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

Abstract

Coronary artery calcification (CAC) is a powerful indicator of cardiovascular disease. Cardiac CT angiography (CCTA) has significant advantages in detecting CAC. However, since the image quality of CCTA can be compromised by cardiac motion or imaging equipment, and the contrast between CAC and surrounding tissue is low, accurate assessment of CAC remains a significant challenge. To address this issue, we propose a model (CAC-Net) for the comprehensive evaluation of CAC to fully exploit the characteristics of clinician annotations. First, inspired by the clinical annotation process, where doctors determine the subject based on boundaries, we propose a cross-frequency regulator module. This module models the interaction between high and low frequencies to distinguish the CAC body and its edges, thereby enhancing edge perception. Then, building on clinicians’ anatomical prior knowledge that CAC is confined within coronary arteries, we introduce a geometric prior module to encode their topological relationship, effectively reducing false positives. In experiments, our proposed method is compared with existing state-of-the-art methods on two CAC datasets. The results demonstrate that: (1) our method significantly improves CAC segmentation performance, as evidenced by a higher Dice score compared to U-Net (0.731 vs. 0.659); and (2) it ensures consistency in clinically relevant indicators, including calcium scores.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3950_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{JiaWei_Coronary_MICCAI2025,
        author = { Jiang, Weili and Li, Yiming and Luosang, Gadeng and Peng, Gang and Peng, Yong and Yao, Yijun and Yi, Zhang and Wang, Jianyong and Chen, Mao},
        title = { { Coronary Artery Calcification segmentation by using cross-frequency conditioner and geometric priors Learning } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},
        page = {100 -- 110}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The primary contribution of the manuscript is the development of CAC-Net, a novel deep learning model for coronary artery calcification (CAC) segmentation in coronary CT angiography (CCTA). CAC-Net integrates two key innovations:

    A Cross-Frequency Conditioner (CFC) module that models interactions between high- and low-frequency features to enhance CAC body and edge detection, mimicking clinicians’ boundary-focused annotation process. A Geometric Prior (GP) module that enforces topological constraints, ensuring CAC predictions align with coronary artery anatomy, reducing false positives. These components improve segmentation accuracy (e.g., over 10% Dice improvement) and clinical relevance (e.g., consistent Agatston scores), potentially streamlining automated CAC assessment and reducing reliance on non-contrast calcium scoring CT (CSCT) scans, thus lowering radiation exposure.

  • 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 CFC module is an original approach, decomposing CCTA features into high- and low-frequency components via 3D FFT, then using frequency-aware attention to prioritize CAC edges and body. The feature alignment step with deformable convolutions ensures spatial coherence, addressing contrast agent interference and noise. This mimics clinical annotation, where radiologists focus on boundaries to delineate CAC. By explicitly modeling frequency interactions, it tackles CCTA’s low contrast and motion artifacts, a persistent challenge not fully addressed by prior methods like U-Net or Transformers. The GP module encodes the anatomical rule that CAC occurs within coronary arteries, using binary masks and a geometric-prior loss (harmonic mean of precision and sensitivity) to reduce false positives. Unlike generic segmentation models (e.g., refs 6, 12), this explicitly incorporates domain knowledge, aligning predictions with vascular topology. This is compelling for clinical applicability, as false positives in CAC segmentation can mislead diagnosis. The model predicts Agatston scores, a key clinical metric for CVD risk, with Bland-Altman analysis showing reasonable agreement (e.g., +3.05 deviation on CAC-CTA, -5.16 on Orcascore). The potential to eliminate CSCT scans (noted in Conclusion) reduces radiation, a practical benefit for patients. The evaluation compares CAC-Net to state-of-the-art (SOTA) methods (nnU-Net, nnFormer, CACer) on two datasets, reporting sensitivity (SE), precision (PE), and Dice scores. Table 1 shows superior Dice (e.g., >10% improvement), and cross-validation on Orcascore confirms generalization. The ablation study (Table 2) quantifies CFC and GP contributions, reinforcing their value. Using coronary artery masks to define geometric prior accuracy (Tprec) and sensitivity (Tsens) is a creative way to leverage anatomical constraints, guiding the loss function to prioritize topologically valid CAC 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.

    The Conclusion introduces “InstanceViT” and “instance-aware guided semantic learning in the Fourier domain,” which are not mentioned in the Abstract, Introduction, Methods, or Experiments. These appear unrelated to CFC or GP modules, creating confusion about the model’s actual components. The 3D U-Net encoder-decoder backbone is standard, used in prior CAC segmentation works like Wolterink et al. (2016, ref 1) and Zhang et al. (2020, ref 27). While CFC and GP are novel, the reliance on U-Net reduces the overall architectural originality. The evaluation focuses on SE, PE, Dice, and Agatston scores but omits other clinical metrics like volume overlap error or false positive rate per patient, which are relevant for CAC (see Santini et al., 2021, Medical Image Analysis). Figure 4 claims better fine CAC preservation, but qualitative evidence is limited to one case. Table 1 reports higher Dice scores, but no statistical tests (e.g., t-tests, Wilcoxon) confirm significance over SOTA (nnU-Net, nnFormer). Prior CAC studies (e.g., Lessmann et al., 2017, IEEE TMI) use such tests to validate improvements. While cross-validation on Orcascore shows generalization, the manuscript doesn’t analyze dataset differences (e.g., scanner types, patient demographics) or why CAC-CTA scores deviate (+3.05) vs. Orcascore (-5.16). Prior work (e.g., van Velzen et al., 2020, Radiology: AI) emphasizes robustness across imaging conditions.

