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

Abstract

Rheumatoid Arthritis (RA) markedly increases the risk of cardiovascular disease (CVD), with coronary artery calcification (CAC) serving as a pivotal biomarker for early risk stratification. RA patients exhibit significantly elevated CAC-related mortality rates, making opportunistic CAC screening on routine chest CT scans highly valuable in clinical practice. However, to the best of our knowledge, automated CAC detection has not been systematically investigated in the RA cohorts, where calcifications often present as minute and low-contrast. To fill this critical gap, we construct the first high-quality CAC dataset for RA patients, comprising 415 cases with professional annotations. We propose Risk-aware Anatomical Prior Optimization (RAPO), a reinforcement learning framework that addresses the challenges of detecting minute and low-contrast lesions. RAPO incorporates a Chan-Vese Anatomical Prior that enforces geometric compactness to improve localization precision, and an Intrinsic Risk Penalty that modulates predictions based on uncertainty to reduce false negatives. Extensive experiments demonstrate that RAPO achieves superior performance, substantially outperforming state-of-the-art methods. The resources will be available at https://github.com/Jiashu-Xu/RAPO.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1391_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Jiashu-Xu/RAPO

Link to the Dataset(s)

N/A

BibTex

@InProceedings{XuJia_RAPO_MICCAI2026,
        author = { Xu, Jiashu AND Zhou, Haipeng AND Ma, Jianda AND Zou, Yaowei AND Dai, Lie AND Zhu, Lei},
        title = { { RAPO: Risk-Aware Anatomical Prior Optimization via Reinforcement Learning for CAC Detection in Rheumatoid Arthritis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes RAPO, a reinforcement-learning framework for coronary artery calcification (CAC) detection in rheumatoid arthritis (RA) patients from routine non-contrast chest CT. The authors claim to construct the first RA-specific CAC detection dataset with 415 patient cases and professional bounding-box annotations. The method is built on Qwen2.5-VL-7B and uses a two-stage training curriculum: Direct Preference Optimization (DPO) for initializing CAC discrimination, followed by risk-aware Group Relative Policy Optimization (GRPO) for localization refinement. The GRPO stage combines three reward components: an IoU-based localization reward, a Chan-Vese anatomical prior reward, and an intrinsic-confidence-based asymmetric risk penalty intended to penalize high-confidence false negatives more strongly. The authors compare RAPO with conventional detectors, general MLLMs, medical MLLMs, and task-specific grounding/RL models. On the private RA-CAC test set, RAPO reports mIoU=0.6042, ACC=0.7660, and F1=0.7311, outperforming the listed baselines. The work is clinically motivated, but the current submission leaves several essential aspects under-specified, especially the 3D CT-to-2D image formulation, the construction of DPO preference pairs, the interpretation of the risk-aware reward, and the screening-oriented evaluation protocol

