U-Net Wins Brain MRI Challenges With Efficient Models

Researchers led by Pedro M. Gordaliza report a U-Net–based approach that ranked first in the SSL3D and FOMO25 brain MRI tracks at MICCAI 2025. Their models incorporate anatomical priors and neuroimaging domain knowledge, training 10–100× faster while being roughly 10× smaller than competing transformer-based methods. Models and code are publicly available (arXiv submission Jan 19, 2026).