Health system learning enables generalist neuroimaging models.

Area of Science:

• Medical imaging and artificial intelligence
• Clinical informatics
• Foundation models in healthcare

Background:

• Frontier AI models trained on public data struggle with private clinical data, particularly neuroimaging due to privacy concerns.
• Neuroimaging data (MRI, CT) is underrepresented in public datasets, limiting AI model performance in clinical settings.
• Existing AI models underperform on specialized neuroimaging tasks.

Purpose of the Study:

• To introduce 'health system learning,' a paradigm for training AI on routine clinical data.
• To develop and evaluate NeuroVFM, a visual foundation model for clinical neuroimaging.
• To demonstrate the capability of health system learning to create high-performance, generalist medical AI.

Main Methods:

• Trained NeuroVFM, a visual foundation model, on 5.24 million clinical MRI and CT volumes using a scalable volumetric predictive architecture.
• Employed a 'health system learning' approach, utilizing uncurated data from routine clinical care.
• Integrated NeuroVFM with open-source language models for radiology report generation.

Main Results:

• NeuroVFM achieved state-of-the-art performance across multiple clinical neuroimaging tasks, including diagnosis and report generation.
• The model learned comprehensive representations of brain anatomy and pathology, embedding MRI and CT scans into a shared latent space.
• AI-generated radiology reports using NeuroVFM surpassed frontier models in accuracy, clinical triage, and expert preference, with reduced errors.

Conclusions:

• Health system learning is a viable paradigm for building generalist medical AI, particularly for neuroimaging.
• NeuroVFM demonstrates the potential of foundation models trained on clinical data for advancing diagnostic accuracy and efficiency.
• This approach offers a scalable framework for developing safer and more effective clinical AI tools.