BrainG3N Tokenizer Enables Controllable 3D Brain MRI Generation.

Three-dimensional (3D) brain MRI scans are indispensable in clinical neurology and neuro-oncology, with generative models holding promise for augmenting under-represented patient cohorts, simulating disease progression, and facilitating privacy-preserving data sharing. However, current latent diffusion models face a challenge: their tokenizers must simultaneously retain crucial clinical information for downstream tasks and accurately reconstruct anatomical volumes. Existing reconstruction-focused tokenizers often compromise the clinical utility of the embeddings. To overcome this, researchers introduce BrainG3N, a fully volumetric masked-autoencoder (MAE) based tokenizer specifically for 3D brain MRI latent diffusion. This innovative design decouples the encoder and decoder components: a frozen 3D MAE encoder generates clinically rich embeddings, while a separate CNN decoder reconstructs voxels from a linear projection of these embeddings. The encoder was pretrained on a vast dataset of 35,309 volumes from 18 public cohorts, encompassing diverse modalities, disease categories, and acquisition sites. BrainG3N demonstrates dual utility: its encoder either matches or surpasses state-of-the-art models on 21 out of 23 tasks in a linear-probing benchmark, proving its clinical informativeness. Furthermore, a conditional diffusion transformer (DiT) trained on these embeddings successfully supports conditional generation across six variables and patient-specific longitudinal forecasting, establishing a versatile 3D brain MRI embedding space for both clinical analysis and controllable generation.