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

Abstract

In the radiation therapy of nasopharyngeal carcinoma (NPC), clinicians typically delineate the gross tumor volume (GTV) using non-contrast planning computed tomography to ensure accurate radiation dose delivery. However, the low contrast between tumors and adjacent normal tissues requires radiation oncologists to delineate the tumors with additional reference from MRI images manually. % In this study, we propose a novel approach to directly segment NPC gross tumors on non-contrast planning CT images, circumventing potential registration errors when aligning MRI or MRI-derived tumor masks to planning CT. To address the low contrast issues between tumors and adjacent normal structures in planning CT, we introduce a 3D Semantic Asymmetry Tumor Segmentation (SATS) method. Specifically, we posit that a healthy nasopharyngeal region is characteristically bilaterally symmetric, whereas the presence of nasopharyngeal carcinoma disrupts this symmetry. Then, we propose a Siamese contrastive learning segmentation framework that minimizes the voxel-wise distance between original and flipped areas without tumor and encourages a larger distance between original and flipped areas with tumor. Thus, our approach enhances the sensitivity of deep features to semantic asymmetries. % Extensive experiments demonstrate that the proposed SATS achieves the leading NPC GTV segmentation performance in both internal and external testing.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1607_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiZi_Leveraging_MICCAI2025,
        author = { Li, Zi and Chen, Ying and Chen, Zeli and Su, Yanzhou and Ma, Tai and Mok, Tony C. W. and Zhou, Yan-Jie and Bai, Yunhao and Zheng, Zhilin and Lu, Le and Wang, Yirui and Ge, Jia and Yan, Senxiang and Ye, Xianghua and Jin, Dakai},
        title = { { Leveraging Semantic Asymmetry for Accurate Gross Tumor Volume Segmentation of Nasopharyngeal Carcinoma in Planning CT } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {291 -- 301}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents a novel method for tumor detection by leveraging anatomical symmetry breaking, achieving accurate and robust gross tumor volume (GTV) segmentation of nasopharyngeal carcinoma on non-contrast planning CT. This work provides a more efficient and reliable technological pathway for automated target area delineation in radiotherapy.

  • 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 medical significance is clear and significant for early diagnosis of nasopharyngeal carcinoma.
    2. Introduced a tumor segmentation strategy based on symmetric anomaly detection
    3. Classical multi-task metric learning loss was introduced based on symmetric anomaly regions to further improve the performance ,which is technically sound.
  • 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 segmentation results of the paper need to be set up in such a way that the pCT used has strict symmetry conditions. The authors’ implementation uses the mapping method in [29]. This makes the accuracy of the proposed method completely dependent on the previous method, reducing its technical soundness. Meanwhile, the “supervised ROI-selection method” mentioned in the paper, which is very important for the symmetric condition, is presented very vague in this paper.

    2. The use of symmetry for anomaly detection as well as metric learning loss is not a new technique, and the authors’ contribution is more to apply it to the scenario of early nasopharyngeal cancer detection

    3. Lack of one comprasion result in current experiment setting : External_train –> External_test. That is, the original data division of SegRap2023.

  • 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

    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.

    (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?

    First, the segmentation results presented in the paper rely on enforcing strict symmetry conditions in the pCT data. However, this is achieved by adopting the mapping approach from [29], which makes the accuracy of the proposed method heavily dependent on prior work. Such reliance substantially limits the technical novelty and undermines the robustness of the approach. Moreover, the “supervised ROI-selection method,” which plays a critical role in maintaining the symmetry condition, is described in a vague and insufficiently detailed manner.

    Second, the use of symmetry for both anomaly detection and metric learning loss is not a novel technique. The main contribution of this work appears to be the application of these existing methods to the specific scenario of early nasopharyngeal cancer detection, rather than proposing fundamentally new methodologies.

    I would be willing to raise my score if the authors can adequately address these concerns.

  • Reviewer confidence

    Very confident (4)

  • [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 author addressed my concerns



Review #2

  • Please describe the contribution of the paper

    This paper introduces a novel segmentation method designed to delineate the gross tumor volume (GTV) of nasopharyngeal carcinoma (NPC) directly from non-contrast planning CT (pCT). It leverages the anatomical principle that healthy nasopharyngeal regions exhibit bilateral symmetry, which is disrupted by tumors.

  • 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 concept of leveraging semantic asymmetry to enhance tumor segmentation seems to be both innovative and clinically grounded, particularly for NPC, where asymmetry is a strong indicator of pathology. The use of a Siamese encoder-decoder network combined with a projection head and a voxel-wise contrastive margin loss appears to be an original architectural contribution that enables the network to learn discriminative features specific to asymmetric regions.

