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

Abstract

Quantitative analysis of lymph node volume is instrumental in the diagnosis and treatment of cancer. However, automatic segmentation models for lymph nodes necessitate pixel-level labeling, which is both time-consuming and labor-intensive. The scarcity of pixel-level annotations has thus spurred interest in label-efficient learning as a potential solution. Considering the variance of shapes and locations, and the low-contrast appearance of lymph nodes in computed tomography scans, we propose a new incomplete annotation strategy called orthogonal partial-instance annotation, in which only two orthogonal slices of a small portion of lymph nodes are annotated. To segment as many lymph nodes as possible from such sparse annotations, we propose a prototype-based label-efficient learning framework with a specifically designed loss. Specifically, we extract intra-batch prototypes from the output features of the encoder and store inter-batch prototypes using a momentum-smoothing approach. To re-inject the extracted information from the two kinds of prototypes, we introduce a feature augmentation module that utilizes the extracted prototypes to enhance features. To further complement the predictions generated from enhanced features with those from original features, we design a reliability-based co-teaching strategy based on feature similarity. Experiments demonstrate that our proposed framework outperforms other methods on two mediastinal lymph nodes datasets. Our implementation is available at https://github.com/HiLab-git/WCODE-PIA.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3517_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/HiLab-git/WCODE-PIA

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanLit_ReCoI2P_MICCAI2025,
        author = { Wang, Litingyu AND Ye, Ping AND Liao, Wenjun AND Zhang, Shichuan AND Zhang, Shaoting AND Wang, Guotai},
        title = { { ReCo-I2P: An Incomplete Supervised Lymph Node Segmentation Framework Based on Orthogonal Partial-instance Annotation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15972},
        month = {September},
        page = {507 -- 517}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors proposed a new partial annotation based framework for lymph node segmentation. The innovation is that the partial annotation is both partial objects and partial annotation planes.

  • 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.
    • 3D annotation is a difficult obstacle that prevents wide adoption of 3D segmentation networks. This work aims to address this important problems.
    • The ablation experiment is thorough.
  • 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 writing of the method section is incomprehensible. It only contains low level details, I could not understand why each component exists and how these components work together, also their necessity.
    • Most of the methods compared are from noisy annotation domain, not the incomplete annotation domain. So not sure how this method stands in comparison with general incomplete weakly supervised SoTAs.
    • Given this application, I believe the following work is more suited as a method to save annotation while preserving quality https://arxiv.org/pdf/1806.11137
    • From the ablation Fig 4., seems Inter-batch Prototypes is not necessary?
  • Please rate the clarity and organization of this paper

    Poor

  • 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.

    (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?
    • Paper is not well written, difficult to understand.
    • Comparison experiments missed key general incomplete annotation methods.
  • 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.

    To show inter-batch prototype is effective, there needs to be a trend (e.g. upward trajectory), a single jump wouldn’t be showing that. There’s no need to locate all lymph nodes in a single scan for bounding box labeling, it works in scanning window. It is also a much more efficient labeling scheme time-wise. So the authors’ rebuttal is not convincing enough for me to change the decision.



Review #2

  • Please describe the contribution of the paper

    The main contribution of the paper is the introduction of ReCo-I2P, a prototype-based label-efficient learning framework designed to segment lymph nodes from extremely sparse annotations. This framework includes several strategies such as prototype-based feature enhancement and a reliability-based co-teaching strategy, which together improve segmentation performance on mediastinal lymph node datasets.

  • 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) A novel formulation of using orthogonal partial-instance annotation (oPIA) for more efficient lymph node segmentation. An original approach in leveraging intra- and inter-batch prototypes for feature enhancement, which aggregates useful foreground information across batches. (2) Demonstration of clinical feasibility by showing improved performance over state-of-the-art methods on two publicly available lymph node datasets.

  • 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) Fig. 3 could benefit from clearer visual differentiation between false and true positive regions by using different colors. (2) The presentation of mathematical formulas could be better integrated into the text rather than being presented independently, enhancing readability.

  • 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

    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?

