Glonce

CARDEA vision-language model adds auditable reasoning to coronary angiography

researchThe Glonce editorsHugging Face🔥65Glonce rating65Impact60Novelty65Relevance85Surprise50Players25Credibility85
CARDEA vision-language model adds auditable reasoning to coronary angiography

Researchers developed CARDEA, a large vision-language model built to serve as the reasoning core of an end-to-end pipeline for interpreting invasive coronary angiography (CAG), the gold-standard imaging test for coronary artery disease. The authors note that CAG interpretation varies substantially between observers, and that existing AI systems for the task can improve consistency but lack auditable decision processes and are limited in open-ended assessment, which undermines clinician trust and clinical adoption readiness.

CARDEA was trained solely on public datasets and on 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) that included a CoB reward encouraging the model to use bounding boxes in its reasoning trace. That bounding-box use is how the model exposes auditable spatial evidence behind its conclusions.

The authors evaluated two study-level diagnostic tasks, dominance classification and complexity assessment, against a dedicated classifier and two interventional cardiologists. On dominance classification, CARDEA trailed the dedicated classifier on data similar to its training distribution, but drew level with it under domain shift on an external cohort, reaching 0.91 accuracy (95% CI, 0.86 to 0.95). On complexity assessment, CARDEA's accuracy of 0.90 (CI, 0.82 to 0.97) was comparable to the two cardiologists.

Report generation, the model's open-ended output, was deliberately excluded from training and instead evaluated zero-shot across all three training stages on an external cohort, scored by vessel-severity macro-F1. Only the RLVR stage improved this zero-shot report-generation ability: it raised macro-F1 to 0.686 (CI, 0.664 to 0.707), above the untuned base model's 0.513 and over twice the 0.312 scored by an always-normal floor baseline. The authors conclude that RLVR on verifiable, closed-ended tasks surfaced open-ended reporting ability that supervised imitation training did not.

CARDEA runs the full pipeline end to end, from raw multi-view angiography videos through keyframe selection to study-level diagnosis, exposing the spatial evidence behind each conclusion for audit. The authors state that clinical use requires prospective validation against expert cardiologists.

Key facts

  • CARDEA trailed a dedicated classifier on in-distribution dominance classification but drew level with it under domain shift, reaching 0.91 accuracy (95% CI, 0.86 to 0.95) on an external cohort.
  • On complexity assessment, CARDEA's 0.90 accuracy (CI, 0.82 to 0.97) was comparable to two interventional cardiologists.
  • Only the RLVR training stage improved zero-shot report generation, lifting vessel-severity macro-F1 to 0.686 (CI, 0.664 to 0.707) from the untuned base model's 0.513, over twice the always-normal floor's 0.312.
  • CARDEA was trained solely on public datasets and closed-ended tasks, in three stages: visual feature alignment, a self-distilled Chain-of-Box cold start, and RLVR with a bounding-box reward.
  • The model runs the whole angiography pipeline end to end, from raw multi-view video through keyframe selection to diagnosis, exposing spatial evidence for each conclusion; the authors say clinical use still requires prospective validation against expert cardiologists.

Why it matters

Coronary angiography readings vary between observers, and the authors argue that existing AI aids improve consistency without giving clinicians an auditable decision process, which limits trust and adoption. CARDEA's contribution is that it exposes the spatial evidence, bounding boxes in its reasoning trace, behind each conclusion, and it runs the entire pipeline end to end from raw video to diagnosis rather than solving one narrow sub-task. The finding that only the RLVR training stage, not the earlier supervised stages, unlocked better zero-shot report generation is also a methodological result that could inform other clinical vision-language systems built on verifiable rewards.

Who it affects

Interventional cardiologists and other clinicians who read coronary angiograms, and researchers building clinical vision-language models, are the direct audience. Patients undergoing CAG are the eventual beneficiaries if the approach holds up, but the authors themselves frame CARDEA as a research system, not one ready for bedside use.

How to use it

The paper does not describe a release, code, weights or licence, and does not disclose model size, compute cost or inference speed. It states only that CARDEA was trained solely on public datasets. There is no pricing or deployment information in the source to report.

How solid is it

The evaluation reports accuracy with 95% confidence intervals against two separate baselines, a dedicated classifier and two interventional cardiologists, and tests report generation on an external cohort to check for domain shift rather than only in-distribution performance. That said, the human comparison group is just two cardiologists, and the source gives no dataset names, sizes or institutional provenance, so the size and diversity of the evaluation cohort cannot be assessed from the paper alone.

Risks and caveats

The authors state plainly that clinical use requires prospective validation against expert cardiologists, meaning CARDEA has not yet been tested in a forward-looking clinical trial. The source discloses no author affiliations, no dataset names or sizes, no timeline for that validation, and no explanation of what dominance classification or complexity assessment mean clinically beyond their names, so how the results would translate into a real diagnostic workflow remains unclear.

“RLVR on verifiable closed-ended tasks surfaced open-ended reporting ability that supervised imitation did not.”

— the paper's authors