You get a brain MRI. Maybe your doctor ordered it because of headaches that won’t quit, or numbness in your hand, or a fall that left you dizzy for too long. The scan itself takes 30 to 60 minutes. Then you wait.

You wait for a radiologist to read your images. That wait can be hours. Sometimes days. If you’re in a rural hospital or an emergency room on a busy Saturday night, it might be even longer.

Researchers at the University of Michigan just built something that could change that equation entirely. It’s called Prima, and it reads brain MRIs in seconds.

What Prima Actually Does

Prima is what’s known as a vision language model. Unlike standard AI tools that analyze only images, Prima processes images, video, and text simultaneously. Feed it an MRI scan along with the patient’s clinical history and the reason the doctor ordered the scan, and it produces a diagnosis in seconds.

Not a guess. Not a rough screening. A comprehensive neurological assessment across more than 50 different conditions, with accuracy reaching 97.5%.

The system was developed by a team led by Dr. Todd Hollon, a neurosurgeon at University of Michigan Health. It was published in Nature Biomedical Engineering in February 2026.

The Training Behind It

Here’s what makes Prima different from the wave of medical AI tools that have come and gone over the past few years: the sheer scale of its training data.

The team fed Prima every brain MRI taken at University of Michigan Health since the hospital digitized its radiology records. That’s over 200,000 MRI studies containing 5.6 million individual imaging sequences. They didn’t just give it pictures. They included the patient’s medical history, the physician’s notes on why the imaging was ordered, and the radiologist’s final diagnosis.

In other words, Prima learned the way a radiologist learns: by seeing tens of thousands of cases, understanding the context behind each one, and connecting patterns across an enormous dataset. The difference is that a radiologist builds that knowledge over a decade of training. Prima built it in the time it takes to run a training job.

How It Compares

The research team tested Prima across a year-long study using more than 30,000 MRI studies. Across all 50+ radiologic diagnoses, from strokes to brain hemorrhages to tumors to degenerative conditions, Prima outperformed every other state-of-the-art AI model.

Co-author Samir Harake put it this way: “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.”

But it does something a radiologist can’t: it works instantly, at any hour, without fatigue, and at scale.

Why Speed Matters More Than You Think

When someone walks into an emergency room with stroke symptoms, every minute counts. A clot blocking blood flow to the brain kills roughly 1.9 million neurons per minute. The faster you get a diagnosis, the faster treatment begins, and the more brain tissue you save.

In a well-staffed urban hospital, an emergency MRI might get read within an hour. In a rural hospital or during overnight shifts, that timeline stretches. An AI that delivers an accurate read in seconds and automatically alerts the right subspecialist (a stroke neurologist, a neurosurgeon) could mean the difference between full recovery and permanent disability.

As Dr. Hollon said: “Accuracy is paramount when reading a brain MRI, but quick turnaround times are critical for timely diagnosis and improved outcomes.”

The Bigger Picture

The researchers see Prima as something bigger than a brain MRI tool. They envision it as “ChatGPT for medical imaging,” a foundation that could be adapted for mammograms, chest X-rays, ultrasounds, and other imaging types.

That’s a significant claim, but the underlying logic is sound. The same approach (train on massive real-world clinical data, combine imaging with patient context, optimize for both accuracy and speed) could work across radiology.

Medical imaging is one of the most obvious applications for AI. Radiologists spend their days doing pattern recognition at enormous scale, exactly what machine learning excels at. The field has been waiting for an AI system that’s accurate enough and reliable enough to trust with real clinical decisions.

Prima’s 97.5% accuracy across 50+ conditions, validated on 30,000+ real studies, is the strongest evidence yet that we’re getting close.

What This Means for Regular People

If you’re not in medicine, here’s the practical version:

Faster answers. If Prima or something like it reaches clinical deployment, the wait between “your MRI is done” and “here’s what we found” could shrink from days to minutes.

Better outcomes in emergencies. For strokes, hemorrhages, and other time-sensitive conditions, seconds matter. An AI pre-read that alerts the right specialist immediately could save lives.

More consistent care. A radiologist in a major academic hospital sees more cases and has more support than one in a small rural clinic. AI could level that playing field by giving every hospital access to the same diagnostic quality.

Not replacing your doctor. Prima is designed as a tool, not a replacement. The idea is that AI handles the initial read and triage, flagging urgent cases instantly, while radiologists review the results, handle complex cases, and make the final call.

The Catch

No AI system with 97.5% accuracy has 100% accuracy. That remaining 2.5% matters enormously when you’re talking about brain diagnoses. A missed tumor or a misidentified hemorrhage has real consequences.

The system was trained entirely on University of Michigan data. How it performs on patients from different hospitals, different demographics, different MRI machines, and different clinical workflows is an open question. Medical AI has a history of performing beautifully in the lab and stumbling in the real world.

And there are practical questions: who is liable when an AI misses something? How do hospitals integrate a system like this into existing radiology workflows without disrupting care? How do you update the model as medical knowledge evolves?

These aren’t reasons to dismiss Prima. They’re reasons to watch it carefully as it moves toward clinical use. The technology is impressive. The challenge now is making it work everywhere, not just at the hospital where it was built.

The Bottom Line

An AI that reads brain MRIs in seconds, with 97.5% accuracy, trained on 5.6 million real clinical imaging sequences, tested on 30,000 real patients over a full year.

That’s not a research demo. That’s a tool knocking on the door of clinical medicine. And for anyone who’s ever waited days for MRI results while anxiety ate them alive, that door can’t open fast enough.