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

Abstract

Chest X-ray (CXR) image examination is a primary tool for assessing thoracic abnormalities. It is widely utilized for initial diagnosis and screening of diseases due to its cost-effectiveness and low radiation dose. Segmentation of ribs in CXR images (CXR rib segmentation) facilitates rapid determination of lesion types and locations, thereby alleviating the workload of medical professionals. Deep learning-based methods have achieved significant progress but still face some challenges in CXR rib segmentation, such as the occlusion challenge caused by artifacts and the interlace challenge caused by the spatial overlap of ribs. Therefore, it can be observed that the topological knowledge of ribs is crucial for CXR rib segmentation but neglected in existing methods, including the connectivity and interactivity of ribs. To address these challenges, we propose a novel learning framework that integrates explicit topological priors into segmentation networks for precise CXR rib segmentation. In particular, we introduce two modules including the connectivity prior embedding module and the interactivity prior embedding module. These modules are designed to explicitly encode the continuity and interactivity of ribs into deep learning models for end-to-end training. Both modules are plug-and-play and can be integrated into various networks. We conduct extensive experiments on VinDr-RibCXR and CXRS datasets to evaluate the segmentation accuracy of each rib using multiple metrics. Evaluation and visual results show that our method exhibits strong adaptability, seamlessly integrating with diverse architectures and enhancing performance across various networks. Our code is publicly available at https://github.com/XWei98/LTSeg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0313_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/XWei98/LTSeg

Link to the Dataset(s)

VinDr-RibCXR dataset: https://vindr.ai/datasets/ribcxr

BibTex

@InProceedings{ZhaXia_Learning_MICCAI2025,
        author = { Zhao, Xiaowei and Li, Chenglong and Tang, Jin and Li, Chuanfu},
        title = { { Learning with Explicit Topological Priors for Chest X-ray Rib Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15975},
        month = {September},
        page = {302 -- 311}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper improves the CXR rib segmentation performance by introducing rib prirors w.r.t. 1) continuity of single ribs, and 2) the overlapping regions of rib masks.

  • 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 connectivity & interactivity idea is straightforward and intuitive. The modules proposed can be plug-and-played into networks with ease. 2) The paper includes comprehensive experiments.

  • 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.
    • Sec.2.2 continuity prior module:
    • The morphological operation in continuity prior module could make the pipeline less robust, i.e., too much dilation will make KR_p less impactful, while under dilation could make the CP_m less constrain, authors should clarify 1) how the kernel size is determined to ensure a reasonable DR_p, and 2) how sensitivity the pipeline is w.r.t. the dilation.

    • The author should justify the dilation and mask overlap operation, as an easier and more robust approach could be: directly predicting binary segmentation for each rib, calculating the GT overlapping regions of the rib, and adding continuity constraint to the rib prediction.

    • Sec.3.2: The detailed setting for baseline models should be clearly explained: 1) Does the models perform binary segmentation for each individual ribs or instance segmentation? 2) Are the pre-trained baseline models finetuned for comparison?

    • Table 1: With the proposed modules, most models perform worse in terms of mSen, authors should analyze the result in detail, as the improvement over other metrics seems marginal.

    • Implementation details: 1) Authors should detail the experiment settings for better reproduction, e.g., parameters for kernal, baseline experiment setting, etc. 2) How does the pieline handle the cases where some of the ribs are missed, e.g., floating ribs.

    • Detailed analysis on results should be included, e.g., scenarios in Fig.4 should be explained, how the proposed method helps?

    • More related works on rib segmentation should be discussed. e.g., disscusion of benchmarks on this topic, justify the choice of 2D image segmentation instead of 3D volume (as the overlapping issue could be avoided in 3D segmentation)
  • 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?

    Please find my major concers of this paper as follows:

    1) As a paper focusing on technical contribution, the performance boost is marginal, and most architectures w. proposed method perform worse on mSen metric.

    2) The proposed method could be sensitive due to the morphological operation, yet the paper doesn’t include sufficient details to justify the approach.

