List of Papers Browse by Subject Areas Author List
Abstract
Intelligent Coronary Artery Disease (CAD) guidance requires comprehensive reasoning over heterogeneous visual attributes and risk profiles aligned with the clinical diagnostic workflow. However, existing image-to-prediction methods lack systematic case-level modeling and evidence-grounded reasoning, limiting their clinical reliability. In this context, we introduce, for the first time, a Clinically Structured Chain-of-Thought (CS-CoT) framework for intelligent CAD guidance based on coronary CT angiography images. Leveraging the semantic understanding capabilities of Large Language Models (LLMs), the proposed vision-language model enables interpretable, clinically aligned, and interactive case-level diagnosis and management. Specifically, CS-CoT transforms heterogeneous anatomical, lesion, pericoronary, risk-related features, and diagnostic case prior into unified textual-space embeddings and organizes them into a structured, traceable chain-of-thought reasoning process aligned with the clinical diagnostic workflow. Extensive experiments on 2,067 clinical cases demonstrate that CS-CoT outperforms state-of-the-art LLMs, thereby providing reliable clinical decision support at the case level. The implementation details and prompt engineering are publicly available at the code-link (https://github.com/PerceptionComputingLab/CS-CoT).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0056_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{MaXin_Clinically_MICCAI2026,
author = { Ma, Xinghua AND Chu, Yuetan AND Wang, Xiaodi AND Qiu, Xingyu AND Li, Fanding AND Wang, Kuanquan AND Luo, Gongning AND Gao, Xin},
title = { { Clinically Structured Chain-of-Thought Reasoning for Intelligent Guidance in Coronary Artery Disease } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes CS-CoT, a vision-language framework for case-level coronary artery disease (CAD) guidance from CCTA, structured around the CAD-RADS 2.0 reporting workflow. Two modules compose the system: (i) a multi-aspect coronary encoding module that extracts heterogeneous visual semantics (anatomical structure, lesion characteristics, pericoronary region) together with patient risk profiles, standardizes them into a shared textual embedding space, and retrieves similar prior cases via graph-based matching; and (ii) a structured diagnostic reasoning module built on a LoRA-fine-tuned Qwen-2-7B-Instruct that executes a four-stage chain-of-thought aligned with clinical workflow — anatomy, patient-level disease attribution, ACS risk assessment from perivascular features, and final case-level recommendation with retrieved-case grounding. The authors validate on 2,067 retrospective cases with CAD-RADS 2.0 reports, comparing against five text-only LLM baselines (Qwen-2, Med-Alpaca, Claude 3.7 Sonnet, Clinical Camel, GPT-4o) on automated metrics, BERT embedding similarity, and radiologist scoring of clinical credibility and logical consistency. Ablations disable individual embeddings and forward/backward reasoning intervals to probe module contributions.
- 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 clinical framing is unusually strong. Aligning the CoT structure explicitly with CAD-RADS 2.0 and generating supervision by decomposing structured reports into staged reasoning trajectories is a principled way to bring LLM reasoning under clinical discipline rather than hoping the model learns the workflow implicitly. The dataset (2,067 cases, five-radiologist annotation, CAD-RADS 2.0 standardization, 2020–2024 window) is large by medical VLM standards and the evaluation protocol is multi-faceted: automated metrics stratified by reasoning stage (anatomy/attribution/risk/case), semantic similarity via BERT, and expert scoring on two clinically meaningful axes. Providing the text-only baselines with the same standardized textual space is the right methodological choice, since it isolates the contribution of the structured reasoning from the contribution of the visual feature extraction. The ablation over forward- versus backward-interval CoT disabling is a thoughtful experiment design and produces interpretable conclusions about which stages matter most. The retrieval-augmented element (top-K=5 similar cases) is clinically reasonable and mirrors how radiologists actually reason by comparison.
