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

Abstract

Accurate 3D–2D liver registration, which aligns preoperative 3D models to partial, view-dependent intraoperative surface observations, is critical for AR-guided laparoscopic surgery but remains challenging due to severe occlusion, limited visibility, and the lack of 3D ground-truth supervision. Existing landmark-free approaches perform partial-to-complete geometric alignment, yet robust self-supervision under extreme partial visibility remains difficult. We propose Vis2Reg, a visibility-aware registration framework that explicitly constrains deformation using mask-consistent visible regions. We introduce a visibility-aware self-supervision that derives a visible-domain 3D supervision signal from intraoperative masks, enabled by differentiable point rasterization and mask-guided back-projection. This formulation improves robustness under severe occlusion while maintaining fully self-supervised learning. Vis2Reg combines a robust geometric rigid initialization module with an implicit neural deformation field for stable alignment. Experiments on real intraoperative datasets achieve Dice score of 0.926 and Chamfer Distance of 1.434 mm, with 111 ms per frame inference time, demonstrating both accuracy and practical efficiency.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5305_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/aimsgroup-Leeds/Vis2Reg.git

Link to the Dataset(s)

N/A

BibTex

@InProceedings{FenJia_Vis2Reg_MICCAI2026,
        author = { Feng, Jiaming AND Zhang, Xukun AND Farid, Shahid AND Ali, Sharib},
        title = { { Vis2Reg: Visibility-Aware Landmark-Free Geometric 3D–2D Registration for Liver Laparoscopy } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16889},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes Vis2Reg, a visibility-aware, self-supervised framework for liver registration in laparoscopic surgery. The method combines a rigid initialization module with an implicit non-rigid deformation field and introduces a visibility-aware supervision strategy derived from intraoperative masks. The approach is evaluated on real intraoperative datasets and shows improvements over prior learning-based 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 motivation of the paper is clearly articulated and well grounded in the challenges of laparoscopic liver surgery, particularly the issues of occlusion, partial visibility, and lack of supervision.
    • The overall organization of the paper is reasonably clear, making the pipeline and high-level contributions relatively easy to follow despite the technical complexity.
    • The proposed method is novel, especially the idea of visibility-aware self-supervision. The combination of: rigid initialization, implicit deformation fields, and mask-guided supervision is interesting and relevant.
    • The design choices for rigid and non-rigid registration modules are conceptually appealing, and the integration of different components (GeoTransformer-like matching, implicit fields, and rasterization-based supervision) is promising.

    [1] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. PREDATOR: Registration of 3D point clouds with low overlap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4267–4276, 2021

  • 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.
    • Lack of methodological details in several key components: Many parts of the method are insufficiently described, which makes it difficult to fully understand or reproduce the approach. This includes architectural details, training specifics, and design motivations.
    • The paper frames the task as 3D–2D registration. However, the intraoperative input is a point cloud extracted from RGB-D data, which is inherently 3D. This raises confusion about whether the problem is truly 2D–3D or rather 3D–3D (partial-to-complete) registration. If the method is intended to be 2D–3D, the process used to estimate depth from laparoscopic images is not explained.
    • The formulation defines the intraoperative input as the union of multiple point clouds across views and frames, yet these observations likely correspond to different temporal states and deformation configurations of the liver. Treating them as a single coherent point set may introduce inconsistencies and could negatively affect both rigid and non-rigid alignment.
    • This is my main concern: The non-rigid module is conditioned only on the preoperative point cloud extracted features, not on intraoperative observations: –> This suggests that the predicted deformation may be independent of the intraoperative input, potentially leading to identical deformations for different intraoperative point clouds. –> The reported performance on non-rigid registration may be influenced by low variability of deformation in intraoperative data and by the use of Chamfer Distance, which measures proximity rather than true correspondence accuracy.
    • The proposed rigid initialization consists of multiple steps (GeoTransformer features, PROSAC sampling, ICP refinement), but: –> The motivation for this multi-stage pipeline is not clearly justified, especially compared to a GeoTransformer-only approach. Given that the proposed approach and GeoTransformer share multiple common features. –> Although the method claims superior performance, the differences with GeoTransformer are not clearly explained, which makes the reported large gain difficult to interpret.
    • The deformation-related losses (L_def, L_topo, L_smooth) are used but they are not defined, and they are not included in the ablation study, which makes their contribution unclear.
    • The evaluation focuses mainly on learning-based methods, which weakens the experimental validation. It would be important to compare against classical registration methods, such as ICP, ICP+RANSAC, or simulation-based approaches (e. g. , FEM-based methods), especially given the claim of robustness.
    • The rigid pipeline combines ideas from both GeoTransformer and PREDATOR [1] (e. g. , robust estimation), yet: PREDATOR [1] is not included in the comparison, which is a notable omission given its relevance.
    • The related work section lacks coverage of important classes of methods, including: simulation-based approaches, shape matching techniques such as functional maps, and broader categories of non-learning-based registration methods.
    • Reproducibility: The paper does not provide sufficient information to reproduce the results: –> Important hyperparameters and implementation details (e. g. , GeoTransformer configuration) are missing, –> Some loss formulations are not fully specified. –> There is no mention of code release, which further limits reproducibility.
  • 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.