    The recommendation is based on the methodological innovation, the strong empirical results supporting the efficacy of the proposed framework, and the potential for broad applicability within the medical imaging community.

  • 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 did not respond to my feedback during the rebuttal phase. I strongly recommend that the program committee or area chair encourage the authors to implement the suggested improvements in the final version. This would ensure the paper meets a higher standard of clarity and scholarly presentation.



Review #3

  • Please describe the contribution of the paper

    To address the high cost and effort associated with fully annotating lymph nodes, the authors propose a novel semi-supervised segmentation framework that leverages orthogonal partial-instance annotations. The framework integrates two key components into a standard V-Net backbone: a Prototype-based Feature Enhancement module that refines feature representation, and a Reliability-based Co-Teaching strategy that improves learning from noisy or sparse labels. The proposed method is evaluated on two publicly available lymph node segmentation datasets and tested under two settings: (1) the conventional partial-instance annotation scenario and (2) the newly introduced orthogonal partial-instance annotation setting. Comprehensive experiments demonstrate that the method consistently outperforms seven state-of-the-art semi-supervised approaches in both settings, across both overlap-based and distance-based metrics. An ablation study further highlights the individual and combined effectiveness of the prototype and co-teaching modules.

  • 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 paper addresses a clinically relevant and practical challenge: reducing annotation costs in complex medical scenarios such as lymph node segmentation. The proposed orthogonal partial-instance annotation (oPIA) strategy presents a meaningful step toward more efficient labeling workflows, with potential applicability to other domains (e.g., Crohn’s disease).
    2. The study provides a thorough validation of the oPIA strategy in the context of lymph node segmentation, showing its effectiveness in real-world settings.
    3. The proposed semi-supervised framework demonstrates strong performance, outperforming several state-of-the-art methods across two public datasets and under both traditional partial-instance and the proposed oPIA annotation settings.
    4. An ablation study is included to evaluate the contribution of each component (prototype-based feature enhancement and co-teaching strategy), offering clear insights into the effectiveness of the framework’s design.
  • 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. The paper lacks a detailed explanation of how partial annotations—particularly orthogonal partial-instance annotations (oPIA)—are derived from the original datasets, especially in cases where full annotations are available. This raises concerns about reproducibility and clarity. The authors should provide a more comprehensive description of the annotation generation process for both fully and weakly annotated datasets.
    2. While the oPIA strategy is relevant and useful, it is limited in novelty. Similar orthogonal annotation strategies have been explored in prior work (“orthogonal annotation benefits barely-supervised medical image segmentation”), and the current paper mainly adapts the concept from a fully annotated target-level setup to an incompletely annotated one (labelled all lymph nodes vs labelled only a fraction of lymph nodes). This adaptation, while practical, should be positioned more clearly in terms of novelty.
    3. In the CT Lymph Node dataset experiments, the lower-bound model (presumably trained on a limited subset of annotations) achieves surprisingly strong performance—often outperforming several of the compared semi-supervised methods. This raises important questions about the practical effectiveness of the competing methods. A discussion or interpretation of this result would be valuable to clarify whether this reflects limitations in prior methods or dataset-specific characteristics.
  • 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 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
    • In “Comparison with SOTAs”: “LowerBound results of the LNQ2023” should probably be “UpperBound”
    • The instance-preserving analysis is a particularly strong aspect of the paper. The observation that oPIA achieves a Dice score of 57.30 compared to 58.26 for full partial-instance annotation (PIA) is compelling. It highlights the efficiency of the annotation strategy—labeling only two slices per lymph node still yields nearly equivalent performance. Moreover, even labeling only half of the lymph node instances achieves a respectable Dice score of 56.41. I recommend emphasizing these findings more prominently, as they strongly support the practicality and clinical relevance of your approach.
    • To further strengthen the real-world impact of your method, I suggest including a discussion—perhaps in the conclusion—on the potential time savings introduced by the oPIA strategy. If direct annotation time is difficult to quantify, an estimate based on the number of annotated voxels could serve as a reasonable proxy.
    • As a suggestion for future work, I believe it would be valuable to investigate the trade-off between labeling completeness within each instance (PIA) and maximizing the number of distinct instances (oPIA). Intuitively, a higher instance-preserving ratio may expose the model to more spatial and morphological variability, which could be more beneficial than fully annotating a smaller number of instances. An experiment analyzing this trade-off would provide useful insights for optimizing annotation strategies in other domains as well.
  • 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 paper addresses a highly relevant and impactful problem—reducing annotation costs for complex medical segmentation tasks such as lymph node identification. This has clear clinical value and practical significance. The proposed framework is well-motivated and demonstrates consistent performance gains over several state-of-the-art semi-supervised methods across two datasets and under multiple annotation scenarios.