    3) Important details of implementation is missed (e.g., dilation in Continuity prior Embedding Module, baseline experiment settings), which makes the reproduction challenging and the analysis less convincing.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    This work primarily contributes a new rib segmentation framework for chest X-rays that explicitly integrates topological priors into deep learning models, namely rib continuity and spatial interaction. This is made possible by two plug-and-play modules that improve the model’s comprehension of rib structure and relationships: the continuity prior embedding module and the interactivity prior embedding module. These modules are end-to-end trained and can be incorporated into current segmentation networks. Extensive assessments on the VinDr-RibCXR and CXRS datasets show that the method successfully handles typical issues such occlusion and rib overlap, resulting in more accurate rib segmentation.

  • 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.

    structure by explicitly including anatomical knowledge, such as rib continuity and spatial interaction, into the segmentation process. The continuity prior embedding module and the interactivity prior embedding module are two plug-and-play modules that make it simple to incorporate into pre-existing deep learning models without having to completely revamp the architecture. Improved Handling of Complex Challenges: By including topological restrictions into the learning process, the method successfully tackles typical rib segmentation problems such occlusion and overlap.

  • 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 approach is restricted to 2D topological relationships obtained from chest X-rays, which are projections of the 3D anatomy of the chest. This could decrease segmentation accuracy, particularly in complicated anatomical regions, as crucial spatial context and depth information are lost and not used. Previous modules’ interaction and continuity depend on predefined topological information, which could not be entirely flexible enough to accommodate changes in patient anatomy or clinical situations like overlapping lesions or rib abnormalities. The method’s efficacy is largely dependent on the proper balancing of the primary segmentation loss (LSeg) and the topological prior loss (LCoIn). Despite the fact that the problem stems from a 3D anatomical structure, the framework does not make use of any reconstruction methods or 3D imaging modalities (such as CT) that could provide richer spatial priors.

  • 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.

    (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?

    Topological Prior Integration: The approach increases model comprehension of rib structure by explicitly including anatomical knowledge, such as rib continuity and spatial interaction, into the segmentation process. The approach is restricted to 2D topological relationships obtained from chest X-rays, which are projections of the 3D anatomy of the chest. This could decrease segmentation accuracy, particularly in complicated anatomical regions, as crucial spatial context and depth information are lost and not used. Fixed Prior information Assumption: Previous modules’ interaction and continuity depend on predefined topological information, which could not be entirely flexible enough to accommodate changes in patient anatomy or clinical situations like overlapping lesions or rib abnormalities.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #3

  • Please describe the contribution of the paper

    The authors propose a new chest rib segmentation method in x-ray modality. It explicitly integrates the rib’s topological priors into the segmentation network for precise rib segmentation in case of spatial overlap of ribs. The two modules that focus on continuity and interactivity, respectively, are introduced. The authors demonstrate that the proposed method can be easily integrated into various networks and effectively improves rib segmentation performance.

  • 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 proposed prior modules are well-defined and effective, with strong Interpretability.
    2. These modules can be easily integrated into existing segmentation networks without requiring architectural modifications. The authors also validate their effectiveness across different models.
    3. The visualization results indicate that the proposed modules significantly mitigate the occlusion challenge in CXR segmentation.
  • 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. From Fig. 3, it appears that in the Continuity Prior Module KR_p assumes that discontinuities along a rib segment do not exceed five or fewer pixels. Is this prior too strong? Such a hard constraint may not generalize well to cases with bigger gaps.
  • 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 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
    1. The authors also validate the effectiveness of the MedSAM backbone. Does this mean that MedSAM was fine-tuned using the proposed topological prior loss? In addition, how was the prompt specified when using MedSAM?
  • 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?

    The proposed novel method is presented in good quality. It is sufficient for a MICCAI publication.

  • 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.

    Good work. The authors has almost addressed the concerns of mine.



