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

Abstract

Liver landmarks provide crucial anatomical guidance to the surgeon during laparoscopic liver surgery to minimize surgical risk. However, the tubular structure properties of landmarks and dynamic intraoperative deformations pose significant challenges for automatic landmark detection. In this study, we introduce TopoNet, a novel topology-constrained learning framework for laparoscopic liver landmark detection. Our framework adopts a snake-CNN dual-path encoder to simultaneously capture detailed RGB texture information and depth-informed topological structures. Meanwhile, we propose a boundary-aware topology fusion (BTF) module, which adaptively merges RGB-D features to enhance edge perception while preserving global topology. Additionally, a topological constraint loss function is embedded, which contains a center-line constraint loss and a homology persistence loss to ensure homotopy equivalence between predictions and labels. Extensive experiments on L3D and P2ILF datasets demonstrate that TopoNet achieves outstanding accuracy and computational complexity, highlighting the potential for clinical applications in laparoscopic liver surgery. Our code is available at https://github.com/cuiruize/TopoNet.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0755_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/cuiruize/TopoNet

Link to the Dataset(s)

L3D dataset: https://github.com/PJLallen/D2GPLand P2ILF dataset: https://p2ilf.grand-challenge.org/

BibTex

@InProceedings{CuiRui_TopologyConstrained_MICCAI2025,
        author = { Cui, Ruize AND Zhang, Jiaan AND Pei, Jialun AND Wang, Kai AND Heng, Pheng-Ann AND Qin, Jing},
        title = { { Topology-Constrained Learning for Efficient Laparoscopic Liver Landmark Detection } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15969},
        month = {September},
        page = {585 -- 594}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors developed a deep learning based framework that is able to detect prominent edges (e.g. anatomical borders) on endoscopic images of the liver. The authors propose a series of different developments (Snake CNN-Dual Path Encoder with RGB and depth image as input, Boundary Aware Topological Fusion, And specific tailored Loss functions) called TopoNet and evaluate them against existing developments and these experiments look promising.

  • 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.
    • Approach could be valuable for Augmented Reality
    • A comparison to state of the art has been performed which seems to be in favour for the method
    • Ablation study has been performed
  • 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 title says “landmarks”, however, I think the most common understanding of a landmark is rather a 2D or 3D point in space, not an edge. Therefore, this needs to be clarified and changed.
    • Ground Truth: I am not so familiar with “contour/edge detection” on laparoscopic images, however, I feel like the Ground Truth needs more explanation because from looking at the examples provided (Fig 1, Fig 2, Fig 4) it is not clear to me why certain edges where highlighted and others not where there seems to be a clear border. Furthermore, it seems like there are different category of edges (red, green, blue) and its not clear what they mean and how these different categories were considered in the method.
    • Missing Medical Background: Since the method is very much tailored to liver surgery, the authors should provide much more medical background knowledge to this use case.
    • Evaluation: It is not clear from how many different surgeries the frames used for training and evaluation are coming from (data split on the level of patients/surgeries not frames). Please provide detailed description to clarify whether current developments rather show evaluation on same-surgery cases, which is a clear disadvantage for generalization purposes.
    • Low number of cases: In general , the approach was trained and evaluated on a low number of frames (167 training, 17 testing; 124 training, 43 testing) and no crossvalidation was performed.
    • Evaluation: No standard deviation is provided, and no significant values are provided to show that own method is superior to other approaches
  • 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

    Please provide the source code.

  • 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 — could be rejected, dependent on rebuttal

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

    Description of the paper needs to be improved. Evaluation needs to be clarified.

  • 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 highly appreciate the authors answers to the points raised, especially clarifications on the term “landmarks” have been provided. I still find the evaluation in its current form a bit weak, e.g. only two patients for testing in one cohort and four patients for testing in the other cohort. In such cases I think crossvalidation would be adequate to avoid selection bias, but of course, such novel results are not allowed according to the MICCAI regulations.



