Abstract

Precise visual perception is a cornerstone of safety and efficacy in embodied robot-assisted surgery. However, the intricate operations of microsurgical anastomosis are often plagued by severe occlusions of frequent instrument-tissue interactions, which severely hinder the visibility of micro-scale structures like suturing threads. To mitigate this challenge, we introduce the Camera Pose-Guided Geometry-Aware Transformer, a novel framework for occlusion-aware reconstruction by leveraging a wrist-mounted camera configuration for surgical robots. Specifically, we curate a multi-view dataset with well-calibrated wrist camera poses. By introducing an implicit camera pose encoding mechanism, our proposed method effectively fuses geometric priors into the framework, facilitating robust cross-view alignment and feature aggregation. Furthermore, our architecture exhibits superior robustness to diverse occlusion ratios and numbers of multi-view inputs, enabling reconstruction with complementary wrist perspectives. Our method establishes a new multi-view benchmark for suturing thread recognition in microsurgical anastomosis and significantly outperforms the current state-of-the-art approaches. This work emphasises the importance of the wrist-camera pose in multi-view fusion, providing a transformative approach to occlusion recovery and fine-grained visual perception for surgical robotics.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3592_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)

https://github.com/Zoeric80/Multiple-Wrist-view-Microsurgical-Anastomosis-Occlusion-Dataset

BibTex

@InProceedings{ZhoXin_Wrist_MICCAI2026,
        author = { Zhou, Xinyao AND Liu, Yuxuan AND Luo, Yating AND Luan, Yunfei AND Yang, Guang-Zhong AND Guo, Yao},
        title = { { Wrist Camera Pose-Guided Multi-view Fusion for Occlusion Reconstruction in Robot-Assisted Microsurgical Anastomosis } },
        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 authors propose an interesting experimental Approach for studying occlusion effects on suturing thread Segmentation from different surgical views

  • 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.
    • novel learning approach for occlusion Management during suturing, establishing a multi-view experimental Setup with reliable suturing thread, ground truth from wrist views
    • multi-path learning Approach efficiently combining Images from several views
  • 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.
    • Storyline of the paper could be improved, especially the introduction could benefit from a revision where the higher-level Motivation and detailed and various technical problems are restructured, some Details are too fine-grained, while the Overall Motivation lacks clarity and conciseness. Most importantly, the Need for high-quality segmentation of small structures needs to be motivated
    • Even if quantitative results demonstrate improvements in most metrics, the qualitative results and clDice show that your Approach is overestimating the suture thread Diameters and/or fails to estimate the thread centerline. Please comment on that. How far could this be related to inaccuracies in camera calibration/registration?
    • Please visually and textually emphasize more clearly what the Input/Outputs of your framework are. Your visualizations and important details in Fig. 1 and Fig. 2 are rather small and not easy to follow. The target of suturing thread segmentation is not visible at first hand
    • The experimental data acquisition and Ablation study on the number of wrist views lacks detail. How did you enable more than 2 wrist views? Did you position the suturing thread for a total of 12 static top views and thread positions and move the wrists around to acquire the 17,000 images from multiple views?
  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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
    • What do you refer to as the vanilla model?
    • Are more than 3 views common in those surgical Scenarios?
    • wording Segmentation of surgical sutures might be misleading - I’d expect that you focus on segmenting suturing threads, not the suture itself?
  • 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 proposes a novel multi-view approach for occlusion management in suturing thread segmentation, showing promising quantitative improvements. While the motivation and presentation need clarification, and some qualitative results reveal limitations in thread diameter estimation, the method and experimental setup offer valuable contributions. With revisions, this work merits acceptance.

  • 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 authors carefully commented on the highlighted criticisms. Comments on potential overfitting from less appropriate metrics or unrealistic occlusion simulation have been discussed and explained clearly. Details on the vanilla model, etc., to ensure clarity should be entered in the final version.



Review #2

  • Please describe the contribution of the paper

    The main contribution of this paper is the Pose-Guided Geometry-Aware Transformer, a framework for occlusion-aware reconstruction leveraging wrist-mounted camera configurations in surgical robots. The core technical contributions are:

    1.Pose-Encoding Integrated Attention Mechanism: This novel method integrates camera intrinsics and extrinsics directly into Transformer attention mechanisms via Rotary Positional Embedding (RoPE). By formulating relative projection matrices to embed 3D spatial-geometric priors into 2D patch embeddings, it enables cross-view alignment under arbitrary 6-DoF movements, facilitating the reconstruction of occluded sutures with geometric fidelity compared to existing baselines.