- 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 claim of being the “first vision-language model for intelligent CAD guidance” is overstated and should be softened. Cardiac VLMs and LLM-based CCTA reporting pipelines have appeared in recent literature, and CAD-RADS-aligned report generation is an active area. The contribution is more precisely framed as the first CAD-RADS-structured CoT framework coupling multi-aspect visual encoding with staged clinical reasoning which is novel enough on its own and does not need the stronger claim. The contribution list in page 3 is numbered (1), (3). A missing (2). The comparison with SOTAs, while large in surface area, has an important design issue. All baselines are text-only LLMs given the pre-computed textual embeddings; none is a vision-language model actually processing CCTA. This makes the comparison a test of reasoning over a fixed feature set, not of end-to-end clinical VLM performance, which means the headline claim of outperforming SOTA VLMs on CAD guidance is not directly supported. The comparison would be substantially strengthened by including at least one genuine medical VLM baseline (LLaVA-Med, Med-Flamingo, or a recent cardiac VLM) operating on the CCTA volumes. Relatedly, giving GPT-4o and Claude 3.7 Sonnet only the pre-structured embeddings without the four-stage CoT prompts that CS-CoT uses is not an apples-to-apples comparison: the baselines should either receive the same staged CoT prompting or the authors should show that CS-CoT’s gains persist against baselines given the same prompting scaffolding. Otherwise the improvement could be attributed to prompt engineering rather than the architectural contribution. The “graph-based” retrieval approach is named but its graph construction, edge definition, and similarity function are not described. The “diffusion parameter” set to 0.2 is mentioned without context of what it modulates. The CoT supervision is said to be generated by “decomposing structured CAD-RADS reports into four staged reasoning trajectories” but the decomposition procedure is not detailed, and this matters because CoT supervision quality largely determines output quality. The paper also does not address how the system should be deployed or how its recommendations would be presented to clinicians. Minor issues: the model is called both CS-CoT and SC-CoT in different places, which should be unified.
- 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.
- 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?
The conceptual contribution is genuinely novel and clinically well-motivated, and the ablation over forward/backward reasoning intervals provides useful insight into how structured CoT actually works. However, the paper’s strongest claims (outperforming SOTA VLMs, providing reliable CAD guidance) are not fully supported by the current experimental design: the baselines are text-only rather than vision-language, the baselines do not receive equivalent staged prompting, hallucination and clinical-error rates are not quantified, and external validation is absent. A revision that adds at least one genuine medical VLM baseline, matches prompting conditions across baselines, reports stage-wise confusion matrices and hallucination/factual-error rates, provides inter-rater agreement for the radiologist scores, and ideally adds temporal or external validation would turn this into a strong paper.
- 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 presents a clinically structured chain-of-thought framework (CS-CoT) for case-level intelligent guidance in coronary artery disease from coronary CT angiography (CCTA). The method first encodes heterogeneous case information, including coronary anatomy, lesion characteristics, perivascular features, patient risk profiles, and retrieved similar prior cases, into a standardized textual space. A Qwen-based large language model then performs staged reasoning aligned with the clinical workflow, from anatomical understanding and disease attribution to ACS risk assessment and final case-level management guidance. The paper aims to improve interpretability, clinical alignment, and interactivity compared with direct image-to-prediction CAD models.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a clinically meaningful problem. Moving from direct image-to-label prediction toward case-level reasoning and management guidance for CAD is important and potentially impactful. 2.The proposed framework is well motivated from a clinical perspective. The decomposition into anatomy, lesion, perivascular risk, patient risk profile, and prior similar cases is intuitive and aligns with how clinicians often synthesize CAD findings. 3.The manuscript proposes a relatively clear staged reasoning pipeline. The idea of structuring the reasoning process according to the diagnostic workflow is interesting, and the intermediate reasoning stages improve interpretability compared with purely end-to-end black-box prediction. 4.The evaluation covers multiple aspects, including anatomy analysis, disease attribution, risk assessment, case-level diagnosis, embedding similarity, and expert scoring. The inclusion of ablation studies is also valuable. 5.The case illustration and the workflow figure help communicate the intended practical use of the framework and highlight the authors’ emphasis on clinically structured reasoning rather than flat report generation.
