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

Abstract

High-fidelity modeling of the pulmonary airway tree from CT scans is critical to preoperative planning. However, the granularity of CT scan resolutions and the intricate topologies limit the accuracy of manual or deep-learning-based delineation of airway structures, resulting in coarse representation accompanied by spike-like noises and disconnectivity issues. To address these challenges, we introduce a Deep Geometric Correspondence Implicit (DGCI) network that implicitly models airway tree structures in the continuous space rather than discrete voxel grids. DGCI first explores the intrinsic topological features shared within different airway cases on top of implicit neural representation(INR). Specifically, we establish a reversible correspondence flow to constrain the feature space of training shapes. Moreover, implicit geometric regularization is utilized to promote a smooth and high-fidelity representation of fine-scaled airway structures. By transcending voxel-based representation, DGCI acquires topological insights and integrates geometric regularization into INR, generating airway tree structures with state-of-the-art topological fidelity. Detailed evaluation results on the public dataset demonstrated the superiority of the DGCI in the scalable delineation of airways and downstream applications. Source codes can be found at: https://github.com/EndoluminalSurgicalVision-IMR/DGCI.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/0885_paper.pdf

SharedIt Link: https://rdcu.be/dVZiP

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72378-0_51

Supplementary Material: https://papers.miccai.org/miccai-2024/supp/0885_supp.pdf

Link to the Code Repository

https://github.com/EndoluminalSurgicalVision-IMR/DGCI

Link to the Dataset(s)

https://drive.google.com/file/d/1RyiA7dRmXHRirtqWsgncX_BN3VB2kPys/view

BibTex

@InProceedings{Zha_Implicit_MICCAI2024,
        author = { Zhang, Minghui and Zhang, Hanxiao and You, Xin and Yang, Guang-Zhong and Gu, Yun},
        title = { { Implicit Representation Embraces Challenging Attributes of Pulmonary Airway Tree Structures } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15001},
        month = {October},
        page = {546 -- 556}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces the Deep Geometric Correspondence Implicit (DGCI) network, a novel approach that leverages implicit neural representation (INR) for modeling the pulmonary airway tree structures in high fidelity using CT scans. DGCI effectively addresses challenges like spike-like noise and discontinuities by exploring intrinsic topological features and employing a reversible correspondence flow along with implicit geometric regularization. This method not only enhances the smoothness and fidelity of airway structure representation but also improves downstream applications such as skeletonization and breakage repair in airway modeling.

  • Please list the main strengths of the paper; you should write about 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 DGCI network introduces a novel method of using implicit neural representations (INRs) to model the pulmonary airway tree structures. Unlike previous methods that rely on discrete voxel grids, DGCI operates in continuous space, which allows for more precise and smooth structure delineation.
    2. The paper employs an innovative approach to regularization by incorporating implicit geometric regularization techniques. This approach encourages the gradients of the implicit signed distance function (SDF) to have a unit norm, which aids in achieving smooth and high-fidelity surface reconstructions without compromising the structural details.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    1. The author should clearly present the motivation and clinical significance of airway smoothness. For radiologists, airway segmentation with slight discontinuity or noise could not disturb their anatomical understanding and surgical planning. The author should include more introductions to their motivation and more demonstrations of the clinical significance after employing their proposed approach.
    2. The paper should include more clinical validation, for example, the training environment setting, and how much time for one case to run for one case. Moreover, there is not enough discussion about directly applying their approach in clinical practice. Only validating their method on public datasets with only a few cases is not convincing.
    3. The author should compare their methods with other SOTA methods. The current included methods mainly only focus on 3D smoothing and are not so applicable to airway continuity and smoothness.
    4. The evaluation metrics lack enough introduction: how are they calculated, and the extent to which they can represent the quality of outcomes. This is important for readers who are not familiar with this area because the metrics are not commonly known.
  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    1. The author should clearly claim the clinical significance of their method, and how they can help in clinical practice. The author should also include more clinical settings and applications.
    2. The authors should give more introductions to the evaluation metrics.
    3. The author should include some failure cases because claiming the current approach is applicable to all the clinical cases is not convincing.
    4. The author should include an introduction to the computational complexity because Theapproach utilizes implicit geometric regularization and a reversible correspondence flow, might lead to increased computational demands.
    5. The author should introduce how they get the input and corresponding ground truth, including the airway skeleton.
  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

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

    The proposed method and their technical contribution could be well presented, but the motivation and clinical significance are suspicious. The experiment section is not also well presented.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • [Post rebuttal] Please justify your decision

    The author claims that some supplementary materials will be added in the revised version; however, as far as I know, such operations are not allowed. The rebuttal cannot address my concerns, so I would like to maintain my original opinion.



