cs.CVSep 7, 2026

CARDEA: Auditable Reasoning Grounded in Spatial Evidence for End-to-End Coronary Angiography Interpretation

Authors: Jia-Jen Lee, Shih-Yen Hou, Kee Koon Ng, Wei-Chun Wang, Shih-Sheng Chang

Organizations: Artificial Intelligence and Robotics Innovation Center, China Medical University Hospital, Taichung, Taiwan · Division of Cardiovascular Medicine, Department of Internal Medicine, China Medical University Hospital, Taichung, Taiwan · Department of Neurology, China Medical University Hospital, China Medical University, Taichung, Taiwan · Neuroscience and Brain Disease Center, China Medical University, Taichung, Taiwan · School of Medicine, China Medical University, Taichung, Taiwan

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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-F1F_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-F1F_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.

Figures & tables

Figure 1: End-to-end inference pipeline of CARDEA. A first pass filters non-diagnostic frames in parallel via keyframe selection and view classification; a second pass applies Chain-of-Box (CoB) reasoning over the curated multi-view keyframes to produce Dominance Classification, Complexity Assessment, and zero-shot Report Generation.
Figure 2: Cold-start data preparation for Stage 2 training. The Stage 1 model generates vessel and stenosis bounding boxes for each keyframe; a teacher model synthesizes CoB reasoning traces conditioned on these detections; only traces reaching the correct diagnostic conclusion are retained for generative fine-tuning.
LevelTaskMetricCARDEABaseline / Expert
FrameStenosis Det.F1F_{1} @IoU ≥\geq 0.50.37 (0.32–0.41)0.36±0.080.36\pm 0.08 a
F1F_{1} @(IoU ≥\geq 0.5 or IoP ≥\geq 0.6)0.60 (0.56–0.65)0.60 f
Vessel Det.F1F_{1} @IoU ≥\geq 0.5, 25 seg.0.50 (0.48–0.53)0.47 (0.45–0.50) b, †\dagger
F1F_{1} @IoU ≥\geq 0.5, 25 seg.0.50 (0.48–0.53)0.46 f
F1F_{1} @IoU ≥\geq 0.5, 11 seg.0.58 (0.54–0.61)0.57 (0.54–0.61) b, †\dagger
View Cls. (3-class)Accuracy0.99 (0.98–1.00)—
Table 1: Trained Closed-Ended Performance. Bold marks the better value in rows where the two are statistically distinguishable: the comparator’s 95% CI does not overlap CARDEA’s 95% bootstrap CI (a baseline reported without a CI is treated as a zero-width interval, so its point estimate must fall outside CARDEA’s CI). These intervals are not adjusted for multiple comparisons; the resulting distinctions should therefore be read as exploratory rather than as confirmatory hypothesis tests.
Task & ModelAccuracyTarget F1F_{1}Macro F1F_{1}
RCA Stenosis Classification — Full
Adaptive Feature Fusion a0.720.81 †\dagger—
CARDEA0.80 (0.77–0.82)0.85 (0.83–0.87) †\dagger0.77 (0.73–0.80)
RCA Stenosis Classification — Valid-views
CARDEA0.81 (0.78–0.84)0.86 (0.84–0.88) †\dagger0.78 (0.75–0.81)
View Classification — Full
Table 2: Zero-Shot Closed-Ended Performance on the External AngioCAD Dataset. Bold marks the better value in rows where the baseline’s point estimate lies outside CARDEA’s 95% bootstrap CI. These intervals are not adjusted for multiple comparisons; the resulting distinctions should therefore be read as exploratory rather than as confirmatory hypothesis tests.
 DiagnosticText similarity
StageVS- F1F_{1} (2-class)VS- F1F_{1} (3-class)ROUGE-LBERTScore
Full cohort ( n=412n{=}412 )
Naive (always-normal)0.312 (0.298–0.325)0.208 (0.199–0.217)0.810.77
Base Model (Qwen3-VL) [ 25 ]0.513 (0.481–0.544)0.361 (0.335–0.389)0.220.66
Stage 1 (Align)0.452 (0.427–0.478)0.309 (0.287–0.332)0.330.73
Stage 2 (Cold start)0.373 (0.351–0.396)0.253 (0.235–0.274)0.340.69
Table 3: Stage-Wise Evolution of Zero-Shot Report Generation Quality on AngioCAD. Bold marks CARDEA (Final) on the full cohort, the result reported throughout the paper.
Figure 3: Ablation of cold-start initialization and the CoB reward. A, Validation CoB usage across RLVR training. Without a CoB reward, usage briefly rises then collapses to ∼\sim 5%; both cold-start configurations climb to near 1.0 within ∼\sim 30 steps; without the cold start, usage stays near zero for ∼\sim 130 steps before rising. B, Distribution of in-trace box areas on the held-out test classification tasks (box-and-whisker, normalized to full-frame area). Without the cold start, boxes degenerate to near-whole-frame; all cold-start variants emit compact, vessel-scale boxes.
RLVR initializationAlignCold startCold startCold start
CoB rewardCond.NoneUncond.Cond.
Validation
Dominance0.800.820.770.84
Complexity Assess.0.710.750.710.76
Test
Dominance (Real)0.81 (0.75–0.86)0.81 (0.75–0.86)0.73 (0.68–0.79)0.76 (0.70–0.81)
Table 4: Ablation over the Cold Start and the CoB Reward. Bold marks the last column, the configuration selected for CARDEA.
Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

