Source-linked AI summary
CARDEA: Auditable Reasoning Grounded in Spatial Evidence for End-to-End Coronary Angiography Interpretation
Jia-Jen Lee, Shih-Yen Hou, Kee Koon Ng, Wei-Chun Wang, Shih-Sheng Chang
TL;DR
CARDEA addresses variable CAG interpretation and limited auditability by providing an end-to-end large vision-language pipeline with spatially grounded reasoning. Trained on closed-ended tasks, it remained competitive on study-level diagnosis, while only RLVR improved zero-shot report generation; clinical use still requires prospective validation.
Problem
CAG interpretation varies among observers, while existing AI systems lack auditable decision processes and comprehensive open-ended assessment.
Method
CARDEA is a unified LVLM that selects keyframes and performs study-level Chain-of-Box reasoning, trained through visual alignment, self-distilled cold start, and RLVR with a CoB reward.
Results
CARDEA remained competitive with task-specific baselines and cardiologists on study-level diagnoses, while only RLVR improved zero-shot report generation, reaching vessel-severity macro_F1 0.686 versus 0.513 for the untuned base model.
Takeaways & Limitations
CARDEA exposes auditable spatial evidence behind study-level CAG conclusions, and RLVR on verifiable closed-ended tasks surfaced open-ended reporting ability that supervised imitation did not.
Takeaways & Limitations
Clinical use requires prospective validation against expert cardiologists, and current evaluation lacks blinded comparison with physician-authored reports.
Abstract
from arXiv · showhide
Invasive coronary angiography (CAG) is the gold standard for diagnosing coronary artery disease, but interpretation varies substantially among observers. Existing AI systems can improve consistency but lack auditable decision processes and are limited in comprehensive open-ended assessment, undermining clinician trust and clinical adoption readiness. We developed CARDEA, a unified large vision-language model that serves as the inference core of a CAG pipeline. It was trained solely on public datasets and closed-ended tasks in three stages: visual feature alignment, a self-distilled Chain-of-Box (CoB) cold start, and reinforcement learning with verifiable rewards (RLVR) with a CoB reward encouraging bounding-box use in the reasoning trace. We assessed its two study-level diagnoses, dominance classification and complexity assessment, against a dedicated classifier and two interventional cardiologists. Report generation was excluded from training and evaluated zero-shot across stages on an external cohort using vessel-severity macro-$F_1$. CARDEA trailed the classifier on in-distribution dominance but drew level under domain shift (accuracy, 0.91 [95% confidence interval (CI), 0.86 to 0.95]) and was comparable to the cardiologists on complexity assessment (accuracy, 0.90 [CI, 0.82 to 0.97]). Only RLVR improved zero-shot report generation, raising its vessel-severity macro-$F_1$ (0.686 [CI, 0.664 to 0.707]) above the untuned base model (0.513) and over twice the always-normal floor (0.312). CARDEA runs an end-to-end CAG pipeline from raw multi-view videos through keyframe selection to study-level diagnosis while exposing auditable spatial evidence behind its conclusions. RLVR on verifiable closed-ended tasks surfaced open-ended reporting ability that supervised imitation did not. Clinical use requires prospective validation against expert cardiologists.
1 Introduction
Coronary angiography is clinically important but difficult to interpret consistently because 3D anatomy is projected into 2D views and different views must be mentally integrated. CARDEA addresses these limitations with an end-to-end model that combines study-level diagnosis with spatially grounded reasoning.
- 77.4% overall agreement was reported among three experienced cardiologists reading the same angiograms.
- 2D projection causes vessel overlap and foreshortening, which can lead to underestimation of lesion severity.
- Existing pipelines can compound errors across chained modules, while discriminative foundation models typically output labels or spatial coordinates rather than open-ended reasoning.
- Bounding boxes embedded in reasoning traces provide spatial anchors that let humans audit which anatomical structures support a conclusion.
