Health system learning achieves generalist neuroimaging models.

Area of Science:

• Medical Artificial Intelligence
• Neuroimaging Analysis
• Foundation Models

Background:

• Frontier AI models trained on public data lack clinical data access.
• Neuroimaging data is underrepresented in public datasets due to privacy concerns.
• This limits the performance of current AI in clinical medicine.

Purpose of the Study:

• To evaluate frontier AI model performance on neuroimaging tasks.
• To introduce and validate a new paradigm: health system learning.
• To develop a high-performance, generalist neuroimaging AI model.

Main Methods:

• Developed NeuroVFM, a visual foundation model trained on 5.24 million clinical MRI and CT volumes.
• Utilized a scalable volumetric joint-embedding predictive architecture.
• Employed lightweight visual instruction tuning to pair NeuroVFM with language models.

Main Results:

• NeuroVFM achieved state-of-the-art performance across neuroimaging tasks, including diagnosis and report generation.
• The model demonstrated emergent neuroanatomic understanding and interpretable visual grounding.
• NeuroVFM-generated reports surpassed frontier models in accuracy, triage, and expert preference, with reduced errors.

Conclusions:

• Health system learning is a viable paradigm for creating generalist medical AI.
• NeuroVFM offers a scalable framework for clinical foundation models.
• This approach enhances AI safety and clinical decision support in healthcare.