VMR likely should read "83.3+-8.2." \#\#\#\# Some notes after reading the codebase - Environment & agent code are complete, but no EG3D synthesis(this may be partially available), perception front-end, or 20-case clinical data(is the mesh refers to the clinical data?) are included, so "from angiography" claims remain untestable. - Observations are purely SOFA 3D states (Tracking3D, Target3D, etc.), confirming inputs are R^3 despite the paper's 2D phrasing. - Both training and evaluation run fully inside SOFA with TrackingOnly; there is no interface to live fluoroscopy, so "real clinical cases" are actually simulated on imported meshes. - The simulator is static (SofaBeamAdapter with fixed vessel meshes); no cardiac-motion or displacement-field model is present despite the "dynamic" narrative. - No implementations of domain baselines ([4-6]) or perception references [12, 26] exist, so comparisons and pipeline diagrams in the paper cannot be reproduced.
  • 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?

    I would recommend acceptance of this paper IF the authors can address the significant issues of overclaiming and omission/misleading of critical experimental details. That said, the underlying technical/resources contributions, upon these issues resolved, are appreciated.

  • 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



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 submission is well written. However, reviewers note that some claims are too strong (e.g., “the first world model”), as key components like TD-MPC are largely based on prior work. The contribution appears to be mainly at the application level. Further comments can be found in the reviewers’ feedback.



back to top