Medical imaging has a bottleneck, and it's not the scanner. Brain MRIs can sit in a queue for hours before a radiologist reviews them. In a stroke, that wait is the difference between a good outcome and a catastrophic one.
Researchers at the University of Michigan have built Prima — a vision-language model trained to read brain MRI scans across 52 neurological diagnoses, from acute strokes and hemorrhages to less urgent findings. In a year-long evaluation of nearly 30,000 MRI studies, it achieved a mean AUC of 92%, with accuracy reaching 97.5% on certain diagnoses. When it detects something urgent, it doesn't just flag the scan — it routes an alert to the appropriate specialist. A suspected stroke goes to the stroke neurologist. A possible hemorrhage goes to the neurosurgeon.
A Co-Pilot, Not a Replacement
The Michigan team is explicit that Prima works alongside radiologists, not in place of them. In practice, this means it acts as a triage layer: catching urgent cases early, reducing the time between scan and specialist contact, and handling the volume of routine reads that would otherwise compete for a radiologist's attention.
This co-pilot framing has become standard for medical AI systems — partly for regulatory reasons, partly because it's accurate. The performance numbers are strong, but a 2.5% to 8% error range matters enormously in a clinical setting. Radiologist oversight, for now, is not optional.
Why This Is Harder Than It Looks
Building an AI that reads MRIs accurately is only part of the challenge. Getting it integrated into clinical workflows, trusted by physicians, and compliant with hospital IT and regulatory standards is where most of these systems stall. The Michigan team evaluated Prima over a full year and nearly 30,000 studies — not a controlled lab benchmark, but real operational data — which makes this more credible than the typical medical AI announcement.
No commercial timeline has been announced. Full details are at Michigan Medicine.
Commentarii · 0