  • 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

    Discuss how Dice improvements or Agatston deviations affect diagnosis (e.g., misclassifying CVD risk), as MICCAI values medical impact.

  • 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?

    The manuscript has notable strengths: the CFC and GP modules are novel, creatively mimicking clinical annotation and anatomical constraints, and the evaluation shows promising Dice improvements and clinical feasibility via Agatston scores.

    Conclusion Mismatch: The unexplained “InstanceViT” reference suggests a major error or oversight, eroding trust in the manuscript’s integrity, a dealbreaker for MICCAI reviewers. Limited Novelty in Backbone: Relying on 3D U-Net, already common in CAC segmentation (refs 1, 27), reduces architectural innovation compared to Transformer trends (e.g., Chen et al., 2021). Incomplete Evaluation: Missing statistical tests and narrow metrics (no volume errors, limited visuals) weaken SOTA claims, especially against rigorous baselines like nnFormer. Clarity Issues: Dense Methods and shallow Results discussion, coupled with the Conclusion’s confusion, hinder accessibility for MICCAI’s broad audience. Generalization Gaps: Limited analysis of dataset variability (e.g., scanner differences) questions robustness, critical for clinical deployment.

  • 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.

    Authors correctly defend using U-Net as a clinically validated, widely accepted baseline, and acknowledge that their novelty lies in domain-informed module design (CFC, GP), not the backbone.

    They also commit to exploring alternative backbones in a future version

    Authors agree to include additional clinical metrics (e.g., volume overlap error, false positives).

    They acknowledge the importance of statistical tests and will explore them further (with references).

    They explain differences in datasets (e.g., vendor consistency, demographics) and their effects on results

    Several figures and equations will be updated for clarity (e.g., frequency masks, deformable conv, post-processing).

    Misinterpretations (e.g., “>10% Dice improvement”) will be clarified



Review #2

  • Please describe the contribution of the paper

    we propose a model (CAC-Net) for the comprehensive evaluation of CAC to fully exploit the characteristics of clinician annotations

  • 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) propose a cross-frequency regulator module to model the interaction between high and low frequencies to distinguish the CAC body and its edges, thereby enhancing edge perception. 2) introduce a geometric prior module to encode their topological relationship to reduce false positives.

  • 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 presentation of the paper needs much improvement. There are some gaps exist which makes understanding the method difficult.

  • 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 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

    1) Eq (1) needs some explanation please. 2) Eq. (4), What is deformable convolution is not clearly defined and relevant references are not cited. 3) What does the number (in line 2 of Section 3.1, page 6) in the bracket mean? 4) There are some gaps in the presentation of the paper. The illustration of the methods showed in Fig. 2 and Fig. 3 need to provide some detailed description, e.g., the masks used in Eq. (5)(6) should be explained. 5) Why sensitivity metrics presented in Table 1 and Table 2 do not change for pre- and post-process options. 6) In conclusion, “leveraging InstanceViT….” Which is first-time appears, please provide reference.

  • 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 proposed method appears to be interesting and ablation study has shown the effectiveness of the proposed losses. Clarification of the paper needs to be improved.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 deep learning model for automated coronary calcium scoring in coronary CT angiography (CCTA). The model is based on an encoder-decoder architecture and introduces a cross-frequency regulator module that processes high- and low-frequency image components in parallel branches, explicitly modeling their interactions. Additionally, a geometric prior module is incorporated to encode topological relationships between anatomical structures, aiming to improve the spatial coherence and accuracy of the calcium scoring predictions.

  • 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 cross-frequency conditioner module is an interesting and well-motivated component, as it explicitly models the interaction between high- and low-frequency features. This design is particularly appropriate for coronary calcium scoring, where calcifications appear as small, high-frequency regions embedded within lower-frequency anatomical context.
    • The geometric-prior loss is also a valuable contribution, as it incorporates spatial and topological relationships between coronary arteries and calcifications. This prior knowledge can enhance anatomical consistency and potentially reduce false positives in calcium detection.
  • 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.
    • In clinical practice, coronary artery calcium scoring is typically evaluated based on risk categories derived from the Agatston score, with inter-method agreement quantified using metrics such as Cohen’s kappa. To facilitate clinical interpretability and validation, the authors should report how well their model performs across clinically relevant Agatston risk strata, especially in distinguishing patients with zero calcium (Agatston = 0), which is critical for ruling out disease. It would also be helpful to assess the rate of small false positive detections in this subgroup.
    • The evaluation includes a comparison between model performance with and without post-processing; however, the manuscript does not provide a clear explanation of what the post-processing entails. Details about the specific operations used (e.g., connected component filtering, morphological operations, anatomical constraints) should be included in the Methods section to ensure reproducibility and interpretability of the reported results.
    • In the abstract, the authors claim that their method achieves more than a 10% improvement in Dice score. However, this reported improvement is not reflected in the quantitative results presented in Table 1.
    • There appears to be an error in the Conclusion section, where the authors state: “We propose a novel CAC segmentation model, leveraging InstanceViT to assess token connectivity and learn instance-specific attention patterns.” This statement seems to have been copied from a different paper—specifically IARCaC: Instance-aware Representation for Coronary Artery Calcification Segmentation (MICCAI). The current manuscript does not introduce or use an InstanceViT model, and this sentence should either be removed or replaced with a statement accurately reflecting the methods described in the paper.
  • 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.

