Synthetic 3D MRI Generation
3D Brain MRI Synthesis using DiT-3D | Graduate Student Researcher, UCSD Health
- Developed a conditional latent diffusion pipeline by integrating a VAE for latent space compression with a 3D diffusion transformer (DiT-3D), enabling the generation of high-fidelity, class-specific synthetic 3D brain MRIs with FID of 16.9
- Augmented the real brain MRI dataset with synthetic data, leading to substantial improvements in F1-score, accuracy, AUC-ROC, and AUCPR for a temporal-lobe epilepsy classifier trained using an EfficientNet-V2 backbone