Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Endoscopic examination of gastrointestinal organs increasingly relies on 3D computer vision techniques such as depth estimation, surface reconstruction, and missing region detection. Yet rigorous method development for endoscopy remains challenging, as ground-truth data cannot be readily acquired during clinical procedures. We present EndoX, a GPU-accelerated synthetic data generation (SDG) framework and interactive simulator for medical endoscopy that enables efficient generation of fully annotated datasets, addressing the data scarcity gap in algorithm development. EndoX automates anatomically accurate 3D mesh generation from computed tomography (CT) data and employs physics-based rendering to produce RGB imagery together with geometric modalities: depth, surface normals, optical flow, occlusion, and camera coverage maps, with precise ground-truth geometry. For validation, we constructed organ-specific datasets by navigating a virtual capsule endoscope through anatomically accurate CT-derived organ meshes, with sensor and lighting parameters configured to approximate real capsule endoscopes. Quantitative evaluations on depth estimation, pose estimation, and 3D Gaussian Splatting demonstrate that models trained on EndoX data achieve competitive accuracy, and qualitative assessment confirms transferability to clinical endoscopy. The framework is designed for scalable deployment on GPU nodes via containerization. A sample dataset comprising three video sequences (12,000 frames) with paired ground-truth annotations and 3D models is available. The source code of the extension for NVIDIA Omniverse, documentation, and the sample dataset are available at https://beatricelam111.github.io/EndoX/.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1584_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/1584_supp.zip

Link to the Code Repository

https://github.com/beatricelam111/EndoX

Link to the Dataset(s)

EndoX dataset: https://huggingface.co/datasets/bealam111/EndoX_Dataset

BibTex

@InProceedings{LamKwa_EndoX_MICCAI2026,
        author = { Lam, Kwan Tung AND Ho, Ping Shu AND Phurtivilai, Patt AND Cheung, Ka Chun},
        title = { { EndoX: A GPU-Accelerated Endoscopic Perception Simulation Framework } },
        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

    This paper presents EndoX, a GPU-accelerated simulator for generating synthetic data for 3D endoscopy reconstruction for capsule endoscopy. The framework converts CT-derived anatomy into 3D meshes using an automated pipeline, imports them into NVIDIA Omniverse to render endoscopic RGB images and 3D geometry ground truths, including depth, surface normals, optical flow, occlusion maps, camera poses, and vertex-wise coverage maps using physics-based rendering. The authors claim to provide a sample dataset with 3 sequences & 12K frames. The authors evaluate the generated data on depth estimation, pose estimation, and 3D Gaussian Splatting reconstruction models. The problem is highly relevant because real endoscopy lacks reliable ground-truth geometry, which limits the development of depth, SLAM, reconstruction, and coverage-estimation algorithms.

  • 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 addresses an important and clinically motivated bottleneck: lack of 3D ground-truth depth, pose, flow, and coverage maps in real GI endoscopy. 2.The framework is practically useful and is methodologically plausible. A pipeline that generates RGB & pixel-wise 3D ground truth could support several endoscopic 3D vision tasks. The end-to-end integration of CT-to-mesh reconstruction, centerline-based navigation, Omniverse physically based rendering (PBR), and GPU-based multimodal GT generation into a single containerized pipeline is a meaningful engineering contribution. 3.The authors promise to release the code and the dataset publicly.

