Abstract

Accurate disparity estimation from stereo laparoscopic images is essential for 3D surgical navigation but remains difficult in occlusion regions, where conventional cost volumes produce unreliable matching costs. To address this, we propose Depth Prior Embedded Disparity Estimation Network that leverages monocular depth cues—inherently less affected by occlusion—to refine disparity predictions. Our framework consists of two key components: a Coarse Disparity Estimation (CDE) module that generates an initial disparity prior from an affine-transformed depth map, and a novel Depth Prior Embedded Cost Volume (DPCV). The DPCV encodes this prior into a probability distribution over disparity levels and iteratively fuses it with the traditional correlation volume via a cross-attention mechanism, thereby improving matching robustness in occluded/unmatched regions. Evaluations on the SCARED dataset show that DEDENet achieves improved performance over several state-of-the-art methods, with clear gains observed particul

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4567_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiuZit_Occlusionaware_MICCAI2026,
        author = { Liu, Ziteng AND He, Dongdong AND Zhang, Chenghong AND Wang, Weirong AND Gao, Wenpeng AND Fu, Yili},
        title = { { Occlusion-aware Disparity Estimation via Depth Prior Embedded Cost Volume for Stereo Endoscopic Images } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16892},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes an iterative refinement process for the cost volume. This process includes converting a coarse depth estimate into a probability distribution and utilizing a cross-attention mechanism to refine the previous cost volume using the generated distribution. The evaluation results in Table 1 show that the proposed method achieves a strong balance between estimation error and inference time.

  • 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 idea of converting a coarse depth map into a probability distribution (Eq. 2) and using this distribution to iteratively refine the cost volume (Algorithm 1) is both interesting and well-motivated.

  • 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.

    Methodology: Parameter Settings: The procedure for defining or tuning $\sigma$ in Eq. 2 is currently unclear and requires further explanation. Unsubstantiated Claims: The authors state that refining the cost volume helps resolve occluded regions in stereo matching. However, the evaluation section lacks specific experiments or ablation studies designed to validate this claim.

    Writing & Presentation: Incomplete List: In the final paragraph of the Introduction, the authors state that their contributions are “fourfold,” but they only enumerate three points (“First. . . “, “Second. . . “, “Third. . . “).

  • 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 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 core contribution iterative refinement for the cost volume is interesting.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper focuses on disparity estimation for endoscopic images. It particularly pays attention to occluded or unmatched regions. It proposes a depth-prior embedded framework, DEDENet. The model incorporates monocular depth cues into stereo matching. The method mainly includes two parts: a coarse disparity estimation module that generates an initial disparity prior from inverse depth, and a depth prior embedded cost volume that encodes this prior as a probability distribution and fuses it with the original cost volume through cross-attention and iterative refinement. Experiments on the SCARED dataset show generally improved performance, with clearer gains on the keyframe evaluation subset and in challenging occlusion regions.

  • 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.Clinical and technical relevance: The paper studies stereo disparity estimation for endoscopic images, which is an important problem for 3D reconstruction and surgical navigation. The focus on occluded or unmatched regions is practically meaningful, since these areas are a common source of failure in laparoscopic stereo matching. 2.Reasonable methodological design: The proposed framework introduces monocular depth priors into stereo matching through a coarse disparity estimation module and a depth prior embedded cost volume. Encoding the depth prior as a disparity probability distribution and fusing it with the original cost volume via cross-attention is a sensible design, especially for improving robustness in occluded regions. 3.Solid empirical validation: The paper includes quantitative comparison against several recent baselines on the SCARED dataset, together with ablation studies on the main components such as DPCV, occlusion masking, auxiliary loss, and iterative refinement. These experiments help support that the proposed modules contribute to the final performance.

  • 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 paper is of generally good quality, but several aspects would benefit from further clarification and discussion. 1.Novelty: The overall idea of incorporating monocular depth priors into stereo matching is reasonable, but the methodological novelty appears somewhat limited. The framework mainly combines a coarse depth-to-disparity prior, probability-based prior encoding, cross-attention fusion, and occlusion-aware refinement, all of which are built on top of an existing stereo backbone. The paper would benefit from a clearer explanation of what is fundamentally new relative to prior cost-volume refinement and fusion-based stereo methods. 2.Occlusion-specific evidence: The paper is motivated by improving disparity estimation in occluded or unmatched regions, but the experimental evidence for this claim is still somewhat indirect. Most results are reported on overall metrics, while more dedicated quantitative analysis specifically within occlusion regions would make the central claim better supported. 3.Clarity and technical details: Some parts of the presentation could be improved. For example, the naming of the proposed module is not fully consistent in the method section (DPCV vs DECV in section 2.2, maybe they are the same thing?), and several implementation choices, such as the design of the probability distribution prior and the effect of iteration number, could be explained more clearly. This would make the method easier to follow and assess.