Author Feedback

We sincerely thank the Reviewers (R1, R2, R3) for their constructive comments and for recognizing our work as “easy to plug-and-play (R1, R2, R3), comprehensive experiments (R2), well-defined and effective, with strong interpretability (R3), significantly mitigate the challenges in CXR segmentation (R1, R3)”. We explain major concerns as follows: (R1) Limited to 2D: In 3D image, rib discontinuities still occur due to artifacts such as fraying or shadow-like noise, which makes our continuity prior embedding module still applicable. As noted in our conclusion, incorporating richer 3D priors and addressing challenging structures such as floating or pathological ribs remains a promising direction for future research. (R2) Choice of CXR: We chose 2D CXR segmentation because CXR remains the most accessible and commonly used imaging modality in global clinical practice due to its low radiation and cost-efficiency. Although the rib overlap issue can be avoided in 3D imaging, it remains a prevalent challenge in CXR. (R2) Sensitivity, (R1, R3) flexibility of the CEM: We appreciate the question on our CEM. We employ full connectivity convolution to perform the dilation. Regarding the sensitivity of the convolution kernel selection, details in Section 3.3 and Table 2-(right), we evaluate five kernels of varying sizes (3×3,5×5,7×7) and shapes (1×5, 5×1), with the 5×5 kernel (covering 25 pixels) achieving the best result. Since this dilation cover effects accumulate across pixels, this setting sufficiently covers most discontinuous pixels. We also agree that tailoring designs to patient-specific anatomy offers greater flexibility, and we are inspired to explore anatomy-aware dynamic kernels in future work to improve adaptability. (R2) Performance Gain and mSen drop: Our CEM penalizes false negatives (FN) where regions belonging to current rib are mistakenly assigned to background and helps the model reduce FN. Our IEM penalizes false positives (FP) where regions belonging to adjacent ribs are mistakenly assigned to the current rib and helps reduce FP. Some of these regions are true positives (TP) for neighboring ribs, penalizing them makes the model slightly conservative, recognizing rib pixels as background and leading to a slight increase in FN. Therefore, our CEM increases TP and decreases FN, leading to higher mSen. In contrast, adding IEM reducing FP but slightly increases FN, leading to a decrease in mSen. These variations are reflected in Table 2-(left); However, our method consistently improves five evaluation metrics across diverse architectures (CNN, Transformer, Mamba, SAM) on two datasets, especially in IoU and DSC, which show substantial gains. We further explore balancing strategies between IEM and CEM to mitigate this trade-off in future work. (R2) Justification of the dilation and overlap: As shown in Fig. 2, our framework directly performs binary segmentation on each individual rib (the visualization combines different rib with distinct colors for comparison). The disconnected region within each prediction do not occur solely in ribs’ overlapping area; they are also common in areas affected by occlusion, noise, or shadowing along the ribs. Therefore, constraining only the overlapping regions is incomplete. We recover these discontinuous regions through dilation, capture and emphasize such regions by applying mask overlap operation between the inverted and dilated predictions, which cannot be addressed by simply computing GT overlapping. (R2, R3) Lack of baseline implementation details, clarification on MedSAM: We apologize for the lack of details due to the page limit. We provide full implementation details on GitHub upon release to ensure reproducibility. Additionally, both MedSAM and MedSAMours are fine-tuned using adapters, and MedSAM uses standard loss, while MedSAMours uses our topological loss. We all adopt a prompt-free strategy for both by setting prompt encoder input to “None”, avoiding reliance on expert prompts.



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.

    Reject

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

    The paper proposes an interesting idea to deterministically find the connectivity-critical and interactivity-critical regions and add more penalties there through the loss function. ([A] is closely relevant)

    However, the experimental results are not solid. (1) All the results in Table 1 are much worse than the SOTA in the original paper ([18-19]); (2) while the improvement from the proposed method is only around 1%. Thus, it’s unclear if the proposed method can improve upon the current SOTA. As the method is conceptually straightforward and lacks theoretical guarantees, stronger empirical results are crucial to demonstrate its practical value and justify its acceptance.

    [A] Maximin affinity learning of image segmentation.



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

    The approach is restricted to 2D topological relationships obtained from chest X-rays, which are projections of the 3D anatomy of the chest. This could decrease segmentation accuracy, particularly in complicated anatomical regions, as crucial spatial context and depth information are lost and not used. Previous modules’ interaction and continuity depend on predefined topological information, which could not be entirely flexible enough to accommodate changes in patient anatomy or clinical situations like overlapping lesions or rib abnormalities.

    The method’s efficacy is largely dependent on the proper balancing of the primary segmentation loss (LSeg) and the topological prior loss (LCoIn).

    Important details of implementation is missed.



back to top