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

AI summary

Overview

Research area: Medical computer vision and clinical AI — specifically large vision-language models (LVLMs) applied to invasive coronary angiography (CAG), combining visual grounding, reinforcement learning with verifiable rewards (RLVR), and zero-shot report generation.

Technical level: Advanced. The paper assumes familiarity with vision-language models, supervised fine-tuning, reinforcement learning, bounding-box grounding, and standard detection/classification metrics.

Scope: The paper introduces CARDEA, a single large vision-language model that runs an end-to-end coronary angiography pipeline — from raw multi-view videos through keyframe selection to study-level diagnosis — while exposing auditable bounding-box "Chain-of-Box" evidence behind each conclusion.

What This Paper Is About

Invasive coronary angiography is the gold standard for diagnosing coronary artery disease, but reading angiograms is subjective: one cited study reports only 77.4% overall agreement among three experienced cardiologists reading the same images. Existing AI systems either output only closed-ended labels (discriminative models) or generate free-text narratives with no spatial anchors a clinician can inspect, so their decisions cannot be audited. CARDEA's goal is to be a single inference core that produces study-level diagnoses while grounding each conclusion in inspectable bounding boxes inside its reasoning trace — and to test whether training only on closed-ended, verifiable tasks can also surface open-ended reporting ability.

Key Contributions

1. A unified LVLM pipeline for CAG with auditable grounding. CARDEA performs the whole study itself in two passes — single-view keyframe selection and view classification, then study-level multi-image Chain-of-Box (CoB) reasoning over the curated keyframes — emitting bounding boxes in its reasoning trace that link conclusions to specific regions in the source images. The authors state no prior CAG system does this.

2. A three-stage training recipe built without human-annotated reasoning traces. The stages are (i) visual feature alignment on single-view tasks, (ii) a self-distilled Chain-of-Box cold start on dominance classification, where the untuned base model acts as the teacher generating CoB traces conditioned on Stage 1 boxes, and (iii) RLVR with a CoB reward that is only granted conditional on a correct final answer at the study level.

3. Evidence that RLVR on closed-ended tasks surfaces open-ended reporting ability that supervised imitation does not. Report generation was excluded from training entirely, yet it declined under the two supervised stages and only recovered after RLVR.

4. Public release of model weights and inference code (Hugging Face and GitHub links given in the data-sharing statement), with training performed solely on public datasets.

Main Findings

• Competitive on trained closed-ended tasks. CARDEA reached 0.37 on stenosis detection (vs. 0.36 for DCA-YOLOv8) and 0.50 on 25-segment vessel detection (vs. 0.47 for DeepCoro's Algorithm 4). Under the 11-segment convention it scored 0.58 vs. 0.57. It reached a Macro F1 of 0.99 on both the 3-class and 2-class view classification tasks (the 2-class vs. 1.00 for YOLOv8x-cls), and a mean frame distance D_k of 0.60 on keyframe selection, placing its predicted frame within one frame of the expert-annotated usable range on average — though this split is only 48 videos from 5 patients.

• The only statistically distinguishable shortfall was in-distribution dominance. On the Real Distribution dominance subset CARDEA scored accuracy 0.94 and Macro F1 0.88 vs. 0.97 and 0.94 for a 2D ConvNeXt. Under Domain Shift the two drew level (accuracy 0.91 vs. 0.89; Macro F1 0.89 vs. 0.88).

• Comparable to interventional cardiologists on complexity assessment. CARDEA reached accuracy 0.90 vs. 0.90 (Expert 1, 10 years) and 0.88 (Expert 2, 3 years), with a slightly lower Macro F1 point estimate of 0.80 vs. 0.83 and 0.81.

• Matches a prior LVLM under its own relaxed criterion. On stenosis detection, CARDEA's F1@(IoU ≥ 0.5 or IoP ≥ 0.6) of 0.60 matched Jiang et al.'s 0.60, up from 0.37 at the standard IoU threshold. Its vessel F1@IoU ≥ 0.5 of 0.50 exceeded their 0.46.

• Zero-shot generalization to the held-out AngioCAD cohort. CARDEA surpassed the Adaptive Feature Fusion baseline on RCA binary stenosis classification (Lesion F1 0.85 vs. 0.81; accuracy 0.80 vs. 0.72) but trailed VGG19+LSTM on view classification (RCA F1 0.90 vs. 0.95; accuracy 0.94 vs. 0.97). On the reference-only valid-views subset (n = 326) both rose (Lesion F1 0.86; RCA F1 0.95).

• Only RLVR improved zero-shot report generation. Vessel Severity Macro-F1 (VS-F1, 2-class) was 0.312 for an always-normal dummy, 0.513 for the untuned base model, then 0.452 (Stage 1) and 0.373 (Stage 2) under supervision, before RLVR raised it to 0.644 with thinking disabled and 0.686 (95% CI 0.664–0.707) with native thinking. The valid-views subset reached 0.716. The three-class score mirrored this trajectory (0.508 final).

• Text-similarity metrics can be misleading. The always-normal report outscored the final model on ROUGE-L (0.81 vs. 0.35) and BERTScore (0.77 vs. 0.74) despite its low VS-F1 of 0.312, because 74% of graded AngioCAD sub-segments were normal. Label extraction by MedGemma-27B-IT was stable: standard deviation of VS-F1 below 0.003 across five runs at temperature 0.5.

• The cold start governs box quality, not accuracy. Removing the cold start delayed CoB adoption to about 130 steps (vs. about 30), and without it in-trace boxes degenerated to a median 0.76 of the frame area versus about 0.01 with a cold start. Removing the CoB reward let grounding collapse to about 5% while accuracy stayed comparable. Making the reward unconditional was statistically indistinguishable at test, so the conditional form was retained as the stricter rule.