Review #2

  • Please describe the contribution of the paper

    The paper presents a method called TopNet for detecting landmarks in laparoscopic images on the liver. The authors conduct experiments on two public datasets for liver landmarks, L3D and P2ILF and compare IoU, Dice, Assd, and inference speed and floating point operations per second. Their method outperforms several models in most metrics on the task at hand. In some metrics their method outperforms the compared methods a lot! Further, they conducted an ablation study on the 3 key components of their method on the evaluation set of L3D. Their method consists of snake topology acquisition blocks to extract depth topological structures and a CNN to extract detail features. Then a boundary aware topological fusion module merges RGB-D features. As their method relies on RGB-D data, they estimate depth as a preprocessing step with AdelaiDepth. The paper gives a detailed description of the method in the form of several formulas. However they do not indicate if they want to release source code.

  • 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.

    I regard the very detailed description of the method in mathematical formulas as one of the major strengths. While other authors might prefer being given access to a implementation, I think the paper lists all details to reimplement the method. Further, the comparison to 12(!) other methods is impressive. Most of all, the results speak for themselves. If the experiment was conducted correctly (which I assume so), the author’s method outperforms the state of the art in all metrics, in some metrics by a lot. Further the number and quality of references is more than adequate.

  • 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.

    I come from a slightly different background, therefore please take these comments with a grain of salt. I found the paper very hard to follow, as the authors expect that the reader is an expert on landmark detection on liver from laparoscopic images, familiar with the state of the art in 2025. I think that not only these experts should be addressed as audience, but also researchers with a different background (e.g. clinicians, medical computer graphic experts etc.) should be able to understand the important bits. All the images are way too small. In Fig1 I could not even see the circle the caption mentioned. Furthermore, the language in some parts of the paper is imprecise or colloquial. Another weakness is a lack of implementation details. While I agree that the method could probably be implemented in any framework, I’d like to know your setup to compare the performance metrics. E.g. what programming language and version was used, which libraries and backends and their versions? Also more details on the hardware (e.g. CPU, VRAM) Some of the metrics for the comparison seem unusual for me (again, this is not my area of expertise!), especially their units. I would have assumed that Assd. is measured in some length unit, e.g. mm, however it is reported in pixels. Furthermore, is GFLOPs not measuring computations per second, why should it be vastly different between methods, should this not be a characteristic of the hardware? I know that the page restrictions for MICCAI are very strict, but I’m not a fan of the references to be shortened. Almost all conferences and journals are only listed by their abbreviation not full title.

  • 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.

  • 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

    Here are several minor comments: Abstract: I find the notion that “landmarks” provide “guidance” a bit odd and especially that this “minimizes surgical risk”. What landmarks are the authors referring to as “tubular”? In the beginning of the abstract it is not yet clear that the modality in which landmarks are detected from is laparoscopic images. CNN,RGB and RGB-D abbreviations should be introduced.

    Keywords: Is RGB-D fusion fitting? Depth is estimated anyways, and I’m not clear what is being fused with RGB-D

    Introduction: Reference 12 is not fitting. Reference 12 is about knee surgery with an HMD. I think this reference is a bit far fetched. For AR liver surgery I recommend for example the early work of N. Haouchine, J. Dequidt, I. Peterlik, E. Kerrien, M. -O. Berger and S. Cotin, “Image-guided simulation of heterogeneous tissue deformation for augmented reality during hepatic surgery,” 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), Adelaide, SA, Australia, 2013, pp. 199-208. Please describe in greater detail what kind of landmarks the authors try to detect. What are the “tubular” landmarks? Some anatomical aspects, like blood vessels? What are “marked forms of liver landmarks”? “D2GPLand [16] integrates depth geometric priors”. Should it maybe be “geometric depth priors”? I’m a bit overwhelmed, about the snake topology acquisition without any explanation here. Could the authors explain this a little bit more, by either giving a short explanation, giving a reference here, mentioning that they will write later about this… I don’t consider this to be basic background that can be expected from any reader.