  • 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 paper’s central claims about sim-to-real transfer and clinical generalization are not sufficiently supported by the experiments. The paper repeatedly suggests that EndoX minimizes the sim-to-real gap and generalizes to clinical capsule endoscopy. However, the only real-data evaluation is qualitative depth prediction on Kvasir-Capsule, which has no ground-truth depth. There is a point to be made about the lack of ground truth for in vivo cases, but validation against phantom-based or synthetic datasets would demonstrate the dataset’s generalization properties. 2.The depth estimation, pose estimation, and 3D GS experiments showcase internal geometric consistency. But this doesn’t support the conclusion of helping “generalize to clinical capsule endoscopy. “ 2a. The depth estimation is trained and tested on the same synthetic dataset. This showcases that the RGB-depth pairs are ‘consistent’ throughout the dataset. To validate the generalizability claim, validation against independent phantom-based datasets with ground-truth geometry, such as C3VD, would better test whether EndoX-trained models generalize beyond their own simulator. Also, low error values for the in-domain testing setup could simply be because the train and test data are very similar. 2b. Similarly, the 3D GS and pose estimation show that the synthetic scene is geometrically consistent (pose, depth, RGB images, & 3D shape) and reconstructable. It does not validate clinical realism or sim-to-real transfer. The pose estimation and 3D GS experiments lack baselines. - i. e evaluation on other datasets for relative comparison of errors. 3.The validation dataset lacks diversity: The sample dataset contains only 3 sequences / 12,000 frames, and the quantitative evaluations appear focused on small intestine data. The paper claims scalable and diverse anatomy generation, but does not show experiments across multiple organs, CT scans, anatomy types, trajectories, or imaging conditions. It is also unclear whether the 3 sequences represent different anatomical shapes. 4.The dataset lacks challenging, realistic conditions, such as debris, blood, bubbles, tools, and tissue deformation. These artifacts are currently the biggest gaps between synthetic endoscopy datasets and in vivo colonoscopy, and their absence would reduce the sim-to-real ability. These limitations should be acknowledged in the paper. 5.The authors claim the advantage of the proposed dataset over VR-CAPS & SimIntestine in terms of real-time performance and computational costs. However, there is no quantitative data on frames/second, GPU memory usage, system requirements, scaling with the number of GPUs, etc. , compared against Unity/Blender-based alternatives. 6.The paper’s novelty is moderate - it’s in engineering integration and not an algorithmic contribution. The primary technical differentiation from VR-Caps and SimIntestine is platform choice (NVIDIA Omniverse over Unity/Blender). This is not transformative, but still a meaningful contribution. Depth representation is unclear: Section 2 describes depth as normalized using near/far clipping planes, while Section 4 describes depth as distance from the camera plane clamped to 0–100 mm and encoded as 16-bit grayscale. Please update for consistency.

  • 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.

  • 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

    I was unable to access the sample dataset via the provided anonymized link on April 15; and the entire code is not provided (which is mentioned in the abstract).

  • 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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    This is a useful and well-motivated framework paper with a credible engineering contribution. The multimodal 3D GT and GPU-accelerated synthetic data pipeline could be valuable for endoscopic 3D vision research. However, the experimental validation is not strong enough for the paper’s central claims. The most serious issue is that sim-to-real transfer is not quantitatively demonstrated. The depth experiments are within-dataset evaluations, the real clinical evaluation is qualitative, the pose 3DGS experiment lacks baselines. The validation dataset size & diversity are limited. The paper also lacks scalability benchmarks despite emphasizing GPU acceleration and high-throughput generation. If additional experiments are not feasible at rebuttal stage, the authors should at minimum weaken the sim-to-real claims, provide computational/scaling details, and explicitly discuss limitations.

  • Reviewer confidence

    Very confident (4)

  • [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 validation and results supporting the claims are relatively weak.



Review #2

  • Please describe the contribution of the paper

    This paper presents a GPU-accelerated synthetic endoscopy data generation framework that uses physically-based rendering to generate optically accurate renderings with variable lighting, material, trajectory, etc. enabling the generation of large, diverse synthetic datasets with ground truth depth, camera pose, etc. The value of this framework and its datasets is demonstrated via comparison against other synthetic datasets using depth estimations from multiple models.

  • 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.
    • Realistic endoscopic data generation along with ground truth depth, camera pose, etc. is extremely valuable for both training and evaluation of depth and camera pose estimation methods.
    • Method is evaluated on both synthetic and real datasets showing successful sim-to-real transfer, which is often not the case for synthetic datasets.
    • Clear presentation of methodology and results
  • 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.
    • Some details necessary for reproducibility are missing and should be added (or authors should clarify why these are not mentioned), e.g., thresholds defined in the Methods section, number of frames used per dataset for evaluation with depth estimation models, etc.
  • 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.

  • 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
    • Are thresholds such as \epsilon and \tau variable by the user? If not, what are these set to?
    • For evaluation with depth estimation models, how much data was used for each dataset (i.e., number of frames used for training, test, etc.)?
    • Presentation of Table 1 could be improved. If possible, space out the columns further, and/or bold best performing dataset per model for each evaluation metric, and/or reduce from 5 decimal places to 4.- Consider increasing font sizes in figures (axis labels, legends, etc.). These are not easily legible without zooming in considerably.
    • An interesting next step could be to fine-tune using real data a model trained using EndoX data and evaluate performance.
  • 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 presentation of this paper is clear and this work contributes an important framework in the development and evaluation of depth estimation methods. Some details are missing and the paper can be accepted with the addition of these details.