Review #2

  • Please describe the contribution of the paper

    The paper proposed a Deep Geometric Correspondance Implicit Network to model the pulmonary airways in continuous space. The intrinsic topologic regularization was used for implicit network design and implicit geometric regularisation was used to promote smoothness.

  • Please list the main strengths of the paper; you should write about 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 paper domonstrated a novel idea of adapting the implicit neural representation to model the airway in continuous space instead of discrete grids. 2) The authors provide open-source code to promote reproducibility.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    1) The technical innovation of DGCI is not sufficently clear, especially regarding to the difference with DeepSDF. 2) The paper is difficult to follow, particularly when the contribution is embedded in a lengthy last paragraph of the Introduction. Similarly, the technical contribution is difficult to identify in the methodology 2.1 and Fig 2 because no subtitle is provided for each technical contribution. 3) The workflow is difficult to follow, especially how section 2.2 connected with 2.3. 4) In terms of the evaluation, a major concern is the lack of ablation study of each technical innovation (regularisations in DGCI). 5) disconnectivity is not a new issue in airway segmentation and modeling. But this paper does not have 1) sufficient analysis in the introduction and 2) experimental comparison with the existing methods. An example is provided here: “Zhang, Minghui, and Yun Gu. “Towards Connectivity-Aware Pulmonary Airway Segmentation.” IEEE Journal of Biomedical and Health Informatics (2023).”

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    Great reproducibility via open-source code.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    Please find the major concerns in the limitation section. Other concerns include:

    1) The font size in all figures is very small. 2) Fig.3 should be moved to a place near description (section 2.3) to improve readability. 3) It is also not clear what is the major innovation in the proposed fine-tune framework. It should be elucidated explicitly. 4) lack of mathematic details or citation for the evaluation metricsmetrics. 5) Figure 1 effectively illustrates the challenge addressed in this work

  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

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

    The paper domonstrated a novel idea of adapting the implicit neural representation to model the airway in continuous space, and provided open-source code to facilitate reproducibility. However, the major concerns include 1) the lack of clarify on technical contribution of DGCI when compared with DeepSDF, 2) the lack of clarify in writing when demonstrated the contributions in introduction and methodology, 3) missing important evaluation such as ablation study of proposed regularisations and experimental comparison with other breakage-fixing algorithms. These concerns need to be addressed before the acceptance of the paper.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    I raised the rating because the authors sufficiently addressed my concern about the novelty of DGCI compared with DeepSDF. However, I would still suggest to improve the clarify about novelties in the manuscript. Additionally, authors did not provide convincing arguments for not including ablation study, which is important to validate the contribution of each regularisation.



Review #3

  • Please describe the contribution of the paper

    This paper introduces the Deep Geometric Correspondence Implicit (DGCI) network to model high-fidelity pulmonary airway tree structures from CT scans.Key contributions are: 1) Implicit modeling in continuous space to address the drawbacks of voxel-based methods. 2) Reversible correspondence flow for learning shared topological priors. 3) Implicit geometric regularization for precise modeling of fine structures. 4) Proven excellence in airway segmentation and tasks like skeletonization and breakage repair.

  • Please list the main strengths of the paper; you should write about 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.

    S1:DGCI leverages intrinsic topological consistency across different cases to regularize shape modeling through a reversible deep correspondence flow, a novel approach for topological regularization.

    S2:The network incorporates implicit geometric regularization to ensure smooth surface reconstruction while capturing fine details, utilizing the Eikonal term penalty on SDF gradients.

    S3: Extensive evaluations demonstrate leading performance in airway reconstruction, achieving a 98.6% tree detection rate and a 91.2% Dice score.

    S4: The approach also shows significant downstream advantages in skeletonization and breakage repair, with a unique data refinement scheme that prevents overfitting to fracture patterns.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    W1: The hyperparameter settings and some implementation details are missing, which affects reproducibility. For example, the specific architecture of the implicit encoder/decoder and hyper-networks is not provided.

    W2: Ablation studies on the impact of key components, like the correspondence flow and geometric regularization, would provide better insights into their individual benefits.

    W3: The paper demonstrates results on airway modeling, but it’s unclear how well the approach would generalize to other tree-like anatomical structures like vessels. Testing on some additional datasets would strengthen the impact.

    W4: The breakage repair application is an interesting contribution but somewhat preliminary. More details on the prevalence of the breakage problem, range of cases tested, and validation against intra-operative data if available would strengthen this part.

  • Please rate the clarity and organization of this paper

    Very 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    C1: Provide additional context in related work on how the method compares to other recent implicit modeling approaches, especially ones applied to similar tree structures. Highlight the key novelties.

    C2: Include details on the specific architectures used for the implicit encoder/decoder and hypernetworks. And specify key hyperparameters used.

    C3:Present ablation experiments to analyze the individual impact of the correspondence flow and geometric regularization. This would provide useful insights.