    (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 several concerns are raised above, my primary issue lies with the conditioning of the implicit deformation field. In its current formulation, the non-rigid module is conditioned only on preoperative features, without explicitly incorporating intraoperative observations. This suggests that the predicted displacement field may not adequately adapt to variations in the intraoperative data, raising fundamental questions about its ability to perform meaningful non-rigid registration. In my view, this is a critical limitation that undermines the core contribution of the paper and constitutes the main reason for my decision to reject it.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    While the conditioning on the intraoperative point cloud appears to be a typo, the authors still did not address my concern (also raised by R2) regarding the significant performance discrepancy reported for GeoTransformer in rigid registration. Specifically, the rigid registration results are inconsistent with those reported in the original dataset paper by Zhou et al. [21]. In the paper, the authors report GeoTransformer to achieve an RRE of 31.3° and an RTE of 78.10 mm, whereas the dataset paper reports values below 1° for RRE and below 2 mm for RTE [21]. In contrast, the reported results for Self-P2IR (RRE 0.21 and RTE 1.32) are in line with the dataset paper. Although differences in random seeds, dataset splits, or hyperparameter settings may lead to minor performance variations, the discrepancy observed for GeoTransformer is unusually large and cannot reasonably be explained by these factors alone. This issue remains unresolved.

    Furthermore, the authors did not address another important concern regarding the problem formulation. The intraoperative input is defined as the union of multiple point clouds collected across different views and frames. However, these observations likely correspond to different temporal states and deformation configurations of the liver. Treating them as a single coherent point introduces inconsistencies.

    Therefore, I maintain my rejection.



Review #2

  • Please describe the contribution of the paper

    The paper addresses a challenging setting in laparoscopic liver registration, namely non-rigid complete-to-partial registration under low overlap, where only the visible intraoperative region is informative for registration. The proposed method is landmark-free and self-supervised, and its main contribution is a visibility-aware self-supervision strategy designed to exploit only the visibility-consistent part of the intraoperative observations. To enable this, the method introduces a pipeline based on rasterization and back-projection to extract the intraoperative features needed for supervision. The overall framework estimates both a displacement field to deform the complete preoperative point cloud so that it matches the partial intraoperative observations, using only geometric information from the point clouds.

  • 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.

    One major strength of the paper is that it focuses on a relevant and difficult problem, namely self-supervised complete-to-partial non-rigid registration under low overlap, where only the visible organ surface should contribute meaningful supervisory information. The paper identifies this challenge clearly and proposes a method specifically aimed at handling the visibility-consistent region. The main strength of the method is the proposed visibility-aware self-supervision, which, in my reading, is the most original part of the work. The idea of constructing a mask-consistent visible-domain 3D supervision signal from intraoperative masks and camera parameters seems interesting, and although it is built using standard components such as rasterization and back-projection, the overall formulation appears novel. The framework is also well structured. It contains a feature extraction stage, a rigid initialization module, and a displacement-field estimation module. The non-rigid part is based on a SIREN-based MLP that predicts 3D displacements and deforms the point cloud locally, while the total deformation field is obtained by combining this with the rigid transformation. The overall design is coherent and technically reasonable. I also found the experimental validation convincing overall. The experiments are well organized, and the ablation study is useful because it supports the main implementation choices. In general, the validation gives confidence that the proposed visibility-aware supervision is beneficial.