    Methodology Is “frozen AdelaiDepth” the right formulation? Why “frozen”? The abbreviation “STA” was introduced in 1 before. I had to look up” tortuous”. Maybe “twisting” is easier? 2.2 and 2.3 is very nicely detailed!

    Experiment “cutting edge” sounds too sensational. It is not mentioned what underline and bold in the table means. how is the impressive improvement in Assd in L3D explained? Why is it not in P2ILF? “Since only the training set can be available” is a weird formulation. While I guess its ok to empirically set these values, how did the authors get here? Are the values in line with what others used, does it completely break, when we slightly alter them etc. Why are some numbers in the text bold? “improvement in the Assd metric is tremendous” sounds too colloquial and sensational. “GLOPs” -> “GFLOPs” Can the authors give any explanation, why TopoNet is so efficient? Table 2: Why are the values for TopoNet so different compared to Table 1?

    Conclusion The conclusion is very similar to the abstract. Is there anything else to say?

    References: Accidentally I noticed that reference 25 seems to list the wrong year. While the authors list the stated year on their github, the publisher lists a different year, and in this case I’d go with the publisher. I recommend looking through the references and verifying the information is correct.

  • 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?

    While I’m not a specialist in this particular subfield, I consider myself part of the broader target audience. I found the paper challenging to follow at times, as it assumes a high level of familiarity with niche-specific concepts. A clearer explanation of the real-world motivation and practical applicability of the task would help broaden the paper’s accessibility.

    The writing is generally okay, though occasionally imprecise or overly casual—for instance, some formulations feel a bit sensational or confuse metrics with units. After reading the paper, I trust that the proposed method works as described, but I struggled to grasp the underlying reasons for its effectiveness or how the approach might transfer to other domains.

    That said, the method itself seems to be described in enough detail for an expert to implement, and the results are compelling. Based on that, I’m leaning toward a Weak Accept.

  • 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.

    My raised issues were addressed and the authors promise to make the corresponding changes to the paper. Especially after reading the other reviews, which confirm my perceived strengths of the paper and reduce the perceived weaknesses, I am now more confident to recommend an “acceptance”.



Review #3

  • Please describe the contribution of the paper

    The paper addresses liver landmark detection in laparoscopic videos and introduces TopoNet, a novel topology-constrained learning architecture built upon state-of-the-art methods and modules. The proposed framework includes: (1) A snake-CNN dual-path encoder, designed to capture both RGB information and depth-informed topological structures, and (2) a Boundary-aware Topology Fusion (BTF) module for adaptively merging RGB-D features, enhancing edge perception while preserving global topology. In addition, the authors introduce two novel loss functions - a centerline constraint loss and a topological persistence loss - to guide and improve the training process.

  • 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 positive aspects of the paper are as follows: (1) The paper is generally well-written and easy to follow; (2) Although it incorporates existing modules (e.g., ResNet-34, DSConv, BTF), the design and efficient training of the TopoNet architecture are non-trivial and demonstrate thoughtful integration. (3) The experimental results showcase state-of-the-art accuracy, along with competitive inference speed and computational efficiency (measured in GFLOPS) compared to existing methods. (4) The selection of competing approaches for the evaluation presented in Table 1 is appropriate and relevant. (5) The ablation studies indicate the relevance of each proposed module.

  • 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 negative aspects of the paper are: (1) λd, λcl, and λper were empirically set. However, the paper does not include a sensitivity analysis or discussion on how these hyper-parameters impact model performance. (2) There is no discussion or evaluation of how the inferred landmark detection contributes to downstream applications, which limits the understanding of its practical relevance.

  • 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.

  • 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.

    (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?

    Liver landmark detection in laparoscopic videos is a timely and relevant research topic, which this paper addresses through the proposed TopoNet architecture that demonstrates state-of-the-art performance. In my opinion, the paper mainly lacks a discussion regarding its applicability to downstream tasks.

  • 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.