- 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 fairness of the comparison with foundation models is questionable. According to the manuscript, GPT-4o, Claude, Med-Alpaca, and other LLM baselines do not directly process CCTA images; instead, they are all given the standardized textual space produced by the authors’ own upstream module. This means the comparison does not isolate the contribution of the reasoning framework itself and may substantially advantage the proposed method. 2.The claim of being the “first” vision-language model for intelligent CAD guidance is not sufficiently substantiated. The manuscript does not clearly define the scope of this claim or discuss enough related work in multimodal medical reasoning, report generation, or clinically guided decision support to justify such a strong novelty statement. 3.The dataset and evaluation protocol are insufficiently characterized for a high-impact medical imaging venue. The paper reports 2,067 retrospective cases from 2020–2024, but does not provide enough detail on class distribution, patient demographics, disease prevalence, institutional diversity, acquisition variation, or whether splits were performed at the patient level with safeguards against leakage. There is also no external validation set, which limits claims of clinical reliability and generalizability. 4.The expert evaluation is under-described. While five radiologists reportedly scored clinical credibility and logical consistency, the manuscript does not provide enough information about the scoring rubric, blinding, inter-rater agreement, variance, or statistical significance. This weakens the strength of the human evaluation evidence. 5.Reproducibility is limited. The paper states that implementation details and prompt engineering will be made publicly available, but no anonymized code or resource link is currently provided. Important details for reproducing the structured prompts, report decomposition strategy, graph construction choices, and case retrieval procedure are still missing or underspecified. 6.There are multiple issues that raise concerns about technical rigor and manuscript reliability. For example, the paper alternates between “CS-CoT” and “SC-CoT”; the contribution list skips from (1) to (3); and, more importantly, several references appear mismatched with the cited systems. For instance, the paper uses Qwen-2-7B-Instruct but cites Qwen-Image as [22]; and the reference for Claude 3.7 Sonnet appears unrelated. These issues significantly reduce confidence in the experimental and scholarly accuracy of the submission.
- 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
I appreciate the authors’ effort to move CAD analysis beyond direct image-to-prediction models toward clinically structured and explainable case-level guidance. I believe the overall direction is interesting and relevant. However, for a stronger MICCAI submission, the paper would benefit from substantially improved experimental rigor and manuscript accuracy. In particular, I encourage the authors to (1) clarify exactly what is novel relative to prior multimodal medical reasoning and CAD analysis work; (2) redesign or better justify the baseline comparisons so that the comparison is methodologically fair; (3) provide more complete dataset statistics, split details, and external validation; (4) report inter-rater agreement and statistical testing for expert evaluation; and (5) carefully audit the manuscript for terminology inconsistencies and reference mismatches. Addressing these issues would considerably strengthen the credibility and impact of the work.
- 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?
My recommendation is mainly driven by concerns about methodological rigor and evaluation credibility rather than by a lack of relevance. The clinical problem is important, and the idea of structuring CAD reasoning with a clinically aligned CoT pipeline is interesting. However, several core issues prevent me from supporting acceptance in its current form. Most importantly, the baseline comparison is not fully convincing because the competing LLMs are not evaluated on raw multimodal input but on the authors’ own standardized textual representation, which may confound where the gains actually come from. In addition, the submission makes a strong novelty claim (“first” vision-language model for intelligent CAD guidance) without sufficiently precise positioning against related work. The validation is also limited to a single retrospective internal dataset with insufficient characterization and no external testing. Human evaluation lacks enough detail on scoring protocol and agreement. Reproducibility is incomplete, as no anonymized code/resources are provided. Finally, multiple manuscript issues, including terminology inconsistency and several apparently mismatched references for major baseline systems, reduce confidence in the technical care and reliability of the submission. Taken together, these issues place the paper below the acceptance bar for MICCAI.
- 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.
After reading the authors’ rebuttal, I am updating my recommendation from weak reject to accept. The rebuttal addresses several of my main concerns in a substantive way. In particular, the authors clarified that the fine-tunable baselines (Qwen2-7B-Instruct, Med-Alpaca, and Clinical Camel) were trained on the same dataset, which improves the fairness of the comparison relative to what was described in the original manuscript. They also provided important missing details about the expert evaluation, including the number of evaluated cases, the scoring rubric, inter-rater agreement (κ=0.83, ICC=0.87), and statistical significance. In addition, the rebuttal clarified the dataset composition (three centers, five scanner models, patient-level splitting) and added key reproducibility details for the graph-based retrieval module, including the use of BioClinicalBERT similarity, sparse graph construction, and sensitivity analysis over α and K. These clarifications substantially improve my confidence in the experimental design and make the submission more convincing than it appeared in the original version. I still think some limitations remain: the lack of a true 3D CCTA vision-language baseline means that the “first vision-language model” framing should be stated cautiously, and the absence of external validation remains an important weakness. I also still encourage the authors to carefully audit citations and terminology inconsistencies in the revision. However, taken together, I believe the rebuttal resolves enough of my major concerns that the paper now sits slightly above the acceptance threshold. Therefore, I support acceptance, while encouraging the authors to make the revised manuscript more precise in its novelty claims and more explicit in its methodological details.