  • 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 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 think the paper is relevant and technically reasonable, especially in its attempt to improve stereo disparity estimation in occluded regions for endoscopic images by incorporating monocular depth priors. The whole framework is well motivated, and the results on the SCARED dataset, particularly on the keyframe subset, are promising. However, my overall score is mainly limited by concerns regarding the clarity of the novelty claim and the lack of more direct quantitative evidence specifically demonstrating improvement in occluded regions. I would like to higher the score after the authors’ feedback and explanation. Anyway, it is a good work.

  • 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.

    I am willing to support acceptance. The response clarifies several points, and the method is technically reasonable. I still have some concerns about the occlusion-specific evidence, but I think these can be handled by making the claim more modest and discussing the limitation more clearly in the final version.



Review #3

  • Please describe the contribution of the paper

    The paper addresses the critical challenge of disparity estimation in stereo laparoscopic images for 3D surgical navigation. The core contribution is a framework that utilizes monocular depth cues to regularize and refine disparity in regions where traditional stereo matching fails, particularly in occluded areas. The authors propose a Depth Prior Embedding Cost Volume (DECV), a Coarse Disparity Estimation (CDE) loss, and an iterative refinement process using occlusion masking to improve the accuracy and robustness of depth perception in surgical environments.

  • 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.
    • Well-Structured Methodology: The integration of DECV, CDE loss, and iterative refinement forms a cohesive and logical pipeline. The architectural choices directly target the identified problem of occlusion in endoscopic scenes.
    • Ablation Validation: The effectiveness of each proposed module (DECV, CDE loss, masking) is clearly demonstrated through a systematic ablation study, proving their individual contributions to the overall performance.
    • Efficiency: The model appears to balance the trade-off between inference speed and accuracy effectively, which is vital for real-time surgical navigation.
    • Clear Motivation: The use of monocular cues as a prior to “fill in” the gaps where stereo correspondence is physically impossible (occlusions) is a sound and intuitive approach for the surgical domain.
  • 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.
    • Missing State-of-the-Art (SOTA) Baselines: The review of current stereo matching literature is incomplete. The paper lacks a performance comparison with modern high-performance architectures like RAFT-Stereo and IGEV-Stereo, which currently define the SOTA for iterative refinement and cost volume construction.

    • Specific Impact on Occlusions: While the paper claims to solve occlusion-related issues, there is no isolated quantitative analysis for these regions. Can the authors provide error metrics (e.g., EPE or D1-all) specifically measured inside the occlusion masks? This would verify if the performance gain is truly coming from the proposed occlusion-handling mechanism.

    • Monocular Uncertainty: Monocular depth estimation is generally less accurate than stereo matching in non-occluded areas. How does the system prevent the lower-fidelity monocular priors from degrading the high-precision stereo-matched regions? A discussion or experiment on the “noise” introduced by the monocular prior is needed.

    • Domain Specificity: The proposed modules seem generally applicable to natural images. Does the design include specific considerations for the unique properties of endoscopic images (e.g., specular reflections, textureless tissue, or fluid presence) compared to standard outdoor/driving datasets?

    • Discrepancy in Latency Reporting: In Table 1, the inference speed is reported as 25.65ms, yet Section 4.1 mentions 45.23ms.
    • Does the 25.65ms figure only account for the proposed refinement modules, excluding the backbone?
    • For a fair assessment of real-time viability, the total end-to-end processing time (45.23ms) should be the primary metric reported.
  • 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 does not provide sufficient information for 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?

    See the weakness part.

  • 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.

    The reviewers address some of my concerns in the rebuttal. Based on my initial rating (weak accept) and the authors’ response, I lean on accept for this paper rather than the reject.