    2.Multiple Wrist-view Occlusion Benchmark: The authors curated a new dataset comprising approximately 17,000 wrist-camera images with well-calibrated poses and fine-grained suture segmentation masks. Rigorous calibration ensures reprojection errors less than 0.5 pixels, establishing the first benchmark for micro-scale flexible object reconstruction under multi-view occlusion in microsurgical anastomosis 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.Implicit Geometric Pose Integration: The novel formulation incorporates relative camera pose matrices directly into Transformer attention via Rotary Positional Embedding, allowing the network to perceive 3D spatial-geometric priors within 2D patch embeddings. This is interesting because it diverges from standard multi-view fusion which often assumes fixed viewpoints or requires explicit depth estimation. Instead, it implicitly learns cross-view correspondences under arbitrary 6-DoF wrist movements, effectively bridging the gap between 2D feature extraction and 3D geometric consistency.

    2.Practical View Count Optimization: Comprehensive ablation studies validate the impact of view count, revealing that a dual-view setup achieves peak performance while preventing informational noise. This finding is critical as it aligns with mainstream embodied robotic hardware configurations and demonstrates that simply adding cameras does not guarantee improvement without pose-guided alignment.

  • 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.Reliance on Simulated Occlusion Patterns: The training strategy employs proactive masking on occlusion-free images rather than capturing real instrument interference. While this ensures label fidelity, it may not fully account for complex optical phenomena inherent to surgical tools, such as specular reflections or depth-of-field blurring, potentially creating a domain gap between synthetic training data and real-world intraoperative scenarios where physical properties of instruments affect visibility.

    2.Mask Width Discrepancy: Qualitative analysis indicates that while structure continuity is improved, the predicted masks are noticeably thicker than ground-truth annotations. This observation suggests a potential trade-off in the loss function design where topological preservation may come at the cost of precise boundary localization, warranting further explanation or refinement.

  • 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

    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.

    (5) Accept — should be accepted, independent of rebuttal

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

    This paper is notable for its novel Pose-Guided Transformer, which integrates camera poses into attention mechanisms to handle arbitrary wrist movements, and the creation of a new Multiple Wrist-view Occlusion Dataset with calibrated poses, offering value to the research community.

  • 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.

    The authors adequately addressed my concerns regarding mask thickness by justifying the prioritization of topological consistency for robotic tasks and explaining how their simulated occlusion strategy supports generalization to real-world scenarios. I am satisfied with these clarifications and recommend acceptance.



Review #3

  • Please describe the contribution of the paper

    1.The authors propose a cross-view alignment module, leveraging camera pose priors to inject a 3D prior into the patch embedding. 2.Curate the first multiple wrist-view occlusion dataset 3.Investigate the potential of integrating camera pose encoding within transformer architectures to facilitate robust wrist-view fusion

  • 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.A novel surgical task is formulated with multi-view fusion for surgical robotic scenes, demonstrating potentials for embodied AI in surgical domain. 2.A novel dataset and benchmark for the proposed problem;

  • 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.

    Major issue: 1.It is unclear whether the compared baselines also undergo and benefit from pretraining. Please clarify the training protocol for all methods to ensure a fair comparison. 2.Please clarify the distinct technical contribution different from pROPE [4]. 3.The paper adopts a two-stage training pipeline with feature-level pre-training followed by segmentation fine-tuning, and the motivation of learning cross-view alignment before the downstream task is understandable. However, the paper does not provide an explicit ablation comparing the full model with and without the pre-training stage, weakening the convincingness to some degree

  • 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

    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.

    (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?

    The evaluation is somewhat insufficient (as noted in the weaknesses), and the technical difference between pROPE and this work is not clear.

  • 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.

    Appreciate the authors’ efforts to clarify all the issues mentioned. All concerns have been addressed



Author Feedback

We thank all the reviewers for their time and feedback. Some expressions and paragraphs will be modified in the camera-ready. Here, we clarify the main issues raised by the reviewers.

