BrainG3N: A Dual-Purpose Tokenizer for Controllable 3D Brain MRI Generation

The article introduces BrainG3N, a dual-purpose tokenizer designed for the generation of controllable 3D brain MRI using a volumetric masked-autoencoder (MAE) architecture. This approach decouples the encoder, which generates clinically informative embeddings from a pretrained model on 35,309 volumes, and a CNN decoder for voxel reconstruction, achieving superior performance on a 23-task linear-probing benchmark compared to state-of-the-art models. This development is significant for practitioners as it enables both enhanced clinical task performance and the capability for conditional generation and patient-specific forecasting in neuroimaging applications.