From Augmentation to Synthetic Evidence: Evaluating Generated Chest Radiographs at Dataset Intake

Evaluating synthetic chest radiographs requires distinguishing between augmentation and medical evidence via automated quality checks and origin classifiers. Using MedGemma for quality assessment and DINOv3 features for origin detection, selecting top-tier images improves quality pass rates from 12 to 19 for Cheff-identified candidates but fails to improve simultaneous passage of both metrics.