Abstract

Geometric estimation including depth estimation and scene reconstruction is a crucial technique for colonoscopy which can provide surgeons with 3D spatial perception and navigation. However, geometric ground truth in colonoscopy is difficult to obtain due to narrow and enclosed space of the colon, while there is a large feature gap between simulated data and realistic data caused by artifacts and illumination. In this paper, we present CoGE, a novel framework for online monocular geometric estimation during colonoscopy. Firstly, we propose an illumination-aware supervision module based on the Retinex theory to address illumination diversity in different colonoscopy scenes. Moreover, a structure-aware perception module is proposed based on wavelet decomposition to extract common structural and local features of the colon. Both quantitative and qualitative results demonstrate that the proposed model solely trained on simulated data achieves state-of-the-art performance in geometric estimation for both simulated and realistic scenes.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2522_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/2522_supp.zip

Link to the Code Repository

N/A

Link to the Dataset(s)

C3VD dataset: https://durrlab.github.io/C3VD/ C3VDv2 dataset: https://durrlab.github.io/C3VDv2/ SimCol dataset: https://doi.org/10.5522/04/24077763

BibTex

@InProceedings{ShaLia_CoGE_MICCAI2026,
        author = { Shao, Liangjing AND Cui, Beilei AND Ren, Hongliang},
        title = { { CoGE: Sim-to-Real Online Geometric Estimation for Monocular Colonoscopy } },
        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 primary contribution of this manuscript is the introduction of CoGE, a sim-to-real framework for online monocular geometric estimation in colonoscopy. By training exclusively on simulated data, the proposed method effectively circumvents the scarcity of real-world geometric ground truth while enhancing domain adaptation to real clinical scenes. The technical approach is well-motivated, integrating three core components: structure-aware feature extraction, illumination-aware supervision, and a memory filtering mechanism for temporal inference. Empirical results demonstrate that the framework achieves commendable performance in both depth estimation and 3D reconstruction across simulated and real colonoscopy benchmarks.

  • 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.

    -A major strength of this work is that the methodological design is well aligned with the main challenges of colonoscopy videos, rather than being a loosely combined set of generic components. Specifically, the Structure-Aware Perception (SAP) block is designed to capture both structural and local texture information, the Illumination-Aware Supervision (IAS) module explicitly addresses unstable illumination and specular artifacts, and the memory forgetting mechanism is intended to reduce redundant information in long temporal sequences. Overall, these components form a coherent and task-specific framework for sim-to-real geometric estimation in colonoscopy.

    • The evaluation covers one simulated dataset and two real-world datasets, with comparisons against both colonoscopy-specific approaches and recent 3D geometric estimation baselines. The reported quantitative and qualitative results are consistently competitive. In monocular depth estimation, the proposed method outperforms strong baselines including Spann3R, MonST3R, CUT3R, and STream3R across SimCol, C3VD, and C3VDv2.In 3D reconstruction, the paper further reports improved reconstruction accuracy over STream3R, CUT3R, and MonST3R. The fact that these results are obtained while training only on simulated data is an important practical strength of the work.
  • 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 ablation study on the memory forgetting mechanism is currently not sufficient to fully justify the specific design in Eq. (6). In particular, the filtering rule depends on two empirically chosen thresholds: a memory token is retained only if at least 5% of feature tokens assign it an attention score above 5×〖10〗^(-4). However, the current experiments only compare the model with and without the memory forgetting module, without any sensitivity analysis for these thresholds. As a result, Table 2 supports that the module is helpful, but does not yet establish that this particular filtering rule is robust or well justified. It therefore remains unclear whether the observed gains come from a generally effective design or from a relatively fragile heuristic choice.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

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

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

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

    I assigned this paper a moderately positive rating because it addresses an important and practically relevant problem: sim-to-real online monocular geometric estimation for colonoscopy under the realistic constraint that dense geometric ground truth is difficult to obtain in clinical scenes. The paper is coherent overall, and the proposed framework is clearly motivated by the characteristics of colonoscopy videos. In particular, the use of simulated-only training, together with structure-aware encoding, illumination-aware supervision, and memory-based temporal modeling, makes the work practically meaningful. The empirical results on both simulated and real datasets are also strong, and the inclusion of online 3D reconstruction further increases the practical relevance of the paper. I did not assign a higher score mainly because one of the claimed core contributions—the memory forgetting mechanism—is not yet fully validated experimentally. The forgetting rule in Eq. (6) depends on two empirically chosen thresholds, yet the paper does not provide a sensitivity analysis for them; the current ablation only shows the effect of including or removing the module. Therefore, while the overall method appears promising, the evidence is still not sufficient to establish that this specific design choice is robust. A secondary limitation is that the method’s runtime, although still online, is lower than several compared baselines, and the paper itself notes the additional computational cost introduced by the structure-aware perception blocks. Overall, I view this as a strong and relevant submission, but not yet one for which all central design choices are equally well justified.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper proposes CoGE, a DUSt3R/CUT3R-style online reconstruction framework augmented with a structure-aware wavelet module, an illumination-aware supervision scheme, and a memory forgetting mechanism for redundant token pruning. Trained solely on simulated data, the method is evaluated on SimCol3D, C3VD, C3VDv2, and limited in-house qualitative cases, and is reported to achieve state-of-the-art depth estimation and improved 3D reconstruction over several recent foundation-model-based baselines.