  • 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 paper targets a clinically relevant problem: opportunistic CAC screening in RA patients using routine non-contrast chest CT. RA patients are a high-risk cardiovascular population, so improving CAC detection in this group could be clinically meaningful.
    • The paper attempts to study a disease-specific RA-CAC cohort rather than only using generic natural-image or general medical VLM benchmarks. If carefully validated, such a dataset could be useful for studying cardiovascular risk screening in autoimmune populations.
    • The framework is organized around a coarse-to-fine training idea: DPO-based discrimination initialization followed by GRPO-based localization optimization. This is a reasonable high-level strategy for stabilizing RL training
    • The method includes an anatomical prior intended to suppress anatomically implausible predictions and confusion with structures such as ribs or non-coronary calcifications.
    • The paper provides ablation experiments for DPO, GRPO, the anatomical reward, and the risk-aware reward, which helps isolate the claimed contributions at least at a preliminary level.
    • The comparison covers several categories of methods, including conventional detectors, general MLLMs, medical MLLMs, and task-specific grounding/RL methods.
  • 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 task formulation is not sufficiently clear. The paper discusses routine chest CT, which is naturally a 3D volume, but the method and figures suggest 2D slice-level bounding-box detection. The paper does not clearly state whether the input is a full CT volume, selected 2D slices, maximum-intensity projections, cropped cardiac regions, or another representation. It also does not explain how predictions across slices are merged into a scan-level or patient-level CAC result. This is a major validity and reproducibility issue.
    • The two-stage design appears to be a training curriculum rather than a clearly defined clinical pipeline. The paper says DPO is used to learn discrimination, but it does not show that inference first performs slice/scan classification and then localization. If the first stage is only a DPO initialization step, the paper should not imply that it is a clinically interpretable two-stage screening workflow.
    • The first DPO stage is seriously under-specified. DPO requires preference pairs, but the paper does not explain how chosen and rejected responses are constructed, whether they are based on presence/absence labels, bounding-box quality, Positive/Negative/Review decisions, human preference, automatic rules, or model sampling. It is also unclear whether DPO is performed before or after SFT and whether bounding boxes are included in the DPO responses.
    • The paper should clarify that the RL framework is not annotation-free. The IoU-based localization reward explicitly depends on radiologist-annotated bounding boxes, and the risk-aware reward also requires knowing whether a prediction is false negative, false positive, or correct. Therefore, the method is better described as reward-based supervised optimization rather than unsupervised learning. Only the Chan-Vese anatomical prior is relatively closer to an unsupervised image-derived constraint.
    • The RA-specific imaging motivation is not convincingly established. RA is well motivated as a high-risk clinical population, but the paper does not provide evidence that RA itself causes CAC lesions to be smaller, lower-contrast, or less well defined on CT compared with non-RA CAC. Many of the stated challenges, such as small lesions, partial volume effects, motion artifacts, limited spatial resolution, and confusion with ribs or vascular calcification, are general challenges of early CAC and non-gated chest CT rather than clearly RA-specific properties.
    • The risk-aware penalty is clinically intuitive but methodologically heuristic and only weakly novel. Penalizing false negatives more than false positives is a standard screening consideration already reflected by sensitivity/recall-oriented evaluation. The paper does not sufficiently justify why this reward formulation provides a substantive methodological advance beyond cost-sensitive training or recall-oriented thresholding.
    • The evaluation protocol is insufficient for clinical CAC screening. The paper reports mIoU, accuracy, F1, and recall in the ablation table, but does not report lesion-level sensitivity, false positives per scan, FROC curves, AP/mAP, precision-recall curves, scan-level sensitivity/specificity, or patient-level CAC detection performance. Accuracy is hard to interpret without class balance and decision thresholds.
    • The ablation results do not fully support the claim that the full method best reduces false negatives. In Table 2, the +Risk variant has higher recall than the full RAPO model (0.7100 vs. 0.6995), while the full model improves mIoU and ACC. This suggests a trade-off between localization, accuracy, and recall, but the paper does not discuss this trade-off or report false-negative counts.
    • The dataset is small, private, and single-center. The 415 cases with a 7:1:2 split leave roughly 83 test cases. The paper does not provide external validation, confidence intervals, statistical testing, lesion-count distribution, CAC severity distribution, scanner/slice-thickness/reconstruction details, or inter-reader variability. Therefore, the claims of state-of-the-art performance and clinical effectiveness are not strongly supported.
  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    The paper states that the proprietary dataset was collected from an anonymous hospital with ethical approval, which is appropriate in principle. However, it does not provide enough information about the ethics approval identifier, consent or consent waiver, de-identification procedure, data governance, or whether the dataset includes sensitive patient metadata. Because the method targets clinical screening, the paper should also discuss safety risks, especially false negatives, model calibration, scanner/protocol bias, subgroup performance, and the intended role of human review.

