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

Abstract

Medical diagnosis using limited-angle computed tomography (LACT) is a beneficial approach for patients due to advantages such as faster scanning times and lower radiation doses. However, images reconstructed from LACT contain limited information, leading to significant artifacts and making an accurate diagnosis more challenging. Although various methods have been proposed to reconstruct LACT images into full-angle computed tomography (CT) images, they primarily focus on improving image quality and operate independently of lesion segmentation models, neglecting critical lesion-related information. In this paper, we propose TransSino, a transformer-based medical image segmentation model that operates in the sinogram domain of LACT. TransSino learns an auxiliary task to reconstruct the unmeasured regions in the sinogram domain for robust segmentation performance. Specifically, it analyzes the sequential nature of the sinogram using the transformer from language models and reconstructs features for the unmeasured regions by using prior sinogram patterns. Moreover, we introduce a contrastive abnormal feature loss to enhance the contrast between abnormal and normal feature regions. Experimental results confirm that TransSino outperforms existing medical segmentation methods on LACT images. The code is available at https://github.com/jhyoon964/TransSino.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0384_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{YooJae_TransSino_MICCAI2025,
        author = { Yoon, Jae Hyun and Lee, Yeong Jong and Yoo, Seok Bong},
        title = { { TransSino: Prior Sinogram Pattern-Based Transformer for Limited-Angle CT Image 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 = {594 -- 604}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents a novel segmentation method for Limited-Angle Computed Tomography (LACT) based on acquired sinogram data. It introduces a contrastive abnormal feature loss using paired CT data generated through an inpainted data augmentation strategy. Additionally, the method employs prior pattern embedding to compute the query features of a prior pattern Transformer.

  • 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 proposed method achieves state-of-the-art segmentation performance. The authors conduct a thorough analysis and evaluate all components of the network. The ablation study is comprehensive, and each component is well analyzed. The paper is clearly written and easy to follow. Furthermore, the problem is well formulated, making it accessible even to readers who may not be deeply familiar with the domain.

  • 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.

    A significant weakness is the absence of quantitative comparisons with other state-of-the-art methods for image reconstruction (e.g., mean PSNR, SSIM). While the authors claim state-of-the-art performance, no such metrics are presented. Since the method includes a reconstruction branch, comparisons with existing reconstruction methods are essential.

    Additionally, the lack of 3D segmentation and 3D reconstruction results such as those provided in DOLCE is a drawback. Including 3D evaluations would better demonstrate the effectiveness and generalizability of the proposed method.

  • 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.

  • 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?

    The method achieves strong segmentation performance and is supported by thorough analysis. However, the lack of quantitative comparison for image reconstruction limits the evaluation. The authors should provide reconstruction metrics such as PSNR and SSIM to substantiate their claims. Furthermore, 3D segmentation and reconstruction results would strengthen the paper. I would support acceptance if these points are addressed in the rebuttal.

  • 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 paper proposes a Transformer-based medical image segmentation model called TransSino, specifically designed for the segmentation of limited-angle CT (LACT) images. Operating in the sinogram domain, the model achieves robust segmentation performance by learning auxiliary tasks to reconstruct unmeasured regions within the sinogram domain.

  • 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.

    A Transformer-based segmentation model is proposed to address the time-dependent characteristics of sinograms in LACT. The model incorporates a prior pattern Transformer module, which uses prior sinogram patterns to reconstruct unmeasured projections. Additionally, a contrastive anomaly feature loss is introduced to enhance the contrast between anomalous and normal regions in the feature space.

  • 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 lack of validation on more diverse datasets. It should involve testing TransSino’s generalization ability across a wider range of datasets, especially with real-world LACT data.

    2. Provide a detailed discussion of TransSino’s computational resource requirements, such as training time and memory consumption, to better assess its feasibility in real-world applications.

    3. “However, existing models do not fully account for these sequential and sinusoidal characteristics,” whereas sequential and sinusoidal features have already been widely explored in CT reconstruction.

    4. The article frequently mentions language models; however, in reality, language models are not utilized in the approach. This inconsistency should be addressed, as the reference to language models creates an expectation that they play a role in the methodology, yet they are not actually incorporated.

  • 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

    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?

    The lack of validation on diverse datasets, especially real-world LACT data, and the absence of discussion on computational resource requirements. Additionally, the mention of language models without their actual use creates inconsistency in the methodology.

