Radiologists are expensive, tired, and in short supply. A brain MRI study can wait hours for a qualified physician to review it - hours that are catastrophic when the scan shows a hemorrhage or an ischemic stroke where every minute of delay means more dead neurons.
Researchers at the University of Michigan have built a system called Prima that cuts that wait to seconds. Published this month in Nature Biomedical Engineering, the study covers over 30,000 MRI evaluations. Across more than 50 distinct radiologic diagnoses, Prima outperformed every other advanced AI model tested.
What Prima Actually Does
Prima is a vision-language model - the same architectural family as GPT-4o or Gemini, but purpose-built for clinical imaging. It processes images, video, and text simultaneously. What makes it different from earlier medical AI isn't just scale, though the training set is extraordinary: 200,000 MRI studies and 5.6 million imaging sequences collected since radiology records were digitized at UMich Health.
The key distinction is context. Earlier MRI models were narrow - trained to find lesions, or estimate dementia risk, or measure tumor volume. One task per model. Prima integrates the patient's clinical history and the reason the scan was ordered in the first place. It works the way an experienced radiologist works: weigh the image against the patient story, then render a judgment.
"Prima works like a radiologist by integrating information regarding the patient's medical history and imaging data to produce a comprehensive understanding of their health." - Samir Harake, data scientist, UMich
For routine diagnoses, Prima delivers results instantly after imaging completes. For emergencies - strokes, brain hemorrhages - it doesn't just flag the case. It automatically routes an alert to the appropriate subspecialist: stroke neurologist, neurosurgeon, depending on what it finds.
The Second-Order Effect Nobody Is Talking About
The obvious headline is "AI beats radiologists at reading brain scans." That framing misses the point almost entirely.
The actual shift is structural. Right now, the bottleneck in emergency neurology is physician availability - specifically the radiologist who needs to sign off on a read before the treatment team can act. In overnight hospital settings, that radiologist may be covering multiple departments remotely. An urgent scan gets queued.
Prima doesn't replace radiologists. It eliminates the queue for the cases that can't afford to wait. Routine reads get processed instantly, freeing radiologists to focus on complex or ambiguous cases. Emergencies get routed before a human even looks at the worklist.
This is a workflow transformation disguised as a diagnostic accuracy paper. The 97.5% accuracy number will dominate the coverage. The 24/7 autonomous triage capability is the thing that actually restructures hospital operations.
What Comes Next
The team notes that demand for MRI globally is rising faster than the supply of trained radiologists. That gap is sharpest in lower-income countries and rural regions within wealthy ones. A model that can run at the point of imaging, without a specialist on call, changes the access equation significantly.
The harder question is liability. When Prima flags a case as non-urgent and a physician agrees without independent review, who is responsible if that call is wrong? Medical AI has historically struggled to get from "impressive benchmark" to "routine clinical deployment" precisely because the legal and regulatory frameworks haven't kept pace with the technical capability.
Prima's publication in Nature Biomedical Engineering - one of the highest-credibility venues for this kind of work - positions UMich for FDA Breakthrough Device designation, which would accelerate the regulatory path considerably. Whether that path leads to widespread deployment or a longer slog through liability negotiations is the story to watch.
The model works. The bottleneck just became the institutions.