Learning neuroimaging models from health system-scale data.

Area of Science:

• Artificial Intelligence in Medicine
• Neuroimaging Analysis
• Machine Learning for Healthcare

Background:

• Increasing demand for magnetic resonance imaging (MRI) strains healthcare systems, leading to delays and physician burnout.
• Disparities in healthcare access disproportionately affect patients in low-resource and rural areas.
• Existing AI models lack comprehensive capabilities for real-world clinical neuroimaging data.

Purpose of the Study:

• To develop and validate Prima, the first vision language model (VLM) designed as an AI foundation for clinical neuroimaging.
• To leverage a large academic health system's data for training a robust AI model.
• To assess Prima's performance in supporting real-world, clinical MRI studies.

Main Methods:

• Trained Prima on over 220,000 clinical MRI studies using a hierarchical vision architecture.
• Conducted a 1-year prospective, health system-wide study involving 30,000 MRI studies.
• Evaluated Prima's diagnostic performance across 52 radiologic diagnoses for major neurologic disorders.

Main Results:

• Prima achieved a mean diagnostic area under the ROC curve of 90.1 ± 5.0% across diverse neurological conditions.
• Outperformed state-of-the-art general and medical AI models in diagnostic accuracy.
• Demonstrated algorithmic fairness across sensitive groups, showing potential to mitigate health system biases.

Conclusions:

• Prima represents a transformative advancement in AI-driven neuroimaging, offering explainable diagnoses and workflow prioritization.
• The VLM can help reduce health system biases, improving turnaround times for underserved populations.
• Prima has the potential to significantly enhance AI-driven healthcare delivery in neuroimaging.