• Deployment trade-offs. Median inference time was about 6 seconds per study even under tensor parallelism on 4× NVIDIA B200 GPUs. CoB usage was 99.5% on Real Distribution dominance and 100% on complexity assessment, but fell to 65% on Domain Shift dominance and 75% on report generation; repeated sampling across seven rollouts raised these to 92% and 98.5% without materially changing task performance.

• Not yet clinically validated. There was no blinded comparison of CARDEA's reports with physician-authored reports, and the authors state prospective validation against expert cardiologists is a prerequisite for clinical use.

Methodology in Plain English

The team built CARDEA on top of an existing open-weight vision-language model (Qwen3-VL-30B-A3B-Thinking) and taught it to read coronary angiograms, using only publicly available data and only tasks with checkable right-or-wrong answers.

The system works in two passes. The first pass looks at each raw angiographic video and throws away non-diagnostic frames, picking a representative keyframe and labeling its view as left coronary artery, right coronary artery, or other. The second pass takes the curated multi-view keyframes from a whole study and reasons across them, producing a dominance classification, a complexity assessment, and a written report.

Training happened in three stages. Stage 1 aligned the model to angiography by fine-tuning it to box vessels and stenoses. Stage 2 taught it to use bounding boxes inside its reasoning trace (the Chain-of-Box behavior), which the base model does not do natively — since hand-writing reasoning traces is expensive, the researchers used self-distillation: the untuned base model acted as a teacher, generating reasoning traces conditioned on the Stage 1 boxes, and only traces that reached the correct conclusion were kept. Stage 3 used reinforcement learning with verifiable rewards to push accuracy on study-level tasks and reinforce the box-using behavior, with the box reward granted only when the final answer was correct.

Critically, they deliberately left open-ended report generation out of training, because such tasks have no rule-based verifier and would need a reward model prone to reward hacking. Report generation was therefore tested zero-shot to see whether it improved as a side effect of the other training.

Evaluation used public datasets reformulated into instructional tasks (ARCADE for vessel and stenosis annotation, CADICA for keyframe selection, CoronaryDominance and CardioSyntax for the study-level labels), with AngioCAD held out entirely as an external zero-shot benchmark. Baselines were published dedicated models evaluated on the same test sets; for complexity assessment the comparison was two interventional cardiologists with 10 and 3 years of experience. Confidence intervals came from 5,000 case-level bootstrap resamples, and the authors state their distinctions are exploratory because the intervals are not adjusted for multiple comparisons.

Why This Matters

Impact on research. The paper makes the case that reinforcement learning on verifiable closed-ended tasks can surface generative, open-ended clinical reporting ability that supervised imitation does not. In their experiments the two supervised stages actively eroded zero-shot report quality, and only RLVR reversed the decline — a finding with direct implications for how medical foundation models are trained. It also argues for treating spatial grounding as an auditability mechanism rather than an accuracy booster: removing the box reward collapsed grounding to about 5% while accuracy stayed comparable.

Real-world applications.

• Diagnostic decision support in the catheterization lab, where a cardiologist could inspect which coronary regions a model's conclusion was based on rather than trusting a black-box label.
• Second-reader consistency checks, given the reported 77.4% agreement among three experienced cardiologists reading the same angiograms.
• Triage and workload distribution across multi-view studies, using study-level dominance and complexity assessments to flag high-complexity cases.
• Structured report drafting on external cohorts, since CARDEA generated reports zero-shot across the four major branches (LM, LAD, LCX, RCA) without ever being trained on the task.

Industry relevance. The system runs from raw multi-view videos to study-level output without a cascade of separate models, which points to a single-model deployment pattern for angiography analysis. That said, the reported median of about 6 seconds per study on 4× NVIDIA B200 GPUs sets clear expectations: this cannot serve sub-second response applications. The measured drop in grounded-reasoning usage from 99.5% in-distribution to 65% under domain shift is a deployment warning — vendors should validate against the expected data distribution before release. Model weights and inference code are publicly released, lowering the barrier to replication and downstream work.

Future Directions

1. Prospective clinical validation. The authors explicitly name validation against expert cardiologists, with blinded comparison of CARDEA's reports to physician-authored reports, as a prerequisite before clinical use.

2. Closing the out-of-distribution grounding gap. Since CoB usage fell to 65% on Domain Shift dominance and 75% on report generation, the authors suggest that closing it fully may require further supervised fine-tuning on rejection-sampled trajectories from out-of-distribution data.

3. Testing the RLVR hypothesis beyond a single run. The paper's central explanation — that verified answers force the model to internalize multi-view synthesis — is described as hypothetical and resting on a single run per stage, so multi-run replication and cross-stage comparisons are open.

4. Broadening findings and temporal modeling. Current training data covers mainly vessels and stenoses, leaving other findings insufficiently learned, and study-level training uses at most 10 static keyframes per study without the temporal dynamics of contrast flow. Evaluating box faithfulness — not just box frequency, box area, and final-answer correctness — is also listed as an unmet need.

Target Audience

This paper is most useful to medical AI and clinical machine learning researchers working on vision-language models, reinforcement learning from verifiable rewards, or visual grounding in healthcare imaging. It will also interest interventional cardiologists and cath-lab informatics teams evaluating AI decision support, regulators and hospital procurement groups assessing auditability requirements for clinical AI, and ML engineers looking for a concrete worked example of combining self-distillation with verifiable-reward RL in a high-stakes medical domain. Readers without a background in LVLM training or detection metrics should expect a steep learning curve, though the clinical motivation and the central result — that RLVR, not supervision, unlocked zero-shot reporting — are stated in accessible terms.

Authors’ 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-$F_1$. 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.