The abstract is read at the publisher; the summary is JClub's.
ReCaRL improves multi-label chest radiograph classification, reaching internal macro AUROC 0.821 and macro AUPRC 0.437, modestly above the masked/weighted baseline.
- Why it matters: Chest X-ray labels from reports can be noisy and co-occurrence patterns shift across settings, while rigid label hierarchies may mix categorical membership with empirical associations.
- What they did: ReCaRL uses an image encoder for initial finding scores and a training-label association graph with reliability-weighted, confidence-gated updates, using observation masks for uncertain annotations.
- The result: Internal results were 0.821/0.437 versus 0.812/0.423 for the masked/weighted baseline, and external testing gave AUROC 0.799 and AUPRC 0.409; post-hoc calibration reduced Brier score and expected calibration error.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.