Conditional Mammography Synthesis & Classification

This repository reproduces the experiments from:

Utility of AI-Generated Images in Disease Classification: A Study in Breast Cancer Mammography

It implements a full pipeline to:

• Construct a binary EMBED dataset (benign vs malignant)
• Train a class-conditional DDPM
• Generate synthetic mammograms (classifier-free guidance)
• Train an Oracle classifier (ConvNeXt-Tiny)
• Perform synthetic augmentation experiments
• Run two-phase training (real + synthetic → real fine-tune)
• Evaluate Balanced Accuracy, Sensitivity, Specificity, and FID