    (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 cross-frequency conditioner module and the geometric-prior loss represent novel and well-motivated contributions. These components offer valuable insights and are worth further discussion during the conference
  • 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.

    Despite its limitations in evaluation, such as CVD risk categorization, I find the cross-frequency conditioner to be a novel contribution that warrants further discussion at the conference.



Author Feedback

We thank all reviewers for their valuable and insightful reviews. They described our method as “notable strengths” (R1), “well-motivated and valuable contribution” (R3), and “novel and interesting” (R1&2&3). Here we address their main concerns:

  • Clarification on “InstanceViT” in Conclusion (R1&2&3) We thank the reviewers for pointing this out. The mention of InstanceViT was a citation/editing mistake. We will remove the sentence and avoid confusion.
  • Response Summary (R1) 1) Our main contribution lies in leveraging clinical insights—specifically, clinicians’ annotation habits and anatomical priors—to design the CFC and GP modules, rather than relying on the backbone itself. We thank R1&2&3 for recognizing the novelty of these modules. U-Net remains a strong baseline for medical image segmentation, as shown in nnU-Net [Ref 13], and is widely adopted in CAC segmentation (e.g., [Ref 1, 27]). Recent Transformer-based methods (e.g., nnFormer [Ref 17], U-Transformer [Ref 26]) also build upon U-Net. We agree that the backbone choice may influence performance and will explore alternatives in an extended version; 2) A higher Dice indicates better agreement between predicted and ground-truth, which typically correlates with lower volume overlap error and reduced false positive rate per patient. We will slightly revise Tab.1 to include these clinical metrics, following [a], to improve quantification completeness. Additionally, Fig. 4 will be slighted updated with more representative cases to enhance qualitative support; 3) We agree with R1 on the importance of statistical analysis (e.g., t-tests, Wilcoxon)) to validate the significance of improvements over SOTA methods. This will be further explored in an extended version, following [b]; 4) We will make minor revisions to analyze dataset differences. CAC-CTA was acquired using uniform Siemens dual-source CT with consistent protocols and known demographics (mean age 64.45 ± 11.44), while Orcascore spans four vendors and lacks demographic data. These differences likely explain the Agatston gap (+3.05 vs. −5.16), consistent with van [c] on protocol sensitivity.
  • Response Summary(R2) 1) Eq. (1) defines the frequency-domain masks for low-frequency (El ​) and high-frequency (Eh​) components. We retain only the central 50% of the spectrum for El ​, and assign the remaining outer frequencies to Eh​​, based on the 3D FFT of the encoded feature E. 2) Deformable convolution in Eq. (4) will be explicitly defined with a citation [d]. 3)The value “(5.36 ± 4.92 per patient)” represents the mean and standard deviation of CAC instances per patient in the CAC-CTA dataset. 4) We will slightly revise Fig. 2 and 3 to align with Eqs. (5) and (6), adding clearer annotations for the masks and variables to improve clarity. 5) The post-processing step applies an element-wise multiplication between the predicted CAC and the coronary artery to suppress false positives outside vessels. Since true positives within arteries remain unchanged, sensitivity (TP / [TP + FN]) is unaffected in Tab.1 and 2.
  • Response Summary(R3) 1) We agree with R3 on the importance of Agatston-based risk stratification in clinical CAC scoring. In our study, the public Orcascore dataset provides Agatston annotations from both CSCT and CTA. In the extended version, we will compare CAC predictions from CTA with CSCT-derived Agatston scores to assess risk category agreement, including Agatston = 0 cases, and report Cohen’s kappa. 2) The post-processing step applies a coronary artery mask as an anatomical constraint via element-wise multiplication, suppressing false positives outside the vessel regions. 3) The “>10% Dice improvement” refers to the maximum gain over baselines like Unet (0.731 vs. 0.659) and Vnet (0.731 vs. 0.651). We will be slightly edited to clarify. [a] Santini et al., 2021, Medical Image Analysis [b] Lessmann et al., 2017, IEEE TMI [c] van Velzen et al., 2020, Radiology: AI [d] Dai et al., ICCV 2017



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.

    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’

    N/A



back to top