  • 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.

    My main weakness concerns the clarity of the novelty with respect to prior work. From my understanding, the key novelty is the masked visible domain used for supervision. This idea seems new to me, even though it is constructed using standard tools such as rasterization and back-projection. However, the paper does not explain clearly enough how this differs from the closest prior work. In particular, Self-P2IR appears to be the most similar reference, since it also shares the idea of a rigid module plus a non-rigid module, as well as a self-supervised strategy based on 2D masks and visible regions. Although there are implementation differences, I think the authors should explain much more explicitly what is truly novel/different compared with that work and discuss on why the proposed approach achieves better performances.

    I also have some doubts regarding Table 1.In the synthetic part, I would like to understand whether GeoTransformer was retrained under the same conditions as Vis2Reg, because the reported performance gap is very large and somewhat surprising, especially considering that the rigid initialization module of Vis2Reg itself is based on GeoTransformer. I would have expected a smaller difference. Whatever the reason for this improvement, I think the paper should explain it explicitly. Moving to the real intraoperative setting in the Table 1, I find the Dice score much more informative than the Chamfer distance. Chamfer is very different from a classical target registration error and is therefore less informative in this context. For example, GeoTransformer shows a rotation error of 31 degrees and a translation error of 78 mm, yet still obtains a Chamfer distance of 3 mm. This suggests that the relevance of Chamfer as a registration metric should at least be discussed, and one could even consider removing it.

    In Sect. 3.1, I would also ask the authors to clarify how the Dice score is computed from the point clouds. It would be useful to explicitly state whether the point clouds are projected into masks or images before computing Dice.

  • 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

    Some minor comments: I think the explanation of the problem setup in Section 2.1 could be improved. At first, the formulation is somewhat confusing because the F views are introduced together with the point clouds, making it unclear how they relate to the actual model inputs. Only later, in Section 3.1, I understood that these F views are linked to the masks used for self-supervision. Even with that clarification, however, I still do not fully understand why the intraoperative point clouds need to be merged into Q. If my interpretation is correct, I believe the setup would be more intuitive if it were presented as: one preoperative point cloud and one intraoperative partial point cloud as inputs to the model, together with one mask from the intraoperative image used for supervision. Related to this, the paper implicitly assumes that the intraoperative information is already available as a point cloud, but it never clearly states how this point cloud can be obtained during a real surgery. For clarity, especially for readers less familiar with the problem, it would be useful to add a sentence explaining that intraoperatively we have laparoscopic images, and that a combination of segmentation and reconstruction methods can be used to extract the point cloud of the organ of interest.

    In Figure 2, I found the visualization somewhat unintuitive. Using liver-shaped point clouds instead of generic square-like clouds would make the figure easier to read. Also, at first sight it is not obvious whether the two encoders are arranged in series or in parallel, because the arrows are not very “visible” and the pipeline does not start in the usual top-left position. I eventually understood the figure by noticing the color coding matching Figure 1, but I still think the current layout is not very intuitive. Reorganizing the rows and making the arrows more visible could improve readability.

    In Sect. 2.2 you should immediately mention GeoTransformer and SIREN-based MLP.

  • 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?

    I am overall positive on this paper. The problem addressed is important and technically challenging, especially because the registration is complete-to-partial and only the visibility-consistent region should contribute useful supervision under low-overlap conditions. The proposed visibility-aware self-supervision appears to be the main contribution, and in my view it is an interesting idea. Although it is built from standard operations such as rasterization and back-projection, the overall supervision strategy seems meaningful and potentially novel. The results and the ablation study support the design choices, and overall the validation is convincing.

    My main reservations are mostly about clarity and positioning, rather than the overall quality of the work. In particular, I think the paper should explain more explicitly how it differs from the closest related approaches, especially Self-P2IR, and should clarify some aspects of the setup and presentation. However, these concerns do not outweigh the strengths of the paper for me.

