Neuroscientists at the University of Michigan have introduced Prima, a groundbreaking vision language model (VLM) designed to revolutionize neuroimaging diagnostics. Published in Nature Biomedical Engineering, the research demonstrates how this AI foundation model can identify more than 50 neurological disorders—ranging from tumors to strokes—in mere seconds. By achieving a remarkable 97.5 percent accuracy rate, Prima signals a shift toward more efficient, data-driven healthcare where rapid diagnosis can save lives.

Unlike previous diagnostic tools that rely on highly curated datasets, Prima excels by processing large-scale, uncurated clinical data. The model was trained using a massive data engine comprising 5.6 million three-dimensional sequences from 220,000 MRI studies. This extensive training allows the system to function as a general-purpose solution, integrating clinical context and imaging sequences to produce a comprehensive digital representation (vector representation) of a patient's brain health.

Beyond simple identification, Prima acts as a critical triage tool by prioritizing the severity of conditions. It is capable of sending automated alerts to clinicians when it detects acute issues like intracranial hemorrhages or strokes, ensuring a faster response time than traditional human review. Future iterations of the framework aim to incorporate electronic health records and even generate automated medical reports, potentially expanding its utility to other modalities like ultrasound and CT scans.

Neuroscientists at the University of Michigan have introduced Prima, a groundbreaking vision language model (VLM) designed to revolutionize neuroimaging diagnostics. Published in Nature Biomedical Engineering, the research demonstrates how this AI foundation model can identify more than 50 neurological disorders—ranging from tumors to strokes—in mere seconds. By achieving a remarkable 97.5 percent accuracy rate, Prima signals a shift toward more efficient, data-driven healthcare where rapid diagnosis can save lives.

Unlike previous diagnostic tools that rely on highly curated datasets, Prima excels by processing large-scale, uncurated clinical data. The model was trained using a massive data engine comprising 5.6 million three-dimensional sequences from 220,000 MRI studies. This extensive training allows the system to function as a general-purpose solution, integrating clinical context and imaging sequences to produce a comprehensive digital representation (vector representation) of a patient's brain health.

Beyond simple identification, Prima acts as a critical triage tool by prioritizing the severity of conditions. It is capable of sending automated alerts to clinicians when it detects acute issues like intracranial hemorrhages or strokes, ensuring a faster response time than traditional human review. Future iterations of the framework aim to incorporate electronic health records and even generate automated medical reports, potentially expanding its utility to other modalities like ultrasound and CT scans.