Marco Parola

Raw generative augmentations frequently degrade medical image classifiers, but filtering synthetic samples for visual realism and class fidelity reliably yields up to +2.38% accuracy gains across CNN and ViT backbones.