  • 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 adequately addressed my concern regarding the novelty of the proposed method with respect to Self-P2IR in their response to point 2.Their explanation is reasonable, and I now consider the core idea to be novel.

    However, my concern regarding the large performance gap with GeoTransformer remains unanswered, and this issue was also raised by Reviewer 1.In fact, the authors’ response to point 4 raises additional questions rather than resolving the concern. They state that the full initialization pipeline, composed of GeoTransformer, PROSAC, and trimmed ICP, is essential for the final registration performance. This implies that a non-negligible part of the reported improvement may come from the initialization strategy rather than from the proposed visibility-aware component itself.

    This distinction is important because it directly affects the interpretation of the paper’s contribution. How much of the final performance gain is attributable to the initialization pipeline, and how much to the proposed visibility-aware logic? For example, if approximately half of the improvement comes from the initialization and the other half from the visibility-aware strategy, then both components should be explicitly presented and evaluated as separate contributions.



Review #3

  • Please describe the contribution of the paper

    The paper proposes a visibility-aware, self-supervised framework (Vis2Reg) for 3D–2D liver registration in laparoscopic surgery. The key idea is to restrict supervision only to visible regions using mask-consistent learning built from differentiable rasterization and back-projection. This is combined with a rigid initialization stage and an implicit neural deformation field, enabling registration under severe occlusion and partial visibility without requiring ground-truth correspondences.

  • 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 paper addresses an important clinical problem: robust AR-ready liver registration under strong occlusion.
    • The visibility-aware supervision is well-motivated and makes sense because it avoids enforcing constraints on unseen regions.
    • The pipeline is practical, especially the mix of rigid initialization followed by non-rigid implicit deformation.
    • The ablation studies are solid and clearly show that each component (visibility masking, rigid seed, and loss terms) actually contributes to performance.
    • Experimentally, the method outperforms baselines on Dice score and Chamfer Distance.
  • 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.

    -A main concern is that the novelty feels incremental. Most components (feature matching, rigid refinement, implicit deformation fields, differentiable rendering) already exist, and the main contribution seems to be how visibility-aware supervision is designed and integrated rather than a fundamentally new registration framework.

    • There are also concerns about robustness and generalization. The evaluation is mainly on one benchmark (plus related synthetic data), so it is unclear how well the method would transfer to other surgical settings or datasets. A missing direction is evaluation on tumor registration. It would be interesting to test the method on public dataset with 3D preoperative liver and tumor models, together with 2D laparoscopic images and known camera intrinsics, to better demonstrate clinical relevance. Public dataset (Rabbani, Navid, et al. “A methodology and clinical dataset with ground-truth to evaluate registration accuracy quantitatively in computer-assisted laparoscopic liver resection.” Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization 10.4 (2022): 441-450.)

    • The method is quite complex, with many interacting modules. This makes it harder to understand what is truly essential beyond what is shown in the ablation studies. A clearer explanation of the rigid initialization stage and more discussion of failure cases (especially under heavy occlusion) would help.

    -How many views 𝐹 are typically required for stable performance?

    • It is not clear how the loss is chosen? The final loss combines several terms (Chamfer variants, silhouette loss, deformation regularizers). How were the loss weights chosen?

    • Figure 3 has bad image resolution and is difficult to interpret, which reduces readability of the results.

    • A key practical question is how the model obtains intraoperative partial point clouds 𝑄𝑣. In real surgery, where only 2D laparoscopic images are available, how 𝑄𝑣 will be estimated? is there recent state-of-the-art method on this.

    • Another key issue is the liver mask 𝑀𝑣. In real settings, this would need to be obtained from automatic liver segmentation on laparoscopic images? how?

    • “Vis2Reg runs at 111.38 ms per forward pass …. supporting near-real-time intraoperative use”. if many pre-steps are needed, how to guarantee real-time use?
  • 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?

    The paper is well written, however the methodological novelty is moderate, and the evaluation is somewhat limited.

  • 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.