- CARDEA runs an end-to-end multi-view CAG pipeline and evaluates whether closed-ended training generalizes zero-shot to open-ended report generation.
2 Methods
CARDEA uses a two-pass, multi-stage training and evaluation design that converts public angiography datasets into an end-to-end pipeline. Its stages combine frame curation, vessel and stenosis grounding, self-distilled CoB reasoning, and RLVR across diagnostic and report-generation tasks.
- Pipeline: The first pass filters non-diagnostic video frames and classifies views, while the second performs study-level CoB reasoning over curated keyframes.
- Datasets and tasks: Five public datasets are reformulated into instructional tasks spanning vessel detection, stenosis detection, view classification, complexity assessment, dominance classification, and report generation.
- Training: CARDEA trains through two supervised fine-tuning stages followed by RLVR, with self-distillation creating CoB traces from Stage 1 bounding boxes and diagnostic guides.
- Metrics: Report generation is evaluated with vessel-severity macro-F1, while classification uses accuracy and Macro F1 and detection uses instance-level F1@IoU≥0.5.
- Baselines: Comparisons use published dedicated baselines, except complexity assessment, which is evaluated against two interventional cardiologists.
- Evaluation caveats: 95% bootstrap confidence intervals support exploratory distinguishability judgments, with small test sets and non-head-to-head comparisons limiting interpretation.
3 Results
CARDEA was competitive on trained closed-ended diagnostics, generalized to held-out AngioCAD tasks, and gained zero-shot report-generation ability only after RLVR. Ablations showed that CoB cold-start data preserves fine-grained grounding, while the CoB reward sustains grounding without materially changing accuracy.
- Closed-Ended Diagnostics: 0.91 accuracy under Domain Shift matched the dedicated classifier’s 0.89, whereas Real Distribution dominance remained CARDEA’s only statistically distinguishable shortfall.On Real Distribution, CARDEA reached 0.94 accuracy versus 0.97 for the 2D ConvNeXt; on Domain Shift, it reached 0.91 versus 0.89.
- Closed-Ended Diagnostics: 0.90 accuracy on complexity assessment matched the two cardiologists’ 0.90 and exceeded the third comparison of 0.88.CARDEA’s Macro F1 point estimate was slightly lower than the cardiologists’ values.
- Zero-Shot Generalization on AngioCAD: 0.85 Lesion F1 on held-out RCA stenosis classification exceeded the Adaptive Feature Fusion baseline’s 0.81, while CARDEA’s 0.90 RCA F1 trailed VGG19+LSTM’s 0.95.On the valid-views subset, CARDEA’s corresponding scores were 0.86 and 0.95, but that filter was not applied to baselines.
- Report Generation Across Training Stages: 0.686 two-class VS-F1 after RLVR exceeded the untuned base model’s 0.513 and more than doubled the always-normal floor of 0.312.Supervised stages reduced VS-F1 to 0.452 and 0.373; RLVR raised it to 0.644 without thinking and 0.686 with native thinking.
- Ablation Studies: About 0.76 median frame area without cold start versus about 0.01 with cold start showed that distilled fine-grained boxes determine grounding scale.Removing cold start slowed near-full CoB usage from about 30 to about 130 steps, but the reward eventually restored usage and accuracy remained comparable.
- Ablation Studies: About 5% grounding after removing the CoB reward, despite comparable accuracy, indicates the reward sustains spatial grounding rather than driving prediction accuracy.An unconditional reward left accuracy statistically indistinguishable from the selected conditional configuration.
4 Discussion
CARDEA’s discussion centers on a unified, auditable CAG pipeline whose spatial grounding depends on cold-start distillation and whose zero-shot report-generation gains emerged only after RLVR. Deployment remains bounded by inference time, out-of-distribution grounding gaps, limited validation, and the need for prospective clinical testing.
