Abstract

Laparoscopic augmented reality (LAR) enables real-time visualization of internal organ anatomy, effectively reducing surgical risks. Rigid point cloud registration aligns the spatial position of the preoperative image point cloud with the intraoperative laparoscopic video point cloud, playing a pivotal role in the virtual-real fusion visualization for LAR. However, the limited field of view in laparoscopic surgery results in only partial visibility of the organ. This leads to an incomplete video point cloud that exhibits low overlap with the image point cloud, rendering registration highly susceptible to local optima. Moreover, the smooth and texture-deficient organ surface makes popular superpoint matching methods based on feature similarity ineffective. Inspired by the highly consistent morphology of the video and image point clouds at organ bottom edges, we propose an edge guidance (EG) mechanism to address the challenge of sparse surface features in laparoscopic scenes. The EG mechanism identifies edge points by calculating the standard deviation of correlations among neighboring points, prioritizes edge alignment, and subsequently guides the matching of other points. We leverage this mechanism to develop an edge-guided rigid point cloud registration network, EG-Net. Compared with the state-of-the-art method PARE-Net, EG-Net achieves at least a 7% improvement in accuracy and an 11% increase in speed across three laparoscopic datasets: the public DePoLL dataset, a pig liver surgery dataset, and a human liver surgery dataset. With its high accuracy, fast speed, and strong generalization, EG-Net holds significant potential for clinical applications in laparoscopic surgery. The code is available at: https://github.com/FDC-WuWeb/EG-Net.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1510_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/FDC-WuWeb/EG-Net

Link to the Dataset(s)

https://github.com/FDC-WuWeb/EG-Net/releases

BibTex

@InProceedings{WuWen_EGNet_MICCAI2025,
        author = { Wu, Wenbin and Gao, Yifan and Wang, Yixiu and Zhang, Jiayi and Zhao, Yiming and Gao, Xin},
        title = { { EG-Net: An Edge-Guided Network for Rigid Registration of Laparoscopic Low-Overlap Point Clouds } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15969},
        month = {September},
        page = {158 -- 167}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper aims to align points cloud with partial overlap for laparoscopic applications.

  • 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 method uses a reasonable set of known building blocks KPConv, Res(Net) Blocks and Pooling as well as Densepoint (optimal transport) and Superpoint (cross-attention) matching. While the edge guidance is not a novel idea but using the standard deviation of correlations to identify edges seems to be a good contribution Overall there is a good evaluation provided. Metrics seem to fit well and some ground truth is available.

  • 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.

    Why are these specific loss functions chosen? Superpoints are not well defined - are those pooled points. I found the paper had an overly complicated description of architecture. Also the experiments not described very well, in particular whether the augmentation was just performed on one case? Is it really one case used for training and the other 52 for testing Why was this choice made? Why is such an extensive augmentation necessary? It seems (also considering the batch-size=1, and just 80 epochs/iterations) that the model is fitted to a single test case based on all those potential initialisations, right?

    While Predator is a good baseline, some more recent approaches from computer vision/graphics should be discussed. EYOC (extend your own correspondences) CVPR 2024 seems to use a similar idea that recomputes neighbourhood interaction based on previous alignment.

    ICP is non-competitive, instead coherent point drift (CPD) should be employed. At MICCAI / MedIA the Free Point Transformer was often used for medical datasets: https://doi.org/10.1016/j.media.2021.102231 which is missing here.

  • 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.

  • 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/2025/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.

    (2) Reject — should be rejected, independent of rebuttal

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

    While in principle the approach is sound and the results promising, but the paper is let down by an overly complicated method description that fails to convey any substantial novelty. There only limited ablation studies to provide meaningful insight.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    I acknowledge that the authors aimed to address some concerns and promised to release code for reproducibility. However, the points about fairness of comparisons remain a bit uncertain. The methodological contribution of the edge weighing is intuitive but not groundbreaking. The poor results of CPD are puzzling. I remain somewhat critical - since some claims are not entirely well backed up. I would consider the paper borderline. Nothing is fundamentally wrong, but by submitting two papers with substantial overlap I have the feeling the authors simply hedged their chances.



Review #2

  • Please describe the contribution of the paper

    This work proposed a point cloud registration network for registration between the point cloud generated from the preoperative image and the point cloud extracted from each video frame. The main contribution is the proposed edge guidance mechanism to filter edge points increase their weights for point-point matching.

  • 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 task is interesting and seems to be chanllenging due to the low overlapping, while the proposed method is effective and robust enough to perform well on three datasets.

    2. The evaluation is also sound and convincing

  • 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. In the introduction, they said “organ surfaces are smooth and texture features are sparse”. What is the meaning of sparse features? Dose that indicate there are almost high frequency information? If yes (that means almost no edges in the surface), how can we fiter those “edge points”, as mentioned in Section 2.2.

    2. In Section 2.1, I recommend using “residual KPConv block” instead of “Res block”.

    3. In Eqn. 1, what are W_i and W_i,k? In Figure 3, it seems there are two convolution weights: one for the center point, and one for the neighbors.