  • 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 #3

  • Please describe the contribution of the paper

    This paper proposes EndoX, a framework for synthetic endoscopic data generation using gastrointestinal meshes reconstructed from CT data and GPU-based physically based rendering (PBR). The framework enables the generation of multimodal ground truth, including not only RGB images but also depth, surface normals, optical flow, occlusion, and coverage maps. The system is implemented on NVIDIA Omniverse and supports containerized deployment, allowing scalable large-scale data generation. Furthermore, the authors evaluate the generated data through downstream tasks such as depth estimation, pose estimation, and 3D reconstruction (3D Gaussian Splatting), demonstrating competitive performance compared to existing synthetic datasets (e.g., ES-U and SimIntestine).

  • 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 fact that it consistently generates depth, normals, optical flow, occlusion, and coverage maps is a significant strength; in particular, the generation of coverage maps and occlusion maps is limited in existing datasets, making it highly useful for task design.
    • Table 1 shows that Metric3D v2 achieves good results with an Abs Rel of 0.053 and a1 = 0.987, demonstrating a clear improvement compared to SI in particular. The fact that relatively stable performance is achieved with Monodepth2 and EndoDAC is also commendable.
    • For pose estimation, an ATE of 3.92 mm is reported, which falls within a sufficiently realistic range for synthetic data. Furthermore, results of PSNR 40.27 dB and SSIM 0.978 on 3DGS are presented, demonstrating to a certain extent the ‘effectiveness of geometry supervision’. -The pipeline from CT to mesh to rendering to GT output has been streamlined; the fact that it is GPU-scalable and containerised is of high practical value.
  • 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 main claim of this paper is that it offers ‘PBR + scalability + multimodal ground truth’ compared to VR-Caps and SimIntestine, but the core architecture (virtual endoscope + mesh space + rendering) remains largely unchanged. The improvements lie primarily in ‘integration and scaling’, and the novelty in terms of algorithms and models appears limited.
    • The ground truth referred to as ‘multimodal’ in this paper centres on geometric information such as depth, normals, optical flow and occlusion; verification of the realism of the RGB appearance (textures, specular reflection, optical properties of the mucosa, etc.) is limited to a qualitative comparison with Kvasir-Capsule. Although the adoption of PBR rendering is claimed, no quantitative evidence is provided to demonstrate how closely the generated RGB images resemble real endoscopy. It would be desirable to clarify in the text that the term ‘multimodal’ primarily refers to geometric modalities, and to mention the verification of visual realism as a future task.
    • As this framework is based on CT-derived meshes, it is generally considered to be strong at reproducing normal structures. On the other hand, in actual endoscopic data, abnormal findings such as polyps and ulcers are crucial, and there is no description of how these can be generated or controlled. Considering its use as a dataset for clinical applications, this constitutes a significant limitation.
  • 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.

  • 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

    I propose that the reasoning behind the description of the results when applied to real data be clarified further.

    In this paper, regarding the model trained on EndoX, it is stated that: “produce smoother, higher-contrast depth maps with better structural detail preservation … indicating that EndoX closely approximates the visual characteristics of real endoscopic imagery and supports sim-to-real transfer.”

    However, this claim is based on qualitative observations of Kvasir-Capsule, and it feels as though there is a slight logical leap in asserting that it ‘reproduces the characteristics of actual endoscopic images’.

    With regard to the ‘reproduction of real endoscopic images’, it would be desirable to explicitly discuss the limitations (e.g. the evaluation being qualitative; the fact that actual images consist of a wider variety of components, meaning they cannot be described as an exact reproduction of endoscopic images; the lack of guaranteed depth accuracy; and dependence on rendering conditions) as such.

    Furthermore, given that endoscopic examinations such as this one are performed for the purpose of detecting disease, it would be beneficial to discuss limitations regarding the reproduction of pathological data and other elements not currently included in the simulator.