  • 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/2026/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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I recommend weak rejection. The paper addresses an important clinical problem and proposes a potentially interesting risk-aware RL framework, but the current submission does not provide enough methodological clarity or experimental evidence for a MICCAI-level acceptance. The most serious issue is the unclear task formulation: CAC detection is described on routine chest CT, but the method appears to operate on 2D images/slices without explaining how full CT volumes, multiple slices, multiple lesions, and scan-level decisions are handled. The DPO stage is also under-specified, and the risk-aware reward is heuristic, only weakly justified, and not supported by sufficient calibration or clinical-error analysis. The RA-specific clinical motivation is plausible, but the paper does not convincingly show that RA-CAC has unique imaging difficulty compared with general CAC on non-gated chest CT. The reported improvements are promising, but they are based on a small private single-center dataset without external validation, confidence intervals, statistical testing, sufficient screening-oriented metrics, or fully transparent baseline protocols. Therefore, the claims of state-of-the-art performance and clinical effectiveness are not yet sufficiently supported.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    After carefully considering the actual technical content of the paper, I recommend rejection. The paper addresses an important clinical problem and proposes an interesting Qwen2.5-VL-based framework for CAC detection. The method combines DPO initialization with risk-aware GRPO, and the GRPO stage integrates an IoU localization reward, a Chan-Vese-based anatomical prior, and an asymmetric risk penalty for high-confidence errors. The reported results on the authors’ RA-CAC dataset are promising, with RAPO outperforming several conventional detectors, general MLLMs, medical MLLMs, and task-specific grounding methods in mIoU, ACC, and F1.However, the actual technical formulation is much narrower than the clinical motivation. Although the paper is framed as opportunistic CAC screening on routine chest CT, the method appears to operate at the 2D slice level and outputs 2D bounding boxes. The manuscript does not demonstrate a complete 3D CT, scan-level, or patient-level CAC screening pipeline. It does not clearly explain how full CT volumes are processed, how multiple slices and multiple lesions are handled, or how slice-level predictions are aggregated into a clinically meaningful patient-level decision. This is a major limitation because opportunistic CAC screening is clinically a volume-level or patient-level task, not only a single-slice localization task. The current evaluation mainly reports mIoU, accuracy, F1, and recall, but does not provide key screening-oriented metrics such as patient-level sensitivity/specificity, lesion-level sensitivity, false positives per scan, FROC analysis, or scan-level CAC detection performance. Therefore, the clinical effectiveness claim is not sufficiently supported by the presented experiments. In addition, some methodological components remain insufficiently justified. The risk-aware reward is clinically intuitive, but still heuristic, and its advantage over standard cost-sensitive training, threshold calibration, or uncertainty-aware detection is not fully established. The RA-specific motivation is also somewhat overstated: the technical challenges described in the paper, such as small lesions, low contrast, motion artifacts, and non-gated CT acquisition, are general challenges of CAC detection on routine chest CT rather than clearly RA-specific imaging properties. Overall, the paper has a potentially interesting methodological idea and promising preliminary results, but the current study does not yet provide enough evidence for a MICCAI-level acceptance. The work would be stronger if it were more clearly framed as a 2D slice-level CAC detection method, or if it included full scan-level/patient-level evaluation, stronger clinical screening metrics, clearer reproducibility details, and more rigorous validation of the risk-aware component.



Review #2

  • Please describe the contribution of the paper
    • The paper introduces a novel reinforcement learning framework for CAC detection in non-contrast chest CT scans. They introduce a risk-aware anatomical prior optimization (RAPO) framework that is a risk aware GRPO mechanism with two parts: an anatomical prior to enforce geometric compactness and risk-aware penalty to reduce false negatives. They have very robust comparisons (against 12 SOTA methods) and strong ablation studies showing that each module in their methodology contribute to their strong 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.

    Major strengths include: novel risk-aware GRPO using anatomical priors and risk-aware penalties. Novel private dataset with non-contrast chest CTs. Novel application as it seems like an opportunistic screening for CAC in routine CTs. Very strong evaluations against other SOTA methods, 4 different families of methods with each having multiple models (2-3 each) for a total of 10 methods to compare against. As well as an ablation study that is very robust; each test adds modules that markedly improve the performance, showing that the proposed method as a whole is necessary to achieve their performance.

  • 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.
    • No external evaluation: the model was evaluated on private institutional data and was not validated on any external datasets. There are other opensource datasets, like the RAD-ChestCT dataset that is also a non-contrast chest CT dataset with CAC labels that could have been used to evaluate against other SOTA methods.
    • Doesn’t provide any code or repository to replicate their experiments.
    • Discussion of how metrics were calculated: It was never mentioned if the models were evaluated at a 2D at a slice level or 3D volume/patient level and if it was calculated on the whole chest CT or not. If the metrics were calculated at the slice level on the whole chest CT, the evaluation metrics would be skewed since most slices would not have CAC.
  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/en/REVIEWER-GUIDELINES.html
    • This paper is very strong in its formulation, evaluation, and ablation studies. However, without any external validation, it is difficult to evaluate how the model would truly perform in real-life situations. Since the model can memorize a lot of the internal data, its performance on different scanner types, acquisitions, hospitals, populations, and annotation styles (one annotator) is a major question.
    • Having the confidence intervals, standard deviations, or any statistical significance for table 1 and 2 metrics (mIoU, ACC, and F1) would be great as the differences between the models may not be significant.
    • if I understood it correctly, ablation ID 1 from table 2 and Qwen2.5-VL-7B from table 1 should be the same (similarly for ID 6 and “Ours”) but there’s some discrepancy in the metrics given, why not report F1 scores in the ablation study?
    • More metrics of evaluation: Since accuracy, F1 and recall were shown in tables 1 and 2; at minimum, why not show precision as well? Seems like there was plenty of room to add those columns.
  • 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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Lack of reproducibility and external validation. Also lacks measures of uncertainty, confidence or significance.