    Thank you for the clarifications provided in the rebuttal. The explanation that the non-rigid module is conditioned on both preoperative and intraoperative inputs addresses part of my main concern, and the clarification regarding the reconstruction of intraoperative point clouds from depth and camera intrinsics improves the overall understanding of the framework.

    I also appreciate the additional discussion comparing the method with Self-P2IR and clarifying the role of the visibility-aware supervision.

    That said, I still believe the contribution is somewhat incremental, and some implementation details remain insufficiently explicit for full reproducibility. In particular, the conditioning mechanism of the deformation field and the evaluation protocol could be described more clearly.



Author Feedback

We thank the reviewers for recognizing (all) clarity and well-grounded research, (all) original visibility-aware supervision, (R2/R3) structured and (R1) promising design, (R2/R3) solid validation/ablations, and (R3) AR relevance. We apologize for compact descriptions that are clarified below. We will release the code upon acceptance. 1.(R1) Non-rigid deformation conditioning. Sec. 2.2/Fig. 2 may misleadingly imply the non-rigid module uses only preoperative features (F_P^r), this is a presentation issue as it does incorporate intraoperative information. In our model, the registration network takes a pair (P,Q), with Q (intraoperative point-cloud) influencing both the rigid initialization (sR_{P,Q},t_{P,Q}) and the non-rigid features. The deformation is thus Δ_i = g_φ([PE(x_i), f_i^r(P,Q)]) and x̂i = sR{P,Q}(x_i + Δ_i) + t_{P,Q}, so it also depends on Q. We will revise to use pair-conditioned notation F_{P|Q}^r and clarify that P denotes the source points where displacements are queried. 2.(all) Problem formulation Sec. 2.1.The clinical goal is 3D-to-2D AR overlay, while the network itself performs partial-to-complete 3D point-cloud registration by aligning P to reconstructed intraoperative Q (obtained from depth and intrinsics). For each view v, the liver mask M_v (as in Self-P2IR [21]) is applied to the depth map (DepthAnything), and intrinsics K_v are used to reconstruct Q_v. (all) We use three views (F=3) as a short local, non-temporal window. Thus, Q_v,M_v,K_v provide view-specific 3D/2D supervision losses for Vis2Reg; this will be clarified. 3.(R2,R3) Novelty Vs Self-P2IR. Self-P2IR is the closest prior work, but the key difference lies in how masks define 3D supervision under partial visibility. Unlike its symmetric rendered-mask consistency, Vis2Reg uses mask-gated visible regions for alignment with Q_v via L_vis and applies one-way 3D loss (L_3D) from Q_v to the warped model avoiding constrain on unobserved model regions, which is supported by clear performance drops in Table 2 when these components are removed (e.g., 92.60/1.43 to 74.29/3.12 w/o L_vis).  4.(all) Rigid initialization and baselines. All baselines used the same folds and evaluation/protocol, and the reported rigid result of Vis2Reg corresponds to the full initialization pipeline (GeoTransformer with mutual filtering, PROSAC scoring, and trimmed ICP), not a standalone matcher. This initialization is essential for the final registration (Dice/CD drops from 92.60/1.43 to 69.32/3.64 when weakened).  (R1) While adding PREDATOR/ICP/RANSAC-ICP could broaden rigid-only evaluation, our main contribution is visibility-aware self-supervision for partial-to-complete registration.  5.(all) Metrics, losses, and reproducibility. (R1,R2) CD is not TRE and is only a complementary surface-distance metric. Dice is the primary AR-overlay metric, computed by rasterizing the registered silhouette and comparing it with the liver mask. (R1,R3) The omitted regularizers are L_def=mean_i||Δ_i||^2, L_smooth=mean_{i,j∈N(i)}||Δ_i-Δ_j||^2, and L_topo=mean_{i,j∈N(i)}(||x̂_i-x̂_j||-||x_i-x_j||)^2.They control displacement magnitude, local smoothness, and local distance preservation. Due to page limits, we did not describe them as they are well-established losses. Table 2 focuses on visibility supervision, one-way partial-observation loss, and rigid seed. Loss weights were fixed across folds and chosen heuristically; values are in Sec.3.1, and released code will include all experiments. 6.(R3) Evaluation scope and runtime. Evaluation uses the P2I-LReg dataset (21 patients, including data from Navid et al.) with patient-level 5-fold validation to avoid data leakage. We therefore restrict claims to this benchmark only. Tumor/vascular evaluation needs internal-anatomy GT or matched annotations and is outside scope. We will clarify. The 111 ms is forward registration after Q_v/M_v are available; i.e., excluding depth/mask reconstruction time; this will be stated as a limitation.