    While there are some weaknesses, particularly related to the clarity and novelty of the annotation strategy and certain experimental results, these issues appear to be primarily related to presentation rather than methodological flaws.

  • 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.

    I did not change my scoring and still would like to see the paper being accepted.



Author Feedback

We thank all reviewers for their valuable feedback and approval of this task’s value. We are grateful to see the appreciation for our approach’s effectiveness, robust results, and broad applicability (R1, R3) and the comments on future work (R3). Below, we provide our responses to the questions. –Clarification of the key components and their motivation (R2) Our approach consists of L_seg (Eq. 7), PFE (Section 2.1), and ReCo (Section 2.2). L_seg is designed to mine useful information from oPIA. Then, PFE and ReCo cyclically purify the extracted information: PFE uses two types of prototypes to enhance features, which are decoded by the decoder to obtain a second prediction with more reliable foreground information. ReCo fuses two predictions based on confidence to obtain pseudo labels, then learns these pseudo labels based on reliability observations of the other prediction at different spatial positions, enabling the two predictions to learn complementary information from each other. This then improves prototype quality in PFE and ensures PFE performance. –Effectiveness of inter-batch prototype (R2) The intra-batch prototype contains foreground information within the current batch, while the inter-batch prototype aggregates foreground information across batches. Increasing the number of inter-batch prototypes ensures the diversity of retained foreground features. Fig.4 shows the hyperparameter adjustment for the number of inter-batch prototypes. The optimal number is 3, achieving the best performance. These highlight the importance of inter-batch prototypes. –Comparison methods (R2) For works of incomplete label, few methods can be transferred to oPIA due to the specific designs for their annotating strategies. For example, TIA (Ju, J et.al, A Weakly-supervised Multi-lesion Segmentation Framework Based on Target-level Incomplete Annotations, MICCAI 2024) relies on distinguishing slices containing foreground and background, which is not applicable to PIA and oPIA, where only slices with foreground can be distinguished. Moreover, annotating background slices is more costly than foreground slices to ensure that no foreground targets appear. Additionally, regarding the paper mentioned by R2 (arXiv:1806.11137), its second stage uses similar annotations as PIA, but its performance mostly benefits from having detection boxes for all instances in the first stage. Also, this labeling form is less effective, as a significant invisible labeling burden to locate all lymph node instances compared to oPIA/PIA. So, comparing with methods that don’t utilize the same or even equivalent annotating information and cost is unfair. It’s worth noting that we optimized noise learning methods to ensure their performance. For example, since oPIA’s noise only appears on background pixels, the ignored pixels in Co-teaching are selected only from the background. –Source code and the reproducibility of the work (R2, R3) We will release all used codes upon paper acceptance (R2, R3). For generating PIA (R3), lymph node instances were selected based on size for the CT Lymph Node dataset, with larger instances more likely to be retained. The middle slice of the axial and coronal planes was retained for each connected component for LNQ2023 and each instance for CT Lymph Node to get oPIAs. The processing code will be made public along with the source code. –Novelty of the oPIA (R3) oPIA innovates by focusing on targets across multiple regions, reducing locating requirements compared to orthogonal annotation (100% oPIA, single target region). –Performances of SOTAs under PIA on CT Lymph Node dataset (R3) This dataset has more small lymph node instances than the LNQ2023 test set, and our PIA processing method also results in more small instances than the LNQ2023 training set. This requires that the algorithm should not ignore small foreground areas.



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.

    N/A

  • 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