    4. In Section 2.2, the definition of N(p_i). The subscripts of neighbor points should be (i, k) instead of only (k): p_i,1; p_i,2; …; p_i,K.

    5. In Section 2.2, weights for features of filtered edge points will be increased. But how to use these weights?

    6. Will dataset 2 and 3 be released?

  • 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.

  • 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/2025/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 — could be accepted, dependent on rebuttal

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

    The task is interesting and performance is good, with some concerns on the implementation details.

  • Reviewer confidence

    Very confident (4)

  • [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.

    Overall, the manuscript is ok and I recommend acceptance. The authors are recommended to carefully revise the paper for the camera-ready version. Additionally, the code and datasets should be released prior to the start of the conference.



Review #3

  • Please describe the contribution of the paper
    • This work presents a method designed for rigid registration of preoperative image point cloud and intraoperative video point cloud.

    • The method is robust and accurate especially when low overlap exists for the pointcloud pairs by design and incorporating an edge guidance mechanism.

    • The method also shows sound generalization capacity across datasets, and superior performance compared to baseline methods.

  • 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 authors propose a novel weighting mechanism leveraging a fair intuition and demonstrate its contribution for registration tasks with the presented design.

    • The presentation of the work is logically sound to follow. Methodology details are mostly clearly stated and solid. Pipeline figures and visualization of the results are also nicely organized and presented.

    • The effectiveness of the design is properly demonstrated.

  • 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.

    Motivation:

    • The authors present the method as a solution for laparoscopic surgery, e.g. liver and kidneys, whereas the experiments are only conducted on liver datasets. I would suggest either constraining the application context or providing experimental evidence (Intuitively I agree it would also work in kidney registration, due to the clear edges of the organ, but I am a bit suspicious for some other laparoscopic surgeries when the edge is absent).

    Architecture design:

    • Not clear how the obtained superpoint matches are populated to the fine intra matches, not clearly stated but ‘ fed into .. to generate …’. Eg. the number of neighbours to populate a superpoint match? Is the number critical? Hyperparameter tuning involved? Also I would assume giving high weight to the matched superpoints is an important design, whereas the ‘high’ is not clear to me. Could the author elaborate more on these?

    • To understand the edge guidance in Fig.3, I would assume edges captured from image point cloud is more like a ‘curvature’ in the closed 3D model, whereas edges captured from partial video point cloud is the ‘boundary’ of the point cloud? Are they intuitively mostly matched and therefore serve as a sound base for guidance? If they are not mostly matched (as shown in fig of the two sets of red points), will the unmatched ‘dorphan’ superpoints or outliers matches be wrongly weighted relatively high? I am personally not clear regarding this point. Or there is more assumptions here like inlier matches will still dominant as they are significantly more in quantity?

    Experiments and baselines:

    • Overall, the effectiveness of the proposed design is demonstrated, whereas its comparison with other baselines looks a bit weak and unfair (unclear) to me: the method demonstrates decent generalization performance across liver datasets. Can the authors comment a bit regarding its (potential) generalization cross other organs (eg. kidney, gallbladder) and the reasons behind? Will retraining be needed? I asked because I am not clear about its level of generalization (organ-specific?)

    • Baselines method may not be fairly compared to concretely demonstrate the performance gaining of the proposed novel design: In my understanding, the architecture looks a bit similar to the one proposed in GeoTransformer (KPconv based feature extraction; superpoint-level matching; intra-superpoint-level matching; corse-guided-fine design, which is also mentioned several times in the manuscript), while the proposed weighting strategy differs as it highlights the weights for detected edge superpoints rather than leveraging an error-guided weighting in GeoTransformer. Would be curious about 1) The justification for GeoTransformer is not in comparison 2) What’s the expected difference in terms of performance and generalization capacity?

    • According to the manuscript, I did not find information that the baselines are retrained on DePoLL liver dataset as the proposed EG-Net has been trained, are they? If not, I would argue the comparison may not be as fair, considering the surgical data has a significant domain gap compared to the ones in the general world (e.g 3D match).

    • Follows above, thin-plate-spline is adopted to bake deformation to the pointcloud in the training set, I would assume in addition to increase the quantities for training as step (1) and (3), one important function is that it helps the method to be robust for actual surgical dataset when object deformation do exist? If this holds, it should be more critical to train the baselines on augmented DePoLL. As per my knowledge, the baselines are not considering deformation by design or baking it into their training data.

    • Additionally, the “Conical cropping” adopted in data augmentation seems to create artificial and “smooth” edges, which can rarely exist in real surgery. Could the author clarify how this augmentation will influence the generalization of the edge detection in the proposed method?

    • The proposed method presents superior performance in terms of registration using detected edges compared to baseline methods. However, there is no discussion or visualization regarding registration in difficult cases, for example, no true edge information is available, which often happens in real surgery. Will the proposed method yield worse results or fail? Furthermore, there is no discussion regarding the limitations of the proposed method and possible future work.

