A new visual foundation model called NeuroVFM has been trained on 5.24 million clinical MRI and CT volumes, creating a generalist neuroimaging AI that outperforms frontier models on diagnostic tasks. The research, published in Nature Medicine, introduces a paradigm called “health system learning.”
Neuroimaging is underrepresented in public training data because identifiable facial features in MRI and CT scans raise privacy concerns. The team behind NeuroVFM trained directly on uncurated clinical data from health systems, allowing the model to learn comprehensive representations of brain anatomy and pathology.
The model achieves state-of-the-art performance across multiple clinical tasks including radiologic diagnosis and report generation. When paired with open-source language models, NeuroVFM generates radiology reports that surpass frontier models in accuracy, clinical triage, and expert preference.
The researchers found that NeuroVFM produces fewer hallucinated findings and critical errors, offering safer clinical decision support than existing approaches.