Review #3
- Please describe the contribution of the paper
This paper proposes CS-CoT, the first clinically structured vision-language framework for intelligent Coronary Artery Disease (CAD) guidance based on Coronary CT Angiography (CCTA) images. CS-CoT consists of two main modules: (1) a multi-aspect coronary encoding module that extracts heterogeneous features (anatomical structures, lesion characteristics, perivascular features, and patient risk profiles) from CCTA volumes and maps them into standardized textual-space embeddings, supplemented by a graph-diffusion-based similar case retrieval mechanism; and (2) a structured diagnostic reasoning module built upon a LoRA-fine-tuned Qwen-2-7B-Instruct LLM, which hierarchically organizes clinical evidence into a four-stage chain-of-thought process aligned with the CAD diagnostic workflow (anatomical understanding → disease attribution → ACS risk assessment → comprehensive diagnosis and management). The authors validate the framework on 2,067 retrospective clinical cases and report improvements over several state-of-the-art LLMs (GPT-4o, Claude 3.7 Sonnet, Med-Alpaca, Clinical Camel) across automated metrics, embedding similarity, and expert scoring, claiming that the clinically structured CoT design is the key factor behind the observed gains.
- 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.
S1: Clinical workflow alignment. The most notable aspect of CS-CoT is the explicit structuring of its reasoning process to mirror the established CAD diagnostic workflow. Rather than treating diagnosis as a flat prediction task, the four-stage CoT decomposition provides a principled and interpretable framework that radiologists can follow, audit, and potentially intervene upon. This is a significant design choice that distinguishes it from generic LLM-based medical reasoning approaches.
S2: Comprehensive multi-aspect feature modeling. The paper addresses an inherent limitation of prior image-to-prediction CAD approaches by jointly modeling anatomical structures, lesion characteristics, perivascular inflammation markers, and patient risk profiles at the case level. The inclusion of fat attenuation index and perivascular connected components as additional risk signals goes beyond traditional stenosis-only analysis and reflects current clinical understanding of plaque vulnerability.
S3: Graph-diffusion-based case retrieval. The graph-based mechanism for sourcing diagnostic case priors is a creative addition. Compared to naive nearest-neighbor retrieval, the use of diffusion-based similarity propagation over a case graph enables transitive similarity reasoning, potentially surfacing clinically analogous cases that may appear superficially different but share deeper structural or risk patterns. This provides an evidence grounding that enhances clinical trust.
- 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.
W1: No evaluation of information fidelity in the image-to-text conversion. The framework compresses 3D CCTA volumes into discrete textual attributes via upstream networks, yet the paper provides no assessment of how much diagnostically relevant information is lost in this process. The ablation in Fig. 3a only evaluates the importance of entire embedding categories, not the faithfulness of the conversion itself. Without comparing diagnoses made from original images versus their textual surrogates—or analyzing whether diagnostic errors originate from the encoding stage or the reasoning stage—the reliability of this core design choice remains unvalidated.
W2: Critical implementation details of the graph-based retrieval module are missing. The similarity function sim(zi, zj) used to construct the case graph is never specified—it is unclear what encoder vectorizes the structured textual representations and whether it is a general-purpose or domain-specific model, which matters significantly given that generic text encoders typically underperform on medical terminology. The graph construction strategy (sparse vs. fully connected, edge thresholds), and the sensitivity to hyperparameters alpha=0.2 and K=5, are also unreported. More importantly, there is no independent evaluation of retrieval quality: the ablation only shows that removing R_sim hurts performance, but does not assess whether retrieved cases are diagnostically consistent with the query case or whether erroneous priors can mislead downstream reasoning.
