MRI-CORE

MRI-CORE is a vision foundation model trained on over 6 million MRI slices from 110,000+ volumes across 18 body regions to address the challenges of limited labeled data in medical imaging. It significantly outperforms state-of-the-art methods in 13 data-restricted segmentation tasks, image classification, and zero-shot segmentation, demonstrating strong data efficiency and transfer learning potential. The model, released with a permissive license, also includes insights on effective pre-training strategies and how data similarity affects downstream performance.