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 submission received mixed reviews. The majority of the major criticisms were raised by R1; however, additional important concerns were also noted by R2, particularly regarding the lack of meaningful novelty/positioning regarding Self-P2IR, and by R3, who raised issues related to limited novelty, unconvincing experiments, and—consistent with R1— issues with the assumption of 3D point clouds as input data, as explained by R1 (practical and experimental concerns).

    R1 further highlights several critical issues, including insufficient methodological detail, lack of comparison to leading classical methods, inadequate discussion of practical limitations arising from dependence on reconstructed 3D point clouds, and the high complexity of the proposed pipeline, particularly in relation to GeoTransformer. Additional concerns include weak ablation studies that do not clearly demonstrate the contribution of individual components to performance gains over state-of-the-art methods, missing baseline comparisons (notably PREDATOR), and a lack of implementation details—especially regarding hyperparameters and loss function definitions—impacting reproducibility.

    Given these concerns, I recommend that the authors be given an opportunity to rebut the important raised issues as summarized above.

  • 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 reviewers agree that the paper addresses an important and clinically relevant problem in laparoscopic liver registration. The proposed visibility-aware self-supervised learning strategy is widely recognized as the main contribution. Overall, the approach is considered well motivated for handling occlusion and partial visibility in surgical settings.

    Post-rebuttal, the decision remains split, with two reviewers leaning toward acceptance and one maintaining rejection. R2 and R3 both support acceptance, while still noting concerns regarding evaluation clarity, reproducibility, and the interpretation of contributions. In particular, both highlight that the experimental protocol and implementation details are not yet sufficiently transparent to fully disentangle the contributions of individual components.

    R1 maintains a rejection, primarily due to concerns regarding the rigid registration evaluation and the large discrepancy in reported GeoTransformer performance compared to prior work. However, the rebuttal clarifies that the rigid registration result corresponds to the full initialization pipeline (GeoTransformer with mutual filtering, PROSAC, and ICP) and that all baselines were evaluated under a consistent protocol and identical folds. This clarification reduces, but does not fully eliminate, these concerns.

    A second central concern raised by multiple reviewers is the attribution of reported improvements. While ablations suggest that both the visibility-aware supervision and the initialization pipeline contribute to performance, the extent to which each component independently drives the gains remains insufficiently disentangled. This limits the interpretability of the claimed contributions.

    In addition, all reviewers point to limitations in methodological clarity and reproducibility, including incomplete loss definitions, missing implementation details, and limited specification of hyperparameters. The rebuttal improves clarity in some areas, but concerns regarding full reproducibility remain.

    Despite these issues, the majority of reviewers support acceptance after rebuttal, citing the relevance of the problem, methodological novelty, and apparent improvements over prior methods.

    Overall, this is a borderline case leaning toward acceptance. However, camera-ready revision should address key issues of clarity and reproducibility. In particular, using additionally supplementary material and/or public code, the authors should provide:

    • Improved clarity on the evaluation protocol and baseline comparability
    • Complete specification of all loss terms and training implementation details
    • Greater clarification on meaningful results impact of initialization vs visibility-aware learning.

    These improvements would significantly strengthen the interpretability and reproducibility of the results.



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 reviewers showed some enthusiasm for this work being presented at the conference. Several concerns remain but the overall contributions are strong enough to justify acceptance.



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.

    Despite of two positive recommendations the three reviewers share concerns about how much the imporvement come from the proposed method itself or from the initialization and differences in evaluation setup. I feel uncomfortable in recommending acceptance without having this matter clarified by a suitable ablation study that shows beyond reasonable doubt that the core contribution, that is the visibility aware strategy, leads to meaningful imporvements.



back to top