Researchers at the University of Michigan have unveiled an artificial‑intelligence platform that can scan a brain MRI, interpret the findings, and flag urgent cases within a few seconds. In testing, the model correctly identified a wide range of neurological disorders with an overall accuracy of 97.5 %.
How the Prima System Was Evaluated
The team, led by Dr. Hollon, named the technology Prima. Over a twelve‑month period the system examined more than 30,000 MRI studies covering over 50 distinct radiologic diagnoses. Compared with several existing AI solutions, Prima consistently delivered higher diagnostic scores and, crucially, could rank cases by urgency.
When a scan shows signs of a stroke, a brain bleed, or another life‑threatening condition, Prima immediately alerts the appropriate specialist—often a stroke neurologist or neurosurgeon—so that treatment can begin without delay.
What Sets Prima Apart
Prima belongs to the class of vision‑language models (VLMs), which blend image analysis with textual reasoning in real time. Earlier AI tools for MRI were trained on narrow, hand‑picked datasets and focused on single tasks, such as spotting lesions. By contrast, Prima was fed every brain MRI collected at the university’s health system since digital records began, amounting to more than 200,000 scans and 5.6 million individual imaging sequences.
The model also ingests the ordering physician’s notes and the patient’s brief clinical history. This richer context enables Prima to make more nuanced predictions across a broad spectrum of conditions.
Why Speed Matters in Neuro‑Radiology
Millions of brain MRIs are performed worldwide each year, and the demand for neuroradiology expertise is outpacing the supply of trained specialists. The resulting bottleneck can lead to delayed reports, longer hospital stays, and, in worst‑case scenarios, missed therapeutic windows for stroke patients. Prima’s ability to deliver an instant, high‑confidence read could alleviate these pressures by triaging the most critical cases for immediate human review.
Looking Ahead
Although the current results are promising, the researchers stress that Prima is still in an early validation stage. Future work will integrate deeper electronic‑medical‑record data and explore how the system performs on other imaging modalities such as mammograms, chest X‑rays, and ultrasounds.
Dr. Hollon likens Prima to a “ChatGPT for medical imaging,” envisioning it as a collaborative partner that assists radiologists rather than replaces them. By acting as a co‑pilot, the AI could free clinicians to focus on complex decision‑making while ensuring no urgent case slips through the cracks.
Funding and Acknowledgments
The study was supported in part by the National Institute of Neurological Disorders and Stroke (K12NS080223), the Chan Zuckerberg Initiative, the Frankel Institute for Heart and Brain Health, and several private foundations.