- CARDEA aims to unify study-level diagnosis and auditable bounding-box reasoning in one end-to-end CAG inference core.
- View filtering raised all three zero-shot point estimates, suggesting that greater view completeness improves information density beyond computational savings.
- RLVR was the only training stage that reversed the decline in zero-shot report-generation quality, whereas supervised stages eroded it.The interpretation remains hypothetical and rests on a single run.
- Deployment is constrained by about 6-second median inference time and lower CoB usage on domain-shifted imaging and report generation.Repeated sampling increased usage from 65% to 92% for domain-shift dominance and from 75% to 98.5% for report generation.
- Clinical readiness remains unestablished because CARDEA lacks clinical validation, blinded comparison with physician-authored reports, and prospective validation against expert cardiologists.
5 Conclusion
CARDEA provides an auditable end-to-end approach to coronary angiography interpretation and remains competitive with specialized baselines and cardiologists on selected tasks. Its feasibility is demonstrated, but clinical deployment still requires prospective expert validation.
- CARDEA links study-level diagnoses to auditable spatial evidence without human-annotated reasoning traces.
- CARDEA performed comparably to two interventional cardiologists on complexity assessment and generalized zero-shot to the held-out AngioCAD cohort.
- RLVR raised zero-shot report-generation performance above the untuned base model after supervised training had reduced it.
- Prospective validation against expert cardiologists remains essential before clinical deployment.
Ethics Approval
The study protocol received ethics approval with a waiver of informed consent, and analyses used only public, de-identified datasets.
- The Research Ethics Committee of China Medical University and Hospital approved the protocol with a waiver of informed consent.
- All analyses used publicly available, de-identified datasets and no hospital patient data.
Funding
The work was supported by Taiwan’s National Science and Technology Council and China Medical University Hospital.
- The National Science and Technology Council, Taiwan, supported the work through grant NSTC 114-2314-B-039-073.
- China Medical University Hospital provided support through grants DMR-115-087 and DMR-115-108.
Use of Artificial Intelligence
The study reports AI-assisted manuscript and code development while describing CARDEA’s data preparation across public angiography datasets.
- AI Use: The authors used ChatGPT and Claude to improve manuscript language, grammar, style, and parts of the experimental code.They reviewed the generated suggestions and retained responsibility for scientific accuracy and integrity.
- Data Preparation: CARDEA organizes its datasets into single-view and study-level tasks for end-to-end CAG reasoning.The single-view datasets support vessel, stenosis, and keyframe-related processing before study-level analysis.
- Data Preparation: ARCADE provides 3,000 keyframes from 1,500 patients with expert annotations for 25 coronary segments and stenotic-lesion localizations.These annotations are reformulated into vessel-detection and stenosis-detection tasks.
- Data Preparation: CADICA supplies 668 multi-view videos, including 382 expertly curated diagnostic videos, for keyframe-selection training.Videos are resized and long sequences are subsampled to bound computational cost.
- Data Preparation: An offline preliminary LVLM selects diagnostic keyframes, classifies views as LCA, RCA, or OTHER, and retains up to 10 keyframes per study.Each retained study is required to contain valid LCA and RCA views.
S1.1.3 Study-Level Datasets
The study-level datasets support dominance, complexity, keyframe, detection, and report-generation evaluation, using task-specific labels and metrics.
- Study-Level Datasets: CoronaryDominance contains 1,574 studies labeled Left or Right dominance, with Real Distribution and Domain Shift test sets.The Domain Shift set contains 149 studies acquired with distinct imaging equipment.
- Study-Level Datasets: CardioSyntax converts continuous SYNTAX scores into normal-to-intermediate complexity and high-complexity classes for binary assessment.The threshold is 32: scores 0–32 versus scores above 32.
- Study-Level Datasets: AngioCAD is excluded from training and serves as a zero-shot benchmark for view classification, RCA stenosis classification, and report generation.Its report task covers the LM, LAD, LCX, and RCA branches.