  • 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.

    My major concerns were external validation, statistical significance tests, and reproducibility. Their rebuttal states that their results was statistically significant and had external validation so if they add it to the submission for the final accepted version with the publicly available code and possibly data, it should be accepted.



Review #3

  • Please describe the contribution of the paper

    This paper presents the first CAC detection dataset for RA patients. The dataset contains high-quality CAC data from 415 RA patients, annotated by professional radiologists, filling a data gap in this field. A two-stage RAPO training framework is proposed. First, a baseline CAC classification capability is established through direct preference optimization, followed by the introduction of a reinforcement learning mechanism based on population-based relative strategy optimization. The proposed decoupled reward normalization strategy addresses the gradient dominance issue in multi-objective optimization by independently normalizing reward signals at different scales, thereby balancing geometric constraints with clinical safety requirements.

  • 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.

    First, the authors contributed a dataset for a niche field, conducting the first systematic study specifically targeting the RA population (characterized by small, low-contrast calcifications). The experimental results of this work appear promising. It achieves state-of-the-art (SOTA) performance in mIoU (0.6042), accuracy (0.7660), and F1 score (0.7311), with improvements of +9.6% to +31.4% over the best baseline. The risk-aware mechanism explicitly reduces the false negative rate. The proposed method demonstrates good training stability, and decoupled normalization resolves multi-objective gradient conflicts.

  • 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.

    This work simplifies clinical actions into a three-class classification of {Positive, Negative, Re-examination}. It explores whether confidence scores can be used for more granular risk stratification. However, there is a lack of validation regarding the quantification of uncertainty. Although intrinsic confidence is introduced, it has not been systematically compared with standard uncertainty estimation methods such as MC Dropout and Deep Ensemble. The paper defines CAC detection as a fine-grained localization task, with the model outputting pixel-level bounding boxes, yet it employs a text-generating VLM (Qwen-VL) as the policy network. Since VLMs generate discrete tokens, this approach may conflict with the requirements for precise localization in a continuous space. In the experimental section, were the comparison experiments involving YOLO, RetinaNet, and DINO conducted using the same pre-trained weights, input resolutions, and anchor boxes? DINO’s mIoU (0.403) is significantly lower than YOLO’s (0.551), which contradicts performance on public benchmarks. Finally, regarding GRPO’s sample efficiency: is a candidate count of K=4 too low for continuous-space localization tasks? Similar works such as MedGround-R1 use K=8-16, and the paper does not demonstrate the sufficiency of K=4.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Experimental design and workload.

  • 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 author has answered questions about experimental details. Experiments are conducted under reasonable conditions and fully illlustrate the advantages of the proposed method.



Review #4

  • Please describe the contribution of the paper

    This paper proposes RAPO, a reinforcement learning framework for detecting coronary artery calcification (CAC) in rheumatoid arthritis (RA) patients from routine chest CT scans. The contributions are:

    • It constructs the first CAC detection dataset specifically for RA patients, comprising 415 annotated chest CT cases.

    • It introduces a two-stage coarse-to-fine training curriculum combining DPO initialization and a novel Risk-aware GRPO mechanism.

    • It introduces a Risk-aware GRPO mechanism with three decoupled reward signals (Rloc, Ranatomy, and Rrisk) to jointly enforce geometric precision.

  • 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 motivation is well-grounded. The lack of automated CAC detection tools for this cohort is an underexplored gap.

    • The three-reward design is clinically meaningful. Each component targets a distinct failure mode. Rloc handles spatial accuracy, Ranatomy prevents anatomically implausible predictions, and Rrisk explicitly penalizes high-confidence false negatives, which are the most dangerous errors in a screening context.

    • The experimental setup is rigorous. The comparison covers a broad set of baselines across four categories, including conventional detectors, general MLLMs, medical MLLMs, and task-specific models. The ablation study is well-structured and each component is validated individually.