W3: Limitations in the evaluation protocol. Relying on GPT-4o as a judge introduces unverified assumptions about its competence in cardiovascular diagnosis. BLEU and BERTScore measure textual similarity rather than clinical correctness. The expert evaluation lacks essential details such as sample size, scoring criteria, and inter-rater agreement.
W4: Unfair comparison with state-of-the-art LLMs. The comparison in Table 1 is problematic. The baseline models (GPT-4o, Claude 3.7 Sonnet, etc.) are general-purpose LLMs that have not been fine-tuned for the CAD reasoning task, whereas CS-CoT uses Qwen-2-7B with LoRA fine-tuning on the training set and task-specific CoT supervision. This is not an apples-to-apples comparison. A fairer evaluation should include: (a) fine-tuning the same baseline models—or at least the open-source ones such as Qwen-2-7B-Instruct and Med-Alpaca—with the same data and supervision but without the structured CoT; (b) providing proprietary models (GPT-4o, Claude) with few-shot or CoT prompts to better elicit their reasoning capabilities; and (c) comparing against domain-specific medical LLMs that have been fine-tuned for cardiology or radiology tasks.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
The paper addresses a clinically meaningful problem and proposes a well-motivated framework whose structured CoT design and multi-aspect feature modeling are genuinely interesting (S1, S2). However, the experimental evaluation has several significant issues that weaken the paper’s claims. Most critically, the comparison in Table 1 is not fair: the proposed method benefits from LoRA fine-tuning and task-specific CoT supervision, while the baselines are evaluated in a zero-shot setting (W4). This makes it difficult to attribute the reported gains to the CoT design itself rather than to supervised fine-tuning. Additionally, the lack of evaluation on information fidelity in the image-to-text conversion (W1), missing implementation details of the retrieval module (W2), and the limitations of the evaluation protocol (W3) collectively leave the core contributions insufficiently validated.
- 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 authors’ rebuttal clarifies the expert evaluation rubric and extends the dataset to a 3-center cohort, it fails to resolve two fatal methodological flaws. First, compressing 3D CCTA imaging volumes into discrete text attributes introduces a severe information bottleneck; the authors still provide no evaluation of information fidelity or diagnostic loss during this conversion, creating a major clinical liability (W1). Second, the baseline comparison remains methodologically confounded (W4); the proposed model benefits from a compound advantage, yet it is compared against zero-shot commercial LLMs. Without an ablation control—such as comparing against the same Qwen backbone fine-tuned on the same data without the structured CoT—it is impossible to isolate whether the performance gains stem from the architectural novelty or simply from standard data fine-tuning. Because the core design remains unvalidated and the experimental controls are unfair, the paper cannot be accepted in its current form.
Author Feedback
We sincerely appreciate your positive feedback on our work, including its Novelty (“genuinely novel”-R1; “clear…pipeline”-R2; “well-motivated framework”-R3), Significance (“clinically well-motivated”-R1; “well motivated from a clinical…”-R2; “clinically meaningful…”-R3), and Design(“unusually strong”-R1; “idea…is interesting”-R2; “genuinely interesting”-R3). We also thank you for taking the time to read our paper and provide constructive suggestions. We address your concerns as follows:
-Experiment
(R1-1,2; R2-1,2; R3-1,4)
Currently, there is no publicly available VLM that can directly process 3D CCTA volumes. Existing medical VLMs (e.g., LLaVA-Med and Med-Flamingo) are primarily designed for 2D medical images. By testing, when these are applied to CCTA analysis, they are unable to produce clinically meaningful outputs.
The ACE module provides the first feasible pathway toward a CCTA-based VLM; therefore, no directly comparable prior work exists. Accordingly, we select widely recognized LLMs as comparison methods, which effectively demonstrate the benefits of staged CoT reasoning.
The core design of SC-CoT is staged CoT reasoning, with Qwen used solely as the backbone. If SOTAs were also equipped with staged CoT, the comparison would shift from evaluating staged CoT itself to backbone differences. We select Qwen as the backbone because it achieves the best performance among all backbone candidates under staged CoT.
All fine-tunable models (Qwen2-7B-Instruct, Med-Alpaca, and Clinical Camel) were trained on the same dataset, which will be clarified in the revised manuscript.
(R2-4; R3-3)
The expert evaluation used a 100-point rubric, with five experts independently assessing 200 cases across balanced disease severity levels.
