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

CARDEA. Auditable Reasoning Grounded in Spatial Evidence for End-to-End Coronary Angiography Interpretation. By Jia-Jen Lee and four others, from China Medical University Hospital and China Medical University. Published on the seventh of September. Its code repository is public and real: about twenty-five source files.

Invasive coronary angiography, or CAG, is the gold standard for diagnosing coronary artery disease, yet the way it is read varies substantially from one observer to the next. AI systems can make interpretation more consistent. But the existing ones lack decision processes that can be audited, so a clinician cannot see why a conclusion was reached. They are also limited when it comes to comprehensive, open-ended assessment. Both gaps undermine the trust of clinicians, and both hold back how ready such systems are for clinical adoption.

CARDEA is a unified large vision-language model that serves as the inference core of a whole CAG pipeline. Its defining feature is that its reasoning is grounded in spatial evidence. It exposes bounding boxes behind its conclusions, so the path from image to judgement can be inspected rather than taken on trust. That is what the title means by auditable reasoning.

The model was trained only on public datasets and only on closed-ended tasks, in three stages. First comes visual feature alignment. Second comes a cold start on self-distilled Chain-of-Box reasoning, CoB for short. Third comes reinforcement learning with verifiable rewards, RLVR, which includes a CoB reward that encourages the model to use bounding boxes in its reasoning trace. In use, the pipeline runs end to end: from raw multi-view videos, through keyframe selection, to a diagnosis at the level of the whole study. Each stage builds on the last: first the model learns to see the images, then it learns to point at what it sees while reasoning, and finally it is rewarded for getting verifiable answers right.

Two study-level diagnoses were assessed, dominance classification and complexity assessment, against a dedicated classifier and against two interventional cardiologists. On dominance, CARDEA trailed the classifier in distribution, but drew level under domain shift, with an accuracy of zero point nine one and a confidence interval from zero point eight six to zero point nine five. On complexity it was comparable to the cardiologists, at an accuracy of zero point nine, with an interval from zero point eight two to zero point nine seven.

Report generation was never part of training. It was evaluated zero-shot, on an external cohort, after each training stage, scored by vessel-severity macro F one. Only the RLVR stage improved it. The score rose to zero point six eight six, above the untuned base model at zero point five one three, and more than twice the floor of zero point three one two set by always answering normal.

One result concerns where the reporting ability came from. Reinforcement learning on verifiable, closed-ended tasks surfaced an open-ended reporting skill that supervised imitation did not. The model was never taught to write reports, and it got better at writing them anyway. The authors are explicit about the limit: clinical use still requires prospective validation against expert cardiologists.