Addressing the Generalizability of AI in Radiology Using a Novel Data Augmentation Framework with Synthetic Patient Image Data: Proof-of-Concept and External Validation for Classification Tasks in Multiple Sclerosis

Generative adversarial network–based synthetic data augmentation during model training substantially improved performance of artificial intelligence models when applied to external MRI data and may be extended to a variety of clinical tasks to reduce domain shift on unseen data.