C.C. is scored as follows: 81~100: clinically complete outputs suitable for direct clinical reference; 61~80: generally accurate outputs with minor imprecision; 41~60: partially accurate outputs requiring clinician revision due to insufficient assessment; 21~40: outputs with major diagnostic errors or missing critical information; 0~20: outputs with severe diagnostic inaccuracies inconsistent with medical knowledge.
L.C. is scored as follows: 81~100: fully coherent reasoning strictly following the clinical logic chain; 61~80: generally coherent reasoning with minor logical jumps or weak transitions; 41~60: partially coherent reasoning with broken chains or unclear stage transitions; 21~40: reasoning with major inconsistencies or disorganized logic; 0~20: reasoning lacking explicit structure or exhibiting largely unstructured generation.
The evaluation showed substantial inter-rater agreement (κ=0.83, ICC=0.87), indicating reliable expert assessment. Statistical significance testing showed that CS-CoT achieved significantly higher credibility compared with SOTAs (p<0.05).
-Dataset
(R2-3)
2,067 cases (age: 55.9±7.5 years; 1,383 males; 28.6% with acute chest pain) were collected from 3 clinical centers using 5 scanner models, with patient-level splitting ensuring no data leakage. In accordance with anonymization rule, additional omitted dataset information will be described in greater detail in the revised version.
-Reproducibility
(R1-3; R2-5; R3-2)
Graph-based retrieval uses sparse graphs, where each case is represented as a node and edge weights are defined based on BioClinicalBERT-derived similarity (Page5,Line14). α controls similarity propagation within the graph (Eq. 2).
In sensitivity analysis, α and K were evaluated over [0.1, 0.5] and [0, 10], respectively, and achieved optimal performance at 0.2 and 5.The retrieved prior cases can be regarded as prompts and do not compromise the decision-making basics.
The CAD-RADS decomposition is formulated as a structured reorganization of the original text into anatomy, lesion, risk, and decision components. The decomposition procedure, standardized text-space, multi-stage prompts, and reproducible project will be released at the code-link.
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.
The reviewers find the problem important and agree that aligning chain-of-thought reasoning with a clinically defined workflow (CAD-RADS) is an interesting direction. The idea of structuring case-level reasoning with multi-aspect inputs and retrieved prior cases is promising, and the overall framework is well motivated. There are some concerns about the method evaluation. The proposed model benefits from fine-tuning, structured CoT supervision, and curated textual inputs, while the baselines are mostly evaluated in weaker settings. This makes it difficult to isolate what is actually driving the improvements. The paper also does not include a vision-language baseline on CCTA. Reviewers also point out gaps in the technical description (e.g. retrieval module, CoT supervision), single-centre retrospective data, no external test set, and insufficient detail in the evaluation. There are also concerns about overstated novelty and some inconsistencies in the manuscript.
- 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 paper addresses an important clinical problem and proposes a clinically structured chain-of-thought framework for coronary artery disease guidance from CCTA. The reviewers generally agreed that aligning the reasoning process with the CAD-RADS workflow is a novel and clinically meaningful contribution, and that the staged reasoning design improves interpretability and clinical alignment. The work is supported by a relatively large multi-centre dataset and comprehensive quantitative evaluation.
During rebuttal, the authors clarified several important concerns regarding evaluation fairness, dataset composition, statistical significance testing, the graph-based retrieval mechanism, etc. These clarifications substantially strengthened confidence in the experimental design and reproducibility of the work.
Some limitations remain, including the absence of external validation, lack of direct comparison with a true 3D CCTA vision-language baseline, and the need for more cautious positioning of novelty claims. Nevertheless, the core contribution is clinically relevant, technically interesting, and sufficiently validated to warrant acceptance. The paper is likely to stimulate further research on clinically grounded multimodal reasoning systems for cardiovascular imaging.
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.
The motivation of the paper is strong, but it does not sufficiently prove that the chain-of-thought structure adds robust clinical value beyond prompt engineering. The comparisons are also weak e.g. GPT-4o and Claude are given the same standardized textual space, so the evaluation mainly tests report-generation ability rather than end-to-end CAD guidance. So I am inclined to a rejection for the current state of the paper.
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.
The reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution especially for what concerns methodological applications in MIC.