  • 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.

    While the authors compare against general-purpose segmentation models (e.g., nnUNet, SwinUNETR), they do not directly compare with prior published works that specifically address NPC segmentation, including the results from SegRap 2023 https://segrap2023.grand-challenge.org/final-ranking/ There are also a number of papers based on the MICCAI 2019 StructSeg challenge (50 NPC patient GTVs) and this might have also been a good comparison? One motivation for the work mentioned is to avoid registration errors from MRI, but perhaps generating synthetic CT from MRI might also be a solution to the registration issues mentioned in the introduction, but could introduce additional uncertainties.

  • 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

    The paper is clearly written, well-structured, and logically organized. The motivation is very good. There are a few typos ( “Alos”, “no-contrast”, )

  • 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?

    I think the main strengths lie in the novel use of semantic asymmetry, effective architectural design, and strong experimental validation. I’m a bit concerned about the lack of comparison of results with other work.

  • 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.

    Increased score to accept post-rebuttal.



Review #3

  • Please describe the contribution of the paper

    The paper proposes a 3D Semantic Asymmetry Tumor Segmentation (SATS) method for accurate gross tumor volume (GTV) segmentation of nasopharyngeal carcinoma (NPC) in non-contrast planning CT (pCT) scans. The main contributions are: (1) To the best of the authors’ knowledge, this is the first work to directly segment NPC tumors from non-contrast CT scans by leveraging the inherent bilateral symmetry of human anatomy. (2) A Siamese contrastive learning framework is introduced, comprising a shared-weight encoder-decoder that jointly processes original and flipped CT images to learn symmetric features. A non-linear projection head and a distance metric module are designed to enhance sensitivity to semantic asymmetries.

  • 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 authors leverage the clinical observation that tumors disrupt the natural bilateral symmetry of anatomical structures, introducing “semantic asymmetry” as a key feature for segmentation. This formulation offers a novel perspective to address the low-contrast challenge between tumors and surrounding tissues in NPC on CT scans. The proposed SATS model combines a Siamese network with contrastive loss and introduces an asymmetrical region selection mechanism, guiding the model to focus on areas with the greatest discrepancy between original and flipped images. By incorporating a projection head and voxel-level margin-based contrastive loss, the model effectively amplifies the feature differences between abnormal and symmetric regions. Extensive experiments on both internal and external datasets demonstrate strong generalization capability, validating the robustness and effectiveness of the proposed approach.

  • 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 core idea of the paper— detecting abnormalities using a symmetry-based network—has prior precedents in medical imaging [1,2]. Thus, from a methodological perspective, the work represents an extension of existing ideas rather than a fundamentally novel innovation. The adaptability of the method to certain challenging scenarios, such as bilateral tumors or multiple lesions, is not discussed. The absence of failure case analysis limits the assessment of the method’s robustness. The paper lacks a deeper discussion. For example, while prior methods for NPC segmentation on MRI and CT are cited, their limitations are merely stated without detailed analysis. Additionally, although quantitative results are extensive, the paper provides limited insight into why the proposed SATS model outperforms others such as nnUNet. More thorough discussions, such as the model’s behavior across tumors of different sizes or locations and feature visualization, would enhance the understanding of its mechanism. Finally, certain sections, such as Asymmetrical Abnormal Region Selection, are not described in sufficient detail.

    [1] Huang J, Li H, Li G, et al. Attentive symmetric autoencoder for brain MRI segmentation[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2022: 203-213. [2] Zeng B, Wang H, Xu J, et al. Two-stage structure-focused contrastive learning for automatic identification and localization of complex pelvic fractures[J]. IEEE Transactions on Medical Imaging, 2023, 42(9): 2751-2762.

  • 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

    Providing a more complete description or releasing code would help ensure reproducibility.

  • 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?

    The paper addresses an important clinical problem by proposing a practical and effective solution for NPC GTV segmentation directly on non-contrast CT scans. Although the core methodological idea builds on existing symmetry-based approaches, the adaptation to NPC segmentation is novel, and the work demonstrates clear clinical relevance. The proposed SATS method is well-designed, incorporating a Siamese network with contrastive learning and asymmetrical region selection, and achieves significant performance gains over strong baselines. The experimental validation is thorough, including internal and external datasets, statistical significance testing, and ablation studies, which strengthen the credibility of the results. Despite some weaknesses—such as limited discussion of failure cases and related work—the strengths outweigh the shortcomings. Given the practical value, strong empirical performance, and reasonable level of innovation, I recommend accepting this paper.

  • Reviewer confidence

    Very confident (4)

  • [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.