    MISC:

    • sec 1, “P2PNet [6], based on kernel point convolution…”: wrong reference

    • sec 2.2, “…the dense point matching block expands it to local dense points I and J based on P and Q,”, but in figure the dense points are P* and Q*

    • The figures are elaborate, but the captions are too brief, in principle, they should be self-explanatory without referring to the main text. considering expanding them with a bit more detail.

  • 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.

  • 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/2025/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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    • The paper presents a practical approach that incorporates with the detected edges from pre- and intraoperative point clouds to improve rigid registration. The method outperforms baseline methods in terms of accuracy and inference time.

    • More clarification and discussions are expected from the reviewer including motivation, method and experiments design.

  • Reviewer confidence

    Very confident (4)

  • [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.

    most of my main concerns have been adequately addressed pretty well by the authors, especially the treatment of the unmatched edge points. Overall, I believe the proposed method is novel and practical in liver rigid registration, the datasets are valuable.




Author Feedback

We warmly thank the reviewers for their positive/constructive comments. They say that our method is “a good contribution”(R1), “novel”(R3), and “interesting”(R2), with evaluation being “good”(R1), “sound and convincing”(R2&R3). Here we address the main points in their reviews.

  1. Loss Function (R1): The proposed EG-Net uses the generic overlap-aware circular loss and point matching loss to optimize superpoint (those pooled points) and dense point matching. They have been explained in SOTA methods P2PNet and PARE-Net. We will clarify this, highlight our innovation, and trim other content in the Method.
  2. Data Augmentation (R1&R3): Our data augmentation includes various types of random cropping, deformation, and transformation. Even with only one original case, the augmented point clouds exhibit significant differences in scale, morphology, or position. The total quantity of 11.2k ensures diversity. Under this augmentation, both the PARE-Net and EG-Net have shown considerable generalization, proving that sufficient augmentation can compensate for the limited original case. Using more cases for augmentation could slightly improve all methods’ performance, as the same data is used for each. It won’t change their relative performance rankings but will reduce the test set’s size and diversity, causing more harm than good. Cone cropping simulates lobe opening states and changes point cloud distribution, benefiting all methods’ generalization. Its artificial edges are unavoidable, but significantly differ from natural edges in the test set, offering no help to EG-Net’s edge guidance generalization.
  3. Additional Experiments (R1&R3): 1) Rigid CPD lacks low-overlap handling and remains prone to local optima (RRE: 21.00°, RTE: 27.68mm, TRE: 33.15mm on DePoLL). Free Point Transformer focuses on non-rigid registration, making comparison unfair. 2) EG-Net significantly differs from EYOC. EYOC uses progressive self-supervised learning with staged training; it uses registration results from near-range, high-overlap point clouds in initial stage to guide registration of far-range, low-overlap point clouds in expansion stage. EG-Net requires no staging; it performs edge identification and registration in one step without intermediate results. Laparoscopic point clouds exhibit no near-far diversity or high-overlap data, making EYOC’s two-stage training unsuitable and direct comparisons difficult. 3) We have compared with P2PNet, known as GeoTransformer, whose error-guided weighting relies on geometric features and struggles in texture-deficient laparoscopic scenes. EG-Net’s improvement by highlighting edge superpoint weights validates our motivation.
  4. Edge and Superpoints (R2&R3): 1) 3D liver models from CT contain only non-closed surfaces in contact with pneumoperitoneum (refer to DePoLL). Both CT and video point clouds have a boundary/edge (red line in leftmost image of Fig. 3’s blue box), which doesn’t conflict with sparse texture features. Edges at the organ base in CT and video point clouds are morphologically similar and overlap, inspiring us to use them for guidance. 2) Non-overlapping edges (e.g., upper edges) vary in shape, causing unmatched edge superpoints with distinct features. EG-Net increases all edge superpoint features’ weights without changing the degree of similarity between them, thus not affecting similarity-based matching. Edge superpoint features are fed into the next Pooling block with all other features, where their weights (proportion) are relatively enhanced by computation. For fairness, the number of neighbors to populate a superpoint match follows SOTA methods P2PNet and PARE-Net. We will elaborate on these if space allows.
  5. Subscripts and Figures (R2&R3): We will revise subscripts, improve figure captions, and discuss other organs/scenes with less distinct edges. We have prepared the code repository. Datasets 2 and 3 will be released together. We sincerely thank the reviewers for their valuable feedback! :-)




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The main issues raised by reviewers were 1) description of the architecture (R1/R3), 2) description of experiments (R1), and 3) training/testing dataset size (R1). The rebuttal did not provide a comprehensive response but only addressed 2) and 3) partially.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    While the concerns about presentation clarity and novelty articulation are valid, the method is technically sound, practically motivated, and received positive support from two confident reviewers. The rebuttal addressed key technical concerns, and no fundamental flaws were identified.



back to top