NeuroVFM: Vol-JEPA-Based Neuroimaging Foundation Model Without Radiology Report Annotation

NeuroVFM is a universal neuroimaging foundation model developed by the University of Michigan, trained on 5.24 million clinical MRI and CT volumes. Its core architecture, Vol-JEPA, extends I-JEPA and V-JEPA to 3D medical images, learning brain anatomy and pathology directly from uncurated clinical images via self-supervised learning, without needing radiology report text labels. This allows efficient use of massive unlabeled clinical data. For AI developers and researchers in medical imaging, NeuroVFM provides a powerful pretrained base for transfer learning or fine-tuning, significantly reducing data and compute costs for training 3D models from scratch and accelerating downstream tasks like brain disease diagnosis and lesion segmentation.