    This paper addresses a clinically meaningful problem and proposes a symmetry-informed segmentation method with practical value. The experimental design is relatively comprehensive, including both internal and external validations, and demonstrates performance improvements over several strong baselines. However, compared to existing work, the method primarily adapts known symmetry modeling approaches to a medical context, without introducing clear algorithmic innovation. In addition, the analysis of failure cases and the explanation of performance gains on small lesions remain somewhat vague. I suggest that the final decision take into account the perspectives of the other reviewers when evaluating the paper’s novelty and overall merit.



Author Feedback

We thank all reviewers for their thoughtful comments and categorize major concerns (C), followed by responses (R).

Reviewer 2,3 C#1: Methodological contribution compared with refs [1,2] mentioned by Reviewer R2. R: While [1,2] consider symmetric properties, how we utilize the symmetric features is significantly different from theirs. [1] utilizes the symmetry of brain structures and introduces an Autoencoder with Symmetric Position Encoding, without explicit constraints on symmetric or asymmetric regions (e.g., via custom losses or modules). [2] tackles the identification and localization of pelvic fractures by leveraging the symmetry of healthy pelvic regions. It focuses on symmetric structures. On the contrary, our method emphasizes the contrast between asymmetric areas. Furthermore, [2] relies on point cloud extraction and reconstruction, while ours provides a general solution for voxel-wise segmentation tasks. C#2: More details of Asymmetrical Abnormal Region (AAR) selection. R: We focus on asymmetrical lesion areas relative to the central sagittal axis, i.e., region B of Fig.1(b). To this end, we perform: 1) head-neck position normalization (bilateral symmetry along the central sagittal axis) of the overall head-neck region by utilizing rigid registration (rotation and translation). 2) The AAR is obtained by subtracting symmetrical lesion regions from the original mask. We would enhance Fig.1.

Reviewer 1 C#3: Lack of comparison methods from SegRap 2023. R: Astaraki, M., et al. (2023) achieved 1st in Task2-GTVs of SegRap2023 using a standard 3D U-Net, reporting a 0.790 Dice score (DSC), whereas our method reaches a 0.816 DSC under the same dataset. C#4: Synthetic CT from MRI. R: Although it is possible to generate CT from MRI using GAN, the synthetic CT may not be accurate enough for downstream tasks, which brings uncertainties, as Reviewer 1 and Wang, Chun-Chieh, et al. (2022) pointed out. We may explore this in the future. C#5: Typos. R: We will revise them.

Reviewer 2 C#6: Adaptability to bilateral or multiple lesions, and failure case analysis. R: (1) NPC does not have multiple lesions by its nature. (2) Even if NPC lesion exhibits exact bilateral conditions, the conventional segmentation loss in our method can identify the lesions (bilateral lesions are usually large [16], which should be easier to identify). Our method fails on extreme outliers, missegmenting symmetric lesions as asymmetric shapes and yielding reduced DSC compared to baseline nnUNet. Integrating ours with nnUNet could form a complementary framework to improve clinical utility. C#7: More discussion of other work (MRI, CT). R: Previous MRI, CT-based methods have explored different architectures or semi-supervised learning, etc. We will add more discussions on their limitations. C#8: Insights on why SATS outperforms others and more detailed results analysis. R: We believe our symmetric learning scheme makes the model more sensitive to identify abnormalities in the nasopharyngeal region than a single conventional segmentation loss (nnUNet). As suggested, we observe that our improvement is more prominent for small lesions as compared to nnUNet. C#9: Reproducibility. R: The code will be made publicly available upon acceptance.

Reviewer 3 C#10: Ext_train->Ext_test results. R: Please refer to R to C#3. We chose the In_train->Ext_test setting to demonstrate the ability of our method against inter-domain variation, which is more difficult and insightful than Ext_train->Ext_test. C#11: Dependent on [29]. R: We employ rigid registration [29] to rotate and translate CT images, symmetrically aligning the head and neck region along the central sagittal axis. Similar to [2] (Reviewer 2), this process globally aligns images with their horizontal flips via rotation and translation. It serves as a preprocessing or intermediate step and is robust—no significant performance difference was observed when tested with an alternative registration method (e.g., ICP).



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”.

    The reviewers have raised concerns regarding methodological novelty and rationale for combining MRI and CT. Please address all reviewers’ comments carefully.

  • 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.

    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’

    The paper tackles gross tumor volume segmentation of nasopharyngeal carcinoma on non‑contrast CT, a clinically meaningful problem given the low tumor‑to‑tissue contrast in these CT. Validation is thorough: the method is tested also on an external dataset from a public challenge and outperforms strong baselines such as nnU‑Net, MedNeXt, and SwinUNETR. The rebuttal is well articulated, and all reviewers have recommended acceptance.



back to top