  • 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 core insight is well motivated: conventional two-stage sim-to-real pipelines, which combine style translation with downstream geometric estimation, are limited by both inference inefficiency and residual domain mismatch. Exploring an end-to-end feature-level domain generalization strategy is therefore a meaningful and clinically relevant direction for colonoscopy.

    2.The experimental evaluation is relatively comprehensive. The method is assessed on both the simulated SimCol3D dataset and the real-world C3VD/C3VDv2 datasets, and the reported results on depth estimation and 3D reconstruction demonstrate competitive performance against several recent foundation-model-based baselines.

  • 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 claimed technical novelty appears limited and somewhat overstated. The method is largely a task-specific extension of an existing online 3D perception pipeline, combining DUSt3R-based initialization, CUT3R-style geometric supervision, and memory designs reminiscent of Spann3R. Its additional components, including wavelet-based feature decomposition, Retinex-inspired illumination modeling, and heuristic memory pruning, seem more like incremental work. 2.The formulation around the low-frequency branch is unclear. The text states that low-frequency information is extracted from the low-low branch using attention, whereas Eq. (2) defines it from hh rather than from the low-frequency feature, which appears inconsistent and may be a typographical error. In addition, several recent 3DGS-based methods, including Wavelet-GS (ACM MM 2025), EndoWave (BIBM 2025), and PCR-GS (ICCV 2025), should be discussed in the related work, or the differences from them should be clarified. 3.The reference list includes Endo3R (MICCAI 2025), a unified online reconstruction method for dynamic monocular endoscopic video, yet this method is absent from the quantitative comparisons. 4.The ablation study in Table 2 provides only limited support for the effectiveness of the proposed modules. For example, on C3VD, removing the SAP block increases Abs Rel only from 0.085 to 0.089, while removing the IAS module increases it to 0.087.It would be helpful to provide a more comprehensive ablation study, including the baseline model and different combinations of the proposed modules, to more clearly assess the individual and joint contributions of each component. 5.The manuscript still contains a number of language and presentation issues. For example, in Figure 1, “Common Structrual Features” should be corrected to “Common Structural Features.” The phrase “to pretend interference of redundant memory” should be “to prevent interference from redundant memory.” Similarly, in Section 1 (Contributions), “to pretend the influence of diverse illumination” should be replaced with “to prevent” or “to mitigate ” The formatting of the piecewise function in Eq. (6) is also non-standard and difficult to read. In addition, there are many missing or incorrect articles (a/an) in the paper.

  • 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

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

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

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

    The paper is well motivated and explores a potentially useful idea, namely that sim-to-real colonoscopy geometry may benefit more from structural and illumination-aware bias than from appearance translation alone. However, the technical novelty is limited, some methodological details remain unclear, and the evaluation is not yet strong enough to fully support the claims. The presentation would benefit from further polishing. Overall, the paper currently fits a weak accept recommendation.

  • 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 presents CoGE, a monocular geometric estimation framework for colonoscopy that is trained on simulated data and generalizes to real-world data. The method introduces an illumination-aware supervision module to handle lighting variation and a structure-aware perception module to better capture colonic geometry and anatomical structure.

  • 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.Using only synthetic data is a promising direction, as it can substantially reduce annotation cost and may still yield strong generalization, consistent with recent findings in related vision tasks. 2.The method shows relatively clear empirical improvements over the compared baselines. 3.The design motivation is reasonable: the illumination-aware and structure-aware modules are well aligned with two key challenges in colonoscopy imaging, namely lighting variation and structural perception.

  • 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.Although the paper reports quantitative results on C3VD and C3VDv2, these datasets are primarily phantom-based rather than true in-vivo clinical datasets. As a result, the evidence for generalization to real clinical colonoscopy remains limited, especially since the in-house clinical examples are only qualitative. 2.The paper describes CoGE as a “foundation model-based framework,” but this framing seems somewhat overstated. In practice, the method is trained on only 24 simulated VR-Caps sequences, initialized from DUSt3R, and optimized for 10 epochs. 3.The paper is motivated by robustness to illumination variation and sim-to-real deployment, yet the evaluation does not systematically examine several clinically important failure cases. This is also consistent with the authors’ own conclusion, which notes performance degradation under motion blur, acute dynamic changes, and lesion tissue.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

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

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

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

    The paper is well motivated, and the idea of learning monocular geometric estimation from synthetic data is promising. However, my confidence is limited by the lack of stronger validation on real clinical data. Since the claimed practical value depends on transfer from simulation to real-world colonoscopy, the current evidence is not yet sufficient to fully establish real-world generalization.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Author Feedback

N/A




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The reviewers agree that the paper is well motivated and methodologically aligned with key challenges in colonoscopy videos, proposing task-specific structure-aware, illumination-aware, and temporal memory mechanisms for end-to-end sim-to-real geometric estimation using only synthetic training data. They also highlight the relatively comprehensive evaluation on one simulated and two real-world datasets, where the method shows competitive or improved performance over recent baselines in monocular depth estimation and 3D reconstruction.



back to top