- Metrics: Keyframe quality is measured by Dk, the minimum absolute index distance from the predicted optimal frame to any expert-annotated usable frame.Lower Dk is better, and zero means the prediction falls within the usable range.
- Metrics: Detection uses one-to-one matching with same-label boxes and IoU≥0.5, reported as F1@IoU≥0.5.For stenosis detection, an additional predicted-area criterion yields F1@(IoU≥0.5 or IoP≥0.6) because box extent can understate localization quality.
- Metrics: Report generation is scored with vessel-severity macro-F1 after branch-level severity extraction from generated reports.Fifteen segments are grouped into four branches, and macro-averaging weights each severity class equally.
- Repeated Sampling: Repeated sampling evaluates CoB coverage and task performance by selecting the first rollout containing a valid bounding box among up to seven rollouts.Domain Shift CoB usage rises from 65.1% at k=1 to 92.0% at k=7, while accuracy changes from 90.6% to 91.3%.
S3.1 Training and Data Limitations
Training data and procedures constrain CARDEA’s coverage of angiographic findings, temporal dynamics, labeling regimes, and training-run variability.
- Data Coverage: Public training datasets mainly annotate stenosis location and severity, so CARDEA may not fully capture other findings and procedural contexts.Examples include calcification, thrombus, dissection, bifurcation involvement, equipment, and implanted devices.
- Temporal Coverage: Training studies were limited to 10 static keyframes, supporting cross-view reasoning but not contrast-flow dynamics.Potential preferences for projection angles or contrast densities were not measured for downstream effects.
- Label Definitions: AngioCAD labels any visible narrowing above 0%, whereas ARCADE uses a ≥50% stenosis threshold.The study did not test whether prior exposure to sub-50% CADICA labels affected AngioCAD performance.
- Training Variability: CARDEA used one training run per stage, so bootstrap confidence intervals exclude variation across training runs.The reported intervals capture test-set sampling variability only.
S3.2 Evaluation Limitations
Evaluation is limited by small or selected test sets, preprocessing and output conversions, parser uncertainty, and coarse report-quality metrics.
- Test-Set Scope: Keyframe selection and complexity assessment use small test sets that may be underpowered and should be considered preliminary.The test sets include 48 videos from 5 patients and 60 studies, respectively.
- Selection Bias: The valid-views subset may introduce selection bias by excluding studies in which CARDEA failed to identify both LCA and RCA views.Results on the complete cohort are treated as primary.
- Label Extraction: Report-generation evaluation depends on MedGemma for label extraction, and its agreement with human annotators on ambiguous reports was not measured.Ground-truth labels and final F1 are otherwise derived deterministically.
- Report Metric: VS-F1 measures branch-level severity agreement after collapsing 15 segments into four branches and omits expression quality and human-expert reference.It can score a report as correct despite wrong segment assignment or omitted additional lesions within a branch.
S3.3 Model and Reasoning Limitations
CARDEA’s reasoning coverage and spatial evidence have important limitations, and its operational performance depends on hardware. Clinical deployment remains out of scope until prospective expert validation is completed.
- Reasoning coverage: CoB coverage was lower under domain shift and during zero-shot report generation, although repeated sampling partly recovered coverage without materially changing task performance.The authors recommend evaluating coverage on data representative of the intended deployment setting.
- Spatial evidence: Generated boxes expose spatial claims for inspection, but their faithfulness to the named coronary structures was not verified.Clinical supervision is required to assess whether boxes accurately localize the structures they name.
- Operational constraints: End-to-end processing took a median of roughly 6 seconds per study on four NVIDIA B200 GPUs, but full CoB decoding may be substantially slower on less capable hardware.The reported timing is hardware-dependent.
- Clinical readiness: CARDEA has not been validated in a clinical setting, so prospective blinded comparison with physician-authored reports is required before clinical use.The proposed validation target is domain experts.