CARDEA Makes Coronary Angiography AI More Auditable

Invasive coronary angiography remains a key reference for assessing coronary artery disease, yet its interpretation can vary across observers. Many AI systems reduce this variation by producing a final label or score. Their weakness is that the route to that answer is usually opaque: clinicians cannot easily tell which vessel segment the model used, whether it overlooked an important region, or whether the explanation was added after the prediction.

CARDEA is designed around that auditability problem. It is a unified large vision-language model for coronary angiography interpretation. Before producing its final answer, the model is encouraged to connect its reasoning trace to several image regions. The authors call this mechanism Chain-of-Box, or CoB. In practice, the model uses bounding boxes as visual evidence alongside textual reasoning. Such localization can make an output easier to inspect, although a highlighted region should not automatically be treated as a clinically correct explanation.

Key points

• Three-stage training: visual feature alignment first, followed by self-distillation to create CoB reasoning traces for cold-start training, and finally reinforcement learning with verifiable rewards. The last stage rewards both task correctness and the use of box-grounded reasoning.
• Public-data training: the model was trained only on public datasets and closed-ended tasks. Report generation was excluded from training, allowing the researchers to test whether the learned capability transfers to an unseen open-ended setting.
• Multiple evaluation targets: the study examines coronary dominance classification and complexity assessment, then evaluates zero-shot report generation on an external cohort using vessel-severity macro-F1.
• Different strengths across settings: on in-distribution dominance classification, CARDEA trails a dedicated classifier. Under domain shift, the gap narrows, with reported accuracy of 0.91 and a 95% confidence interval of 0.86 to 0.95. For complexity assessment, its accuracy is 0.90, with a confidence interval of 0.82 to 0.97, comparable to two interventional cardiologists. For zero-shot report generation, the largest improvement appears only after RLVR: vessel-severity macro-F1 reaches 0.686, compared with 0.513 for the untuned base model.

The broader contribution is a training recipe that treats visual evidence as part of the output behavior rather than as an after-the-fact visualization. If the predicted finding and the selected vessel region align, clinicians may be able to identify missed anatomy or unsupported conclusions more quickly. This also makes the evaluation more demanding: a fluent rationale is not enough if its boxes point to irrelevant image content.

The results should nevertheless be read as research evidence, not as proof of clinical readiness. CARDEA does not outperform a specialized classifier on every task, and a better report score does not establish diagnostic safety. Further work is needed to test localization accuracy, reasoning stability, equipment and acquisition differences, and broader patient populations. Prospective workflow studies would also be necessary before such a system could be considered for routine decision support.

The authors have released model weights and inference code, which creates an opportunity for independent replication and auditing. CARDEA therefore represents a useful step toward evidence-grounded medical vision-language models, while leaving the central question open: whether a trace that is easier to inspect is also reliably correct in real clinical practice.

Source: Hugging Face Daily Papers