  • 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 counterintuitive drop in Accuracy and Recall when adding the Anatomical Prior alone (ID 4 vs ID 3 in Table 2) remains unexplained. This is a notable result that deserves further analysis.

    • The RA-specific framing could be somewhat overstated. RAPO’s design is not fundamentally specific to RA pathophysiology. The method would likely apply to other screening tasks, and the paper does not clearly articulate what makes RA a uniquely challenging case beyond general detection difficulties.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    My current decision is a weak accept. RAPO is a well-motivated paper that addresses a clinically relevant problem with a technically sound framework. Experimental results show that the proposed method outperformed the SOTA baselines. However, the paper does not mention any plans to release the dataset or code. Given that this is the first CAC dataset specifically constructed for RA patients, making it publicly available would be a significant contribution to the community and would allow future work to build on this benchmark. While open-sourcing is not a requirement, if the authors can release either the dataset or the code, I would be willing to raise my score.

  • 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 have addressed my concerns. Their promise to open-source the code upon acceptance is appreciated. Accept.



Review #5

  • Please describe the contribution of the paper

    This paper targets coronary artery calcium detection on routine non-contrast chest CTs in rheumatoid arthritis patients. It builds a 415-case RA-CAC dataset and proposes RAPO, a Qwen2.5-VL-based DPO + Risk-aware GRPO framework. The method uses an anatomical prior reward and an uncertainty-based risk penalty to improve localization and screening reliability for small, low-contrast lesions.

  • 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.RA patients have higher cardiovascular risk, so opportunistic CAC screening on routine chest CT has clear clinical value. 2.the paper introduces a Chan-Vese-based anatomical constraint and a risk-aware reward that assigns a higher cost to false negatives. 3.on the authors’ dataset, RAPO achieves the best mIoU, ACC, and F1 compared with Qwen2.5-VL, InternVL, Lingshu, MedGround-r1, VLM-R1, YOLO, RetinaNet, and DINO.

  • 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 does not clearly state whether the model input is a full 3D CT scan, 2D slices, or manually selected lesion-containing slices. For opportunistic screening, the key question is whether the model can find CAC in a full chest CT, not just ground lesions on preselected slices. It also does not report the positive/negative case ratio, lesion count per case, or number of annotated slices per case 2.It does not report key screening metrics such as patient-level sensitivity/specificity or lesion-level sensitivity 3.The paper highlights decoupled reward normalization, but the ablation study does not compare against standard GRPO with directly summed rewards or non-decoupled normalization. 4.the construction of DPO preference pairs is not explained.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    please see the weaknesses

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    the method is evaluated mostly at the slice level, while the clinical goal is patient-level CAC screening, and the needed screening metrics and validation are still not strong enough