  • 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 is considered to have practical value in the generation of synthetic endoscopy data, as it integrates a geometric multimodal GT with a scalable generative framework. In particular, the quantitative results for depth estimation, as well as the results for pose estimation and 3DGS, demonstrate a certain degree of effectiveness as data for learning geometric information, which is commendable. On the other hand, the differences from existing methods appear to lie primarily in system integration and scalability, and the novelty of the algorithm itself may be somewhat limited. Furthermore, as the evaluation tasks are mainly concentrated on geometric systems, there appears to be scope for future investigation regarding its applicability to clinically important tasks such as anomaly detection. In light of the above, this research is considered to possess a certain value as a foundational technology useful to the community, and is therefore judged as ‘Weak Accept’.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    We commend the authors for reconstructing the clinical framing and stating that they are repositioning this study as a general stereo endoscopic reconstruction framework. On the other hand, this modification requires essential reorganisation across Abstract, Introduction, Methods and Conclusion, and is beyond the scope of what can be addressed at the camera-ready stage. Although the rebuttal may clarify the position of the paper, the initial clinical problem importance will be relatively weakened and, as a result, the importance of validation of the methodological contribution itself will increase. On this basis, the main concerns raised by Meta-Reviewers and multiple reviewers regarding methodological novelty and experimental validation (comprehensive ablation, teacher pseudo-labelling dependence, differentiation from existing methods, inadequate evaluation, etc.) are considered to be more important. The main concerns regarding methodological novelty and experimental validation (comprehensive ablation, teacher pseudo-labelling dependency, differentiation from existing methods, sufficiency of evaluation, etc.) are still being farely resolved. The main concerns have therefore not been sufficiently resolved by the rebuttal, and the original decision to reject is retained.



Author Feedback

We thank reviewers for constructive feedbacks. We are encouraged that EndoX, a Omniverse (OV) extension is regarded as a credible contribution addressing a bottleneck. R1&R3:Sim-to-real wording. We acknowledge our wording caused confusion about our central claim. We make no algorithmic-novelty claim. Our contribution is the synthetic data generation framework: automated CT-to-mesh, real-time PBR with customisable parameters, and scalable deployment, using SOTA 3D platform for improved data quality and generalization. Sim-to-real wording in Abstract and Sec. 3 overstates the qualitative evaluation and will be removed. R1:Real-data validation and experiment scope. No public dataset has in-vivo per-frame capsule depth GT. Clinical capsule datasets (Kvasir, KID) have only classification labels; KID owner didn’t respond to access request. C3VD/v2 depth is itself rendered against phantoms (Sec.2.3 in Ref6), so C3VD comparison is rendered-vs-rendered. Measured-depth (SCARED, SERV-CT) are not capsule datasets, the gap motivates our work. In-domain evaluation (Sec. 3): training (7000) and test (2000 frames) cover spatially disjoint segments. Pose and 3DGS were not claimed for clinical generalisation; pose supports “a suitable training signal for self-supervised visual odometry” (Sec. 3), and 3DGS verifies sequence quality. R1:Scalability and speed. Blender is offline authoring; Unity lacks multi-GPU and headless docker support; OV Kit v106.5 supports multi-GPU near-linear scaling and Docker headless deployment (Sec. 2.2). On an A6000 (48 GB), frame-per-second: 23/32 (RTX/Path-Traced, capturing); 115/116 static. End-to-end generation: 0.134 / 14.87 s/frame for 4 modalities + 5.76 s/frame for occlusion. Warp kernels 2.4×–29.3× over CPU for different modalities. Full metrics at public repo release. R1&R3:Limitations. We will state limitations: a. no capsule dataset with measured per-frame depth, precluding quantitative sim-to-real; b. sample data covers limited anatomy (but pipeline supports any GI organ); c. physical artifacts require manual authoring. R1&R3:Dataset diversity, abnormalities. 3-sequence sample data (SI-train 7000 + SI-test 2000 + colon 3000 frames @512×512) is a demo preview; capsule endoscopy in small intestine is a prioritised use-case as inaccessible to common endoscopy and scarce in public datasets. EndoX accepts any GI organ from CT; trajectories, sensor intrinsics, lights, and post-processing (Sec. 2.2, Fig. 2) are user-configurable for domain randomisation. Public release of dataset will expand coverage across organs and imaging conditions. Debris, blood and tools are manually authorable via projected textures and mesh import (tutorials will be on the public GitHub), enabling anomaly detection task(R3). Polyps require pathology-bearing CT input. Real-time tissue deformation is on our roadmap. R1:Dataset and code access. We apologise: Zenodo link was masked by Anonymous GitHub rewriting and slow GIF/Zenodo loading. Token link (published 2026-02-26, pre-deadline) is now visible on the repo. Full EndoX extension for OV + demo scene is on the repo; depth-model scripts (Metric3D v2, Monodepth2, EndoDAC) are adopted from the original authors’ repos and not duplicated. R2:Reproducibility details. ε (occlusion threshold) = 0.02, τ (coverage tolerance) = 1e-8 are user-tunable. Depth experiments use a 70/7.8/22.2% train/val/test split (per Sec. 3). Resolutions / total frames: EndoX-SI 512×512 / 9000; EndoSLAM-UnityCam 320×320 / 12558; SimIntestine 340×340 / 1514; EndoSLAM-LowCam 640×480 / 3969; EndoSLAM-HighCam 1280×720 / 5049.Minor edits and future directions. We will unify depth encoding (Sec. 2 vs 4) wording(R1), use “geometric modalities”(R3), fix table & figure formating(R2), and acknowledge suggested future directions(R2,R3). We hope EndoX provides community a simulator leveraging SOTA real-time graphics, extending success of OV-based simulators in domains: surgical (ORBIT-Surgical) and underwater (OceanSim) robotics.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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 reviewers commonly note that the paper’s claims of sim-to-real transfer, clinical generalization, and visual realism are not sufficiently supported, as real-data evaluation is only qualitative, the quantitative experiments are mostly in-domain synthetic tests, and the dataset lacks diverse anatomies, clinical abnormalities, and realistic artifacts such as debris, blood, bubbles, or tissue deformation. They also raise concerns about limited algorithmic novelty, missing reproducibility details and quantitative resource comparisons, and unclear or inconsistent descriptions of dataset settings and depth representation, which weaken the strength of the proposed benchmark and its practical clinical relevance.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    This manuscript proposes a synthetic data generation framework. However, the proposed method has the following problems, which were also pointed out by the reviewers.