C1.To R1/R2 on suture thickness, centerline accuracy and threshold sensitivity. The qualitative blurring reflects predictive uncertainty increasing with mask area, significant at 60% but acceptable at 10%, but real-world instrument self-occlusion rarely reaches 60%. Regarding the Precision metric, ours maintains a performance gap of less than 3% compared to AOT-GAN under mask ratios of 10% and 30%. We designed the 60% mask ratios to demonstrate the robustness under extreme conditions. Our <0.5-pixel calibration error has negligible impact on performance. The clDice metric, targeting centerline accuracy, further confirms skeleton fidelity: we lead or tie all baselines and match THread (a topology-specialized method) at 10% occlusion. For uniform tubular targets manipulation in robotics, we emphasize that topological consistency outweighs exact boundary precision, since uniform thickness can be trivially recoverable from the centerline with the width prior via post-processing. On thresholds, each method has been tuned to its own optimum (ODS-style) to ensure fair comparison. Following the order in Table 1, the AP scores at 30% occlusion are 0.55, 0.44, 0.80, and 0.83(ours), respectively. Generative baselines are threshold-insensitive, their errors arise from incorrect generated geometry instead of the threshold choices. Small threshold variations do not affect our contributions on connectivity, particularly on clDice.

C2.To R1/R3 on the vanilla baseline and fair comparison. The vanilla baseline shares the same architecture as ours, and implement the same two-stage training strategy, with only the pose-guided encoding removed (RoPE retained). Pretraining is essential for mitigating Transformer overfitting on the limited amount of surgical dataset. We pretrained on the Realestate dataset with a batch size of 4 for 35 epochs. Other baselines are CNN or task-specific architectures without requirements of large dataset pretraining. We verified that pretraining them before fine-tuning yields no meaningful gain over direct surgical-domain training. Consequently the reported results are fair.

C3.To R3 on technical difference with PRoPE. We cite PRoPE as inspiration for combining projective relations with RoPE in attention, differences are on formulation-level: (i) PRoPE uses symmetric self-attention where every token plays an homogeneous role. Differently, we explicitly separate the occluded query view (D_o) from wrist key/value views(D_w) to model the directional relation between them, composing an asymmetric operator D_o^T D_w^(-1) suited for occlusion completion. (ii) Note also that our output transformation in Eq.1 uses D_w (wrist-side) rather than the query-side used in PRoPE to maintain numerical stability and frame consistency of the aggregated features under asymmetric setup. (iii) We transfer to a new embodied setting and emphasize that camera pose can serve as a guiding prior for embodied surgical perception, that we hope draws greater attention.

C4.To R2 on proactive masking vs. real instrument occlusion. Specular reflections and depth-of-field blur can be treated as occlusion variants, degraded regions are simply masked and reconstructed via the same pipeline. Although we do not implement real-world instrument occlusion, our large-area feature-level masks pose a more challenging case, supporting the framework’s potential for severe occlusion recovery in clinical scenes.

C5.To R1 on dataset acquisition and multi-wrist-view setup. Wrist-view data is captured via the robot’s dual arms recording static-scene videos, ensuring consistent targets within each sequence. Views beyond two are obtained by sampling temporally distant frames with large viewpoint displacement from the same stream, simulating multi-view images of the same scene.




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.

    This paper proposes a method for semantic segmentation of thin suturing thread occlusions from multiple wrist-mounted cameras on a surgical robot. Due to the very specialised set-up, a dataset with groundtruth labels is curated as part of the paper contribution. The method is evaluated in comparison to baselines aimed at inpainting occlusions.

    Reviewers acknowledge that the method is methodologically sound and that the generated dataset is valuable.

    A few issues are identified:

    • R1, R2 note that the proposed method over-segments the needle threads in comparison to some of the baselines. This raises questions on how much the parameters of all these methods can be tweaked to over/under segment and influence quantitative metrics. In addition, I would also like ask whether the present fixed evaluation metrics based on final predictions with predefined thresholds are enough to validate the method (this is perhaps the main parameter influencing oversegmentation). Should alternative metrics also be reported that may be less sensitive to individual method’s choice of prediction threshold? For example in edge detection literature it’s common to report metrics with data-adjusted thresholds (OIS,ODS) or that aggregate results from different thresholds (AP).
    • R3 notes that experiment fairness needs to be clarified with further training details about the baselines
    • R3 requests clarification regarding technical novelty of this paper compared to [4]
    • R2 asks whether the generated dataset has enough realism to be representative of real surgery conditions and R1 requires clarifications on how the results from Fig. 4 are generated.

    All the above issues are appropriate to respond in a rebuttal for a more informed final decision on this paper

  • 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 authors have carefully answered to the reviewer concerns, and there is now consensus to accept this paper.



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.

    Accept

  • Please justify your recommendation.

    Remaining issues include over-segmentation and related choice of metric, experimental fairness, technical novelty, and dataset realism. All post-rebuttal reviews commend the author’s response, saying that the rebuttal addresses the concerns, and recommend acceptance.



back to top