Author Feedback

We thank all reviewers for the constructive feedback. We address the main concerns below. General Response 1.Code/data availability (R2, R4) If accepted, we will release the code. we will strictly de-identify the data and provide controlled access to qualified researchers. 2.Task definition(R1, R5) RAPO is a 2D slice-level CAC detection method, not 3D detection. The model takes CT slice as input, predicts CAC presence/absence, and outputs bboxes for positive slices. And our setting is not limited to lesion-containing slices. The evaluation set contains 5,821 positive and 112,439 negative slices, with 13.7 lesions per case on average. These statistics will be added to the revised manuscript. 3.DPO pairs (R1, R5) DPO is only used to initialize CAC presence/absence discrimination. For CAC-positive slices, positive responses are chosen and negative responses rejected; for CAC-negative slices, the construction is reversed. We will clarify this construction in the revision. 4.Statistical significance (R2) We performed paired bootstrap analysis based on existing evaluation outputs. Compared with SFT-only, DPO-SFT, GRPO-base, R_anat-only, and R_risk-only, all paired comparisons show statistical significance with p<0.001.5.External generalization (R1, R2) External COCA dataset evaluation shows that RAPO achieves mIoU = 0.709, outperforming YOLO (0.664) and Lingshu (0.596). We will clarify this in the revision.. R1-1.Two-stage design We will clarify that the two-stage design is a training curriculum, not a clinical inference workflow. DPO initializes presence/absence discrimination, while GRPO optimizes localization and risk-aware behavior. R1-2.Annotation-free We did not claim RAPO is annotation-free. We only intended to emphasize that R_anat​ does not require additional pixel-level annotations. R1-3.RA-specific motivation (R4) We agree that the RA-specific framing should be refined. Small lesions etc. are common challenges in non-gated chest CT CAC detection. The RA-specific motivation of this work mainly lies in the clinically high-risk cohort and screening scenario. R1-4.R_risk vs. threshold tuning R_risk combines asymmetric clinical cost with VLM intrinsic confidence and specifically penalizes high-confidence false negatives. After replacing R_risk with post-hoc thresholding based on bbox-token log-probability, the mIoU drops to 0.5565.R1-5.Evaluation metrics (R2) We will add Recall to Table1 and F1 to Table2 to avoid interpreting performance only from ACC. R1-6.Higher recall of +Risk Using R_risk alone encourages the model to avoid missed detections more aggressively, leading to higher recall but potentially lower localization quality. R3-1.Intrinsic confidence vs. uncertainty Intrinsic confidence is a training-time risk signal for identifying over-confident false negatives. Replacing it with uncertainty estimation drops mIoU to 0.5605.R3-2.VLM localization RAPO outputs bbox coordinates (x1,y1,x2,y2), not pixel-level masks. The coordinates are generated as discrete tokens, which is consistent with grounding-style VLM detection. R3-3.Baseline fairness All baselines use officially released pretrained weights and the same train/evaluation setting. The lower performance of DINO may be related to the small size, low contrast of CAC targets. R3-4.GRPO candidate number K Existing training runs with different K values show that larger K improves mIoU (0.5726/0.6042/0.6156/0.6217 for K=2/4/8/10) but also significantly increases sampling cost; therefore, K=4 is used as a trade-off. R4-1.Acc/Recall drop after adding +Anatomy R_anat improves anatomical plausibility and geometric compactness, but alone may be overly conservative and filter out low-contrast or small CAC regions, which may reduce recall and ACC. R5-2.Decoupled normalization Non-decoupled normalization achieves mIoU = 0.5940, while decoupled normalization achieves 0.6042, suggesting that it helps balance reward signals with different scales.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This paper addresses a clinically relevant problem and proposes a potentially interesting risk-aware RL framework for CAC detection in RA patients. Reviewers appreciated the motivation, dataset effort, broad comparisons, and ablation studies. However, there is substantial concern about the core task formulation, especially whether the method operates on full 3D CT volumes, selected 2D slices, or another representation, and how slice-level outputs translate into scan- or patient-level CAC screening. Additional concerns include under-specified DPO preference construction, lack of external validation, limited reproducibility, missing screening-oriented metrics, and insufficient evidence that the RA-specific framing introduces imaging challenges beyond general non-gated CAC detection. Overall, the work is promising but currently not sufficiently clear or validated for acceptance unless the rebuttal can resolve these fundamental concerns.

  • 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.

    The rebuttal clarifies that the method is a 2D slice-level detection approach, and I think this setting is acceptable as long as the final paper clearly states this scope and avoids overstating it as a complete patient-level screening system.

    Given the reported improvements, external validation, statistical testing, and planned code release, I recommend acceptance, with the remaining limitations on 3D/patient-level evaluation clearly discussed.



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.

    After reviewing the feedback, I recommend a reject despite 3 vs 2 votes in favor. As the reviewers noted, there is a fundamental mismatch between the claimed clinical goal (patient-level CAC screening on routine chest CT) and the actual implementation (2D slice-level bounding box detection). The evaluation lacks essential patient-level metrics (sensitivity/specificity, FROC, false positives per scan) and a clear slice-to-volume aggregation strategy. The former is a serious presentation/clarity problem that hiders me from recommending acceptance. A revision with proper task framing and volumetric validation could be interesting perhaps for one of the MICCAI workshops.



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.

    This paper proposes a reinforcement learning framework for coronary artery calcification detection in rheumatoid arthritis patients from routine non-contrast chest CT. It received three Accept and two Reject. The primary concern raised by the reviewers is that the method focuses exclusively on slice-level analysis, and does not match the claimed clinical screening situation. In the AC’s view, this concern mainly reflects an issue of overstatement rather than a fundamental flaw in the methodology. The work constitutes a solid step toward automated CAC detection, and the proposed framework is both interesting and effective. Therefore, AC leans toward acceptance, and asks authors to more accurately define the setting and its clinical implications.



back to top