    Liver landmark detection in laparoscopic videos is a timely and relevant research topic, which this paper effectively addresses through the proposed TopoNet architecture, demonstrating state-of-the-art performance. The authors have confirmed that they will incorporate additional discussions in the final documents to address the concerns I raised in my initial review. Given this commitment, the other reviewers and the overall feedback provided, I have decided to accept the paper.



Author Feedback

We appreciate valuable comments from all reviewers, and we are encouraged by positive comments including valuable (R1) and non-trivial (R3), detailed describe method (R2) and state-of-the-art (R1-R3). We promise to release our code upon acceptance. Below we address specific concerns. R1: Landmarks&GT: The definition of liver landmarks is consistent with previous works [1][2]. Unlike the traditional concept of defining landmarks as points, liver landmarks are defined as silhouette (green), ridge (red), and falciform ligament (blue). The continuous landmarks used have been validated and can provide effective liver registration. We will add more explanations. [1] Ref.16 [2] Ref.1 Background: Thanks for your suggestion. As mentioned at Para.1 of Sec. 1, liver landmarks are important for preoperative-to-intraoperative registration and AR navigation. We will add more background descriptions in final version. Dataset: As mentioned in Sec. 3.1, we use two public datasets. L3D [1] includes 1,146 frames (921 train, 126 validation, 109 test) from 39 patients (32 train, 3 validation, 4 test). P2ILF [2] includes 183 frames, but we can only access the training set with 167 frames (124 train, 43test) from 10 patients (8 train, 2 test). We will add patient details. Evaluation: Table 1 shows the superior performance of our model. We additionally compute significant values, indicating our results are statistically significant (p < 0.05). We will report standard deviation and p values. R2: Paper accessibility: Thanks for constructive advice. Based and beyond [1][2], we will include more related fundamental knowledge including definitions and applications. [1] Ref.16 [2] Ref.1 Implementation details: We use Python 3.9.19 and PyTorch 2.1.2 for coding. The CPU is Intel Xeon Gold 5218R and VRAM is 48GB. We will report in the revision. Other libraries will be released with our code. Metrics: The units of DSC and IoU are %. In line with related works, we calculate Assd on frames, so it is measured in pixels. We require camera parameters to transfer to real-world length units like mm. GFLOPs (lowercase ‘s’) measures the total number of floating-point operations of an algorithm. Manuscript problems: We thank R2 for detailed comments, including small figures, missing explanations, reference, etc. We will refine them. Minor comments: Guidance: As consistent biomarkers, landmarks can guide surgeons to locate critical anatomical structures, leading to more precise surgeries and lower risk. ‘tubular’: The datasets we used [1][2] annotate liver landmarks (silhouette, ridge, falciform ligament) as thin and twisting regions, which we regard as tubular structures. Indeed, vessels also have tubular structures. ‘marked forms’: Annotation forms of landmarks like point, region, etc. RGB-D fusion: Fusion of RGB and depth features. ‘frozen’: Model tuning parameters are kept static during training to reduce cost. Gain in Assd: We have analyzed in Sec. 3.3, line 7 that this mainly because less false positive predictions. As two datasets have different data distribution and scales, our model has different sensitivity on them. But we still perform better and the improvement is outstanding on P2ILF. Efficiency: We design the TopoNet in a lightweight structure (fewer parameters & GFLOPs). Table 2 vs. Table 1: In Sec 3.4, line 3, we have claimed that we use the validation set of L3D for ablations in Table 2 and the test set of L3D in Table 1. Effectiveness: As explained in Sec. 1, Para. 2, TopoNet enhances performance by embedding topology constraints to learn tubular structures. It can also be transfer to other tubular detection tasks. R3: Hyper-parameters: Actually, we have previously ablated several values and applied the optimal ones. We will add discussion. Applications: Thanks for constructive comment. Detected landmarks are applied for intraoperative decision making and accurate detection contributes to 2D-3D landmark-based liver registration. We will discuss more.



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.

    Reject

  • 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’

    N/A



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’

    N/A



back to top