SettingValue
SFT: Align and Cold Start
AdaptationLoRA (rank 32, α=64\alpha=64 , dropout 0)
OptimizerAdamW
Learning rate1×10−41\times 10^{-4} for Align; 5×10−55\times 10^{-5} for Cold Start
Global batch size64
Training duration2 epochs per stage
Appendix
Table S1: Stage-specific training configuration.
Figure S2: Repeated-Sampling CoB Coverage and Task Performance. For each study, the first rollout with a valid in-trace bounding box among the first kk was selected; if none had one, rollout 1 was retained. A, CoB usage@ kk , the percentage of selected rollouts containing a box. B, Task performance@ kk , measured by exact-match accuracy for the classification tasks and two-class VS- F1F_{1} for report generation. Curves show CARDEA with thinking enabled at sampling temperature 0.5 on the official CoronaryDominance Real ( n=400n{=}400 ) and Domain Shift ( n=149n{=}149 ) test sets, the CardioSyntax Complexity test set ( n=60n{=}60 ), and zero-shot report generation on the full AngioCAD cohort ( n=412n{=}412 ); k=1k{=}1 matches the corresponding single-pass results in the main text.
Figure S3: Qualitative CARDEA inference example for coronary dominance classification. The example shows the multi-view input, the Chain-of-Box (CoB) reasoning trace with spatially grounded coronary structures, and the final dominance prediction.
Figure S4: Qualitative CARDEA inference example for anatomical complexity assessment. The example shows the multi-view input, the Chain-of-Box (CoB) reasoning trace with spatial grounding of coronary vessels and lesions, and the final complexity category.
Figure S5: Qualitative CARDEA inference example for structured report generation. The example shows the multi-view input, the Chain-of-Box (CoB) reasoning trace with vessel-level spatial grounding, and the resulting branch-level JSON report.

Explore similar work

CardsList

  1. DeepCORO-CLIP: A Multi-View Foundation Model for Comprehensive Coronary Angiography Video-Text Analysis and External Validation

    Mar 18, 2026Sarra Harrabi, Yichen Wu, Geoffrey H. Tison +6Coronary AngiographyMedical Imaging Foundation Models

    Division of Cardiology, Department of Medicine, Montreal Heart Institute, 5000 Belanger Street, QC, H1T 1C8 Canada · HeartWise.ai, 5000 Belanger Street, QC, H1T 1C8 Canada · McGill, Faculty of Engineering, Montreal, Quebec, Canada. +7

  2. ARIADNE: A Perception-Reasoning Synergy Framework for Trustworthy Coronary Angiography Analysis

    Mar 19, 2026Zhan Jin, Yu Luo, Yizhou Zhang +5Medical Image AnalysisCoronary Angiography

    School of Mathematical Sciences, Ocean University of China, Qingdao, 266100, Shandong, China. · Department of Cardiology, Qilu Hospital (Qingdao), Cheeloo College of Medicine, Shandong University, No. 758 Hefei Road, Qingdao, 266000, Shandong, China.

  3. CARDIAG: A Dense Segment Classification Benchmark of Deep Learning Architectures for Coronary Angiography

    Jul 24, 2026Dominik Bernard Lau, Hubert Malinowski, Jerzy Szyjut +5Coronary AngiographyMedical Image Segmentation

    Gdańsk University of Technology · NASK National Research Institute · Medical University of Gdańsk