Author Feedback

We sincerely thank the Area Chair and all reviewers for the constructive comments. 1.SOTA Comparison Rationale (R#2, R#3): We agree with the reviewers that the selection of SOTA baselines in the current experiments is relatively limited,a comprehensive comparison with RAFT-Stereo, IGEV-Stereo, and other SOTA methods related to CV construction​is planned for the extension paper. Regarding the chosen baselines, HSM, DeepPruner, and MSDESIS represent the state-of-the-art for surgical scenarios, while ACVNet​represents the SOTA in CV construction, which is the primary focus of our improvement; further ablation studies will be conducted. GMStereo​is a multi-task framework that incorporates monocular depth estimation, but it does not explicitly use the monocular prior to guide the stereo matching process. 2.Occlusion Validation (R#1, R#2, R#3): We acknowledge the reviewers’ concern that the occlusion validation in the current manuscript is insufficient. A quantitative analysis (EPE, D1-all, Bad3)focused on occlusion will be presented in the extension paper. In the current version, qualitative analysis (Fig. 3n) demonstrates that our method yields lower error in the left-image occlusion regions (caused by camera baseline disparity) compared to the backbone (Fig. 3m). This visually proves the efficiency of our approach in occluded areas. 3.Novelty (R#2, R#3): Our fundamental novelty lies in addressing the precision degradation of CV-based methods at disparity discontinuities. Monocular depth estimation does not suffer from these specific failure modes. Thus, we probabilize the monocular depth prior​to enhance the CV, providing geometric guidance where traditional matching fails. This differs from prior hard features or directly connection concatenation based works. 4.Technical Details and Implementation (R#1, R#2): R#1:The standard deviation σ in Eq.2 controls the spread of the PDF; we empirically selected σ=0.2.Ablation study is planned for the extension paper. R#2: a) We have standardized the module naming to DPCV​throughout the manuscript. b)The normal distribution(σ=0.2) for the probability prior was empirically selected; we plan to explore adaptive parameter selection in future work. c)Regarding the iteration number, our ablation study tested various values, and the results indicate that 1 iteration​ currently offers the most balanced trade-off between performance and efficiency. 5.Reproducibility and Resources (R#1, R#2, R#3): We confirm our commitment to releasing all code upon acceptance. 6.Monocular Uncertainty(R#3):​We acknowledge that monocular depth estimation is generally less accurate than stereo matching in non-occluded areas. Thus, we restrict CV updates only to masked regions to safely leverage the monocular prior(Alg.1). Regarding noise, the CDE implicitly suppresses noise by leveraging the intrinsic geometric relationship between stereo pairs to output robust coarse inverse depth. Notably, although LeRes-ResNet50(applied in this work) and LeRes-ResNeXt101​exhibit different noise characteristics, we observed similar disparity estimation results during training. This result empirically demonstrates that our method is robust to noise in the depth prior. 7.Domain Specificity(R#3): The proposed modules demonstrate effectiveness on the FlyingThings3D dataset during pre-training. This confirms that our method generalizes beyond surgical scenarios to other domains. DPCV targets stereo-occluded regions​where the monocular depth prior is assumed to be more accurate than stereo matching. Consequently, the network’s performance in specular highlight regions depends on whether the underlying monocular depth estimator can effectively handle such artifacts. 8.Latency(R#3):​ The 25.65ms refers to the added overhead of our modules. In the revised manuscript, we report the end-to-end time (45.23ms) as the speed metric . 9.Typo Corrections (R#1): R#1:We have corrected the “fourfold” to “threefold”.




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 paper has received 3 bordeline positive reviews. All reviewers agree on the interest of the approach to iterative refine the cost volume via a prior depth estimate. However all 3 identify key limitations in terms of:

    1) Limited compairson to SOTA disparity methods. It is crucial for the selected comparison baselind to be fully justified - Rev #3 2) Validation on occlusions which is the main selling point - all 3) Clarity on introduced novelty of utilizing prior depth information - all 4) Key information on method’s development details - Rev #2, #3 5) Unclear reproducability without a commitment for public release of code and models. This diminishes the paper contributions- all

    The recommendation is for the paper to be invited for rebuttal.

  • 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.

    All 3 reviewers are overall satisfied with the rebuttal and recommend acceptance. However the main criticism (SOTA comparison, validation under occlusions) is not addressed or explained and is only left as extension work. Without these points fully addressed the paper should follow the suggestion from R2 and make their claims more modest and discussing the limitation more clearly in the final version.

    I recommend acceptance



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 rebuttal effectively addressed the reviewers’ concerns, providing sufficient clarification and additional evidence where requested. Following the rebuttal, all reviewers expressed support for acceptance, indicating a clear consensus that the paper meets the standard for acceptance at MICCAI.



Meta-review #3

  • 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.

    After reading the paper and the comments, I felt that the reviewers have identified the major weakness of the paper. 1) using depth prior for stereo matching is not new, which can been seen from recent foundation models in computer vision, such as foundation stereo. Therefore the technical novelty is limited. This work did not compare with the method. 2) Following this comments, this paper failed to catch the sota in the domain, such as foundation stereo, Monster, etc. 3) One of the main selling point about occlusion, but the reviewers are not convinced to this point, but suggest to “making the claim more modest and discussing the limitation more clearly in the final version”. Overall, I suggest to reject the paper.



back to top