3D Multi-Contrast Self-Attention GAN for Brain MRI Synthesis

A research paper proposes 3D-MC-SAGAN, a generative adversarial network for synthesizing missing brain MRI modalities from a single T2-weighted input. The model uses a multi-scale 3D encoder-decoder with residual connections and a Memory-Bounded Hybrid Attention block to capture long-range dependencies while preserving tumor characteristics. Training employs a WGAN-GP critic and auxiliary domain classification. The work addresses the practical limitation of acquiring all MRI contrasts (T1c, T1n, T2w, T2f) for every patient due to scan time and cost. The paper is available on arXiv under ID 2604.00070.