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

Researchers have introduced new frameworks for 3D brain MRI and liver MRI report generation, leveraging advancements in artificial intelligence and latent diffusion models.

A team of researchers has developed BrainG3N, a dual-purpose tokenizer for controllable 3D brain MRI generation, pre-trained on 35,309 volumes from 18 public cohorts^[1]. BrainG3N decouples encoder and decoder functions, with a frozen 3D MAE encoder producing clinically informative embeddings. Meanwhile, a conditional diffusion transformer (DiT) trained on these embeddings supports both conditional generation and patient-specific longitudinal forecasting. In a related development, researchers have proposed PIRTA, a system that retrieves clinically similar 3D DWI/ADC volumes using a pretrained 3D vision encoder and leverages paired clinician-authored reports to ground large language model-based report generation^[2]. Additionally, a new framework called MRI2Rep has been introduced for liver MRI report generation, achieving 76.0% case-level sensitivity and 82.4% liver-level accuracy. MRI2Rep uses a Report-to-Label Canonicalization (RLC) module to convert free-text reports into structured diagnostic sequences and outperforms adapted medical vision-language baselines with 76.0% case-level sensitivity compared to 8.3%^[3]. Two radiologists rated 75% and 70% of AI-generated reports from MRI2Rep as clinically acceptable.