    C4: Discuss limitations and potential negative impacts. For example, could the method have failure modes on highly abnormal airways or low-quality scans? Acknowledging limitations builds trust.

    C5: Proofread the paper to fix minor typos, such as “2-norm” in page 4.

  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

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

    The method offers a new and efficient way to model airways accurately using implicit representations with geometric constraints. It outperforms previous voxel-based and implicit methods, with original contributions like correspondence flow and geometric regularization that improve topological and smoothness priors specifically for airway structures. Its technical innovation is well-recognized.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Author Feedback

We thank all reviewers (R1,R3,R4) for their comments. The main concerns are grouped and clarified as below: Q: Technical Contribution and Method Discrimination(R4) We introduce DGCI to model high-fidelity airway in the continous space. Our contribution are: 1) Proposing intrinsic topological consistency within the same class to regularize implicit shape modeling via a reversible deep correspondence flow. 2) Utilizing implicit geometric regularization to promote smooth and high-fidelity representations. 3) Generating airway shapes with state-of-the-art topological fidelity, beneficial for downstream applications (skeletonization and breakage repair). Unlike DeepSDF, which focuses solely on implicit surface reconstruction, DGCI addresses underlying topological constraints within the same category, particularly for fine-scaled airway structures. Our reversible correspondence flow leverages the inherent topological consistency of airway structures, which are single-connected tree structures. In addition, while DeepSDF only constrains SDF values, DGCI incorporates implicit geometric regularization to ensure smooth surface reconstruction without sacrificing fidelity, allowing precise modeling of fine structures. Q: Ablation Study (R1&R4) A: The proposed DGCI is inspired by DeepSDF, further enhanced with the novel reversible correspondence flow and the implicit geometric regularization. Without these modules, our method degrades to DeepSDF. Thus, the results of DeepSDF serve as an important ablation study. Tab.1 reports proposed modules boost more than 10% improvement of TD,BD, and Dice. Fig.4 qualitatively presents the improvement in airway modeling. Q: Clinical Significance and Experimental Setting(R3) A: For clinical significance, automatic preoperative planning for the navigation of endobronchial interventions can save significant effort and provide valuable reference for radiologists. The discontinuity disturbs automatic preoperative path planning algorithms, leading to interrupted trajectories and wrong results for radiologists. This motivated us to design DGCI for smooth, high-fidelity representations of fine-scaled airways. For training, we trained on a Linux workstation (Intel Xeon Gold 5119T CPU @ 1.90GHz, 2 NVIDIA Geforce RTX 3090 GPUs). During inference, our method only needs ~10s / per-case, revealing that the computational complexity is not heavy. Q: Framework Workflow(R3&R4) A: In this work, we introduce the DGCI to model fine-scaled airways in continuous space, achieving high-fidelity performance. DGCI not only enhances the continuity of airway modeling but also aids in its skeleton extraction and breakage repair. Hence, Sec 2.1 and 2.2 detail the design and optimization of DGCI, while Sec 2.3 discusses its use in downstream applications. For the airway breakage repair, we design the implicit fine-tune framework. The main innovation of this fine-tune framework lies in a unique data refinement scheme that prevents overfitting to fracture patterns. While disconnectivity is not a new issue in airway segmentation, our approach addresses it from a novel continuous perspective. We have checked the results from the paper suggested by R4, our method achieved better perfomance on the same dataset, with an increase of over 2% in TD and more than 8% in BD, and also achieved competitive DSC. This can be attributed to the topological regularizations proposed in DGCI. Q: Evaluation Metrics(R3&R4) A: A non-manifold vertex (NM-V) is a vertex connected to three or more non-coplanar faces. A non-manifold edge (NM-E) is an edge shared by three or more faces. A non-manifold face (NM-F) refers to faces arranged in such a way that a closed and continuous body cannot be formed. Fewer non-manifold properties indicate better connectivity and smoothness of reconstructed shapes. TD and BD measure the topological completeness of reconstructed airways, defined in [25]. Detailed description of the metrics will be added in the revised version.



Meta-Review

Meta-review #1

  • 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 proposed method to handle airway tree topology is insteresting and worth discussion in our community. Authors provides decent response to address reviewers comments. Still, authors are encouraged to further clarify the items raised by reviewers for the final version.

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    The proposed method to handle airway tree topology is insteresting and worth discussion in our community. Authors provides decent response to address reviewers comments. Still, authors are encouraged to further clarify the items raised by reviewers for the final version.



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 authors have addressed the major concerns raised by the reviewers. Positive reviews outweigh negative opinions. The authors should revise the paper as promised.

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    The authors have addressed the major concerns raised by the reviewers. Positive reviews outweigh negative opinions. The authors should revise the paper as promised.



back to top