    Limitation of the target organ: Although the manuscript states that the method targets endoscopic examinations, the method is specialized for the intestine and cannot be applied to other endoscopes, such as bronchoscopes or laparoscopes, which image other organs.

    Scalability to anatomical structural variations: The value of this method depends on whether it can generate images based on various anatomical structures, but this point is not explained.

    Quality of rendered images: The quality of the rendered images appears outdated, and the method does not support the generation of clinical abnormalities or realistic artifacts. Without the ability to generate these, the value of the synthetic data is limited.

    Limitation of the evaluation: The pose estimation results shown in Fig. 4 were obtained only for a very short section of part of the intestine. Therefore, the method is unlikely to produce practical results when applied to the actual intestine.



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.

    The main criticism raised by the reviewers concerns the lack of experiments demonstrating effective sim-to-real generalization. In their rebuttal, the authors argue that this limitation stems from the absence of suitable real-world datasets with ground-truth annotations that would enable a quantitative evaluation of sim-to-real transfer. While this argument is partially valid, I find the rebuttal somewhat overly defensive. There are several avenues through which additional evidence could have been provided. For instance, evaluating depth estimation performance on datasets such as SCARED, or on other synthetic datasets including SimIntestine and C3VD, would have offered useful insights into the robustness and generalization capabilities of the proposed approach. Likewise, experiments on real datasets without ground truth could still have provided valuable qualitative evidence of sim-to-real transfer. Despite this weakness, the overall reviewer consensus leans toward acceptance. I therefore recommend acceptance, provided that the authors appropriately tone down their claims regarding sim-to-real generalization and explicitly discuss the current limitations of the evaluation in the final version of the paper.



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.

    The initial reviewers’ concerns were primarily regarding quantitative sim-to-real transfer validation and limited diversity within the evaluation dataset, raising generalizability concerns. The rebuttal addressed these concerns, with author maintaining that the provided numerical results were conducted with original experiments but omitted due to space constraints, as well as, explained limitations of using datasets like C3VD for sim-to-real evaluation due to rendered ground truth. A key concern with diversity was addressed to a certain extent as authors noted that framework provides the ability to incorporate clinical abnormalities though not explicitly stated in the manuscript. In the post-rebuttal phase, R1 still noted that the results are not strong enough to support the claims, while R2 and R3 leaned towards acceptance, nothing the presented pipeline would enable generation of practical and representative synthetic data. While the concerns raised by R1 are valid, AC agrees with other reviewers that the pipeline would be valuable to the community and could be contributed/built upon and thus recommends acceptance. However, this decision is contingent on authors incorporating the points mentioned in the rebuttal as well as adding a limitations section.



back to top