  • 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 #3

  • Please describe the contribution of the paper

    This paper proposes TransSino, a transformer-based medical image segmentation model that operates in the sinogram domain of Limited-Angle CT (LACT), leveraging the time/angle-dependent characteristics in the sinogram. It employs a prior pattern transformer module to inject prior knowledge of sinusoidal patterns. It also adopts auxiliary reconstruction branch to enhance its performance. Finally, it further proposes a contrastive abnormal feature loss for improved segmentation. Generally, this paper is novel and well-written, the performance is promising, and I felt that I learn something.

  • 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. This paper proposes TransSino, a transformer-based medical image segmentation model that operates in the sinogram domain of Limited-Angle CT (LACT). This is inspired by the time/angle-depedent sequential characteristics in the sinogram.
    2. It designs a prior pattern transformer module. It could capture the prior sinusoidal pattern cues of the sinogram to help the model reconstruct unmeasured projections.
    3. It takes reconstruction as an auxillary task to enhance its segmenation performance.
    4. It further adopts contrastive abnormal feature loss to maximizes the differences between normal and abnormal regions in the feature space for improved reconstruction. In conclusion, this paper is novel and achieves promising results, I like this paper.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    I still have some issues listed as follows.

    1. In the section of “Prior Pattern Transformer”, you mentioned about disentangled representation learning. I dont think this module is related to disentangled learning. It just incorporates the prior knowledge of sinogram into your framework to improve the generated missing sinogram in LACT. I think you’d better change the presentation.
    2. In your total loss of Eq.4, the weights to each term are all equal to 1. Is that appropriate? It would be better to verify the weights choice or provide some illustrations.
    3. In Fig.4, you evaluate the reconstruction performance within mask region. I think you’d better evaluate the reconstruction performance from whole image. The reconstrcution module is also very important to your framework.
    4. You only evaluate your model in simulated dataset. I think it would be better if you verify the model performance in real-world data. You could train on simulated dataset and test on real-world data.
    5. It would be better if you could provide visual results of your ablation study.
    6. There are some typos. (1) In the second paragraph of “Introduction” section, you use Figure 1(b), please change as Fig. 1(b) to be consistent with other hyperrefs. (2) In “Implementation Details” section, you mentioned “It sets N to 16”, however, in your “Ablation Study”, you verify N=32 is the best. Please check.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    Furthermore, I also watch your supplementary video, and the following comments should not affect the review results. There are too much AI generated pictures in your video, which I think is not good and I dont like it. And the majority is introducing the background rather than your method. I think you should put more attention on method illustration in your video. Hope you can improve it.

  • 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?

    This paper propose a novel method named TransSino, a transform-based medical image segmentation model that operates in the sinogram domain of LACT. It incorpoated several key ingredients for its performance, including prior sinogram transformer, auxiliary reconstruction, and contrastive abnormal feature loss. This paper is well-written and novel enough. I think this paper is a good job and would like to accept it.

  • 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



Author Feedback

We are deeply grateful for the reviewers’ positive assessment and constructive comments. (Quantitative comparison for reconstruction image) We already presented not only qualitative results but also quantitative evaluations of the average PSNR at each angular range in Figure 4. Moreover, we included both the metrics for the masked region (defined as R-PSNR) and the whole region of the image (defined as PSNR) in Figure 4. (Extensive evaluations) Due to the page limit of the main paper and constraints on the supplementary material, we selectively included only the most critical components of our analysis in the manuscript. We regret that we were unable to present extensive evaluations on more diverse datasets, visual results, or ablation studies on weight choice. (Evaluation for real LACT dataset) Although experiments on real LACT data are currently limited due to the lack of publicly available datasets, we will focus on such evaluations in future work, as noted in the Conclusion section. To this end, we aim to collaborate with a hospital to enable validation on real LACT data. (Language model) Although we do not directly utilize natural language in our framework, we adopted the transformer architecture inspired by language models such as Llama3 [6], treating the projection views in CT as analogous to word tokens. However, to avoid any confusion or expectation that our method incorporates natural language processing, we will revise the manuscript by clarifying this point. (Some typos) The usage of “Figure 1(b)” in the Introduction section was intentional, as we aimed to avoid starting a sentence with an abbreviation. However, in response to the reviewer’s comment for stylistic consistency, we will revise the sentence structure accordingly. Additionally, we appreciate the comment regarding the N, and we will update the manuscript to ensure consistency and clarity. Once again, we sincerely appreciate the reviewers’ valuable feedback and will refine the manuscript according to the insightful comments.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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



back to top