An assistive, non-autonomous multi-reader system for adult PA/AP chest radiographs, with honest per-finding operating points and a commercial-clean data-provenance path.
Retrospective validation only · not cleared for clinical use · decision support, confirmed by a qualified radiologist.
Findings
15
reader panel, each with a named operating point
Validation breadth
6
countries · 15 public datasets · 2 labelling regimes
Runs on
CPU
~6 s/scan · no GPU required
The headline
Featured panel — pleural effusion, pneumothorax, cardiomegaly, atelectasis, and a normal/abnormal triage gate; leave-site-out AUC 0.84–0.91, each shipped with a stated operating point.
It generalises — the featured panel holds at 0.82–0.95 AUC on cohorts never seen in training (a US community hospital and an independent multi-site benchmark).
Commercial-clean — a build trained only on openly-licensed data is available with no research-only encumbrance; several readers already run on those clean weights in production.
Honest by design — every finding reports sens / spec / PPV / NPV and its validation basis; single-source or weak findings are labelled as such, never hidden.