Meet NeuroVFM: A New Neuroimaging Foundation Model Trained With Vol-JEPA on Uncurated Clinical MRI and CT Volumes

NeuroVFM is a foundation model trained on millions of clinical brain scans (MRI and CT volumes) to learn patterns in neuroimaging without needing paired radiology reports or labeled data. It uses a method called Vol-JEPA, which predicts learned representations of masked regions in 3D medical images rather than reconstructing pixels, making it more efficient than traditional approaches. The model matters because general AI models perform poorly on brain imaging tasks, since MRI and CT scans with identifiable facial features are rarely available in public training datasets. Beyond diagnosis, NeuroVFM enables downstream applications including report generation, clinical triage, and cross-modal transfer between imaging types, though developers frame it as decision support rather than autonomous screening.